IMFstaffpapers - IMF eLibrary€¦ · ttt t (6) Empirical Dealer-Level Tests—Lyons (1995): ()* eIID it it it t=− − +μ α γ (1) Open Macroeconomics Theory—Parity conditions,

IMFstaffpapersRobert Flood

Editor and Committee Chair

Eswar S. PrasadCo-Editor

Sean M. CulhaneSenior Editor

Nicolas AmorosoResearch Assistant

Rosalind OliverAdministrative Coordinator

Editorial CommitteeReza Baqir Eduardo LeyTito Cordella Donald J. MathiesonGiovanni Dell’Ariccia Gian Maria Milesi-FerrettiEnrica Detragiache Jorge RoldosAndrei Kirilenko Miguel A. SavastanoLaura Kodres Sunil SharmaAyhan Kose Antonio Spilimbergo

The objective of IMF Staff Papers is to publish high-quality research produced by IMF staffand invited guests on a variety of topics of interest to a broad audience including academicsand policymakers in the member countries of the Fund. The papers selected for publicationin the journal are subject to an extensive review process using both internal and external ref-erees. IMF Staff Papers also welcomes outside comments, criticisms, and interesting replica-tions of published work. The views presented in published papers are those of the authorsand should not be attributed to or reported as reflecting the position of the IMF, its ExecutiveBoard, or any other organization mentioned herein.

Subscription: US$99.00 a volume or the approximate equivalent in the currencies of mostcountries. Three issues constitute a volume. Single copies may be purchased at $30.00. Individualacademic rate to full-time professors and students of universities and colleges: $58.00 a volume.Subscriptions and orders should be sent to:

International Monetary FundPublication Services

700 19th Street, N.W. Washington, D.C. 20431, U.S.A.

Telephone: (202) 623-7430Fax: (202) 623-7201

E-mail: [email protected]: http://www.imf.org

©International Monetary Fund. Not for Redistribution

http://www.imf.org

International Monetary Fund

IMFstaffpapersVolume 52 Number 3

2005


EDITOR’S NOTE

The Editor invites from contributors outside the IMF brief comments (not more than1,000 words) on published articles in IMF Staff Papers. These comments should beaddressed to the Editor, who will forward them to the author of the original article forreply. Both the comments and the reply will be considered for publication.

The data underlying articles published in IMF Staff Papers (where available) maybe obtained from the journal’s website (http://www.imf.org/staffpapers). Readers areinvited to use these data to expand on the material in the articles, and the journal willconsider publishing such work.

© 2005 by the International Monetary FundISBN 1-58906-475-5

International Standard Serial Number: ISSN 1020-7635

This serial publication is catalogued as follows:

International Monetary FundIMF staff papers — International Monetary Fund. v. 1– Feb. 1950–

[Washington] International Monetary Fund.

v. tables, diagrs. 26 cm.

Three no. a year, 1950–1977; four no. a year, 1978–

Indexes:Vols. 1–27, 1950–80, 1 v.

ISSN 1020-7635 = IMF staff papers — International Monetary Fund.1. Foreign exchange—Periodicals. 2. Commerce—Periodicals.

3. Currency question—Periodicals.

HG3810.15 332.082 53-35483


http://www.imf.org/staffpapers

International Monetary Fund

ContentsVolume 52 Number 3

2005

Why Are Asset Markets Modeled Successfully, But not Their Dealers?Rafael Romeu • 369

Real Exchange Rates in Developing Countries: Are Balassa-Samuelson Effects Present?Ehsan U. Choudhri and Mohsin S. Khan • 387

The Internal Job Market of the IMF’s Economist ProgramGreg Barron and Felix Várdy • 410

Banking on Foreigners: The Behavior of International Bank Claims on Latin America, 1985–2000

Maria Soledad Martinez Peria, Andrew Powell, and Ivanna Vladkova-Hollar • 430

Assessing Early Warning Systems: How Have They Worked in Practice?Andrew Berg, Eduardo Borensztein, and Catherine Pattillo • 462

Does SDDS Subscription Reduce Borrowing Costs for Emerging Market Economies?John Cady • 503

IndexVolume 52 • 539

Special Data Section

Domestic Debt Markets in Sub-Saharan AfricaJakob Christensen • 518


Underline

Underline

Underline

Underline

Underline

Underline

Underline

Underline

This page intentionally left blank


369

IMF Staff PapersVol. 52, Number 3

© 2005 International Monetary Fund

Why Are Asset Markets Modeled Successfully,But Not Their Dealers?

RAFAEL B. ROMEU*

Market-level microstructure models of asset pricing succeed where dealer-levelmodels do not. This study addresses this empirical difficulty in the context of for-eign exchange dealers. New evidence is presented rejecting the latter models’specifications of how information asymmetry and inventory accumulation affectdealer pricing. This rejection is consistent with those of other dealer-level empir-ical studies. A new modeling avenue may be to reconsider optimal price settingwhile relaxing assumptions that specify incoming orders as the only componentthrough which dealer inventories evolve. This approach is consistent with inventoryevolution data and with market-level models’ assumptions about currency markets.[JEL F3, F4, G1]

High-frequency data combined with recent microstructure models have deliv-ered empirical success. For example, exchange rate models that reflect infor-

mation gathering and risk sharing in their currency-trading processes outperform arandom walk.1 In these models (often referred to as micro exchange rate models),the exchange rate depends not just on tracked statistics of economic aggregates,such as inflation or investment, but also on other variables that reflect the market’s

*Rafael B. Romeu is an Economist in the Caribbean II Division of the IMF’s Western HemisphereDepartment. The author thanks Roger Betancourt, Michael Binder, Juan S. Blyde, Martin Evans, JonFaust, Robert Flood, Andrei Kirilenko, Richard K. Lyons, José Pineda, John Rogers, Jorge Roldos,Carmen M. Reinhart, Dagfinn Rime, Francisco Vázquez, Jonathan H. Wright, and two anonymous refer-ees for helpful comments, as well as seminar participants at the Bank of Canada, the Board of Governorsof the Federal Reserve, Citibank FX, the IMF Research Department, the University of Maryland EconomicsDepartment and the R. H. Smith Department of Finance. Thanks also to Jushan Bai for code.

1Out of sample, in the sense of Meese and Rogoff (1983). See Evans and Lyons (2002).


Rafael B. Romeu

370

view of economic conditions. One can partition micro exchange rate models intomarket-level (ML) models and dealer-level (DL) models. ML micro exchange ratemodels study how a market-wide consensus of asset values is achieved. ML modelsfocus on how the entire market builds such a consensus and settles on an exchangerate. These models can explain more than 50 percent of exchange rate movements.2DL micro exchange rate models abstract from the market as a whole and focusinstead on price setting and risk management by individual currency market par-ticipants, or dealers. This study explores a rift between ML models and DL mod-els. First, it shows new empirical rejections of some DL model predictions. Next, itshows that a basic DL assumption is inconsistent with ML models and with thedata. This may be why some DL model predictions are routinely rejected both inthis and in previous studies of equity and other asset markets.

Before explaining the difficulties with DL models put forth here, it is useful tomap exactly where they lie in the literature of exchange rates. Figure 1 partitionsthe research on exchange rates into six broad categories. Traditional models ofexchange rates, which face well-known empirical difficulties, are represented byBox (1) in Figure 1. In these models, a handful of parity conditions are assumed tolink macroeconomic activity across countries. One such condition is purchasingpower parity (PPP). PPP relates the difference in inflation rates across countries totheir exchange rate depreciation. Although empirical predictions of macroeconomicmodels are generally inconsistent with exchange rate data, parity conditions areconsistent. For example, Flood and Taylor (1996) show that long-run data supportPPP and other parity conditions, as denoted in Box (2) in Figure 1.3 The upshotof their study is given in equation (1). Exchange rate depreciation between twoperiods of time (t denotes time; Δe denotes exchange rate depreciation) depends onpublicly observable fundamental macroeconomic variables (denoted by F ) and an“unexplained” component (denoted by U):

Instead of assuming that parity conditions govern exchange rate evolution,micro exchange rate models consider factors that drive currency market partici-pants’ price setting. Empirical ML micro exchange rate models, such as Evans andLyons (2002), are represented in Box (4) of Figure 1. In these models, market par-ticipants receive economic information through order flow that they cannot learnfrom public macroeconomic statistics. Order flow results from partitioning totaltraded volume into either buyer-initiated transactions or seller-initiated transactionsand taking their difference. Order flow plays an important role in estimating theexchange rate because it captures changes in expectations and risk preferences thatare absent from publicly tracked economic aggregates. The resulting exchange ratedepreciation equation (2) is almost identical to equation (1). The interest rate dif-ferential change (denoted Δ(rt − r*t )) represents the fundamental variable, and the

Δe F Ut t t= + . ( )1

2Evans (2002).3Also confirmed by Sarno and Taylor (2002).


WHY ARE ASSET MARKETS MODELED SUCCESSFULLY?

371

(5) Dealer-Level Models— Lyons (1995); Madhavan and Smidt (1991 and 1993):

*( )it it ite I Iμ α= − −

(3) Market-Level Microstructure Theory—Lyons (1997):

t t te F xΔ = Δ + Δ

(4) Empirical Market-Level Microstructure—Evans and Lyons (2002):

*( )t t t te r r xΔ = Δ − + Δ

(6) Empirical Dealer-Level Tests—Lyons (1995):

*( )it it it te I I Dμ α γ= − − +

(1) Open Macroeconomics Theory—Parity conditions,

such as PPP: *t t tp e p=

(2) Empirical Macro—Flood and Taylor (1996), and

others: Δet = Ft + Ut

e = exchange rate (i indicates that it is set by dealer i, t indicates at date/trade t).p = price level (* indicates foreign). I = inventory of foreign exchange (* indicates desired or optimal inventory level). μ = the dealer’s best guess of the full-information value of the currency.x = order flow.r = interest rate (* indicates foreign). F = publicly observable measures of economic fundamentals, for example, interest rates, price levels. U = exchange rate variation “unexplained” by publicly observable measures of economic fundamentals.D = indicator that is 1 if x > 0, and is –1 if x < 0.γ , α > 0.

Notes:

Figure 1. Partitions in the Exchange Rate Literature

Figure 1 shows the disconnect between dealer-level (DL) and market-level (ML) microstructure,which is explored here. The exchange rate literature is partitioned into broad categories (each indicated bya numbered box), with arrows indicating theoretical/empirical support among areas. This paper shows thatDL microstructure models predict a pricing equation (in Box (5)) that is rejected by DL empirical studies(and hence the broken link to Box (6)). Furthermore, DL microstructure models are inconsistent with MLmicrostructure models—represented by Box (3). However, ML microstructure models are empirically sup-ported by microdata (Box (4)), and they are closely related to open macroeconomic empirical studies.These (in Box (2)) support parity conditions from open macroeconomic models using long-run data andthe same estimating equation as predicted by ML microstructure models. Finally, the theoretical link fromopen macroeconomics to ML microstructure (Box (3) and Box (1)) is under development (for example,Evans and Lyons, 2004).


Rafael B. Romeu

372

“unexplained” variable of Flood and Taylor (1996) is the order flow variable(denoted Δxt) in Evans and Lyons (2002):

The theory that yields the empirical specification in equation (2) is based onML models of simultaneous trading in currency markets (see, for example, Lyons,1997—Box (3) in Figure 1). In these models, first exchange rates are simultane-ously set by currency dealers. These dealers must all set prices (simultaneously) atwhich they are willing to buy or sell any amount of currency. Next, market partic-ipants observe everyone else’s exchange rates and submit their orders to the othersin the market. These conditions guarantee that all dealers set the same exchangerate, because any differences would lead to large arbitrage opportunities and unravelthe equilibrium. In equilibrium, all dealers set the same exchange rate and there areno opportunities for arbitrage. For all dealers to know which exchange rate to set,it must be based on publicly available information. Hence, in these models, deal-ers’ exchange rates are common and based on publicly known order flow andmacroeconomic variables.

Actual market participants, however, are constantly changing prices in over-the-counter currency markets.4 That is, since currency trading occurs over thecounter, at any point an individual dealer’s exchange rate may diverge from others’in the market.5 To study price setting in this market, DL models consider an indi-vidual dealers’ exchange rate setting—Box (5) in Figure 1. Dealers in these mod-els set prices as they receive incoming orders from other market participants. Theinitiators of the incoming orders may know more about future asset values than thedealers receiving the orders. In this situation, the incoming orders reflect informa-tion about future asset values and consequently drive currency prices. This is theasymmetric information effect. Also, in these models dealers have a finite inventoryof the asset on which to draw for liquidity provision. As incoming orders drive thedealer’s asset inventory away from her optimal level, she changes prices to inducecompensating orders. This is the inventory effect. The classic DL pricing conjec-ture is given by Madhavan and Smidt (1991)—in Box (5) in Figure 1.

Empirical tests of DL models generally support asymmetric informationeffects;6 they do not, however, find inventory effects.7 One study, Lyons (1995),

Δ Δ Δe r r xt t t t= −( ) +* . ( )2

4See Evans (2002) for evidence of concurrent, unequal prices in foreign exchange markets.5Then one may ask why the assumptions of ML models guarantee that all dealers set the same price.

The return in economic insight to modeling competitive dealers setting different prices concurrently is likelyto be small relative to the cost of overcoming the intractability of competitive market equilibrium, particu-larly in terms of the necessary assumptions. See O’Hara (1995, Chapter 2) on precisely this intractability.

6For example, Hasbrouck (1988, 1991a, and 1991b) and Madhavan and Smidt (1991 and 1993) inequity markets; Lyons (1995), Yao (1998), and Bjønnes and Rime (2005) for foreign exchange markets,among others.

7Madhavan and Smidt (1991) do not find inventory effects. Madhavan and Smidt (1993) allow a chang-ing optimal inventory level and find evidence of inventory management with a half-life of more than sevendays, suggesting quite different effects from theoretical predictions. Furthermore, they reject the hypothesisof intraday inventory management, whereas Madhavan, Richardson, and Roomans (1997) argue that if there



373

finds direct evidence of asymmetric information and inventory management pre-dicted by DL inventory theory—Box (6) in Figure 1.

This study reconsiders the use of traditional dealer-level pricing specifica-tions, and, specifically, this study reexamines the Lyons (1995) result. Evidence ofparameter instability and model misspecification in Lyons (1995) is presented.When estimated over the full data set, that study’s DL pricing equation containsbreaks. In subsamples where no breaks are present, the results do not fully supportDL model predictions. Specifically, asymmetric information and inventory effectsare not present simultaneously in subsamples; hence, although they do not rejectthe presence of asymmetric information or inventory effects in the data, the mod-els’ specifications of these effects are rejected. This is discussed further below andis indicated by the broken link between Box (5) and Box (6) in Figure 1. Then,Section II discusses an underlying assumption in DL models’ pricing specifica-tion that may be behind their persistent empirical difficulties. Basically, theassumption that inventory accumulation is driven only by incoming order flow isquestionable. This assumption is shown to be in contradiction to both the inven-tory data and ML micro exchange rate theory. This is indicated by the broken linkbetween Box (3) and Box (5). Relaxing this assumption is a promising avenue forfurther DL modeling. Section III concludes.

I. Reconsidering the Lyons (1995) Result

This section reconsiders the Lyons (1995) DL exchange rate model (for details,see that study). Equation (3) gives the Lyons (1995) DL specification for how thedealer’s price changes at each incoming trade (denoted by subscript t). Intuitively,the change in the exchange rate is a function of the incoming order size, directionof trade (that is, purchase or sale), and current and past inventory levels.

with predicted signs: {β1, β3, β4 > 0}, {β2, β5 < 0}.

Pt: The price of the dealer at which an incoming sale or purchase occurred.Qjt: The incoming quantity demanded by the opposite party, that is, order flow.It: This is the dealer’s inventory at the time of (but not including) the incom-

ing quantity Qjt.Dt: The indicator that picks up the direction of trade, positive for purchases,

negative for sales.

ΔP Q I I D D mat jt t t t t= + + + + + + ( )2 3 − −β β β β β β0 1 1 4 5 1 1 , (( )3

is inventory management, it occurs toward the end of the day. Hasbrouck and Sofianos (1993) find very slowinventory adjustment as well, although they confirm that specialists are able to adjust inventory quicklyduring large exogenous shocks if they choose to. Hence, inventory levels are voluntary, not due to volumeconstraints, and must reflect long-term positions. Manaster and Mann (1996) find strong evidence that spe-cialists do not control inventory as models would predict; rather, the exact opposite occurs. Furthermore,Madhavan and Sofianos (1998) also find that dealers do not change quotes to induce trades as theoreticallypredicted, but rather participate selectively in markets to unwind undesired positions. The general empiricalfailure of inventory model predictions described above for equity markets is borne out in foreign exchangemarket studies by Yao (1998) and Bjonnes and Rime (2000). Neither study can find the inventory manage-ment results predicted by the Madhavan and Smidt (1991) model.


Rafael B. Romeu

374

Equation (3) predicts increasing prices with purchase orders and larger laggedinventory, and decreasing prices with sale orders, and larger current inventory.8The Lyons (1995) estimates of this equation are presented in Table 1.9 The esti-mates are consistent with model predictions and significant at better than 1 per-cent. The robustness of these estimates is the subject of this section.

Figure 2 shows evidence of parameter instability in equation (3). In each graph,the abscissa indexes the incoming trades. The top two panels graph the probabil-ity that the trade is a breakpoint, with P-values indicated in the ordinate (both theF-test and the Likelihood Ratio test are reported). As the graphs show, the nullhypothesis of no break is rejected toward the middle of the sample, as well astoward the end (the left graph uses the Chow breakpoint tests, the right uses Waldtests). This is indicated by the declining P-values throughout the middle of the sam-ple and again at the end. The bottom panels show how the coefficients on equation(3) change as the regression is estimated on a rolling window of 150 transactions(beginning with the transaction indicated on the abscissa). The bottom left panelgraphs the coefficient on incoming order flow (β1) and its t-statistic. The bottomright panel does the same for the contemporaneous inventory coefficient (β2). Whileone would expect some variation in the significance of the estimates owing to asmaller sample, the variation should not be systematic and should reduce the esti-mates’ significance uniformly. One can observe that order flow is significant in the

8The moving average coefficient on the error term in equation (3) is predicted negative.9The data are a one-week (843 observations) data set of a New York currency dealer of the dollar/DM

market from August 3–7, 1992. See Lyons (1995) for an extensive exposition of this data set. The Lyons(1995) model includes a public information signal and specification of equation (3) with an extra regressor—brokered trading, Bt. That study estimates equation (3) both with and without the public signal because ofpoor measurement of the public signal in relation to the measurement of the other variables. Essentially, thebrokered trading variable has measurement error and is zero in 84 percent of the dealer’s transactions. Thissection focuses on estimates without brokered trading; however, a single break is found with it included inthe Sup-F test.

Table 1. Reproduction of Lyons (1995) Original Estimates

Variable Coefficient Std. Error t-Statistic Prob.

C −1.29 0.00 −0.96 0.34Qjt 1.47 0.00 3.17 0.00It −0.92 0.00 −3.38 0.00It−1 0.72 0.00 2.76 0.01Dt 10.30 0.00 4.77 0.00Dt−1 −9.16 0.00 −6.28 0.00MA(1) −0.09 0.03 −2.71 0.01

R-squared 0.22 F-statistic 39.28Adjusted R-squared 0.22 Prob(F-statistic) 0.00

Notes: Table 1 reproduces the baseline DL model estimates of exchange rate changes given inequation (3). (See Lyons, 1995, Table 4, p. 340). All coefficients are multiplied by 105 except themoving average.



375

beginning of the sample, whereas inventory is significant toward the end of thesample. Hence, the DL model predictions of both asymmetric information (signif-icant order flow coefficient β1) and inventory effects (significant inventory coeffi-cient β2) appear to not hold in subsamples. To get a feel for what is occurring atthese points, Figure 3 shows the price set by the dealer. Solid vertical lines showthe end of days of the week, and dashed vertical lines show two breaks consideredin this section. The declining P-values in Figure 2 come at the end of the third dayand close to the end of the sample.

To investigate the possibility of parameter instability in equation (1), Table 2reports the results for the presence or location of (possibly multiple) structuralbreaks.10 A break is found at transaction 449.11 The right column of Table 2 reportsthe starting and ending observations of each of the five trading days from which thedata were recorded. As Figure 3 shows, the break occurs near the end of Wednesday(overnight observations are removed). This break coincides with the end of a trad-ing day; however, with three other day changes, there is no evidence to suggest thatthese alone induce structural breaks. Figure 2 suggests that there is another breaktoward the end of the sample; however, Sup-F tests cannot detect breaks within 5percent of sample endpoints. On the last day of the sample, a $300 million Fedintervention occurred after the close of the European markets.12 This event maycause further parameter instability in the DL model estimates.13 Hence, Table 3reports conventional break tests conducted on the trade at which the interventionbegins. The breaks and price are jointly shown in Figure 3. Given these jointresults, one may conclude that the DL model is subject to two breaks when esti-mated on the Lyons (1995) data.

Table 4 reports estimations of the DL model on the subsamples that result fromsegregating the data at the breaks. Estimates from the subsample prior to the firstbreak (observations 2 to 448) are in the top two lines; this subsample of data repre-sents more than 53 percent of the available observations. The estimates reveal thatthe coefficients for inventory are insignificant at conventional levels, whereas signed

10Sup-F tests are based on Andrews (1993) and Bai and Perron (1998).11The Sup-F tests allow for heterogeneity and autocorrelation in residuals using the Andrews (1991)

method. Separate tests of the Lyons (1995) residuals fail to reject the null hypothesis of no breaks at con-ventional significance levels, although an overnight break for the first day is found at the 10 percent signif-icance level.

12The Federal Reserve confirms a $300 million intervention on that day but does not reveal its inter-vention timing or strategy. The financial press widely report (ex post) the approximate intervention start.The most precise timing is documented by the Wall Street Journal, August 10, 1992: “The Federal ReserveBank of New York moved to support the U.S. currency . . . as the dollar traded at 1.4720” (Linton, 1992).That price corresponds to 12:32 p.m. in the Lyons (1995) data set, and that time is consistent with otherfinancial news reports.

13Models that show how interventions affect trading include Bhattacharya and Weller (1997), Vitale(1999), Evans and Lyons (2001), Dominguez (2003), and others. For example, the Evans and Lyons (2001)model finds evidence of portfolio balance effects from interventions. A late-day and end-of-week interven-tion, one that occurs after other major markets (London and Tokyo) have closed for the weekend, would pre-sumably bring to bear these effects. That is, the dealer would have very little time and fewer marketparticipants (since the entire market would be affected) with which to share the intervention’s portfolioimbalance over the weekend and, hence, would charge a higher premium for liquidity provision than at othertimes.


Ra

fae

l B. Ro

me

u

37

6

Figure 2. Rolling Estimates of Break Tests and DL Pricing Equation

Chow Breakpoint Test

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

50 150 250 350 450 550 650

Observation

F Stat Like Rat

Incoming Order Flow Coefficient

-1.50

-1.00

-0.50

0.00

0.50

1.00

1.50

2.00

2.50

3.00

3.50

1 101 201 301 401 501 601

-1

-0.5

0

0.5

1

1.5

2

2.5

3

QJT T-stat (right axis)

Inventory Coefficient

-3.50

-3.00

-2.50

-2.00

-1.50

-1.00

-0.50

0.00

0.50

1 101 201 301 401 501 601

-5

-4

-3

-2

-1

0

1

Inv T-stat (right axis)

Wald Test, Entire Sample

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

1 101 201 301 401 501

Observation

F-stat Pval LR Pval

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992. Notes: The abscissa indexes observation number of the sample (on all graphs). The top left panel graphs the probability that the observation is a breakpoint, with the

P-value indicated in the ordinate (both the F-test and the Likelihood Ratio test are reported). The top right graphs the same using a Wald test. The bottom left panel graphs the coefficient on incoming order flow using a rolling window of 150 observations (beginning with the observation indicated on the abscissa) and also reports the t-statistic. The bottom right panel does the same for the contemporaneous inventory coefficient.



377

Daily Cumulative QQ

-150

-100

-50

0

50

100

150

1 76 151 226 301 376 451 526 601 676 751 826

Daily Cumulative QJT

-150

-100

-50

0

50

100

1 76 151 226 301 376 451 526 601 676 751 826

Price

1.46

1.465

1.47

1.475

1.48

1.485

1 76 151 226 301 376 451 526 601 676 751 826

Monday Tuesday Wednesday Thursday Friday

End of days Breaks

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.Notes: Figure 3 graphs the price set by the dealer in the top panel. The middle panel graphs

cumulative daily incoming order flow, and the bottom panel graphs the cumulative sum of the unmodeled inventory evolution variable, QQ. The solid vertical lines represent the end of days; the dashed lines represent breaks.

Figure 3. Price, Daily Cumulative Components of Inventory, and Breaks


Rafael B. Romeu

378

order flow (that is, the asymmetric information effect) and the order flow indicatorsare significant and estimated at magnitudes similar to the baseline estimates.

The estimates from the subsample 449 to 794 are reported in the third andfourth lines. The order flow coefficient is now insignificant, and the inventory com-ponents are significant at all conventional levels. These estimates suggest thatasymmetric information is not present in dealer pricing on the last two days of thesample, which is just prior to the Fed intervention.

The third subsample, consisting of approximately 5 percent of the total avail-able observations, likely reflects the effects of the Fed intervention. The only sig-nificant effect (at the 10 percent level) is the asymmetric information effect, and itseems to be an order of magnitude larger than the other subsample estimates. Ingeneral, the model fits this section of the sample poorly.

The bottom two lines shows estimates that result from joining the third sub-sample to the second, essentially ignoring the Fed intervention break. The Sup-F testcannot find this break (because of its proximity to the sample endpoint), but theChow test rejects the null of no break at this point. Estimating these two subsamplesjointly shows order flow and the order flow indicator coefficients significant at the10 percent level but not at 1 percent. The inventory effects are significant, and thesigns of the coefficients are as predicted (which was not the case for the Fed inter-vention subsample alone). However, the proportion of variation explained by the

Table 2. Sup-F Tests for Location and Number of Structural Breaks

Structural Breaks

Significance = 1% End of DayFixed Break(s) Point(s) Monday 181(p=0) 1 449 Tuesday 330

Wednesday 440Thursday 592Friday 843

Notes: Table 2 shows the results of Sup-F tests for multiple structural breaks on equation (3). Thetest finds a break at observation 449 at the 1 percent significance level. The right column showschanges in days in the sample; breaks are not found at changes from one day to the next (overnightobservations are excluded), however, the break date is close to the change from Wednesday toThursday. All estimations and break tests are based on the Lyons (1995) DL specification that excludesBt—brokered trading. Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.

Table 3. Break Test for Fed Intervention

Chow Breakpoint Test: Observation 795

F-statistic 5.8 Probability 0.00Log likelihood ratio 40.5 Probability 0.00

Note: Table 3 shows the results of traditional break tests on the suspected entry point of the Fedin the market.



379

regression falls from 32 percent (without the intervention subsample) to 17 percent(with the intervention subsample). Hence, while the estimation that averages acrossthe two subsamples (that is, ignoring the Fed intervention) recuperates to someextent DL model predictions, adding observations reduces its explanatory power.14

II. A Puzzle of Microstructure Market Maker Models

DL models study the transaction prices that currency dealers set as orders arrivethroughout the trading day. They draw from equity market studies, which considerthe price-setting behavior of a “monopoly” specialist, a single market maker withno other source of liquidity. Consistent with specialists’ inventory managementtheory,15 DL models assume that dealers set prices to control an inventory thatevolves according to equation (4):16

I I Qit it jt+ = −1 4, ( )

14Furthermore, identifying the first break at the first or last observation at which the Chow test p-valuefalls below 5 percent in Figure 2 (observations 392 and 541) does not change the result that the first regimedoes not have inventory effects, and the second has no asymmetric information effects.

15For example, Stoll (1978), Amihud and Mendelson (1980), Ho and Stoll (1981), O’Hara andOldfield (1986), among others. It is useful to note, however, that equity market specialists on the New YorkStock Exchange compete aggressively against a limit order book that they themselves manage and, if nec-essary, can induce orders from the trading floor through moral suasion.

16Equivalently, some models (for example, Madhavan and Smidt, 1991; or Lyons, 1995) conjecture apricing equation consistent with inventory of equations (4) and (5). Prices are assumed to be set accordingto Pit = μit − α(Iit − I*) + γDt, where I* is the dealer i’s desired inventory level, and Dt is one if the trans-action is on the offer (that is, the aggressor purchases), and negative one if the transaction occurs on thebid (that is, the aggressor sells). It picks up the bid-ask bounce for quantities close to zero. Hence, pricesare set according to the best estimate of the full information value and then adjusted to induce inventory-compensating trades.

Table 4. Estimates of DL Pricing Model in Subsamples with No Breaks

C Qjt It It−1 Dt Dt−1 MA(1) Subsample Adj. R2

Coefficient −1.75 1.28 −0.354 0.12 12.60 −8.82 −0.20 2 to 448 0.32Prob. 0.15 0.01 0.20 0.65 0.00 0.00 0.00Coefficient −3.17 0.90 −2.04 1.86 11.00 −11.2 −0.10 449 to 794 0.30Prob. 0.14 0.19 0.00 0.00 0.00 0.00 0.06Coefficient 15.40 14.40 3.22 −2.58 −28.1 −1.65 0.10 795 to 839 −0.05Prob. 0.38 0.06 0.39 0.43 0.30 0.92 0.54Coefficient −0.78 1.73 −1.63 1.45 7.12 −10.1 −0.04 449 to 839 0.17Prob. 0.77 0.04 0.00 0.00 0.07 0.00 0.40

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992. Notes: Table 4 shows estimates of the three subsamples, with breaks at observations 449 and 795.

The first break is given by the Sup-F test. The second break, observation 795, is given by the tradi-tional F-test. The top panel reports the first subsample, observations 1 to 448. The second panelreports estimates from observations 449 to 794. The third panel reports estimates from observations795 to 838. The fourth panel reports the second and third subsamples estimated together. All coeffi-cients are multiplied by 105 except the moving average.


Rafael B. Romeu

380

with Iit dealer i’s inventory at the beginning of period t, and Qjt, the incoming orderflow from other dealers (represented by subscript j), given by:

In equation (5), μit is dealer i’s best estimate of the full information value, vt, at thetime of quoting. Thus, order flow is a scaled deviation of dealer i’s price fromdealer j’s expectation of vt, plus an orthogonal liquidity shock, Xjt.

In the world of equations (4) and (5), price setting is used to control inven-tory imbalances (and reduce inventory risk) owing to incoming orders. Intuitively,the dealer’s pricing strategy reduces the randomness of the order arrival processby balancing incoming purchases with incoming sales. Such assumptions implythat inventory control is achieved by diverting asset prices away from the full-information value, thus discounting the asset to attract inventory-compensatingtrades. The DL model specifications for inventory effects that these assumptionsyield are consistently rejected by the data.

To find a new direction for market maker modeling, one may consider a smallpart of the Lyons (1995) data set, which is shown on Table 5. The first columnindexes the observations according to the order of arrival; the second column showsthe price set by the dealer; the next columns show incoming order flow, the inven-tory at the beginning of the trade, and a variable called QQit that is backed out ofequation (6):

QQit in equation (6) reflects inconsistencies between the data and the inventoryevolution assumed in equation (4). Consider, for example, the third incoming trade,which was a sale to the dealer of $28.5 million. At the time of the trade, the dealerwas long $1 million. If equation (4) held, then the $28.5 million purchase wouldimply a $29.5-million-long inventory at entry four. Instead, the dealer is short$1.5 million at the next incoming trade, which implies that the inventory somehowdeclined by $30.5 million between the third and the fourth trade. This decline is

I I Q QQit it jt it+ = − +1 6. ( )

Q P Xjt jt it jt= −( ) +θ μ . ( )5

Table 5. First Five Entries of Lyons (1995) Data Set

Entry Pit Qjt It QQ

1 1.4794 −1 1 12 1.4797 −2 3 −43 1.4795 −28 1 −30.54 1.4794 −0.5 −1.5 0.255 1.479 −0.75 −0.75 0

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.Notes: Table 5 shows the first five entries of the price (second column), incoming order flow

(third column), and inventory (fourth column) variables from the data set. The last column is backedout from the equation: Iit+1 = Iit − Qjt + QQit. The generated variable QQ captures the part of inven-tory evolution that is not due to incoming order flow.



381

reflected in QQi3. It captures the gap in the inventory evolution that incoming orderflow did not generate.

Figure 3 graphs the daily cumulative incoming order flow and the daily cumu-lative gap, QQ. This variable appears to be synchronized with incoming order flow.This suggests that whatever is driving QQ may balance the asynchronous arrival ofincoming purchases and incoming sales. QQ may, for example, reflect other meth-ods of inventory control available to the dealer.17 In this case, optimal pricing prob-lems based on equation (4) may be misspecified. Furthermore, DL modeling ofQQ may also consider information about asset values contained similar to thosespecified in equation (5) that reflect alternate sources of information available to thedealer.18

According to both inventory management theory and market data, inventory isstrongly managed by dealers (Iit is mean-reverting), implying that E[Iit+1 − Iit] is sta-tionary. According to equation (4), Qjt is then also stationary (which would be con-sistent with price setting that induces a balance between incoming purchases andsales), thereby making QQit noise. However, another possibility is that [−Qjt + QQit]is stationary. This would imply that QQit and Qjt are economically related, and thatQQit may be a good candidate for microstructure modeling. Figure 4 plots kerneldensities of the empirical distribution of these two series (the two peaks in the dis-

17In currency markets, these methods include initiating interdealer bilateral trades, interdealer bro-kered trades, or International Monetary Market Futures trades.

18Ho and Stoll (1983) model inventory management with two dealers and two assets, thereby includ-ing aspects of competitive trading. Romeu (2003) models DL pricing with a dealer that takes into accountmultiple methods of inventory control and multiple sources of information. See footnote 5 for an impor-tant caveat regarding these types of models.

Figure 4. Kernel Density Plots for QQit and Qjt

.00

.02

.04

.06

.08

.10

.12

–20 –10 0 10 20

QJT

Kernel Density (Normal, h = 0.8729)

.00

.01

.02

.03

.04

.05

.06

.07

.08

–60 –50 –40 –30 –20 –10 0 10 20 30

QQ

Kernel Density (Normal, h = 1.4839)

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.Notes: Figure 4 shows Gaussian kernel densities for the empirical distributions of the unmodeled

inventory evolution variable, QQ, and incoming order flow, Qjt . The two peaks in the distribution of Qjtmost likely reflect clustering at the standard order sizes of $10 million.


Rafael B. Romeu

382

tribution of Qjt most likely reflect clustering at the standard order sizes of $10 mil-lion), which appear to be similar. Table 6 gives descriptive statistics, which showthat the means of the distributions are almost equal in magnitude, the pair-wise cor-relation is 0.64, and tests fail to reject the null hypothesis that the variables’ meansare equal. The similarity in distributions suggests that QQit may be a good candi-date for microstructure modeling. Table 7 shows lag selection criteria for a vector

Table 6. Descriptive Statistics for QQit and Qjt

Mean Median Max Min Std. Dev. Skew. Kurt. Correl.

QQ −0.39 0.00 34.45 −66 8.99 −0.55 7.44 0.64QJT −0.39 0.45 20.00 −28 5.24 −0.29 5.44Test for Equality of MeansIncluded observations: 843Method df Value Probabilityt-test 1684 0.00 1.00Anova F-statistic (1,1684) 0.00 1.00

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.Notes: Table 6 shows descriptive statistics for the unmodeled inventory evolution variable, QQ, and

incoming order flow, Qjt. Tests for equality of means fail to reject equality, and the correlation betweenthe series is presented.

Table 7. VAR Lag Order Selection Criteria

Endogenous variables: QJT QQExogenous variables: C

Included observations: 835

Lag LogL LR FPE AIC SC HQ

0 −5353.6 NA 1276.6 12.8 12.8 12.81 −5151.7 402.3 794.8 12.4 12.4 12.42 −5131.1 40.9* 763.8* 12.3* 12.4* 12.3*3 −5128.3 5.6 765.9 12.3 12.4 12.34 −5123.6 9.2 764.7 12.3 12.4 12.45 −5122.0 3.3 769.0 12.3 12.4 12.46 −5120.0 3.8 772.8 12.3 12.5 12.47 −5118.4 3.0 777.4 12.3 12.5 12.48 −5115.2 6.3 778.8 12.3 12.5 12.4

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.* indicates lag order selected by the criterionLR: sequential modified LR test statistic (each test at 5% level)FPE: Final prediction errorAIC: Akaike information criterionSC: Schwarz information criterionHQ: Hannan-Quinn information criterionNotes: Table 7 shows multiple lag selection tests for a vector auto regression (VAR) of the un-

modeled inventory evolution variable, QQ, and incoming order flow, Qjt. Two lags are selected bymultiple criteria.



383

auto regression (VAR) of the variables. All tests select two lags, which are then esti-mated in Table 8. The coefficients are significant at conventional levels and showan inverse relationship between the lags and contemporaneous values of QQit andQjt. Hence, the evolution in time of incoming order flow may be compensated bythe evolution of QQit. Figure 5 shows the impulse responses of each variable to ashock in the other. A shock in Qjt invokes an immediate response in QQit, whichfurther suggests that elements of microstructure models may be useful in explain-ing the evolution of QQit and consequently of inventories and prices.

Finally, DL models that assume equation (4) and ML models such as Lyons(1997) have conflicting inventory evolution assumptions. In ML models, dealers’

Table 8. Vector Auto Regression Estimates

QJT(−1) QJT(−2) QQ(−1) QQ(−2) C

QJT 0.47 [11.84] 0.24 [5.88] −0.47 [−22.69] −0.15 [−5.67] −0.35 [−2.48]QQ 0.68 [8.72] 0.27 [3.38] −0.58 [−14.34] −0.13 [−2.60] −0.30 [−1.07]t-statistics in [ ]R-squared 0.39 Akaike AIC 5.67Adj. R-squared 0.39 Schwarz SC 5.69F-statistic 134.65 Mean dependent −0.38Log likelihood −2377.6 S.D. dependent 5.25

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.Notes: Table 8 shows the results of a two-lag vector auto regression on the unmodeled inventory

evolution variable, QQit, and incoming order flow, Qjt (t-statistics in parentheses).

–4

–3

–2

–1

0

1

2

3

4

5

1 2 3 4 5 6 7 8 9 10

Response of QJT to QQ

–6

–4

–2

0

2

4

6

8

1 2 3 4 5 6 7 8 9 10

Response of QQ to QJT

(Response to Cholesky one S.D. innovations ± 2 S.E.)

Source: Lyons (1995) data: New York–based dollar/DM dealer, August 3–7, 1992.Notes: Figure 5 shows the responses of the unmodeled inventory evolution variable, QQ, and

incoming order flow, Qjt , to a one-standard-deviation shock in the other respective variable.

Figure 5. Impulse Responses for QQit and Qjt


Rafael B. Romeu

384

inventories change not just by incoming orders, but also by outgoing and customerorders. That is, ML dealers (for example, Lyons, 1997—Box (3) in Figure 1)receive incoming orders but also initiate orders with other dealers and trade withcustomers. Hence, these models allow a role for customers and outgoing orders inprice determination. DL models where a dealer’s position is governed by equation(4) only receive incoming orders. They do not incorporate these other tradingvenues into the dealer’s price-setting optimization.19

III. Conclusion

This paper considers the empirical viability of (partial equilibrium) dealer-levelmicrostructure models. It presents new empirical results that reject the specifica-tions of such models. The DL model of currency dealer price setting is found tocontain structural breaks when estimated on a one-week sample of currency trad-ing. In the two relevant subsample estimations, asymmetric information effects arerejected in one, and inventory effects are reflected in the other. That is, they do notoccur simultaneously, as the model would predict. This rejection of the DL modelis consistent with other empirical studies (see footnote 7).

Future work may investigate whether the consistent rejection of dealer-levelmodels stems from assumptions limiting the sources of inventory changes. In therejected dealer models, inventory is assumed to evolve only through incoming pur-chases or sales. This implies that price setting is crucial for controlling inventory.This study suggests, however, that inventory evolution may also depend on otherfactors beyond incoming orders. In particular, evidence is presented of an unex-plained component of inventory evolution that is correlated with incoming ordersand is of similar magnitude. Evidence of causality running in both directionsbetween this unexplained inventory component and incoming orders is presented.Taken together, these suggest that this component may be a good candidate forwhere dealer-level modeling should go next. Furthermore, including this unex-plained component may allow the inclusion of assumptions that condition dealerprices on incoming, outgoing, and customer orders, as in ML models.

BIBLIOGRAPHY

Admati, A., and P. Pfleiderer, 1988, “A Theory of Intraday Patterns: Volume and PriceVariability,” Review of Financial Studies, Vol. 1 (Spring), pp. 3–40.

Amihud, Y., and H. Mendelson, 1980, “Dealership Market: Marketmaking with Inventory,”Journal of Financial Economics, Vol. 8 (March), pp. 31–53.

Andrews, D. W. K., 1991, “Heteroskedasticity and Autocorrelation Consistent CovarianceMatrix Estimation,” Econometrica, Vol. 59 (November), pp. 817–58.

———, 1993, “Testing for Parameter Instability and Structural Change with Unknown ChangePoint,” Econometrica, Vol. 61 (July), pp. 821–56.

19Lyons (1995) controls empirically for outgoing orders and finds that these do not bias the effectsreported in Table 1; however, the underlying pricing relation in that model is rejected here.



385

Bai, J., and P. Perron, 1998, “Estimating and Testing Linear Models with Multiple StructuralChanges,” Econometrica, Vol. 66 (January), pp. 47–78.

Bhattacharya, U., and P. Weller, 1997, “The Advantage to Hiding One’s Hand: Speculation andCentral Bank Intervention in the Foreign Exchange Market,” Journal of MonetaryEconomics, Vol. 39, (July), pp. 251–77.

Bjønnes, G. H., and D. Rime, 2005, “Dealer Behavior and Trading Systems in the ForeignExchange Market,” Journal of Financial Economics, Vol. 75 (March), pp. 571–605.

Cao, H. H., M. D. D. Evans, and R. K. Lyons, 2005, “Inventory Information,” Journal ofBusiness (forthcoming).

Dominguez, K. M. E., 2003, “When Do Central Bank Interventions Influence Intra-Daily andLonger-Term Exchange Rate Movements?” NBER Working Paper No. 9875 (Cambridge,Massachusetts: National Bureau of Economic Research).

Easley, D., and M. O’Hara, 1987, “Price, Trade Size, and Information in Securities Markets,”Journal of Financial Economics, Vol. 19 (September), pp. 69–90.

———, 1992, “Time and the Process of Security Price Adjustment,” Journal of Finance, Vol. 47(June), pp. 577–605.

Evans, M. D. D., 2002, “FX Trading and Exchange Rate Dynamics,” Journal of Finance, Vol. 57(December), pp. 2405–47.

———, and R. K. Lyons, 2001, “Portfolio Balance, Price Impact, and Secret Intervention,”NBER Working Paper No. 8356 (Cambridge, Massachusetts: National Bureau of EconomicResearch).

———, 2002, “Order Flow and Exchange Rate Dynamics,” Journal of Political Economy,Vol. 110 (February), pp. 170–80.

———, 2004, “A New Micro Model of Exchange Rate Dynamics,” NBER Working PaperNo. 10379 (Cambridge, Massachusetts: National Bureau of Economic Research).

Flood, R., and M. Taylor, 1996, “Exchange Rate Economics: What’s Wrong with theConventional Macro Approach?” The Microstructure of Foreign Exchange Markets, ed.by J. Frankel, G. Galli, and A. Giovannini (Chicago: University of Chicago Press),pp. 261–94.

Frankel, J., and A. Rose, 1995, “Empirical Research on Nominal Exchange Rates,” Handbook ofInternational Economics, Vol. 3, ed. by G. Grossman and K. Rogoff (Amsterdam: ElsevierScience), pp. 1689–729.

Glosten, L., and P. Milgrom, 1985, “Bid, Ask, and Transaction Prices in a Specialist Marketwith Heterogeneously Informed Agents,” Journal of Financial Economics, Vol. 14 (March),pp. 71–100.

Hasbrouck, J., 1988, “Trades, Quotes, Inventories, and Information,” Journal of FinancialEconomics, Vol. 22 (December), pp. 229–52.

———, 1991a, “Measuring the Information Content of Stock Trades,” Journal of Finance,Vol. 46 (March), pp. 179–207.

———, 1991b, “The Summary Informativeness of Stock Trades: An Econometric Analysis,”Review of Financial Studies, Vol. 4, No. 3, pp. 571–95.

Hasbrouck, J., and G. Sofianos, 1993, “The Trades of Market Makers: An EmpiricalExamination of NYSE Specialists,” Journal of Finance Vol. 48, pp. 1565–93.

Ho, T., and H. Stoll, 1981, “Optimal Dealer Pricing Under Transactions and Return Uncertainty,”Journal of Financial Economics, Vol. 9 (March), pp. 47–73.

———, 1983, “The Dynamics of Dealer Markets Under Competition,” Journal of Finance,Vol. 38 (September), pp. 1053–74.


Rafael B. Romeu

386

Ito, T., R. K. Lyons, and M. Melvin, 1998, “Is There Private Information in the FX Market? TheTokyo Experiment,” Journal of Finance, Vol. 53 (June), pp. 1111–30.

Kyle, A., 1985, “Continuous Auctions and Insider Trading,” Econometrica, Vol. 53 (November),pp. 1315–35.

Linton, Clifton, 1992, “Dollar Likely to Fall Though Banks May Try to Slow Currency’sDescent,” Wall Street Journal (New York), August 10, p. C13.

Lyons, R., 1995, “Tests of Microstructural Hypotheses in the Foreign Exchange Market,”Journal of Financial Economics, Vol. 39 (October), pp. 321–51.

———, 1996, “Optimal Transparency in a Dealer Market with an Application to ForeignExchange,” Journal of Financial Intermediation, Vol. 5 (July), pp. 225–54.

———, 1997, “A Simultaneous Trade Model of the Foreign Exchange Hot Potato,” Journal ofInternational Economics, Vol. 42 (May), pp. 275–98.

———, 1998, “Profits and Position Control: A Week of FX Dealing,” Journal of InternationalMoney and Finance, Vol. 17 (February), pp. 97–115.

———, 2001, The Microstructure Approach to Exchange Rates (Cambridge, Massachusetts;and London: MIT Press).

Madhavan, A., and G. Sofianos, 1998, “An Empirical Analysis of NYSE Specialist Trading,”Journal of Financial Economics, Vol. 48 (May), pp. 189–210.

Madhavan, A., M. Richardson, and M. Roomans, 1997, “Why Do Security Prices Change? A Transaction-level Analysis of NYSE Stocks,” Review of Financial Studies, Vol. 10,pp. 1035–64.

Madhavan, A., and S. Smidt, 1991, “A Bayesian Model of Intraday Specialist Pricing,” Journalof Financial Economics, Vol. 30 (November), pp. 99–134.

———, 1993, “An Analysis of Daily Changes in Specialist Inventories and Quotations,” Journalof Finance, Vol. 48 (December), pp. 1595–648.

Manaster, S., and S. C. Mann, 1996, “Life in the Pits: Competitive Market Making and InventoryControl,” Review of Financial Studies, Vol. 9 (Fall), pp. 953–75.

Meese, R., and K. Rogoff, 1983, “The Out-of-Sample Failure of Empirical Exchange RateModels,” Exchange Rate and International Macro-economics, ed. by J. Frenkel (Chicago:University of Chicago Press).

O’Hara, M., 1995, Market Microstructure Theory (Malden, Massachusetts: Blackwell PublishersInc.).

———, and G. S. Oldfield, 1986, “The Microeconomics of Market Making,” Journal ofFinancial and Quantitative Analysis, Vol. 21 (December), pp. 361–76.

Romeu, R. B., 2003, “An Intraday Pricing Model of Foreign Exchange Markets” IMF WorkingPaper 03/115 (Washington: International Monetary Fund).

Sarno, L., and M. Taylor, 2002, “Purchasing Power Parity and the Real Exchange Rate,” IMFStaff Papers, Vol. 49 (April), pp. 65–105.

Stoll, Hans R., 1978, “The Supply of Dealer Services in Securities Markets,” Journal of Finance,Vol. 33 (September), pp. 1133–51.

“U.S. Fed Intervenes as Dollar Nears Low Against D-Mark,” Financial Times (London), August8, 1992, p.1.

Vitale, P., 1999, “Sterilized Central Bank Intervention in the Foreign Exchange Market,”Journal of International Economics, Vol. 49 (December), pp. 245–67.

Yao, J., 1998, “Market Making in the Interbank Foreign Exchange Market,” Salomon CenterWorking Paper No. S-98-3 (unpublished; New York: New York University).


387



Real Exchange Rates in Developing Countries: Are Balassa-Samuelson Effects Present?

EHSAN U. CHOUDHRI AND MOHSIN S. KHAN*

There is surprisingly little empirical research on whether Balassa-Samuelsoneffects can explain the long-run behavior of real exchange rates in developingcountries. This paper presents new evidence on this issue based on a panel-datasample of 16 developing countries. The paper finds that the traded-nontraded pro-ductivity differential is a significant determinant of the relative price of nontradedgoods, and the relative price in turn exerts a significant effect on the real exchangerate. The terms of trade also influence the real exchange rate. These results pro-vide strong verification of Balassa-Samuelson effects for developing countries[JEL F31, F41]

The well-known analyses of Balassa (1964) and Samuelson (1964) provide anappealing explanation of the long-run behavior of the real exchange rate in

terms of the productivity performance of traded relative to nontraded goods. Basi-cally, the argument is that as the productivity of traded goods rises relative to thatof nontraded goods, there will be a tendency for the real exchange rate to appreci-ate. Balassa-Samuelson effects are generally thought to be the key source ofobserved cross-sectional differences in real exchange rates (i.e., the same currencyprices of comparable commodity baskets) between countries at different levels of

*Ehsan U. Choudhri is Chancellor’s Professor at Carleton University in Canada. Mohsin S. Khan isDirector of the Middle East and Central Asia Department at the IMF. The authors would like to thankRobert Flood, Aasim Husain, Jean Le Dem, Gene Leon, Gian Maria Milesi-Ferretti, Nkunde Mwase, SamOuliaris, Miguel Savastano, and anonymous referees for helpful comments and suggestions, and MandanaDehghanian and Tala Khartabil for excellent research assistance.


Ehsan U. Choudhri and Mohsin S. Khan

388

income per capita.1 There is considerable empirical research on Balassa-Samuelsoneffects based on time-series data, but this research has been confined to industrialcountries.2 The time-series evidence on the working of the Balassa-Samuelsonmechanism for developing countries has been largely unexplored.3 One reason forthis neglect is that sectoral price and productivity data are not readily available fordeveloping countries. To address this problem, this paper makes use of recentlyavailable data from a number of sources to assemble a suitable data set for devel-oping countries, which is used to obtain new time-series evidence on the operationof Balassa-Samuelson effects in these countries.

Our data set includes time-series data from 1976 to 1994 for 16 countries.4The behavior of the dollar real exchange rate for each country during this periodis shown in Figure 1. The figure also displays the long-run component of the realexchange rate series based on the Hodrick-Prescott filter. As the figure shows, thelong-run component registers large changes over the sample period for a numberof countries. It is, thus, interesting to examine whether Balassa-Samuelson effectshave played an important role in causing these long-term movements. For manycountries, the figure also exhibits large fluctuations around the long-term trend.Some of these movements represent currency crises in response to speculativeattacks. Our empirical analysis attempts to control for the effect of short-rundynamics in order to identify long-run Balassa-Samuelson effects.

Balassa-Samuelson effects can be embedded in a variety of models. Theseeffects are typically derived within a static model, but they can be easily incorpo-rated in the dynamic framework of the new open economy macroeconomic mod-els.5 Using a framework compatible with the new open economy macroeconomicapproach, this paper derives two steady-state relations that capture key channels ofthe Balassa-Samuelson mechanism. The first relation links the real exchange rateto relative prices of nontraded goods at home and abroad. Under certain condi-tions, this relation includes the terms of trade as an additional determinant of thereal exchange rate.6 The second relation explains the relative price of nontraded

1For a review of the evidence and a discussion of alternative explanations, see Edwards and Savastano(1999). See also Bergin, Glick, and Taylor (2004), who point out that although recent data reveal a strongassociation between national price levels and income per capita, this association disappears in historicaldata going back 50 years or more.

2See, for example, Canzoneri, Cumby, and Diba (1999), and Lane and Milesi-Ferretti (2002).3See, however, Ito, Isard, and Symansky (1997), who use time-series data to explore the Balassa-

Samuelson hypothesis for Asia-Pacific Economic Cooperation (APEC) economies that include somedeveloping countries.

4This set includes 14 countries at low- and medium-income levels and 2 high-income economies(Republic of Korea and Singapore) that had lower income levels at the beginning of the sample period.

5These models tend to focus on the short- to medium-term dynamics arising from nominal rigiditiesand have not paid much attention to long-run Balassa-Samuelson influences. Benigno and Thoenissen(2003), however, do use a new open economy macroeconomic model to explore the effect of a productiv-ity improvement in the traded-goods sector on the United Kingdom real exchange rate.

6The relation assumes that the law of one price holds for each traded good in the long run. The realexchange rate for the traded-goods basket, however, need not be stationary and could influence the rela-tion if weights for individual traded goods differ between the home and foreign countries. Our empiricalprocedure accounts for this possibility.


REAL EXCHANGE RATES IN DEVELOPING COUNTRIES

389

Cameroon

0.0016

0.0020

0.0024

0.0028

0.0032

0.0036

78 80 82 84 86 88 90 92 94

Chile

0.0016

0.0020

0.0024

0.0028

0.0032

0.0036

0.0040

0.0044

0.0048

78 80 82 84 86 88 90 92 94

Colombia

0.0008

0.0012

0.0016

0.0020

78 80 82 84 86 88 90 92 94

Ecuador

0.0003

0.0004

0.0005

0.0006

0.0007

0.0008

0.0009

78 80 82 84 86 88 90 92 94

India

0.03

0.04

0.05

0.06

0.07

78 80 82 84 86 88 90 92 94

Jordan

1.2

1.4

1.6

1.8

2.0

2.2

2.4

2.6

78 80 82 84 86 88 90 92 94

Kenya

0.014

0.016

0.018

0.020

0.022

0.024

0.026

0.028

78 80 82 84 86 88 90 92 94

Korea

0.0009

0.0010

0.0011

0.0012

0.0013

0.0014

1976 1976 78 80 82 84 86 88 90 92 94

Real dollar exchange rate (1994 CPI = 100 for all countries)Long-term component (based on Hodrick-Prescott filter)

Source: See Appendix II.

1976 1976

1976 1976

1976 1976

Figure 1. Selected Developing Countries: Real Exchange Rate Behavior, 1976–94



390

Malaysia

0.32

0.36

0.40

0.44

0.48

0.52

0.56

78 80 82 84 86 88 90 92 94

Mexico

0.16

0.20

0.24

0.28

0.32

0.36

78 80 82 84 86 88 90 92 94

Morocco

0.08

0.10

0.12

0.14

0.16

0.18

0.20

78 80 82 84 86 88 90 92 94

Philippines

0.024

0.028

0.032

0.036

0.040

0.044

0.048

0.052

78 80 82 84 86 88 90 92 94

South Africa

0.16

0.20

0.24

0.28

0.32

0.36

0.40

78 80 82 84 86 88 90 92 94

Singapore

0.50

0.52

0.54

0.56

0.58

0.60

0.62

0.64

0.66

78 80 82 84 86 88 90 92 94

Turkey

0.00003

0.00004

0.00005

0.00006

0.00007

0.00008

78 80 82 84 86 88 90 92 94

Venezuela

0.004

0.006

0.008

0.010

0.012

0.014

0.016

78 80 82 84 86 88 90 92 94

Real dollar exchange rate (1994 CPI = 100 for all countries)Long-term component (based on Hodrick-Prescott filter)

Source: See Appendix II.

1976 1976

1976 1976

1976 1976

1976 1976

Figure 1. (Concluded)



391

goods. Following Canzoneri, Cumby, and Diba (1999), we use restrictions on pro-duction technology to derive a simple form of the relation, which makes the laborproductivity differential between traded and nontraded goods the main determi-nant of the relative price of nontraded goods. The technology restriction used toobtain the second relation is not needed to derive the first relation.

An important limitation of the use of labor productivity to represent long-termchanges in technology is that the long-run value of this variable can also be affectedby permanent shifts in demand.7 This problem may not be too serious if technologyshocks are the key source of permanent shocks affecting labor productivity. Testsof the Balassa-Samuelson hypothesis are typically based on a single relation relat-ing the real exchange rate directly to the productivity differential. Such a relationcan be derived by combining our two relations. However, separate estimation of thetwo relations provides additional tests of the Balassa-Samuelson model and is use-ful in identifying the sources of departures from this model.

As the time series for individual countries in our sample are not very long, wepool these series across countries to estimate our relations. Recent panel-dataeconometric techniques are used to identify long-run effects in these relations. Theresults provide strong evidence that the Balassa-Samuelson mechanism operatesin developing countries. Using the United States as the reference country, we findthat U.S.–developing country differences in the relative price of nontraded goodsand the terms of trade are significant determinants of the real exchange rate in thelong run. The differences in the labor productivity differential, moreover, exert asignificant long-run effect on the relative-price differences. One puzzling result isthat the estimated effect of the relative-price variable is greater and that of thelabor productivity variables smaller than the predicted value. We suggest explana-tions based on data problems to account for these discrepancies between estimatedand predicted values.

I. Theoretical Framework

This section outlines a framework to provide theoretical underpinnings for ourempirical analysis. As we are concerned with long-term effects, we do not modelshort-run dynamics but focus on steady-state relations under complete adjustmentof wages and prices. We consider a multicountry framework, with each countryusing fixed endowments of labor and capital to produce traded and nontraded goodsunder perfect competition.8 We focus on two special models of the pattern oftraded-goods production. The first model follows the standard Balassa-Samuelsonformulation and assumes that each country is diversified and produces all tradedgoods. The second model assumes that each country is specialized in the productionof a country-specific traded good, as in Armington’s (1969) model. We discuss

7One way to deal with this problem is to use an index of total factor productivity instead of labor pro-ductivity. Data constraints for developing countries, however, prevent us from using this approach.

8Our framework can be readily extended to incorporate monopolistic competition. As such an exten-sion would make little difference to the long-run relations derived in the paper, we assume perfect com-petition for simplicity.



392

below only the part of the model that is needed to derive the relations used in ourempirical analysis.

Basic Setup

Households in country i supply a fixed amount of labor and maximize the follow-ing expected lifetime utility:

where δ is the discount factor, and Ciτ represents a consumption index for periodτ. The consumption index is defined as

where CTi and CN

i are the subindices for consumption bundles of traded and non-traded goods, γi is the share of traded goods in aggregate consumption, and timesubscripts are dropped for simplicity. The traded-goods basket is also assumed tobe a Cobb-Douglas index of m (> 1) goods:

where CiTj is the amount consumed of traded good j, and θ j

i represents the shareof the good in the basket.

Let Pi denote the consumer price index, and P iT and P i

N the price indices fortraded and nontraded goods. Using equations (1) and (2), we define Pi and Pi

T asthe cost-minimizing prices of Ci and Ci

T, which are given by

The pattern of production for traded goods is characterized by either diversi-fication (with each country producing all traded goods) or specialization (witheach country producing a different traded good). In the case of specialization, weuse the same index for a country and its traded good (i.e., good i is produced bycountry i). Letting Yi

N and YiTj denote outputs of the nontraded and jth traded good,

we assume the following Cobb-Douglas production function for these goods:9

Y A K LiN

iN

iN

iNN N= ( ) ( )α β

, ( )5

P PiT

iTj

j

m ij

= ( )=∏ θ

14. ( )

P P Pi iT

iNi i= ( ) ( ) −γ γ1

3, ( )

C Ci

ij

TiTj

ij

j

m= ( )⎡⎣⎢

⎤⎦⎥=∏ θ

θ

12, ( )

C C Ci iT

iN

i ii i i i= ( ) ( ) −( )( )− −γ γ γ γγ γ

1 11 1, ( )

E U Ctt

it= ( )−

=

∞∑ δτττ ,

9The Cobb-Douglas form of the production function is used below to derive a simple relation betweenthe relative price of nontraded goods and the labor productivity differential. Canzoneri, Cumby, and Diba(1999) discuss more general production conditions, which would also imply such a relation.



393

where K iN and L i

N represent the amounts of capital and labor used in the produc-tion of the nontraded good, while Ki

Tj and LiTj are the corresponding amounts for

the traded good j. If there is specialization, KiTj = Li

Tj = 0 for i ≠ j.Let country 1 be the reference country, and define Si as the exchange rate of

country i (expressed as the price of country i’s currency) with respect to country 1.We distinguish between the short and long run in the present model. The short runis characterized by nominal rigidities in the form of sticky wages and prices. Thelong run, on the other hand, represents steady-state equilibrium with full adjust-ment of wages and prices. In the short run, nominal rigidities can cause departuresfrom the law of one price and the marginal productivity condition for labor. Weassume below that there are no departures from these relations in steady state. Wefocus on the steady-state behavior of variables to derive Balassa-Samuelson effects.A tilde over a variable is used to denote the steady-state value of the variable.

Assuming that the law of one price holds in steady state, we can link steady-state prices of traded goods in different countries as follows:

Also, assume that the marginal productivity condition is satisfied in steady state.Thus, letting Wi denote the wage rate, and using equations (5) and (6), we have

where the second equality in equation (8) holds only for traded good i under specialization.

Key Relations

We now derive key relations in the log-linear form. Using lowercase letters todenote values in logs, we define the consumption-based log real exchange rate as

Next, we use equation (3) to decompose the log real exchange rate as

where qiT ≡ si + pi

T − p1T is the log real exchange rate for traded goods. Using equa-

tion (4), we can express this variable as

q s p piT

ij

i iTj j Tj

j

m= +( ) −⎡⎣ ⎤⎦=∑ θ θ1 1111. ( )

q q p p p pi iT

i iN

iT N T= + −( ) −( ) − −( ) −( )1 1 101 1 1γ γ , ( )

q s p pi i i≡ + − 1 9. ( )

� � � � � � �W Y L P Y L Pi N iN

iN

iN

j iTj

iTj

iTj= ( ) = ( )β β , (8))

� � �S P Pi iT Tj j= 1 7. ( )

Y A K LiTj

iTj

iTj

iTjj j= ( ) ( )α β

, ( )6



394

The traded-goods price in logs can be linked to export and import price indices as

where piX and pi

M are the price indices for goods for which country i is, respectively,a net exporter and net importer, and θ i

X is the share of the export good in the traded-goods bundle.10 Note that in the specialization case, pi

X = piTi and θ i

X = θii.

Let rpi denote the log relative price of nontraded goods to domestically pro-duced traded goods. In the diversification case, rpi = pi

N − piT, since all traded

goods are produced domestically. Thus, for this case, equation (7) and the steady-state versions of equations (10) and (11) imply the following long-run relation forthe real exchange rate:

The Balassa-Samuelson analysis is often simplified by the assumption that expen-diture shares are the same everywhere. In this simple case, θ j

i = θ j1 for all j, γi = γ1,

and equation (13) can be expressed simply as qi = (1 − γ1)(rpi − r p1).In the case of specialization, rpi = pi

N − piTi, since only traded good i is

produced in country i. Using equation (12) and recalling that piTi = pi

X, we obtainrpi = pi

N − piT − (1 − θi

X)(piX − pi

M). Then, letting tti ≡ piX − pi

M denote the log termsof trade and using equation (7) along with equations (10) and (11) for steady state,we derive the following long-run relation for the specialization case:

Note that even if a country has the same expenditure shares as the reference coun-try, the terms of trade differential (tti − tt1) would affect the long-run real exchangerate in addition to the relative-price differential (rpi − rp1). This effect arisesbecause, in each country, the terms of trade influence the price of the traded-goodsbasket relative to that of the traded good produced at home.

The first term on the righthand side of equations (13) and (14) represents thelog real exchange rate for traded goods in steady state, q i

T. This term will not equalzero and may exhibit nonstationary behavior if the composition of a country’straded-goods basket differs from that of the reference country. In the case of het-erogeneous expenditure shares, q i

T represents an additional channel through whichthe terms of trade influence the real exchange rate, regardless of whether there isdiversification or specialization.11 In our empirical analysis based on panel data,

� � � �q p rp ri ij j Tj

j

mi i= −( ) + −( ) − −( )

=∑ θ θ γ γ1 11 11 1 pp

tt ttiX

i iX

1

1 1 11 1 1 1 14+ −( ) −( ) − −( ) −( )θ γ θ γ� � . ( ))

� � � �q p rp ri ij j Tj

j

m

i i= −( ) + −( ) − −( )=∑ θ θ γ γ1 11 11 1 pp1 13. ( )

p p piT

iX

iX

iX

iM= + −( )θ θ1 12, ( )

10Letting Ei and Ii represent sets of country i’s export and import goods, we define pXi ≡ θTj

i pTji /

θXi , θ X

i = θTji , j ∈ Ei, and pM

k ≡ θTki pTk

i /(1 − θXi ), k ∈ Ii.

11Although qTi = (θ j

i − θ j1 ) pTj

1 in equations (13) and (14), we can also relate it to the terms oftrade by using equation (12) to express: qT

i = si + pMi − pM

1 + θXi tti − θ X

1 tt1.j

m

=∑ 1

k∑j∑j∑



395

however, we do not link q iT to the terms of trade; instead, we use time effects to

control for variations in this variable.Next, the relative price of nontraded goods can be related to the productivity

differential between domestically produced traded and nontraded goods. We definethe log labor productivity in the two sectors as

where ω ij is the weight for good j’s labor productivity in the aggregate labor pro-

ductivity index for traded goods. In the specialization case, ω ij equals one for j = i

and zero otherwise. Let lpi ≡ lpiT − lpi

N denote the labor productivity differentialbetween traded and nontraded goods. In defining the diversification labor pro-ductivity index in steady state, we use the same weights as those in the priceindex for traded goods. Thus, let ω i

j = θij under diversification; and ωi

i = 1 for j = iand ω i

j = 0 for j ≠ i under specialization. Using equation (8) and steady-state ver-sions of equations (4), (15), and (16), we can express the steady-state relativeprice as

where ϑ equals in the case of diversification and logβi −logβN in the case of specialization.

II. Empirical Implementation

Data

We use a number of sources to put together a developing economies panel-data setthat includes time series from 1976 to 1994 for 16 countries.12 Traded goods areassumed to consist of manufacturing and agriculture sectors. Nontraded goodsrepresent all other sectors. The United States is chosen as the reference country.The real exchange rate is based on consumer price indices and represents the realvalue of a currency in terms of U.S. dollars.

Although our classification of the traded- and nontraded-goods sectors is sim-ilar to the one used for industrial countries, one potential problem is that a sub-stantial portion of the agriculture sector (and possibly of the manufacturing sector)in developing countries may consist of traditional activities producing nontradedgoods. Another problem is that the quality of labor is likely to vary considerablyacross sectors in developing countries, and our labor productivity measure (basedon employment figures unadjusted for quality changes) does not account for this

θ β βij

j

m

j N=∑ −1 log log

rp lpi i� �= +ϑ , ( )17

lp y liN

iN

iN≡ − , ( )16

lp y liT

ij

j

m

iTj

iTj≡ −( )=∑ ω

115, ( )

12Details of the variables and data sources are provided in Appendix II.



396

variation.13 We are unable to address these issues because of data limitations.However, we explore below certain implications of these measurement problemsfor the estimation of the empirical model.

Empirical Model

To undertake panel-data tests of the Balassa-Samuelson relations, we assume thatlong-run parameters are the same across our developing country set (D).14 Thus,we set θ i

X = θX and γi = γ for i ∈ D. However, to allow for possible differences inexpenditure shares between developing and industrial countries, we do not requireU.S. (country 1) parameters to be the same as those for our developing countrysample.

The following two equations are estimated to test for Balassa-Samuelsoneffects:

where rpdit = rpit − rp1t, ttdit = ttit − tt1t, and lpdit = lpit − lp1t are, respectively, thelog differences in the relative price of nontraded goods, the terms of trade, and thetraded-nontraded productivity ratio between developing country i and the UnitedStates; μi and ψi are country-specific fixed effects while κt and χt are common timeeffects; and uit and vit are error terms. Time effects represent the influence of com-mon time-specific (short- and long-run) factors, and error terms capture the effectsof short-term deviations from steady state (that are not included in time effects).

Equation (18) is derived from equations (13) and (14). Under our assumptionthat θ i

j = θ j for i ∈ D, time effects in equation (18) would control for movementsin q it

T(= (θ ij − θ1

j)p1Tj) arising from parametric differences between developing

countries and the United States. In the presence of time effects, equation (18)nests the diversification and specialization cases with τ = 0 under diversificationand τ = (1 − θX)(1 − γ) > 0 under specialization.15 In both cases, π = (1 − γ) > 0.

j

m

=∑ 1

rpd lpd v i Dit i t it it= + + + ∈ψ χ λ , , ( )19

q rpd ttd uit i t it it it= + + + +μ κ π τ , ( )18

13If intersector labor quality differences are not taken into account, the marginal productivity condi-tion equation (8) would not be satisfied and there would be departures from the relative price equation(19) based on this condition. Another limitation of the data on labor inputs is that employment measuresfor the manufacturing, agriculture, and other (nontraded-goods) sectors come from different sources, andare not fully comparable. Also, note that labor productivity for traded goods is simply measured as theratio of total output to total employment in the traded-goods sector. For the diversification case, thisindex does not fully conform to the theoretical index used in equation (17), since the implicit weightsfor individual traded goods in this index could differ from the weights used in the traded-goods priceindex.

14We later allow these parameters to vary between developing countries at different income levels.15In the estimation of equation (18), if time effects do not fully capture changes in qT

it because of dif-ferences in expenditure shares across countries, τ could also pick up the effect of the terms of trade via qT

itand could be positive even in the absence of specialization.



397

Equation (19) is based on equation (17). In this equation, λ = 1. The absence ofBalassa-Samuelson effects would imply that π = τ = λ = 0.16

Although the long-run parameters in equations (18) and (19)—π, τ, and λ—are constrained to be the same across developing countries, these relations allow theshort-run dynamics (reflected in the time-series behavior of the error terms) to bedifferent across countries. The explanatory variables—rpdit, ttdit, and lpdit—can bestationary, trend-stationary, or nonstationary. In the case of trend-stationary behav-ior, equations (18) and (19) can be modified to include a time trend. Coefficients oftime trends in the two relations would be homogeneous across countries anddepend on the long-run parameters.17 Note that if the explanatory variables are inte-grated or trend-stationary, then qit would also be integrated or trend-stationary. Inthis case, Balassa-Samuelson effects would cause permanent departures from thepurchasing power parity.

As discussed above, our measure for the traded-goods sector (i.e., agricultureplus manufacturing) may be too broad for developing countries and could includenontraded goods. As discussed in Appendix I, the measured relative price of non-traded goods in this case would understate the true relative price and bias therelative-price coefficient upward in equation (18). This measurement problemwould not lead to a systematic bias in the estimation of equation (19), since themeasured value of the traded-nontraded productivity differential would alsounderstate its true value. A more serious problem for estimating equation (19) isthat the labor productivity measure is not adjusted for quality variation. AppendixI also shows that the estimated effect of the measured labor productivity differen-tial would be biased downward if there is a positive association between the aver-age labor quality and the true labor productivity.

III. Results

Estimation

Before estimating equations (18) and (19), we examine whether the variables inthese relations contain a unit root or not. Table 1 shows the results of two tests of aunit root in panel data. In the first test (LL), based on Levin and Lin (1993), the nullhypothesis of a unit root is tested against the alternative of a homogeneous auto-regressive coefficient. The second test (IPS), based on Im, Pesaran, and Shin(2003), tests the unit root null against a more general alternative of a heterogeneousautoregressive coefficient. Both tests indicate that qit contains a unit root (with or

16Tests of Balassa-Samuelson effects could also be based on alternative versions of equations (18) and(19) that exclude U.S. variables—rp1t, tt1t, and lp1t—and are expressed as qit = μ*

i + κ*t + πrpit + τttit + u*

it,and rpit = ψ*

i + χ*t + λlpit + v*

it. However, we estimate relations in the form that includes U.S. variablesbecause this form allows us to explore whether U.S. variables exert an effect additional to their effect viarpdit, ttdit, and lpdit.

17Letting rpdit = g1t + rpd ′it, ttdit = g2t + ttd ′it, and lpdit = g3t + lpd ′it, we can restate equations (18) and(19) as follows: qit = μi + κt + (g1π + g2τ)t + πrpd ′it + τttd ′it + uit, and rpdit = ψi + χt + g3λt + λlpd ′it + vit.



398

without a time trend).18 For the remaining variables, the tests are sensitive towhether a time trend is included or not. In the absence of a trend, the unit roothypothesis is not rejected for rpdit and ttdit by both the LL and IPS tests, and forlpdit by the LL test. However, if a trend is present, both tests indicate that rpdit andlpdit are not integrated, and the IPS test indicates that ttdit is also not integrated.

We first consider the basic form of equations (18) and (19), which does notinclude a time trend. In this case, since there is indication of nonstationary behav-ior for variables in these relations, we also undertake tests for co-integration. Weuse two parametric tests, the panel t-test and the group t-test, suggested by Pedroni(1999). The panel t-test rejects the hypothesis that there is no co-integration for thevector (qit, rpdit), but does not reject this hypothesis for vectors (rpdit, lpdit) and (qit,rpdit, ttdit). The group t-test rejects the no-co-integration hypothesis for all threevectors.19 The group t-test (unlike the panel t-test) does not constrain the first-ordercorrelation in the residuals to be homogeneous under the alternative hypothesis andis more relevant for our model, which allows the short-run dynamics to vary acrosscountries. The test’s failure to reject the hypothesis of no co-integration for theabove vectors supports the Balassa-Samuelson model’s implication that a long-runrelation exists between the real exchange rate and relative prices (and possibly theterms of trade) as well as between relative prices and productivity ratios. We nextestimate Balassa-Samuelson effects in these relations.

We estimate equations (17) and (18) by Dynamic Ordinary Least Squares(DOLS), which is an appropriate framework for estimating and testing hypothesesfor homogeneous co-integrating vectors.20 The relations are estimated in the fol-lowing form:

Table 1. Unit Root Tests

Levin-Lin Test Statistic Im-Pesaran-Shin Test Statistic

Variable Without trend With trend Without trend With trend

qit 0.478 −1.008 −1.513 −1.480rpdit 0.231 −3.730** −0.358 −6.615**ttdit −0.070 −1.327 −0.388 −1.987*lpdit 0.604 −3.297** −2.059* −6.169**

Notes: qit is country i’s dollar real exchange rate in logs, while rpdit, ttdit, and lpdit represent,respectively, log differences in the relative price of nontraded goods, the terms of trade, and thetraded-nontraded labor productivity ratio between country i and the United States.

* indicates significance at the 5 percent level, and ** at the 1 percent level.

18Because of the assumption of homogeneous autoregressive coefficients, the LL test is encompassedby the IPS test. The results of the IPS test, however, are not conclusive. Although the test does not reject theunit-root hypothesis for qit at the 5 percent level, it does indicate rejection at slightly higher levels (p-value= 0.069 with trend and p-value = 0.065 without trend).

19For vectors (qit, rpdit), (rpdit, lpdit), and (qit, rpdit, ttdit), the panel-t test statistic is −1.730*, −1.093,and 0.278, respectively. The corresponding statistic for the group-t test is −2.074*, −1.955*, and −1.959.*An asterisk indicates significance at the 5 percent level.

20See Kao and Chiang (2000), and Mark and Sul (2002) for a discussion of the properties of panel DOLS.



399

where n is the number of lags and leads used for the first-difference terms. Coeffi-cients of these terms capture the short-run dynamics. We allow the short-rundynamics to be heterogeneous (i.e., let ξir, ζir, and ϕir differ across i). We test thenull hypotheses that π = τ = 0 in equation (20) and λ = 0 in equation (21) againstthe alternative hypotheses that these variables are positive.

If a linear trend is included, unit root tests suggest that the explanatory vari-ables in equations (18) and (19) are not integrated. We, thus, also consider thetrend-stationary setting for estimating these relations. DOLS is a useful estimatingprocedure even in this case. Since first-difference terms are included in this pro-cedure, the coefficients of level terms represent long-run effects. Therefore, weestimate equations (20) and (21) with trend variables to identify long-run Balassa-Samuelson influences in the trend-stationary case.

Basic Results

Tables 2 and 3 present DOLS estimates of different variants of the real exchangerate equation with one lag and one lead of the first-difference terms.21 Table 2shows the estimates of the equation for the diversification case excluding the termsof trade variable, and Table 3 for the specialization case including this variable.For both cases, we report the results for homogeneous as well as heterogeneousshort-run dynamics. Regressions 1 and 4 in these tables show estimates of thebasic form of the equation without a time trend. In all of these cases, the effect ofthe relative-price variable is positive and significant. The predicted value of thisvariable’s coefficient equals 1 − γ (which represents the share of the nontraded-goods sector). The estimated value, however, is greater than unity in most cases.The small size of our sample (based on only 19 years of data for each country) isa concern; it could be a source of bias in DOLS estimates. As discussed above,however, the discrepancy between the predicted and estimated values could reflectan upward bias arising from defining the traded-goods sector too broadly.22 Theresults also show that the terms of trade variable exerts a positive and significant

rpd lpd lpd vit i t it ir i,t r itr n

n= + + + + ′+=−∑ψ χ λ ϕ Δ , (( )21

q rpd ttd rpd ttit i t it it ir i,t r ir= + + + + ++μ κ π τ ξ ζΔ Δ dd ui,t r itr n

n+=− ( ) + ′∑ , ( )20

21The short length of each time series makes it difficult to explore the possibility that the short-rundynamics involve higher lags and leads. Indeed, there are not enough degrees of freedom to estimate equa-tion (20) with additional lags and leads in the case of heterogeneous dynamics. In the case of homoge-neous dynamics, however, we did estimate equations (20) and (21) with two lags and leads, and found littledifference in the results.

22The magnitude of the bias depends on the extent to which the share of the traded-goods sector isoverestimated. For our sample, the average share of manufacturing and agriculture in GDP is 35 percent.It is interesting to note that the true share of traded goods does not have to be much below this value toimply that the estimated coefficient of the relative price variable is greater than unity. For example, if about30 percent of manufacturing plus agriculture sectors in fact consist of nontraded goods, so that the actualshare of traded goods is 22.5 percent, then (as shown in Appendix I) the estimated coefficient of rpdit

would equal 1.12 (after setting φ = 0.3 and π = 0.775).


400

Table 2. The Exchange Rate Relation Without the Terms of Trade

Coefficient Estimates

Variable (1) (2) (3) (4) (5) (6)

Homogeneous short-run dynamics Heterogeneous short-run dynamics

rpdit 0.962** 0.962** 0.790** 1.066** 1.066** 0.846**(0.146) (0.146) (0.161) (0.156) (0.156) (0.173)

Trend 0.057 0.071(0.055) (0.060)

rpdit*D 0.329* 0.401*(0.129) (0.156)

Adjusted R2 0.997 0.997 0.997 0.997 0.997 0.997Standard error 0.154 0.154 0.152 0.160 0.160 0.158

of regression

Notes: The dependent variable is qit (see notes to Table 1 for the definitions of variables). Allregressions include country-specific and time-specific dummy variables as well as first differences ofeach explanatory variable at time t, t − 1, and t + 1. Coefficients of the first-difference terms are con-strained to be the same across countries under homogeneous dynamics, and unconstrained underheterogeneous dynamics. White heteroskedasticity-consistent errors are shown in parentheses. D is adummy variable, which equals one for low-income developing countries and zero for others. The num-ber of observations equals 256. * indicates significance at the 5 percent level, and ** at the 1 percentlevel (using a one-sided test for rpdit and a two-sided test for other variables).

Table 3. The Exchange Rate Relation with the Terms of Trade


Variable (1) (2) (3) (4) (5) (6)


rpdit 1.111** 1.111** 0.851** 1.217** 1.217** 0.834**(0.143) (0.143) (0.163) (0.204) (0.204) (0.251)

ttdit 0.300** 0.300** 0.477** 0.332** 0.332** 0.565**(0.091) (0.091) (0.103) (0.129) (0.129) (0.141)

Trend 0.063 0.111(0.054) (0.075)

rpdit*D 0.407** 0.601*(0.143) (0.271)

ttdit*D −0.348** −0.407(0.123) (0.209)


of regression

Notes: The dependent variable is qit (see notes to Table 1 for the definitions of variables). Allregressions include country-specific and time-specific dummy variables as well as first differences ofeach explanatory variable at time t, t − 1, and t + 1. Coefficients of the first-difference terms are con-strained to be the same across countries under homogeneous dynamics, and unconstrained underheterogeneous dynamics. White heteroskedasticity-consistent errors are shown in parentheses. D is adummy variable, which equals one for low-income developing countries and zero for others. The num-ber of observations equals 246. * indicates significance at the 5 percent level, and ** at the 1 percentlevel (using a one-sided test for lpdit and ttdit, and a two-sided test for other variables).



401

effect when introduced in the real exchange rate equation (see Table 3). This find-ing is consistent with the specialization version of the model, in which each coun-try produces a different good.

Table 4 shows the results for estimating the relative-price relation by DOLS.Regressions 1 and 4 in this table estimate the basic form of the relation without atime trend. The effect of the labor productivity index in both regressions is positiveand significant. But the estimated values of its coefficients in the two regressionsare substantially below the predicted value of unity. One possible explanation ofthis result, suggested above, is that measuring employment without adjustment forquality changes leads to a downward bias in the productivity coefficient.23 Otherlimitations of employment data and the small sample size could also have con-tributed to a bias in the estimates of the productivity coefficient.

Tables 2–4 also report the results for the trend-stationary case, in which ahomogeneous linear trend (with the same coefficient across countries) is includedin the two relations. The tables show (see regressions 2 and 4 in each table) that thecoefficient of the trend variable is insignificant in all cases, and the introduction ofthis variable in the regressions makes no difference to the estimates of Balassa-Samuelson parameters. We also introduced heterogeneous trends in the two rela-tions, but this variation made little difference to the results.

Further Analysis

Our empirical model includes time effects to allow the effect of U.S. variables tobe different from that of developing countries variables because of parametric dif-ferences. Time effects are, in fact, significant in both relations. Nevertheless, wealso estimated the two relations without time effects but did not find a substantialdifference in results. We further examined whether the results are sensitive to vari-ation in income levels across countries. To explore this question, we divided thedeveloping country sample into high- and low-income groups, and tested whethercoefficients of Balassa-Samuelson variables differ between the two groups.24

Regressions 3 and 6 in Tables 2–4 show the results of these tests. These regressionsinclude interactions between explanatory variables and a dummy variable for thelow-income group. Thus, coefficients of the variables show the effects for the high-income group, and interaction terms represent the additional effects for the low-income group. Interestingly, the results show that the effect of the relative-pricevariable (in the real exchange rate regressions) is significantly higher for the low-income group, while the effect of the labor productivity differential (in the relative-price regressions) is significantly lower. The departures from predicted values are,

23The downward bias arises because unobserved labor quality is assumed to be positively related totrue labor productivity. It is not clear, however, how much bias would be produced by this relation.According to Appendix I, the magnitude of the bias would depend on the elasticity of labor quality withrespect to true labor productivity (ρ). This elasticity would need to be 2.3 to generate, for example, an esti-mate of the productivity coefficient equal to 0.3.

24The classification of countries in the two groups is based on average income per capita for the sam-ple period. Each group includes eight countries (see Appendix II for the lists of countries).



402

thus, more pronounced for low-income countries. Since data problems are likely tobe more severe for the developing countries at the lower end of the income scale,this finding supports our suggested explanation that the estimates of Balassa-Samuelson effects are biased because of measurement errors. The results also indi-cate that the terms of trade effect is smaller for the low-income group.25

The conventional tests of Balassa-Samuelson effects are based on a singlerelation that links the real exchange rate directly to the labor productivity index.To derive such a relation, we combine equations (18) and (19) to obtain

where μ ′i = μi + πψi, κ′t = κt + πχt, and u′it = uit + πvit. For the purpose of compari-son with the existing literature, we also present results for the single-equation ver-sion of our two relations. Table 5 reports DOLS estimates of six variants ofequation (22), which are similar to those shown in Tables 2–4. Note that the esti-mates of the coefficients of the labor productivity and terms of trade variables inthe DOLS version of equation (22) need not fully conform to the estimates of these

q lpd ttd uit i t it it it= ′ + ′ + + + ′μ κ πλ τ , ( )22

25Thus, the support for the specialization version seems to be weaker for the poorer developing coun-tries. This result may seem paradoxical, as production and exports of low-income countries tend to be lessdiversified. However, specialization could also mean production of goods (e.g., sophisticated manufac-tured products) that are significantly differentiated from goods produced elsewhere. Poor countries may beless specialized in this sense.

Table 4. The Relative-Price Relation


Variable (1) (2) (3) (4) (5) (6)


lpdit 0.287** 0.287** 0.345** 0.302** 0.302** 0.397**(0.042) (0.042) (0.051) (0.048) (0.480) (0.062)

Trend 0.000 −0.004(0.028) (0.028)

lpdit*D −0.152* −0.229**(0.076) (0.086)


of regression

Notes: The dependent variable is rpdit (see notes to Table 1 for the definitions of variables). Allregressions include country-specific and time-specific dummy variables as well as first differences ofeach explanatory variable at time t, t − 1, and t + 1. Coefficients of the first-difference terms are con-strained to be the same across countries under homogeneous dynamics, and unconstrained underheterogeneous dynamics. White heteroskedasticity-consistent errors are shown in parentheses. D is adummy variable, which equals one for low-income developing countries and zero for others. The num-ber of observations equals 256. * indicates significance at the 5 percent level, and ** at the 1 percentlevel (using a one-sided test for lpdit and a two-sided test for other variables).



403

variables in equations (20) and (21) because of the use of different variables tocontrol for short-run dynamics.26 The results indicate that the labor productivitycoefficient in the single-equation version is significant in all cases, but its valuetends to be smaller than the product of the estimates of π and λ (obtained fromregressions of equations (20) and (21)). The terms of trade coefficient also differssomewhat from the estimate of τ based on equation (20) and is significant in allcases except regression (6) in the table. The effect of the two variables is no longersignificantly different between the high- and low-income groups. For the laborproductivity variable, this result (that its coefficient, πλ, does not differ betweenthe two income groups) is consistent with the earlier findings that π is higher andλ is lower for the low-income group.

During our sample period, currency crises involving large exchange rate depre-ciations occurred in a number of countries. Adverse economic conditions duringcrisis times could have caused comovements in exchange rates, labor productivity,and relative prices. This paper uses an estimation procedure that attempts to dis-entangle long-run Balassa-Samuelson effects from short-run correlations produced

26The DOLS version of equation (22) includes first differences of ttdit (which do not appear in equa-tion (21)) but does not include those of rpdit (which enter equation (20)).

Table 5. The Combined Exchange Rate Relation


Variable (1) (2) (3) (4) (5) (6)


lpdit 0.177* 0.177* 0.205** 0.212* 0.212* 0.302**(0.080) (0.080) (0.087) (0.109) (0.109) (0.124)

ttdit 0.357** 0.357** 0.388** 0.432** 0.432** 0.203(0.089) (0.089) (0.102) (0.135) (0.135) (0.184)

Trend −0.007 −0.014(0.047) (0.078)

lpdit*D −0.080 −0.157(0.169) (0.271)

ttdit*D −0.085 0.347(0.120) (0.224)


of regression

Notes: The dependent variable is qit (see notes to Table 1 for the definitions of variables). Allregressions include country-specific and time-specific dummy variables as well as first differences ofeach explanatory variable at time t, t − 1, and t + 1. Coefficients of the first-difference terms are con-strained to be the same across countries under homogeneous dynamics, and unconstrained under het-erogeneous dynamics. White heteroskedasticity-consistent errors are shown in parentheses. D is adummy variable, which equals one for low-income developing countries and zero for others. Thenumber of observations equals 246. * indicates significance at the 5 percent level, and ** at the 1 percent level (using a one-sided test for lpdit and ttdit, and a two-sided test for other variables).



404

by temporary shocks (such as those leading to currency crises). However, to addressthe concern that our method may not have adequately removed the influence of cri-sis shocks, we explore the sensitivity of our results to inclusion of crisis periods.To identify crisis periods, we follow Kaminsky, Reinhart, and Vegh (2004), whodefine a crisis year as a year in which there is a 25 percent or higher monthly depre-ciation that is at least 10 percent higher than the previous month’s depreciation.27

Using their crisis data, we reestimate our basic regressions, excluding the observa-tions for crisis years.28 Note that since our regressions include one lag and one leadof each explanatory variable’s first differences (which are not available for the yearof the crisis and the following year), the exclusion window for these regressions is generally four years for a single crisis.29 Longer periods are excluded for coun-tries with multiple crises. In fact, for three countries—Ecuador, Turkey, andVenezuela—there were not enough observations to estimate country-specificdynamics. These countries were, thus, excluded from regressions with hetero-geneous dynamics.

Table 6 presents the results of basic regressions based on data for crisis-freeperiods for both the two- and one-equation versions of the model (see columns 1–2and 4–5 of the table for the two-equation version and columns 3 and 6 for the one-equation version). As the table shows, the effect of the basic Balassa-Samuelsonvariables—the relative-price and labor productivity indices—remains robust evenafter excluding crisis periods. The effect of the labor productivity variable, in fact,becomes stronger. The terms of trade effect, however, becomes weaker and isinsignificant in most cases. Thus, the results on the influence of the terms of tradeon the real exchange rate are sensitive to whether crisis periods are included or not.Although our regressions generally exclude four years for a crisis, this period maynot be considered long enough to fully remove the effect of a crisis shock.30 To dealwith this concern, we explored additional variations that introduced longer exclu-sion windows or excluded all the data for countries that faced multiple crises withinthe sample period.31 These variations further reduced the sample size but still did

27See Frankel and Rose (1996) for a discussion of the usefulness of this measure of crisis for emerg-ing economies. For industrial countries, Eichengreen, Rose, and Wyplosz (1996) use an alternative mea-sure based on a weighted average of changes in the exchange rate, international reserves, and interest rates.This measure is designed to develop a crisis index that would include unsuccessful speculative attacks(which do not change the exchange rate but lead to a loss of international reserves and/or a rise in the inter-est rate). We need, however, to identify only successful attacks that could cause co-movements betweenthe exchange rate and other variables and potentially bias our results. Thus, international reserves andinterest rates may not be useful indicators for our purposes. For developing countries, moreover, interestrate data are generally lacking and international reserve changes are often an inadequate measure ofexchange market intervention.

28See Appendix II for a list of crisis years for our sample.29For example, if the crisis year is 1982, the period from 1981 to 1984 is excluded from the regression.

A shorter period would need to be excluded if the crisis occurs in the first or last two years of the sample.30Estimates of half-life for shocks to the real exchange rate, for example, typically range from three

to five years.31In the first variation, we also dropped the observations for one year before and one year after the cri-

sis year, which generally extended the regression exclusion window for a crisis to six years. Four countries—Mexico, Ecuador, Turkey, and Venezuela—experienced multiple crises. These countries were excluded fromthe sample in the second variation.



405

not much affect our results about the robustness of the effect of the labor produc-tivity and relative-price variables.

IV. Conclusions

The Balassa-Samuelson hypothesis would seem to be especially relevant for devel-oping countries where relative prices and productivities are likely to be more vari-able. Yet, there is little or no empirical evidence on whether Balassa-Samuelsoneffects can successfully explain long-run movements of the real exchange rate indeveloping countries. This paper presents new time-series evidence for developingcountries on the presence of Balassa-Samuelson effects. To test for these effects,we estimate two long-run relations: relative prices (of nontraded goods) affect thereal exchange rate in one relation, and labor productivity differentials (betweentraded and nontraded goods) affect relative prices in the second relation. Terms oftrade also affect the real exchange rate (in the first relation) under certain conditions.A key finding of this paper is that the labor productivity differential exerts a signif-icant effect on the real exchange rate via its influence on the relative price of non-traded goods.32 The paper also finds that terms of trade are a significant determinant

32Previous work (for example, Lane and Milesi-Ferretti, 2004), using GDP per capita as a proxy forthe labor productivity differential, has not found a systematic effect of the productivity variable on realexchange rates in developing countries. We believe that we are able to identify this effect by using a moreappropriate measure of labor productivity differential based on sectoral data.

Table 6. Basic Regressions, Excluding Crisis Years


Variable (1) (2) (3) (4) (5) (6)


lpdit 0.342** 0.247** 0.337** 0.240*(0.042) (0.093)* (0.044) (0.123)

ttdit 0.152 0.222* 0.130 0.161(0.098) (0.105) (0.196) (0.141)

rpdit 1.153** 1.099**(0.135) (0.192)

Adj. R2 0.861 0.998 0.997 0.877 0.997 0.997Standard error 0.069 0.132 0.150 0.065 0.144 0.140

of regressionNo. Obs. 215 205 205 215 190 190

Notes: The dependent variable is rpdit for regressions in columns (1) and (4), and qit for otherregressions (see notes to Table 1 for the definitions of variables). All regressions include country-specific and time-specific dummy variables as well as first differences of each explanatory variableat time t, t − 1, and t + 1. Coefficients of the first-difference terms are constrained to be the sameacross countries under homogeneous dynamics, and unconstrained under heterogeneous dynamics.White heteroskedasticity-consistent errors are shown in parentheses. * indicates significance at the5 percent level, and ** at the 1 percent level (using a one-sided test).



406

of the real exchange rate. This finding, however, is sensitive to whether the sam-ple includes crisis periods or not.

Although the effect of relative-price and labor productivity variables operatesin the direction indicated by the Balassa-Samuelson hypothesis, the effect of rela-tive prices is stronger and that of productivity differentials weaker than the predictedvalue. The paper also finds that the departures from predicted values are larger fordeveloping countries with lower income levels. We suggest an explanation thatattributes these results to biases caused by measurement problems. These problemsare likely to be more pronounced in countries with lower incomes and, thus, couldaccount for differences in estimated Balassa-Samuelson effects between countriesat low and high income levels.

Our tests of the Balassa-Samuelson explanation are based on two long-runrelations, which are derived from theory under fairly general conditions and can beimplemented empirically for developing countries. One important caveat for ourformulation is that labor productivity is used to capture the effect of permanenttechnology shocks emphasized by the Balassa-Samuelson theory. This measurecould also pick up the influence of permanent demand shocks. Disentangling theinfluence of permanent demand and technology shocks on long-run labor produc-tivity would be an interesting topic for future research. Further theoretical andempirical analysis could also extend the framework considered here and explorethe role of additional factors.33 Such analysis is beyond the scope of this paper.The results of this paper do suggest that the Balassa-Samuelson mechanism is anempirically useful framework for investigating the long-run behavior of the realexchange rate for developing countries.

APPENDIX I

Potential Biases Due to Measurement Problems

Traded-Goods Sector Measure Includes Nontraded Goods

Using a hat over a variable to denote the measured value, let the measured traded-goods pricebe pT

it = φpNit + (1 − φ)pT

it, 1 > φ > 0, where φ is the weight for the nontraded goods that areimproperly included in the traded-goods sector measure. The measured relative price of non-traded goods is then related to the true price as rp it = pN

it − pTit = (1 − φ)rpit. Let the corre-

sponding relation for country 1 be rp1t = (1 − φ1)rp1t, with 1 > φ1 ≥ 0. Using these relationsand letting rpdit = rp it − rp1t, we can express equation (18) in the text as

where κ′t = κt + π[1/(1 − φ) − 1/(1 − φ1)]rp1t and π′ = π/(1 − φ). Thus, if rpdit is used instead ofrpdit in equation (18), its coefficient would be biased upward.

Note that this problem need not introduce a systematic bias in equation (19). For example,if we also have lpT

it = φlpNit + (1 − φ)lpT

it, then lp it = lpTit − lpN

it = (1 − φ)lpit. Using this relation

q rpd ttd uit i t it it it= + ′ + ′ + +μ κ π τˆ ,

33For example, Lane and Milesi-Ferretti (2004) explore the theoretical link between the real exchangerate and net foreign assets, and provide evidence that the net foreign assets position is an important deter-minant of the real exchange rate for developing (as well as developed) countries.



407

and the corresponding one for country 1, we can show that the use of measured values in equa-tion (19) would not bias the estimate of the effect of labor productivity differential.

Measured Employment Not Adjusted for Labor Quality

Express the amount of effective labor in sector Z = T, N, as LZit = EZ

it LZit, where LZ

it is the actual(measured) quantity of labor and EZ

it is the average quality or efficiency of labor. Themeasured labor productivity is related to the true productivity (in logs) as lpZ

it = yZit − l Z

it = lpZit

+ eZit. Suppose that efficiency is positively correlated with true labor productivity. Assume that

this relation takes the simple form eZit = ρlpZ

it, ρ > 0. Recalling that lpit = lpTit − lpN

it, it followsthat lpit = lp it/ (1 + ρ). Let lp1t = lp1t/(1 + ρ1), ρ1 ≥ 0, be the corresponding relation for coun-try 1. Using these relations and letting l pdit = l p it − lp1t, we can express equation (19) in thetext as

where χ′t = χt + λ[1/(1 + ρ) − 1/(1 + ρ1)]l p1t and λ′ = λ /(1 + ρ). Thus, the use of l pdit insteadof lpdit in the text equation (19) would bias the effect of the productivity variable downward.

APPENDIX II

Data Appendix

The data set consists of a number of annual time series for 16 developing countries and theUnited States. All series cover the time period 1976–94. The selection of developing countriesand the choice of the time period are dictated by the availability of data.

Definitions and Data Sources

The U.S. dollar exchange rate (S) and the consumer price index (P) are from IMF InternationalFinancial Statistics (IFS). The export and import price indices (PX, PM) represent the price/unit-value series from IFS or, if IFS data are not available, export and import price defla-tors from the IMF World Economic Outlook database. These indices are used to calculate theterms of trade. The terms of trade data are not available for Singapore for the years 1976–78 andfor Turkey for the years 1985–88.

Measures of the labor productivity differential and the relative price of nontraded goods arebased on sectoral data on output, employment, and prices. Traded goods are represented by man-ufacturing and agriculture sectors, and nontraded goods by all other sectors. Value added in con-stant local currency units is used to measure outputs of traded- and nontraded-goods sectors (YT,YN). Labor inputs in the two sectors (LT, LN) represent the number of persons employed in eachsector. Price indexes for traded and nontraded goods (PT, PN) are price deflators derived fromvalue-added data in current and constant local currency units. For the United States, all of theseseries are from the Organisation for Economic Co-operation and Development (OECD) Struc-tural Analysis (STAN) database. For developing countries, the series, YT, YN, PT, and PN arefrom World Bank World Development Indicators (WDI). The price deflator for services and soon, which accounts for the bulk of the nontraded-goods sector, is used to estimate PN. The dataon total employment in manufacturing are from the World Bank Trade and Production database.34

A short gap in these data for Cameroon was filled by linear interpolation. Employment in agri-

rpd lpdit i t it it= + ′ + ′ +ψ χ λ νˆ ,

34See Nicita and Olarrega (2001) for a description of this database.



408

culture is derived from value added per worker and total value-added series given in WDI. LT isdefined as the sum of employment in manufacturing and agriculture obtained from the abovesources. LN is measured residually as the difference between total labor force (also from WDI)and LT. A limitation of the employment data is that employment in agriculture, manufacturing,and other (nontraded-goods) sectors is not measured on a consistent basis. Labor productivitymeasures for traded- and nontraded-goods sectors equal YT/LT and YN/LN, respectively.

Income Groups

The 16 developing countries were divided into low- and high-income groups according to aver-age GDP per capita (from WDI) for the sample period. Low- (high-) income group representscountries with per capita income smaller (greater) than $2,000 in 1995 U.S. dollars. The coun-tries in each group are listed below.

Low-Income Group High-Income Group

Cameroon ChileColombia Republic of KoreaEcuador MalaysiaIndia MexicoJordan SingaporeKenya South AfricaMorocco TurkeyPhilippines Venezuela

Country Crisis Years

Cameroon 1994Chile 1985Ecuador 1982, 1985–86, 1988Mexico 1976, 1982, 1994Philippines 1984Turkey 1978–80, 1994Venezuela 1984, 1986, 1989, 1994

List of Crisis Years

According to the crisis data used in Kaminsky, Reinhart, and Vegh (2004), crisis occurred in thefollowing years for our sample countries from 1976 to 1994. (Their data set does not includeSingapore, but this country did not experience a crisis in this period according to their criterion.)

References

Armington, Paul S., 1969, “A Theory of Demand for Products Distinguished by Place ofProduction,” IMF Staff Papers, Vol. 16, pp. 159–78.

Balassa, Bela, 1964, “The Purchasing Power Parity Doctrine: A Reappraisal,” Journal ofPolitical Economy, Vol. 72, pp. 584–96.



409

Benigno, Gianluca, and Christoph Thoenissen, 2003, “Equilibrium Exchange Rates andSupply-Side Performance,” Economic Journal, Vol. 113, No. 486, pp. 103–24.

Bergin, Paul, Reuven Glick, and Alan M. Taylor, 2004, “Productivity, Tradability, and theLong-Run Price Puzzle,” NBER Working Paper No. 10569 (Cambridge, Massachusetts:National Bureau of Economic Research).

Canzoneri, Matthew B., Robert E. Cumby, and Behzad Diba, 1999, “Relative LaborProductivity and the Real Exchange Rate in the Long Run: Evidence for a Panel of OECDCountries,” Journal of International Economics, Vol. 47, No. 2, pp. 245–66.

Edwards, Sebastian, and Miguel A. Savastano, 1999, “Exchange Rates in EmergingEconomies: What Do We Know? What Do We Need to Know?” NBER Working Paper No.7228 (Cambridge, Massachusetts: National Bureau of Economic Research).

Eichengreen, Barry, Andrew Rose, and Charles Wyplosz, 1996, “Contagious Currency Crises:First Tests,” Scandinavian Journal of Economics, Vol. 98, pp. 463–84.

Frankel, Jeffrey A., and Andrew K. Rose, 1996, “Currency Crashes in Emerging Markets: AnEmpirical Treatment,” Board of Governors of the Federal Reserve System, InternationalFinance Discussion Paper No. 534 (Washington: Federal Reserve).

Im, Kyung So, M. Hashem Pesaran, and Yongcheol Shin, 2003, “Testing for Unit Roots inHeterogeneous Panels,” Journal of Econometrics, Vol. 115, No. 1, pp. 53–74.

Ito, Takatoshi, Peter Isard, and Steven Symansky, 1997, “Economic Growth and Real ExchangeRate: An Overview of the Balassa-Samuelson Hypothesis in Asia,” NBER Working PaperNo. 5979 (Cambridge, Massachusetts: National Bureau of Economic Research).

Kaminsky, Graciela L., Carmen M. Reinhart, and Carlos A. Vegh, 2004, “When It Rains, ItPours: Procyclical Capital Flows and Macroeconomic Policies,” NBER Working PaperNo. 10780 (Cambridge, Massachusetts: National Bureau of Economic Research).

Kao, Chihwa, and Min-Hsien Chiang, 2000, “On the Estimation of a Cointegrated Regressionin Panel Data,” Advances in Econometrics, Vol. 15, pp. 179–222.

Lane, Philip R., and Gian Maria Milesi-Ferretti, 2002, “External Wealth, the Trade Balance,and the Real Exchange Rate,” European Economic Review, Vol. 46, No. 6, pp. 1049–71.

———, 2004, “The Transfer Problem Revisited: Net Foreign Assets and Real ExchangeRates,” Review of Economics and Statistics, Vol. 86, No. 4, pp. 841–57.

Levin, Andrew, and Chien-Fu Lin, 1993, “Unit Root Tests in Panel Data: New Results,”University of California at San Diego, Discussion Paper No. 93–56 (San Diego: UCSD).

Mark, Nelson C., and Donggyu Sul, 2002, “Cointegration Vector Estimation by Panel DOLSand Long-Run Money Demand,” NBER Technical Working Paper No. 287 (Cambridge,Massachusetts: National Bureau of Economic Research).

Nicita, Alessandro, and Marcelo Olarrega, 2001, “Trade and Production, 1976–99,” WorldBank Policy Research Working Paper No. 2701 (Washington: World Bank).

Pedroni, Peter, 1999, “Critical Values for Cointegration Tests in Heterogeneous Panels withMultiple Regressors,” Oxford Bulletin of Economics and Statistics, Vol. 61, pp. 653–70.

Samuelson, Paul A., 1964, “Theoretical Notes on Trade Problems,” Review of Economics andStatistics, Vol. 46, pp. 145–54.


410

IMF Staff PapersVol. 52, Number 3© 2005 International Monetary Fund

The Internal Job Market of the IMF’s Economist Program

GREG BARRON AND FELIX VÁRDY*

This paper shows how the internal job market for participants in the IMF’s Econ-omist Program (EPs) could be redesigned to eliminate most of the shortcomingsof the current system. The new design is based on Gale and Shapley’s (1962)deferred acceptance algorithm and generates an efficient and stable outcome. AnExcel-based computer program, EP-Match, implements the algorithm and ap-plies it to the internal job market for EPs. The program can be downloaded fromhttp://www.people.hbs.edu/gbarron/EP-Match_ for_Excel.htm. [JEL C78, D73]

The Economist Program is the gate of entry into the IMF for young economistsjoining the Fund soon after graduate school. Each year, 35 to 45 participants,

usually referred to as “EPs,” join the program for a two-year period. During thistime, EPs work in two different IMF departments. After graduating from the Pro-gram, EPs are considered for permanent staff positions.

In the context of the Economist Program, each EP is matched to a departmenton three occasions. First, upon joining the IMF in June or October; then, upon trans-fer to a second-year position one year later; and, finally, when the EP is assigned toa permanent position at the end of the program.

Matching of EPs to second-year assignments and permanent positions takesplace through a decentralized, internal job market system. Matching of incoming

*Greg Barron is an Assistant Professor at Harvard Business School and Felix Várdy is a second-yearparticipant in the IMF’s Economist Program, currently in the Asian Division of the IMF Institute. Theauthors would like to thank David Coe for encouragement to write this paper, and Abdul Abiad, ChrisClarke, Andrew Feltenstein, Mohsin Khan, Jorge Marquez-Ruarte, Dayalini Mendez, Benoit Mercereau,Miguel Messmacher, Jacques Miniane, Saleh Nsouli, Al Roth, Martin Schindler, Sunil Sharma, ClintonShiells, Jay Surti, Marijn Verhoeven, Zhiwei Zhang, an anonymous referee, and the editor for valuablecomments and suggestions. All errors are our own.


http://www.people.hbs.edu/gbarron/EP-Match_ for_Excel.htm

THE INTERNAL JOB MARKET OF THE IMF’S ECONOMIST PROGRAM

411

EPs to first-year positions is done centrally by the IMF’s Human Resources Dep-artment (HRD).

This paper argues that the procedures currently in place exhibit structural short-comings. It proposes an alternative centralized matching mechanism, based on Galeand Shapley’s (1962) deferred acceptance algorithm (DAA), that resolves most ofthese shortcomings. Finally, it introduces EP-Match, an Excel-based computer pro-gram that implements the DAA and applies it to the internal job market for EPs.

I. The Current Situation

Decentralized Matching of Transferring and Graduating EPs

Transferring and graduating EPs are matched to second-year assignments and per-manent positions through a decentralized job market system. However, to promoteequality of opportunities, and to prevent unraveling of the market,1 departmentsand EPs are supposed to abide by strict rules as to when they can first contact eachother and when they can start making and accepting job offers.

In terms of Roth and Xing’s (1994) taxonomy of entry-level labor markets, theinternal job market for transferring and graduating EPs is in “Stage 2.” Here,“Stage 1” refers to an entirely unregulated job market, while “Stage 3” refers to afully centralized market. Their detailed description of the symptoms of a typicalStage 2 market is so recognizable that we quote it here at some length.

In stage 2, a . . . market organization . . . attempts to establish a uniformdate before which offers should not be made, and often an earlier datebefore which interviews should not be conducted, and a later date (or time)before which candidates who have received offers should not be requiredto respond. Sometimes this is hardly successful at all, with many marketparticipants ignoring or circumventing the rules, and those who obey themquickly finding that this puts them at a disadvantage. And even when uni-form dates are successfully established and maintained, the market oftenexperiences a great deal of congestion and chaotic behavior, as the dead-lines for accepting or rejecting offers grows near. A firm is eager to knowif its offers will be accepted in time, so that if it has unfilled positions it mayapproach its most preferred alternative candidates before they have had toaccept any offers they have received. And candidates who have receivedoffers, but not from their first choice firm, are intent upon waiting until thelast allowable moment before accepting any offer, in the hope of receivinga better one. Particularly if, as often seems to be the case, some fraction ofcandidates holds on to multiple offers as the final deadline approaches, thismeans that just before the deadline expires many transactions still remainto be made. Firms whose first choice candidates reject them may now

1By “unraveling” we mean the phenomenon, common in entry-level labor markets, of employers mak-ing job offers earlier from year to year. This is done in an effort to preempt the competition and recruit thebest candidates. Once unraveling has started, it is hard to stop and often spins out of control. In the marketfor federal appellate law clerks, for example, job offers have recently been made almost two years inadvance of employment (see Avery and others, 2001). The same holds for entry-level gastroenterology posi-tions for medical doctors. Interviews for these positions now take place even before students have had theopportunity to explore other subspecialties (see Niederle and Roth, 2003). In markets that experience unrav-eling, applicants typically receive “exploding” offers that must be accepted or rejected on short notice.


Greg Barron and Felix Várdy

412

find that their next dozen candidates have already accepted offers, and can-didates may receive preferred offers moments after making a verbalcommitment to accept an earlier offer. In some markets such verbal com-mitments are virtually always honored, and in others they are sometimesreneged on. In either event, in the aftermath, many firms and candidateshave just missed making connections they would have preferred. Theresult is that the following year witnesses a resurgence of strategic behav-iors designed to avoid being caught short at the end of the market. Oftennew rules are formulated to prohibit the more brazen of these, and newadaptations are made. While some markets have persisted for many sea-sons in this fashion, systems of formalized dates are often abandoned, withthe market either reverting to stage 1 [an unregulated market], or movingon to stage 3 [a centralized market].” (Roth and Xing, 1994, p. 996.)

From personal experience, one of the authors can attest that the internal job mar-ket for transferring and graduating EPs exhibits virtually all the symptoms de-scribed by Roth and Xing. Specifically,

• The IMF’s Human Resources Department tries to prevent unraveling by en-forcing rules on market timing; for example, timing of first contact, interview-ing, making, and accepting offers.

• Departments and EPs routinely try to circumvent these rules, while those par-ticipants who do obey the rules often end up with rather unfavorable outcomes.A prime example is the rule that EPs and departments should not contact eachother before a certain date. This rule is clearly unenforceable.

• When a department makes a job offer to an EP, it often pressures the EP toaccept or reject the offer on the spot, or on very short notice.

• Conversely, if EPs can get away with it, they often try to postpone a deci-sion on a particular offer until the very last moment, in the hope of receiv-ing a better offer.

• Sometimes, the hoarding of multiple offers by a few sought-after EPs leads toa virtual standstill in the market. And once these EPs have made up their minds,a flurry of offers and acceptances takes place in a very short time period. (Thiskind of “congestion” is studied extensively in Roth and Xing, 1997.)

• In response to continuous complaints, HRD has modified the rules governingthe market numerous times, apparently without much success.

In addition to the symptoms mentioned by Roth and Xing (1994), it has beennoted that

• The current system hurts EPs who are traveling—or prevents departmentsfrom sending EPs on business trips during the job market period—because ofthe limited time frame in which EPs and departments are allowed to interact.

• The current system creates a lot of stress and anxiety among EPs, because theyconsider the process to be rather unpredictable and not particularly fair.

Centralized Matching of Incoming EPs

HRD matches incoming EPs to departments on the basis of EPs’ and departments’preferences. These preferences are solicited in the form of rank-order lists (ROLs).Participants who do not respond are assumed to be indifferent.


The process by which HRD translates participants’ preferences into a matchof EPs and departments is heuristic. This means that HRD tries to find the “best”fit without formulating an explicit objective function or a procedure to resolveconflicting preferences. The drawback of a heuristic approach is that the resultingmatch is not guaranteed to be Pareto efficient and/or stable.2 Indeed, we show inthe Appendix that, for a recent cohort of incoming EPs for which we were able toobtain the necessary data, HRD’s match was in fact neither.

Motivated by the shortcomings of the matching systems currently in place, inthe next section, we discuss Gale and Shapley’s deferred acceptance algorithm.

II. The Deferred Acceptance Algorithm

Preliminaries

The internal job market for EPs can be viewed as a “College Admissions Problem”(CAP). The CAP is concerned with two disjoint groups, which we shall refer to asEPs and departments, that have to be matched to each other while taking intoaccount their mutual preferences. The CAP was first formulated by David Gale andLloyd Shapley in a pathbreaking paper in the American Mathematical Monthly(Gale and Shapley, 1962).

Gale and Shapley (GS) argue that a match of EPs and departments, denoted bym, solves the CAP only if it is stable; that is, only if there is no (EP, department) pairthat would prefer to be matched to each other over sticking to their matches as pre-scribed by m. Obviously, a matching process that violates this property can easily beupset by an EP and a department getting together in a manner that benefits both. Suchan (EP, department) pair is called a “blocking pair.” The empirical literature indeedconfirms that stability is a very important property in the design of a matching mar-ket to fix unraveling and associated Stage 2 market failures (see Table 1, p. 417).

GS prove that, for the CAP, a stable match always exists.3 Although there maybe multiple stable matches for a given preference profile, when preferences arestrict, the set of positions filled and EPs employed are the same in all stable matches(Roth and Sotomayor, 1990). Another attractive feature of stability in the CAP isthat it implies Pareto efficiency.4 The reverse, however, is not true; that is, typi-cally, there are many Pareto-efficient matches that are unstable. Thus, stability isa strictly stronger requirement than Pareto efficiency.

In their original paper, GS also provide an iterative procedure for finding a sta-ble match with respect to the participants’ stated preferences. This procedure iscalled the deferred acceptance algorithm. Currently, (a variant of) the DAA is used

2A formal definition of stability is given in the next section.3This result is far from trivial, as is illustrated by the closely related “Roommates Problem.” In the

Roommates Problem, an even number of boys has to be paired up. GS show that, in this case, a stablematching may not exist. Instead, cycles of blocking pairs may cause the set of stable matches to be empty:Consider four boys, A, B, C, and D. Let A rank B first (i.e., B is A’s most preferred roommate), while Branks C first, and C ranks A first. At the same time, A, B, and C all rank D last. Then, regardless of D’spreferences, there can be no stable matching; whoever is matched to D will want to move out, and one ofthe other two will be willing to take him in.

4This follows from the fact that, under strict and “responsive” preferences, bilateral stability impliesgroup stability, and vice versa. See Roth and Sotomayor (1990).


413



414

to match applicants and employers in virtually all medical and dental residencyprograms, pharmacy practices, and psychology internships, both in the UnitedStates and Canada (see, e.g., www.nrmp.org).5

The Algorithm

The DAA is best illustrated in a simplified environment in which each departmenthas exactly one vacancy and the total number of vacancies is equal to the total num-ber of EPs. This simple setup is known in the literature as the “marriage problem.”Subsequently, we can modify the DAA to handle the general case with an arbitrarynumber of departments, vacancies per department, and EPs.

It should be noted that, under the DAA, EPs and departments continue tomake initial contact and arrange interviews in a decentralized manner. But in theend, instead of making offers directly, participants submit ROLs to the marketmaker, HRD, which then runs the DAA. The output of the DAA is a stable match;that is, a stable allocation of EPs across departments.

For ease of exposition, we describe the algorithm as if offers and rejections inthe DAA are made by the departments and EPs. In reality, it is the market maker thatmakes these decisions for them, based on the ROLs submitted by the participants.

For the simple case (marriage problem), the DAA works as follows:

Round 0: EPs and departments rank one another in order of attractiveness.Round 1: Each department makes a job offer to its top-ranked EP. An EP who

receives one or more job offers rejects all but the most preferred among theoffers received. The relatively most preferred job offer is kept on hold.

Round i (i ≥ 2): A department that was rejected in the previous round makes anoffer to its top-ranked EP among those who have not yet rejected it. An EP whoreceives new offers rejects all but the most preferred among the new offers re-ceived and the offer kept on hold from the previous round. Again, the relativelymost preferred offer is kept on hold.

Stop: The algorithm terminates when no new job offers are made. At that pointeach EP has exactly one offer and accepts it.

Note that the roles of EPs and departments in the DAA can be reversed, such thatit is the EPs who make the offers to the departments. While both procedures lead toa stable match, the outcomes are not necessarily identical. In fact, Gale and Shapleyshow that the DAA-generated stable match is preferred over all other stable matchesby the side making the offers, provided that preferences are strict.

The extension of the DAA to the general case is now straightforward. With kdepartments each having a quota of qi positions, n EPs, and EPs making the offers,the algorithm proceeds as follows. First, all EPs apply to the department of theirfirst choice. Department i, with qi positions, then places on its waiting list the qi EPsit likes best, or all applying EPs if there are fewer than qi. It rejects the rest. RejectedEPs then apply to the department of their second choice. Again, each departmentselects the top qi among the new applicants and those on the waiting list, puts these

5The variant used is the Roth and Peranson (1999) algorithm. Its advantage over the standard DAA isthat it can handle couples applying together.


www.nrmp.org

on its new waiting list and rejects the rest. The procedure terminates when all EPseither are on a waiting list or have been rejected by every department to which theywere willing to apply. At this point, each department accepts the offers from every-one on its waiting list and the EP-optimal stable match has been achieved.

Though we have implicitly assumed that all departments rank all EPs (and viceversa), in practice this is not really necessary. Instead, for each of its positions, adepartment should rank sufficient EPs to be reasonably assured of at least one EPaccepting the job.6 This implies that the effort spent on interviewing and rankingEPs under the DAA will not differ much from the effort expended under the cur-rent system. Also in the current system, a department will want to avoid a situationin which, as a result of too many rejections, it has to start interviewing additionalEPs after the market has opened and offers have been made, because, by then, veryfew EPs will be left to choose from.

Finally, rankings do not have to be strict. Indifference between some or allcounterparts is allowed and is resolved by randomization.7 Moreover, departmentscan classify certain EPs as unacceptable, and EPs can do the same for depart-ments. We will get back to these issues later.

Strategic Issues

Strategy-proofness

Roth (1982) proves that there is no stable matching mechanism, including theDAA, that is 100 percent strategy-proof. Even in the case of the DAA, it is possi-ble to cook up a constellation of preferences such that at least one participantwould gain by distorting them.

The catch is that strategic misrepresentation tends to require much more infor-mation about other participants’ preferences than is usually available. Roth andRothblum (1999) show that when participants in the DAA are sufficiently uncertainabout the preferences of others, they cannot gain by reversing the order of their truepreferences. However, if toward the bottom of the ROL a participant is close toindifferent between being matched or not, and the other side makes the offers, aparticipant might gain by shortening the ROL, falsely claiming that the least pre-ferred options are “unacceptable.” In practice, however, such a strategy is unlikelyto be used much, because of its high risk: it might mean that an EP or departmentends up with no match at all.

In any case, truncation clearly is not an option for EPs, unless they are pre-pared to quit and leave the IMF. For departments, truncation is possible only in themarket for permanent positions, since, at the end of the day, all transferring EPshave to be matched to a department.

6For example, in the National Resident Matching Program, each employer interviews and ranks onlya tiny fraction of all potential candidates. This does not interfere with the workings of the DAA.

7When a large number of participants on both sides of the market are indifferent between many counter-parts, a complication may arise. For example, suppose that some EPs and some departments do not care whothey are matched to, or have not submitted an ROL. In that case, the DAA proceeds with randomly gener-ated (strict) preferences. What if on the basis of these randomly generated preferences, an indifferent EPand an indifferent department end up matched to each other? Obviously, neither of them can or will


415



416

Roth and Rothblum’s result of “almost-strategy-proofness in practice” goes along way in explaining the popularity and robustness of the DAA in real-worldapplications such as the National Residency Matching Program for physicians.Indeed, Roth (1984a, 1990, 1991) and Roth and Xing (1994) show that, with veryfew exceptions, centralized matching mechanisms survive if, and only if, they pro-duce stable matches with respect to the participants’ (stated) preferences. This isillustrated in Table 1, taken from Roth and Rothblum (1999).

The DAA and the pressure to precommit

In the current system, departments and EPs routinely flout the rules by commit-ting to one another before the official opening of the market. This raises the ques-tion of whether the situation would be any different under the DAA. Specifically,what prevents departments and EPs from reducing the DAA to a mere formalityby precommitting to top-rank one another?

In the current system, rejection of an offer in the official market is potentiallyvery costly for a department. The reason is that a rejected department may find thatall other attractive EPs have already accepted offers from other departments. Tomitigate this risk, departments routinely put extreme pressure on EPs to precom-mit to accepting a forthcoming job offer. By contrast, rejections are costless in theDAA (at least, from a strategic perspective). The reason is that all acceptances andrejections occur simultaneously. Therefore, a department has no particular interest inmaking sure that its job offer is accepted by the first EP it is offered to. Indeed, thezero rejection cost in the DAA implies that a department can and should top-rank itsfavorite EP even if the probability of the EP accepting the offer is very small. In turn,this undermines the credibility of the exploding nature of a pre-market offer, sincethe department will top-rank its favorite EP, even if the EP refuses to precommit.

These considerations explain why, over time, unraveling and precommitmenttend to occur less under the DAA than in decentralized markets. Nevertheless, itis true that precommitment might still occur on a smaller scale, particularly withrepeated interaction. In this context, Niederle and Roth (2004b) stress the impor-tance of “market culture,” such as norms governing exploding offers.

Departments or positions?

So far, we have assumed that EPs rank, and are matched to, departments. Usually,however, EPs do not have preferences for departments as such, but for specificpositions within departments. For example, even though an EP’s first choice mightbe to work on the China desk in the Asia and Pacific Department (APD), the EPmight very well prefer working on South Africa in the African Department overworking on Tuvalu in APD. In this case, APD can do two things: either list the

object to that. However, there may be other, non-indifferent EPs or departments that would very much liketo be matched to one of these indifferent participants. In that case, we might be able to accommodate oneof them without hurting anybody else. In other words, when the randomization is such that indifferent par-ticipants end up matched to each other, the matching may not be strictly Pareto optimal. In that case, weshould either redo the DAA with newly generated random preferences or manually modify the matching torestore strict Pareto optimality. See the Appendix for a real-world example.


China and Tuvalu jobs separately, thereby allowing EPs to express interest specif-ically in China or specifically in Tuvalu, or lump both jobs together, allowing onlyfor expressions of interest in working for APD, without formal assurance of beingassigned to any particular job.

When EPs have preferences for particular positions instead of for departments,there is a trade-off between ex-ante clarity for EPs and ex-post flexibility for dep-artments. If a department chooses to specify its vacancies in great detail, it lowersthe risk and/or ambiguity for the applicants. In turn, this may help the departmentattract better EPs. On the other hand, if a department chooses to be less specific,it retains full flexibility to assign EPs to one job or another, depending on the needof the moment.

Presumably, competition will force all but the most popular departments to listtheir vacancies separately. From a transparency perspective, separate listings areindeed preferable.

Summary

The DAA solves many of the problems commonly encountered in matching mar-kets. In the DAA, all final acceptances occur simultaneously. This eliminates pre-mature decisions based on incomplete information. Also, undesirable phenomenasuch as unraveling, hoarding of multiple offers, and reneging on a prior acceptancewill no longer occur. At the same time, it eliminates the need to impose rules on


417

Table 1. Stable and Unstable Centralized Matching Mechanisms

Market Stable DAA Still in use

Entry level medical markets Y Y YUS (NRMP) Y Y YEdinburgh (’69) Y Y YCardiff Y Y YCanada Y Y YCambridge N N YLondon Hospital N N YBirmingham N N NEdinburgh (’67) N N NNewcastle N N NSheffield Y Y Y

Other marketsMedical specialties Y Y YCanadian lawyers

Toronto Y Y YVancouver Y Y NCalgary and Edmonton Y Y Y

Dental residencies Y Y YPharmacists Y Y YSororities Y (at equilibrium) N Y

Source: Roth and Rothblum, 1999.


Figure 1. Instructions


418

the participants that are hard to enforce. In particular, EPs and departments cancommunicate with one another as early and as often as they wish. The only hardconstraint is that participants must hand in their rankings of one another on a pre-announced date. The submitted ROLs then determine the matching of EPs todepartments (or positions) through application of the DAA. Operationally, a mar-ket maker running the algorithm can execute all steps of the DAA in real time, oneafter the other, provided that EPs and departments have handed in their ROLs. Thiseliminates the need for repeated communication between EPs and departments, andthereby the need for EPs to be in Washington during any particular time period.

III. The Computer Program

Description

To facilitate implementation of the DAA at the IMF, we have developed an Excel-based computer program, called EP-Match, that executes the algorithm and appliesit to the IMF’s internal job market for EPs. The user interface of the program con-sists of four spreadsheets:1. Instructions (see Figure 1);2. Departments’ Preferences (see Figure 2);



419

Figure 2. Departments’ Preferences

3. EPs’ Preferences (see Figure 3);4. Match (see Figure 4).To find the stable match generated by the DAA, the user, that is, the market maker,sequentially works through the four spreadsheets. The first spreadsheet givesinstructions on how to use the program and requires the market maker to enter thenumber of EPs and the number of departments (or positions) participating in thematching process. Here, the market maker must also choose between the EP-proposing and department-proposing variant of the DAA. In the second spread-sheet, the market maker specifies the number of vacancies in each department andenters the departments’ ROLs in a matrix. The EPs’ ROLs go into the third spread-sheet. The fourth and last spreadsheet presents the resulting stable match.

The program is quite versatile. First, it allows for an arbitrary number of EPs,departments, and positions per department. Second, departments are allowed to be


Figure 3. EPs’ Preferences

Figure 4. The Match


420



421

This means, for example, that D1 likes EP3 best, EP1 second best, and so on.Similarly, the EPs’ ROLs are as follows:

This means, for example, that EP4 prefers the departments in order D3, D1, D2.

indifferent concerning EPs, and EPs concerning positions. Third, a department canbe assured that it will not be matched to an EP it finds unacceptable by not rank-ing that EP at all. Of course, it should only do this if it prefers leaving the positionunfilled to hiring that EP. In principle, EPs can do the same with departments theydo not want to join under any circumstances.

Finally, the program can also handle the following, slightly more complex, sit-uation. Suppose department Dx has multiple positions open, but its ranking of EPsis not the same for all positions. This might happen if a certain position requiresparticular skills, such as fluency in a foreign language or financial expertise. Inthat case, we partition the set of Dx’s positions into r disjoint subsets S1, . . . , Sr,such that, for all positions in a particular subset, Dx’s ranking of EPs is in fact thesame. Then department Dx is replaced by multiple “virtual departments” Dx1, . . . ,Dxr, such that Dxi offers positions Si and ranks the EPs accordingly. Finally, theprogram is run on the modified matching problem.

Examples

The following two examples illustrate how EP-Match works in practice.

Example 1

There are three departments, D1, D2, and D3, and six EPs, EP1–EP6. Each depart-ment has two vacancies. Let the departments’ ROLs be as follows:

Departments’ Rankings of EPs D1 D2 D3

1st EP3 EP1 EP3 2nd EP1 EP3 EP1 3rd EP4 EP4 EP4 4th EP6 EP5 EP2 5th EP2 EP2 EP6 6th EP5 EP6 EP5

EPs’ Rankings of Departments

EP1 EP2 EP3 EP4 EP5 EP6 1st D1 D1 D2 D3 D1 D1 2nd D2 D3 D1 D1 D3 D3 3rd D3 D2 D3 D2 D2 D2



422

To solve for the associated department-optimal stable match using EP-Match,the market maker first enters the number of departments (3) and EPs (6) in theappropriate boxes on the “Instructions” spreadsheet and then checks that the option“Dept. Proposing” is selected. On the “Departments’ Preferences” spreadsheet, thedepartments’ rankings of the EPs are coded by assigning rank-order “1” to the most-preferred EP, “2” to the next-most-preferred EP, and so on. On the same spread-sheet, the number of vacancies (“Positions”) in each department (i.e., 2, 2, 2) isentered. The departments’ preference matrix then looks like the following:

Similarly, on the “EPs’ Preferences” spreadsheet, the market maker codes the EPs’rankings of the departments by assigning rank-order “1” to the most-preferreddepartment, “2” to the next-most-preferred, and “3” to the least-preferred depart-ment. The EPs’ preference matrix then looks like this:

Finally, the “Match” button on the “Match” spreadsheet can be pressed, andthe department-optimal stable match appears on screen as follows:

D1 D2 D3

2 2 2

EP1 2 1 2EP2 5 5 4EP3 1 2 1EP4 3 3 3EP5 6 4 6EP6 4 6 5

Clear All

EP1 EP2 EP3 EP4 EP5 EP6D1 1 1 2 2 1 1D2 2 3 1 3 3 3D3 3 2 3 1 2 2

Clear A ll

EP1 D1 D1 EP1EP2 D3 D1 EP6EP3 D2 D2 EP3EP4 D3 D2 EP5EP5 D2 D3 EP2EP6 D1 D3 EP4


Here, the list on the left is ordered by EP, while the list on the right is ordered bydepartment.

Note that stability implies that no department will be able to lure a more pre-ferred EP from another department. Neither can any EP succeed in “stealing” amore preferred position from another EP.

To calculate the EP-optimal stable match, select the option “EP Proposing” onthe Instructions spreadsheet and press the Match button on the Match spreadsheet.It is easily verified that, in this example, the EP-optimal and departmental-optimalstable matches are identical.

Example 2

In our second example we illustrate how EP-Match can be used to solve morecomplicated scenarios, when not all EPs are acceptable to all departments, whenthere are more (or fewer) positions than EPs, and when departments are indiffer-ent about some EPs, and EPs are indifferent about some departments.

Let us take the setup of Example 1 as our point of departure but assume thatdepartment D1 is indifferent between EP1 and EP4, and also between EP2 andEP5. Moreover, department D2 now considers EP2, EP5, and EP6 to be unaccept-able, while D3 has two additional vacancies (i.e., four instead of two). On the EPside, EP2, EP4, and EP6 are assumed to be indifferent between the departmentthey have ranked second and the department they have ranked third. For the rest,the situation is the same as in Example 1.

In EP-Match, on-screen instructions explain how indifference and unaccept-ability can be coded into the preference matrices. To express that department Di isin fact indifferent between EPx and the EP ranked one lower than EPx, add a “*”to the rank-order number of EPx on the “Departments’ Preferences” spreadsheet.In our example, this means that D1’s rank-order numbers of EP1 and EP2, that is,“2” and “5,” respectively, get a “*.” To express that department Di considers EPyunacceptable, leave cell (EPy, Di) empty. In our example, this means that cells(EP2, D2), (EP5, D2), and (EP6, D2) remain empty.

The departments’ preference matrix then looks like this:


423

D1 D2 D3

2 2 4

EP1 2* 1 2EP2 5* 4EP3 1 2 1EP4 3 3 3EP5 6 6EP6 4 5

Clear All



424

Now, the department-proposing stable match looks like this:

Again, the list on the left is ordered by EP, and the list on the right is ordered bydepartment.

This stable match differs little from the stable match of Example 1. EP5 hasmoved from D2 to D3, filling one of D3’s two additional vacancies, while D2’ssecond and D3’s fourth vacancy go unfilled. The latter reflects the fact that, in ourexample, there are more positions than EPs. The EP-proposing stable match isagain identical to the department-proposing stable match.

IV. Conclusion

In this paper we have argued that adopting the deferred acceptance algorithm tomatch EPs and departments can enhance the well-being of all parties involved.Unlike the current system, the DAA generates an efficient and stable outcome byoptimally using all available information on the preferences of the participants ascontained in their rank-order lists. In addition to its attractive theoretical proper-ties, the DAA has been extensively tried and tested in practice, and has producedexcellent results. To facilitate implementation of the DAA at the IMF, we have

Incorporating EP2’s, EP4’s, and EP6’s indifference between their second- andthird-ranked departments, we get the following EP preference matrix:

EP1 D1 D1 EP1EP2 D3 D1 EP6EP3 D2 D2 EP3EP4 D3 D2 --- not matched ---EP5 D3 D3 EP2EP6 D1 D3 EP4

D3 EP5D3 --- not matched ---

EP1 EP2 EP3 EP4 EP5 EP6D1 1 1 2 2* 1 1D2 2 3 1 3 3 3D3 3 2* 3 1 2 2*

Clear All


developed an Excel-based computer program that executes the algorithm on thebasis of EPs’ and departments’ rankings of each other.

APPENDIX

Matching of Incoming EPs

Introduction

As described in Section I, the IMF’s Human Resources Department (HRD) matches incomingEPs to first-year positions on the basis of EPs’ and departments’ preferences. The question thatarises is whether these matches are stable and/or Pareto efficient. Because HRD tries to find the“best” fit without formulating an explicit objective function or a procedure to resolve conflictingpreferences, this question can only be answered by looking at actual data.

In this appendix we show that HRD’s match was neither stable nor Pareto efficient in thecase of a cohort of incoming EPs for which we were able to retrieve the necessary data. Thesedata consist of the participants’ preferences, given by their rank-order lists (ROLs) (Table 2), andthe match implemented by HRD (Table 3).8

Methodology

Participants (i.e., departments and EPs) who chose not to submit ROLs are assumed to be in-different as to who they are matched to. Similarly, participants who ranked only a subset ofcounterparts are assumed to be indifferent about those left unranked, while strictly preferringranked over unranked counterparts.

To check the stability of HRD’s match, it might be tempting to run EP-Match on the par-ticipants’ preferences and compare the outcomes. However, this would be misguided, becausethe set of stable matches usually contains many elements, while, in principle, EP-Match findsonly two of them; namely, the department-proposing and the EP-proposing stable matches.Thus, discrepancies between HRD’s match and those calculated by EP-Match are not neces-sarily informative.

A second complication arises from the large number of declared and imputed “indiffer-ences” in participants’ preferences. Only when preferences are strict does stability imply Paretoefficiency. Otherwise, stable matches may contain “suspect pairs.” A suspect pair is a match ofan EP and a department, such that both the EP and the department are indifferent betweenremaining matched to each other and being matched to someone else. If the EP and/or thedepartment making up the suspect pair are very sought after among other departments or EPs,it may be possible to make one or more of these other departments or EPs better off by break-ing up the suspect pair, without hurting anybody else. This implies that, with a lot of indiffer-ence on both sides of the market, a stable match is only a candidate solution. To be afull-fledged solution, it must be established that none of the suspect pairs give rise to Paretoinefficiencies.

Therefore, a two-step procedure is followed to assess the optimality of HRD’s match withrespect to the participants’ preferences:

• Check the stability of the match by looking for blocking pairs.• Check whether suspect pairs, if any, give rise to Pareto inefficiencies.

8To protect confidentiality we have made certain changes to the data. In particular, “dummy” EPs,departments, and preferences have been added. However, the gist of the example and the validity of theconclusions are fully retained.


425


Stability

Is HRD’s match stable with respect to the participants’ stated preferences? The answer is no.To see why, note that EP7 and department D1 form a blocking pair: EP7 strictly prefers D1 tohis current match D5, and D1 strictly prefers EP7 to its current match EP2.

In this case, a simple exchange is sufficient to transform HRD’s match into a stable one.That is, we reassign EP7 to D1 and EP2 to D5, while keeping everything else the same. BothEP7 and D5 go from being matched to a counterpart that was not even on their submitted pref-erence list to being matched to their 1st choice. On the other hand, EP2 goes from his 1st choiceto his 2nd choice, while D5 goes from its 3rd choice to its 6th choice.


426

Table 2. Participants’ Preferences

EPs’ Preferences over Departments Departments’ Preferences over EPs

EP participant Preferences1 Department Preferences1

EP1 D12, D11*, D13*, D5*, D1 EP7, EP15, EP17D7*, D4 D2 EP12, EP7, EP25

EP2 D1, D11*, D13*, D5*, D7*, D3 EP19, EP22*, EP11*, EP16D12*, D4 D4 EP22, EP19, EP6, EP11, EP12

EP3 No preference indicated D5 EP15, EP21, EP7, EP6, EP19, EP2EP4 D1, D6, D13, D12 D6 EP15, EP1, EP26EP5 D12, D11 D7 EP17, EP14, EP25EP6 No preference indicated D8 No preference indicatedEP7 D1 D9 No preference indicatedEP8 No preference indicated D10 EP9*, EP19EP9 No preference indicated D11 EP22, EP6, EP19, EP21, EP23, EP7EP10 No preference indicated D12 No preference indicatedEP11 No preference indicated D13 No preference indicatedEP12 D1, D5, D13 D14 E15, EP24EP13 No preference indicatedEP14 No preference indicated

1Preferences are listed in rank order.

EP15 D1, D2, D12*Indicates indifference between EPx and the

EP16 D1, D3, D10, D13, D5EP ranked one lower than EPx.

EP17 No preference indicatedEP18 D13, D1EP19 No preference indicatedEP20 No preference indicatedEP21 D1, D5, D12, D13EP22 No preference indicatedEP23 D5EP24 D14, D1EP25 No preference indicatedEP26 D1, D6, D2

1Preferences are listed in rank order.*Indicates indifference between

department Di and the department ranked one lower than Di.

Source: IMF’s Human Resources Department.


Pareto Efficiency

It is easily established that HRD’s match as given in Table 3 contains the following four sus-pect pairs: (EP3, D12); (EP10, D13); (EP13, D8); and (EP20, D8). Not all of these suspect pairsgive rise to Pareto inefficiencies, however. Only the first two suspect pairs do. By breaking upthe EP3, D12 pair we can make EP5 strictly better off without hurting anyone else, while break-ing up EP10, D13 allows us to make EP18 strictly better off.

Specifically, we perform the following Pareto-improving reassignments:

Table 3. Assignments

EPs’ Departmental Assignments Departments’ EP Assignments

EP Assigned department Department Assigned EP

EP1 D6 D1 EP2EP2 D1 D1 EP15EP3 D12 D1 EP17EP4 D4 D2 EP12EP5 D11 D3 EP16EP6 D11 D3 EP19EP7 D5 D4 EP4EP8 D11 D4 EP11EP9 D10 D4 EP22EP10 D13 D5 EP7EP11 D4 D5 EP21EP12 D2 D6 EP1EP13 D8 D6 EP26EP14 D7 D7 EP14EP15 D1 D7 EP25EP16 D3 D8 EP13EP17 D1 D8 EP20EP18 D11 D10 EP9EP19 D3 D11 EP5EP20 D8 D11 EP6EP21 D5 D11 EP8EP22 D4 D11 EP18EP23 D11 D11 EP23EP24 D14 D12 EP3EP25 D7 D13 EP10EP26 D6 D14 EP24

Source: IMF’s Human Resources Department.

Box 9

1. (EP3, D12) (EP3, D11)(EP5, D11) (EP5, D12)

2. (EP10, D13) (EP10, D11)(EP18, D11) (EP18, D13)


427


In this case, EP5 achieves his 1st choice instead of his 2nd choice, while EP18 achieves his 1stchoice instead of being assigned to a department not even on his preference list. At the sametime, none of the others are worse off than before.

Conclusion

We have shown that HRD’s match of incoming EPs to departments was neither stable norPareto efficient. Application of EP-Match followed by a critical assessment of all suspect pairswould have resulted in a better outcome.

BIBLIOGRAPHY

Avery, C., C. Jolls, R. Posner, and A. E. Roth, 2001, “The Market for Federal Judicial LawClerks,” University of Chicago Law Review, Vol. 68 (Summer), pp. 793–902.

Blum, Y., A. E. Roth, and U. G. Rothblum, 1997, “Vacancy Chains and Equilibration in Senior-Level Labor Markets,” Journal of Economic Theory, Vol. 76, No. 2, pp. 362–411.

Gale, D., 2001, “The Two-Sided Matching Problem: Origin, Development, and Current Issues,”International Game Theory Review, Vol. 3, No. 3, pp. 237–52.

———, and L. Shapley, 1962, “College Admissions and the Stability of Marriage,” AmericanMathematical Monthly, Vol. 69, pp. 9–15.

Kagel, J. H., and A. E. Roth, 2000, “The Dynamics of Reorganization in Matching Markets: ALaboratory Experiment Motivated by a Natural Experiment,” Quarterly Journal ofEconomics, Vol. 115, pp. 201–35.

Niederle, M., and A. E. Roth, 2003, “Unraveling Reduces Mobility in a Labor Market:Gastroenterology With and Without a Centralized Match,” Journal of Political Economy,Vol. 111, No. 6, pp. 1,342–52.

———, 2004a, “The Gastroenterology Fellowship Match: How It Failed and Why It CouldSucceed Again,” Gastroenterology, Vol. 127, No. 2, pp. 658–66.

———, 2004b, “Market Culture: How Norms Governing Exploding Offers Affect MarketPerformance,” NBER Working Paper No. 10256 (Cambridge, Massachusetts: NationalBureau of Economic Research).

Roth, A. E., 1982, “The Economics of Matching: Stability and Incentives,” Mathematics ofOperations Research, Vol. 7, pp. 617–28.

———, 1984a, “The Evolution of the Labor Market for Medical Interns and Residents: A CaseStudy in Game Theory,” Journal of Political Economy, Vol. 92, No. 6, pp. 991–1016.

———, 1984b, “Misrepresentation and Stability in the Marriage Problem,” Journal of Eco-nomic Theory, Vol. 34, pp. 383–87.

———, 1984c, “Stability and Polarization of Interests in Job Matching,” Econometrica, Vol. 52,No. 1, pp. 47–57.

———, 1985, “The College Admissions Problem Is Not Equivalent to the Marriage Problem,”Journal of Economic Theory, Vol. 36, pp. 277–88.

———, 1986, “On the Allocation of Residents to Rural Hospitals: A General Property of Two-Sided Matching Markets,” Econometrica, Vol. 54, pp. 425–27.

———, 1989, “Two Sided Matching with Incomplete Information About Others’ Preferences,”Games and Economic Behavior, Vol. 1, pp. 191–209.

———, 1990, “New Physicians: A Natural Experiment in Market Organization,” Science,Vol. 250, pp. 1524–28.


428



429

———, 1991, “A Natural Experiment in the Organization of Entry Level Labor Markets:Regional Markets for New Physicians and Surgeons in the U.K.,” American EconomicReview, Vol. 81, No. 3, pp. 415–40.

———, 2002, “The Economist as Engineer: Game Theory, Experimental Economics andComputation as Tools of Design Economics,” Fisher Schultz lecture, Econometrica, Vol.70, No. 4, pp. 1341–78.

———, 2003, “The Origins, History, and Design of the Resident Match,” Journal of theAmerican Medical Association, Vol. 289 (February), pp. 909–12.

———, and E. Peranson, 1999, “The Redesign of the Matching Market for American Physi-cians: Some Engineering Aspects of Economic Design,” American Economic Review,Vol. 89, No. 4, pp. 748–80.

Roth, A. E., and U. G. Rothblum, 1999, “Truncation Strategies in Matching Markets: In Searchof Advice for Participants,” Econometrica, Vol. 67, No. 1, pp. 21–44.

Roth, A. E., and M. Sotomayor, 1989, “The College Admissions Problem Revisited,” Econo-metrica, Vol. 57, No. 3, pp. 559–70.

———, 1990, “Two-Sided Matching: A Study in Game-Theoretic Modeling and Analysis,”Econometric Society Monograph Series (Cambridge, United Kingdom: CambridgeUniversity Press).

———, 1992, “Two-Sided Matching,” in Handbook of Game Theory, Vol. 1, ed. by R. J.Aumann and S. Hart (Amsterdam: Elsevier), pp. 485–541.

Roth, A. E., and X. Xing, 1994, “Jumping the Gun: Imperfections and Institutions Related to the Timing of Market Transactions,” American Economic Review, Vol. 84, No. 4,pp. 992–1044.

———, 1997, “Turnaround Time and Bottlenecks in Market Clearing: Decentralized Matchingin the Market for Clinical Psychologists,” Journal of Political Economy, Vol. 105, No. 2,pp. 284–329.

Zhou, L., 1991, “Stable Matchings and Equilibrium Outcomes of the Gale-Shapley’s Algorithmfor the Marriage Problem,” Economics Letters, Vol. 36, No. 1, pp. 25–9.


430


Banking on Foreigners: The Behavior of International Bank Claims

on Latin America, 1985–2000

MARIA SOLEDAD MARTINEZ PERIA, ANDREW POWELL, AND IVANNA VLADKOVA-HOLLAR*

The significant rise in foreign bank claims observed during the 1990s, followingtheir steep decline during the 1980s debt crisis, reignited interest in understand-ing the behavior of these flows. This paper analyzes changes in foreign bankclaims on the Latin American private sector over the period 1985–2000. We findthat banks transmit shocks from their home countries (where banks’ headquartersare located) and that changes in claims on individual host countries (those thatreceive claims) are correlated with aggregate changes in claims on other coun-tries. However, over time, we observe that foreign bank claims have become lessresponsive to external factors. Also, we present evidence that the sensitivity of for-eign bank claims to host factors diminishes, as banks’ aggregate exposure rises.Finally, we find that foreign bank claims react more to positive than to negativehost shocks and are not significantly curtailed during crises. [JEL G21, N26]

*The authors are affiliated with the World Bank, Universidad Torcuato di Tella, and InternationalMonetary Fund, respectively. The authors thank the Bank for International Settlements for providing us dataand, in particular, Jesper Wormstrup at this institution for answering many questions on this data set. We aregrateful for comments and suggestions from Michael Binder, Guillermo Calvo, Linda Goldberg, GracielaKaminsky, Miguel Kiguel, Luc Laeven, Norman Loayza, Enrique Mendoza, Sergio Schmukler, MiguelSebastián, Beatrice Weder, and participants at the Universidad Torcuato di Tella 2002 Summer Camp inInternational Economics and Finance and at the 2002 Latin American and Caribbean Economic AssociationConference. Comments received from Robert Flood and an anonymous referee are also gratefully acknowl-edged. This study was completed with financial support from the World Bank Office of the Chief Economistfor Latin America and the World Bank Research Committee. The findings, interpretations, and conclusionsexpressed in this paper are entirely those of the authors and do not necessarily represent the views of theWorld Bank, the IMF, Executive Directors at these institutions, or the countries they represent.


BANKING ON FOREIGNERS: THE BEHAVIOR OF INTERNATIONAL BANK CLAIMS

431

The 1990s saw a significant increase in foreign bank claims on developing coun-tries. According to the Bank for International Settlements (BIS), between 1985

and 2000, international banks’ total claims on developing countries increasedsteadily from US$545.2 billion in 1985 to US$1,318.8 billion in 2000.1,2 By theend of the 1990s, total claims of BIS-reporting banks (internationally active banksthat report data to the BIS on their overseas claims) on developing countries rep-resented 31 percent of total local credit in the developing world.3 Among countriesin Latin America and in Central and Eastern Europe, foreign bank claims exceeded50 percent of local credit.

Total claims of BIS-reporting banks on developing countries include cross-border claims extended from outside the host countries, along with local claimsbooked with the bank branches or subsidiaries operating in the host countries.4Claims refer primarily to loans and advances but also include holdings of securitiesand equity participations.

Following the steep decline in foreign bank claims during the 1980s debt cri-sis, the rapid increase observed during the 1990s fueled a growing interest in thebehavior of these claims, and so emerged a new literature on multinational bank-ing.5 Because part of the 1990s increase in foreign bank claims went hand in handwith the establishment of foreign bank branches and subsidiaries in developingcountries, most existing studies focus on the performance of foreign bank operationsin these countries and analyze the impact of foreign bank entry on the efficiency andprofitability of domestic banks in developing countries (see Barajas, Steiner, andSalazar, 2000; Claessens, Demirgüç-Kunt, and Huizinga, 2001; Denizer, 2000;Crystal, Dages, and Goldberg, 2001; and Bonin, Hasan, and Wachtel, 2004).

On the other hand, the question of what drives changes in foreign bank claimshas received less attention. There are, however, some insightful studies on thisissue. Goldberg (2002) examines the determinants of U.S. bank claims abroad andfinds that while U.S. economic conditions affect U.S. bank claims abroad, suchclaims are less affected by economic conditions, including crises, in the host coun-tries. Looking specifically at crises periods, Van Rijckeghem and Weder (2003)

1Throughout this period international bank claims refer to those from banks headquartered in Australia,Austria, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Luxembourg,Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, the United Kingdom, and the United States.

2This implies a 51 percent increase in real terms.3Here, total local credit refers to credit provided by all banks (both foreign and domestic) with offices

in the developing world. Source: International Financial Statistics, International Monetary Fund.4“Host country” denotes the country to which a foreign bank extends claims either cross-border or

through its branches and/or subsidiaries in that country. “Home country” refers to the country of origin ofthe foreign bank, that is, the country where the bank’s headquarters are located.

5Previously, the literature on multinational banking focused primarily on the experience of developedcountries (especially the United States) with foreign bank entry and on the internationalization of the activ-ities of banks from these countries during the 1970s and 1980s. For example, Goldberg and Saunders(1981a and b); Cho, Krishnan, and Nigh (1987); and Goldberg and Grosse (1994) investigate the factorsdriving the extent and type of foreign bank presence in the United States, while Fisher and Molyneux(1996) conduct a similar study of foreign bank activities in London. On the other hand, papers such asGoldberg and Saunders (1980); Nigh, Cho, and Krishnan (1986); Goldberg and Johnson (1990); and Buch(2000) examine the operations of German (in the case of the last paper) and U.S. banks abroad.


Maria Soledad Martinez Peria, Andrew Powell, and Ivanna Vladkova-Hollar

432

investigate the role of international banks in transmitting crises and find evidencethat in certain episodes, changes in banks’ exposure to crises countries help predictbank flows in other countries. However, their data capture primarily cross-borderclaims.6 On the other hand, taking into account foreign bank local claims to devel-oping countries, Dages, Goldberg, and Kinney (2000); Peek and Rosengren (2000);Goldberg (2002); and De Haas and Van Lelyveld (2003) provide evidence that for-eign bank claims did not retrench during recent crises in Latin America and Centraland Eastern Europe.

Using a comprehensive data set on foreign bank claims to the private sector inLatin America for the period 1985–2000, we revisit some of the issues examinedby previous papers on the determinants of foreign bank claims. More importantly,we explore new questions associated with this issue. The purpose of this study isnot to compare the behavior of foreign and domestic banks, but rather to understandwhat drives changes in foreign bank claims to developing countries and how for-eign banks respond to different types of shocks, under various circumstances.These are important issues both for countries already relying heavily on foreignbank financing and for those countries considering a greater role for foreign banks.

Like other papers that have examined the behavior of foreign bank claims, weanalyze their reaction to home and host conditions, and, in particular, we investi-gate whether foreign banks retrench during host crises. One contribution of ourpaper vis-à-vis others that have looked at these issues is that we consider the behav-ior of foreign claims for a larger combination of home and host countries, over alonger period of time, including both tranquil and crisis episodes.

Furthermore, we extend the analysis on the determinants of foreign bankclaims in some new directions. First, we examine whether the sensitivity of foreignbanks to external and host shocks is the same across banks from different homecountries. Second, we investigate whether foreign banks respond similarly to posi-tive and negative shocks. Third, we analyze whether foreign bank behavior and theimpact of different types of shocks changed over time. Finally, we study how for-eign bank claims are affected by factors previously overlooked in the literature. Inparticular, we evaluate how the level of foreign banks’ exposure affects theirresponsiveness to host country shocks and whether aggregate movements in claimsto other countries drive changes in foreign bank claims to individual hosts.

Our analysis focuses on Latin America for at least three reasons. First, foreignbanks have had an active presence in the region for an extended period. Second,while for the region as a whole foreign bank claims increased over our period ofstudy, there are still significant differences in the importance of this source of fundsacross countries in Latin America.7 Finally, most countries in the region have been

6Van Rijckeghem and Weder (2003) examine a panel of BIS data on flows to 30 emerging markets dis-aggregated by 11 banking centers, to test the role of bank claims in transmitting currency crises. They findthat changes in bank exposures to a crisis country helped predict bank flows in third countries after theAsian crisis, but to a lesser extent during the Mexican 1994 crisis.

7For example, international bank claims (cross-border claims and local claims in foreign currency) in2000 represented more than 55 percent of domestic credit for Argentina and Peru, but they accounted foronly 19 percent of domestic credit for Brazil.


subject to pronounced economic cycles and several crises, providing us with aunique opportunity to analyze the impact of these factors on foreign bank claims.

We consider the 1985–2000 period an interesting one to study because duringthis time frame foreign financing grew significantly across Latin America and thedeveloping world, in general. Furthermore, over this period, many developingcountries, and certainly most in the Latin American region, liberalized their finan-cial systems, allowing foreign banks to play a greater role in their local financialsectors. As a result, during this period, the nature of foreign bank financing changedconsiderably from almost purely cross-border to a mix that also included locallending through foreign bank branches and subsidiaries in the host countries.

Our empirical estimations allow us to corroborate, for a larger combination ofhome and host countries and over a relatively long period, many of the results foundby previous studies. In particular, like Goldberg (2002), we find that home countryconditions drive changes in foreign bank lending. Also, controlling for host growthand credit ratings, we find that foreign bank claims do not retrench significantlyduring crises in the host countries (as found by Dages, Goldberg, and Kinney, 2000;Peek and Rosengren, 2000; and Goldberg, 2002).

More importantly, our work yields interesting new results. First, while foreignbanks from different home countries appear to react similarly to host countryshocks, their reaction to shocks from their own countries seems to vary by homecountry. Second, claims on individual host countries are positively associated withaggregate changes in claims to other countries. However, foreign banks’ reaction toexternal shocks (with respect to the host) has diminished over time. Third, foreignbanks also respond to host country shocks. However, the higher the aggregateexposure of foreign banks to a given host country, the lower the sensitivity ofclaims to host country shocks. In other words, foreign bank claims become lessprocyclical as exposure rises. Finally, we uncover asymmetries regarding foreignbanks’ response to positive and negative shocks, given that banks appear to reactmore to the former than to the latter.

I. The Data on Foreign Bank Claims to Latin America

Our data on foreign bank claims to Latin America come from the BIS.8 Specifically,the data we obtained are international financial claims on the nonbank private sec-tor as reported in the BIS’s Consolidated Banking Statistics.9 These country-levelstatistics sum the claims extended by the headquarters of foreign banks or by theiroffices outside the host countries (that is, cross-border claims) with the foreign cur-rency claims provided by the affiliates (that is, branches and subsidiaries) of for-eign banks in the host countries. Therefore, in our analysis, foreign bank claimsrefer to international financial claims to the nonbank private sector as defined by

8For a full description of these data see BIS (2003).9The BIS distinguishes among international financial claims directed to the private, public, and bank-

ing sectors. Claims on the latter include those on the central bank and on public and private financial insti-tutions. As a result, we study the behavior of claims on the nonbank private sector only.


433



434

the BIS.10 However, the BIS does not typically publish the disaggregation by sec-tor (public, private, or banking) and by country of origin at the same time, so thesedata are confidential and were provided by the BIS with the authorization of eachof the home/lender country’s central banks.

We specifically focus on the behavior of claims from banks headquartered inseven industrialized countries (Canada, France, Germany, Japan, Spain, the UnitedKingdom, and the United States) on the private sector in 10 Latin American coun-tries (Argentina, Brazil, Chile, Colombia, Costa Rica, Ecuador, Mexico, Peru,Uruguay, and Venezuela) over the period 1985–2000.11 Our choice of home andhost countries is driven by their relative importance as lenders and borrowers to andfrom the Latin American region, respectively. Banks from the seven home countriesaccount for more than 80 percent of all foreign bank claims to Latin America. Inturn, the 10 host countries in our sample receive more than 95 percent of all foreignbank claims to the region.

Our period of analysis was determined by several factors. From a conceptualstandpoint, we wanted to look at the behavior of foreign bank claims during aperiod characterized by greater financial liberalization in the region and by anobserved preference by international banks to increase their operations in develop-ing countries. From a practical standpoint, we are unable to look at the period pre-1985, because the BIS started gathering information on foreign bank claims in thatyear. Also, our analysis ends in 2000 because extending the sample would beworthwhile if we could analyze the case of the 2001 Argentine crisis. However,because during this episode the Argentine government forced the conversion of allforeign currency claims into pesos, we cannot disentangle from the BIS data theimpact of the pesification from a true cancellation of claims on the part of foreignbanks. As a result, we stop our analysis in the year 2000.

Rather than examine the behavior of total—both public and private—foreignbank claims, we focus on private sector claims exclusively, for several reasons.First, in recent years, foreign bank claims on the private sector have come to rep-resent the bulk of foreign bank claims to developing countries and, in particular,to Latin American economies.12 Second, foreign bank claims on the public sectormay reflect the heterogeneous and particular fiscal policies of different govern-ments. Also, changes in public sector claims may not be driven by the voluntaryprofit-maximizing choices of foreign banks but rather may be affected by political

10Our definition of foreign bank claims ignores the local claims in local currencies extended by for-eign banks. The BIS does not report data on these statistics with a sectoral breakdown (that is, there is nodiscrimination between claims held with the private and public sectors). Nevertheless, we feel that the def-inition of claims used here, which focuses on foreign currency claims, might be more representative of theactual exposure or potential losses that foreign banks could face from their operations in developing coun-tries, since in general it will be harder for countries to repay claims in foreign currency, especially if someof those claims go to individuals or firms that do not receive dollar incomes.

11Though the BIS statistics are biannual until 2000 and quarterly thereafter, data availability for theremaining variables in our empirical model leads us to focus on annual, end-of-year changes.

12By the end of 2000, claims to the nonbank private sector represented 53 percent of all claims todeveloping countries and 62 percent of all claims to Latin American countries, with the remaining claimsevenly split between the public and banking sectors in those countries.


considerations and/or moral suasion on the part of governments.13 Finally, claimson the public sector are more likely to take the form of bonds, and public bondmarkets are more liquid than those for private sector debt. As a consequence, end-of-period valuations of foreign bank claims might not be representative of foreignbanks’ exposure over a given period.14

Figures 1 through 4 illustrate the behavior of foreign bank claims to the non-bank private sector in Latin America from 1985 to 2000. In the early to mid-1980s,claims to Latin America accounted for more than one-third of the claims extendedby banks from the seven BIS-reporting countries on non-BIS-reporting countries(see Figure 1). Over the second half of the 1980s, foreign banks diversified awayfrom the region, and claims to the 10 selected Latin American economies in theregion declined in real terms between 1985 and 1990 (see Figure 2). However, overthe 1990s, real claims rebounded, rising rapidly and surpassing the US$100 billionmark by the end of the decade. Thus, exposure to Latin America remained below

13An example is the recent crisis in Argentina, where domestic and foreign banks were coerced intoincreasing their exposure to the public sector through debt swaps.

14Also, derivative markets on public sector bonds are reasonably liquid, and the BIS data may not con-trol well for such operations. While the same objections may be raised with respect to claims on the pri-vate sector, loans tend to be a much higher percentage of the total claims on the private sector.

0

5

10

15

20

25

30

35

40

1985 1990 1995 2000

Source: Authors’ calculations based on Bank for International Settlements (BIS) data.1The seven selected BIS-reporting home contries are Canada, France, Germany, Japan, Spain, the

United Kingdom, and the United States. Latin America here refers to the 10 largest countries in the region: Argentina, Brazil, Chile, Colombia, Costa Rica, Ecuador, Mexico, Peru, Uruguay, and Venezuela.

Figure 1. The Exposure of Banks from Selected BIS-Reporting Countries to the Private Sector in Latin America1

(Claims on Latin America from all banks as percent of these banks’ total private sector claims)


435



436

1980s levels but rose steadily over time, reaching 17 percent of all claims (to devel-oping and developed non-BIS-reporting countries) in 2000.15

U.S., French, German, and, recently, Spanish banks were the most importantsources of bank financing to the region throughout the sample (see Figure 3). U.S.banks almost always held the most claims on the region, accounting for more than20 percent of all claims to the nonbank private sector throughout the entire period1985–2000. The exception is France in 1990, when French banks accounted formore than 35 percent of all claims to Latin America. However, French claims to theregion have dropped, reaching less than 15 percent of all claims to Latin Americain the year 2000. German bank claims on Latin America hovered between 15 and20 percent of all claims to this region. In the mid-1990s, Spain emerged as thecountry with the fastest-growing share of claims to the region, accounting for lessthan 5 percent of claims in 1985 but exceeding 20 percent of total claims to the non-bank private sector in Latin America by 2000.

At the same time, throughout this period, Spanish and U.S. banks had the high-est exposure to this region (see Figure 4). Spanish banks’ exposure averaged 50 per-cent of all their total international private claims on non-BIS-reporting countries,while for the United States this figure was 35 percent. However, the trend in exposure

15As a share of claims on developing countries, claims on Latin America reached 40 percent in 2000.

0

2 0

4 0

6 0

8 0

1 0 0

1 2 0

1985

1986

1987

1988

1989

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

Source: Authors’ calculations based on Bank for International Settlements (BIS) data.1 The seven selected BIS-reporting home countries are Canada, France, Germany, Japan, Spain, the


Figure 2. The Evolution of Claims from Banks from Selected BIS-Reporting Countries on the Latin American Private Sector1

(In billions of constant U.S. dollars)


across these two countries is very different. While U.S. exposure remained fairly con-stant throughout the period 1985–2000, Spanish banks’ exposure increased signifi-cantly from less than 40 percent in the early 1990s to more than 68 percent by 2000.

II. Empirical Methodology

Our general econometric model explaining changes in foreign bank claims is rep-resented by equation (1) below, where j = 1 to 7 identifies the seven BIS homecountries, i = 1 to 10 indicates each individual Latin American host country, and t = 1985 to 2000 refers to the time period considered.16 Equation (1) includes bothhome and host country individual effects, α j

0 and α j1i, respectively, and allows the

coefficients to vary depending on the home country (this explains the j superscripton all coefficients).17 However, since it is possible that banks from different homecountries react similarly to host and even home country shocks, we test different

16The United Kingdom is the exception, where data on private sector claims are available only for theperiod 1993–2000.

17Alternatively, we could estimate a separate regression for each home (lender) country, usingZellner’s Seemingly Unrelated Regressions (SUR), to account for the contemporaneous cross-equationcorrelation in the error terms. As a robustness check, we estimated separate equations for each lender andcompared those results with the results from estimating equation (1). The differences are not significant,and, furthermore, the drawback of the SUR method is that it forces our data into a balanced panel, signif-icantly reducing the number of observations.

0

5

10

15

20

25

30

35

40

45

U . S . S p a i n Germany France Japan Canada U . K .

1985

1990

1995

2000

Source: Authors’ calculations based on Bank for International Settlements (BIS) data.1The seven selected BIS-reporting home countries are Canada, France, Germany, Japan, Spain, the



437

Figure 3. The Importance of Selected BIS-Reporting Countries’ Bank Claimsfor the Latin American Private Sector1

(Percent of total selected BIS-reporting countries’ claims on Latin America’s private sector)



438

restricted versions of equation (1) to arrive at the baseline specification that wereport in the results section.

The dependent variable, %ΔClaimsj,i,t,, is the annual percentage change in realconsolidated international claims from banks in home country j to the nonbank pri-vate sector in host country i between t − 1 and t.18 The empirical model implementedto analyze the behavior of this variable draws on existing studies on foreign bank

% , , , ,ΔClaims Home Factorsj i tj

ij j

j t= + + ′−α α β0 1 1 ++ +

×

′−

′−

λ

δ

ji t

ji t

Host Factors

Host Factors E

,

,

1

1 xxposure

Private Claims on Other

j i t

j

, , −( ) +1

γ %Δ CCountriesj i t j i t( ) +−, , , , . ( )

11ε

18Note that while an increase in claims reflects a rise in foreign bank’s exposure, it is not necessarilyassociated solely with new lending to the region. For example, the acquisition of a domestic bank by a for-eign bank will lead to a rise in claims (as the loan portfolio of the domestic bank is absorbed by the foreignbank), but it may or may not lead to new lending, depending on the actions of the foreign bank followingthe acquisition. Nevertheless, based on some rough calculations using the BIS Locational Statistics, we canestimate that more than 70 percent of the international claims to Latin America are in the form of loans.Also, in the robustness tests we try to explicitly control for the impact of mergers and acquisitions.

0

1 0

2 0

3 0

4 0

5 0

6 0

7 0

8 0

U . S . S p a i n France Germany Japan Canada U . K .

1 9 8 5

1 9 9 0

1 9 9 5

2 0 0 0

Source: Authors’ calculations based on Bank for International Settlements (BIS) data.1The seven selected BIS-reporting home countries are Canada, France, Germany, Japan, Spain, the


Figure 4. The Exposure of Selected BIS-Reporting Countries’ Banks to Latin America’s Private Sector1

(Claims to Latin America as a percentage of each country’s total international private sector claims)


claims (especially Goldberg, 2002) and on the extensive literature on capitalflows.19 These studies estimate reduced-form models that take into account therole of both home or push and host or pull factors. Home or push factors are con-sidered to be exogenous to the host country and refer to structural or cyclical fea-tures of the home countries, which affect banks’ desire to invest abroad. Homecountry interest rates and growth rates have been commonly used to proxy for therole of push factors (see, for example, Calvo, Leiderman, and Reinhart, 1993;Chuhan, Claessens, and Mamingi, 1998; Goldberg, 2002; and Hernandez, Mellado,and Valdes, 2001). On the other hand, pull factors refer to host country characteris-tics that affect the risk-return trade-off of investing in these countries. Countrycredit ratings and host growth rates are among the most frequently used pull factors(see Chuhan, Claessens, and Mamingi, 1998; Goldberg, 2002; and Hernandez,Mellado, and Valdes, 2001).

Following the literature discussed above, to account for host factors, weinclude the real GDP growth, the change in country risk rating, and a dummy cap-turing crisis episodes in each of the Latin American host countries. As home fac-tors, we include the real GDP growth and real interest for each of the seven homecountries.20 Growth and interest rate figures come from the International FinancialStatistics (IFS), published by the International Monetary Fund. The credit ratingsused are those reported by Institutional Investor magazine.21,22 This rating takesvalues between 0 to 100, with higher numbers representing a better repaymentcapacity on the part of the host country. The crisis dummy variable equals 1 duringbanking, currency, or twin crisis periods. A chronology of crises in the region wasobtained from Caprio and Klingebiel (1999) and Bordo and others (2001).23

Aside from the impact of push and pull factors on foreign bank claims to eachhost, we also take into account the role of exposure to each host and the potentialinfluence of movements in foreign claims to other countries. To test the impact ofexposure on how foreign banks react to host country shocks, we interact variablescapturing host factors with an aggregate measure of banks’ exposure. Exposure isthe ratio of home country j’s bank claims on the private sector of host country i tothe total private sector claims extended by country j’s banks worldwide. This ratiois calculated from the BIS’s Consolidated Banking Statistics.

19See Calvo, Leiderman, and Reinhart (1993); Fernandez-Arias (1996); Chuhan, Claessens, andMamingi (1998); and Hernandez, Mellado, and Valdes (2001).

20For example, in modeling the behavior of Canadian claims to Latin America, we allow for CanadianGDP and interest rates to affect changes in these claims, but economic conditions from other home coun-tries are not assumed to enter the regression for Canadian claims.

21Institutional Investor magazine publishes a semiannual survey of country credit ratings. The maga-zine surveys bankers, money managers, and economists around the world on their evaluations of the rela-tive risk of countries to which they lend. On the basis of their responses, the magazine produces a ratingfrom 0 to 100, with higher numbers representing a better repayment capacity. We use end-of-year ratings.

22In alternative specifications that are not shown but are available upon request, we replaced the creditrisk rating for a number of macro variables (government deficit, current account deficit, and real exchangerate appreciation, among others) that serve as proxies for country risk. Given that results were very simi-lar, we prefer this more parsimonious specification, which allows us to examine interaction effects andpositive and negative shocks more easily.

23See Table A.1 for a list of crises in each host country in the period 1985–2000.


439



440

To examine whether movements in claims to other countries spill over to indi-vidual hosts, we include as an explanatory variable the aggregate changes in claimsfrom home country banks to all non-BIS-reporting countries other than that indi-vidual host (%Δ Private Claims on Other Countries). This variable is also con-structed from the BIS’s Consolidated Banking Statistics.

Finally, because foreign bank claims are reported in U.S. dollars, we also con-trol for changes in the exchange rate vis-à-vis the dollar for each country.24

Exchange rate data also come from the IFS. Table 1 summarizes the definition andsources of all the variables included in equation (1).

Foreign banks are less likely to extend claims abroad if the riskiness or thereturns obtained from the host countries worsen. Lower host growth or ratingdowngrades should then lead to a decline in claims to the affected host country.Hence, we expect to find a positive coefficient on the growth and rating variables,consistent with what the literature on capital flows has found.25

A priori, we might expect banking, currency, and/or twin crisis episodes in aparticular host country to be accompanied by a decline in foreign bank claims,since these episodes are typically associated with a fall in the capacity of crises-stricken countries to repay their obligations. On the other hand, foreign banksmight view crises in host countries as an opportunity to expand their operations andincrease their market share locally.26 Also, crises might coincide with a deteriora-tion in economic fundamentals such as GDP growth, making their impact indistin-guishable from that of other cyclical downturns. In other words, it is possible thatthe crisis dummy in our regressions may not be significant because the impact ofthese episodes is being captured by changes in host GDP growth. This, in turn,would suggest that crises are not perceived as different from any other cyclicaldownturn in output.

In principle, given the overall importance of foreign claims to the region,changes in such claims could affect host countries’ right-hand-side variables (forexample, host real GDP growth, timing of crises, and credit rating), implying apotential endogeneity problem. However, we believe that the scope for this is lim-ited, since our estimations focus on bilateral claims (that is, changes in real claims

24Because the BIS data are denominated in U.S. dollars and exchange rates vis-à-vis the dollar havebeen volatile in Latin America, one could be concerned that exchange rate movements are disproportion-ately affecting the measured behavior of foreign bank claims. However, we believe that this should not bea serious issue for two main reasons. First, our analysis focuses on international claims, which includecross-border claims (denominated in any currency) and local claims (that is, those issued by foreign banksubsidiaries and branches) denominated in foreign currency. Thus, since local claims in local currency arenot included in our study, the concern that some of the foreign claims that we analyze might have origi-nated in the volatile host country currency is small in our view. This could occur only if some of the cross-border claims were denominated in the local currency, which seems unlikely. Second, while some of thecross-border claims could have originated in a home currency other than the dollar, some rough estima-tions, using data from the BIS Locational Statistics, indicate that for all countries in our sample, the aver-age share of assets denominated in dollars was close to 80 percent or higher during the sample period weconsider.

25For example, Chuhan, Claessens, and Mamingi (1998) find credit ratings to have a positive impacton portfolio flows to Asia and Latin America. In turn, Hernandez, Mellado, and Valdes (2001) find hostGDP growth to have a similar effect on private capital flows to a larger sample of developing countries.

26This argument is made by Peek and Rosengren (2000).


Table 1. Data Definition and Sources

Variable Definition Source

Dependent variable%Δclaimsj,i,t

Independent—host country—variablesHost country real

GDP growth

Host country real GDP growth ×exposure to host

%Δ(Host country rating)

%Δ(Host country rating) × exposure to host

Host crisis dummy

Host crisis dummy ×exposure to host

%Δ(Host local currency/US$ exchange rate)

Independent—home country—variablesHome j real GDP growth ×

home country j

Home j real interest rate ×home country j

%Δ(Home local currency j/US$ exchange rate)

%Δ(Private real claims on other countries)

Notes: *BIS stands for Bank for International Settlements. **IFS stands for International Finan-cial Statistics, an International Monetary Fund publication.

Percentage change in claims from homecountry j banks on the private sector inhost country i at time t.

Real GDP growth in host country i at timet − 1 where i stands for Argentina,Brazil, Chile, Colombia, Costa Rica,Ecuador, Mexico, Peru, Uruguay, andVenezuela, respectively.

Real GDP growth in host i interacted withhome country j banks’ exposure to i,where i stands for Argentina, Brazil,Chile, Colombia, Costa Rica, Ecuador,Mexico, Peru, Uruguay, and Venezuelaand j refers to Canada, France, Germany,Japan, Spain, United Kingdom andUnited States, respectively.

Percentage change in host country i creditrisk rating, where i is defined above.

Percentage change in host i credit ratingtimes home country j banks’ exposureto i, where i and j are defined above.

Dummy equal to 1 when host country i hasa crisis, where i is defined above. SeeTable A.1 for a list of crisis episodes.

Dummy equal to 1 when host country ihas a crisis times home country j banks’exposure to i, where i and j are definedabove.

Percentage change in the dollar exchangerate vis-à-vis host i’s currency, where iis defined above.

Home country j real GDP growthinteracted with dummy for homecountry j, where j is defined above.

Home country j real interest rateinteracted with dummy for homecountry j, where j is defined above.

Percentage change in the dollar exchangerate vis-à-vis home country j’s currency,where j is defined above.

Percentage change in home country j bankclaims on countries other than host i,where j and i are defined above.

BIS ConsolidatedBanking Statistics*

IFS**

IFS and BIS

Institutional Investormagazine

Institutional Investorand BIS

Caprio andKlingebiel (1999)

Caprio andKlingebiel (1999)and BIS

IFS

IFS

IFS

IFS

BIS


441



442

from banks in home country j on host country i), and no bilateral relationship isimportant enough to warrant such concern. Nonetheless, as a precaution, all right-hand-side variables are lagged one period (one year).27 Also, to mitigate the con-cern that changes in foreign bank claims from different home countries might bedriven by the same events or news (for example, on the health of the world econ-omy), our robustness tests include time dummies to capture such factors.

The impact of host shocks on foreign bank claims might be affected by thedegree to which foreign banks are exposed to that host. On the one hand, the largerthe exposure of foreign banks to a particular country, the more procyclical (that is,the more responsive to host conditions) foreign claims might become. This mightbe due to a lack of diversification. On the other hand, as banks’ exposure to a coun-try grows, banks might have more incentives to learn about host country conditionsand, hence, not to respond as strongly to signals of good or bad future events.28

Alternatively, it could be the case that greater exposure (especially in the case ofbrick-and-mortar operations) might signal a stronger commitment to the host,which also gets translated into a smaller reaction to host shocks.29 Hence, thereare reasons to expect that foreign bank claims might become more stable or lessresponsive to host shocks as exposure rises.

To test the impact of exposure on host factors, we interact host country vari-ables (the change in ratings, the real growth, and the crisis indicator for host coun-try i) with a measure of foreign banks’ exposure to the country. A priori, if indeedhigher exposure is translated into more stable financing, we expect these interactionterms to be opposite in sign to that of the host country shock. For example, weexpect the interaction between host growth (or changes in host rating) and exposureto be negative and the interaction between host crisis and exposure to be positive.

Studies such as Van Rijckeghem and Weder (2003) have shown that there isscope for contagion in international banking. In particular, they show that changesin foreign bank claims on one country might spill over to other countries that holdclaims from the same foreign banks. Furthermore, models of portfolio allocationshow that under standard rules of portfolio choice an unexpected decline in thevalue of one or more assets may provoke a portfolio adjustment across the board.30

Because our data aggregate bank positions at the country level, we cannot conducta strict test of portfolio effects at the bank level. Nevertheless, we seek to verifywhether at least in the aggregate there is evidence that banks’ changes in claims onother countries affect individual hosts. If this were the case, we would expect to findthat the variable %ΔPrivate Claims on Other Countries is positive and significant.

27We also conducted estimations including all regressors contemporaneously and found that our mainresults do not change. These estimations are available upon request.

28Calvo and Mendoza (2000) argue that as investors become more diversified, and hence their aver-age exposures in any particular asset decrease, they have reduced incentives to learn about the fundamen-tals of each asset, and hence react more strongly to signals on expected return or risk. This suggests thatas foreign banks become more exposed to a particular host country, they may react less to changes in hostcountry variables.

29The argument that foreign bank brick-and-mortar presence signals a greater commitment to the hostcountry is made in Palmer (2000) and Peek and Rosengren (2000).

30Schinasi and Smith (1999) discuss optimal portfolio rebalancing as “contagion” after different typesof shocks to expected asset returns and variances.


Home country economic conditions could have a negative or a positive impacton foreign bank lending to host countries. On the one hand, adverse economic con-ditions and a lack of profit opportunities at home could encourage banks to holdclaims abroad. If this were the case, we would expect to find a negative coefficienton home growth. On the other hand, a recession at home could lead to a deteriora-tion in the capital of foreign banks and an overall retrenchment in claims held athome and abroad. Therefore, we remain agnostic regarding the sign of this variable.

Low real interest rates in lender countries tend to signal periods of excess li-quidity, and this might increase banks’ willingness to extend riskier, higher interestrate claims to developing countries. Therefore, we expect home real interest ratesto have a negative impact on the change in claims to countries in Latin America.

To deepen our understanding of the determinants of foreign bank claims underdifferent circumstances, we estimate some modified versions of equation (1). First,we examine whether banks’ responsiveness to shocks depends on the type of shockexperienced, by allowing the coefficients in equation (1) to vary on the basis ofwhether the host country undergoes positive or negative shocks.31 Second, weexplore whether there is evidence that the determinants of foreign bank claims havechanged over time by estimating equation (1) over two subsamples: 1985–94 and1995–2000.32 In particular, we examine whether banks’ sensitivity to host andhome conditions and to aggregate movements in claims changed over this period.Finally, to summarize the relative importance of home and host conditions and ofaggregate shocks in claims, we report the percentage of the variance of claimsexplained by each of these factors, and we study whether it has changed over time.To the extent that home country conditions and aggregate shocks in claims consis-tently dominate host country variables in explaining changes in claims, we wouldconclude that foreign banks have the potential to destabilize host countries by trans-mitting shocks external to these economies.

III. Empirical Results

Starting from a fully unrestricted model where all regressors are allowed to be dif-ferent depending on the banks’ home country, we tested a number of coefficientrestrictions on equation (1) until we arrived at our baseline specifications reportedin Table 2.33 Home and host dummies and exchange rate changes vis-à-vis theU.S. dollar are included in these specifications but are not shown owing to spaceconstraints.

The baseline model accepted in the specification tests reported in the appendixis one where host factors (growth in host GDP, change in credit ratings, the exchange

31We do not investigate positive/negative home growth shocks, because for the seven home countrieswe focus on, there have been virtually no years in which home growth has been negative.

32These two subsamples are selected on the basis of papers such as Crystal, Dages, and Goldberg(2001) and García (2002) that argue that foreign bank presence in Latin America rose after 1995.

33Essentially, the fully unrestricted model is equivalent to estimating a separate equation for eachlender/home country, including its corresponding home factors, a matrix of host factors, and a variablecapturing changes in claims to other countries. The fully unrestricted model is shown in Table A.2 and therestriction tests are shown in Table A.3.


443



444

Table 2. The Determinants of the Percentage Change in Foreign Bank Claims on Latin America

1985–2000 1985–1994 1995–2000 Coefficient Coefficient Coefficient

Variable (t-Statistic) (t-Statistic) (t-Statistic)

Host country variablesHost country real GDP growth 2.120 2.244 1.157

(2.900)*** (2.990)*** (0.730)Host country real GDP growth × −0.210 −0.242 0.050

exposure to host (2.120)** (2.340)** (0.230)%Δ(Host country rating) 1.142 0.062 2.279

(2.460)** (0.210) (1.960)*%Δ(Host country rating) × −0.093 0.029 −0.380

exposure to host (1.490) (0.560) (2.150)**Host crisis dummy −3.409 −6.855 −0.639

(0.640) (1.190) (0.060)Host crisis dummy × 0.760 0.013 0.349

exposure to host (0.660) (0.010) (0.180)

Home country variablesCanadian real GDP growth × 8.544 1.942 −6.327

Canada (2.010)** (0.550) (0.290)French real GDP growth × −1.759 −3.501 −1.953

France (0.860) (1.490) (0.380)German real GDP growth × −2.140 −2.822 −22.063

Germany (0.840) (1.050) (1.600)Japanese real GDP growth × 5.370 −0.809 11.884

Japan (1.670)* (0.250) (2.460)**Spanish real GDP growth × −4.815 −2.185 −4.579

Spain (0.720) (0.300) (0.280)U.K. real GDP growth × U.K. −2.928 −8.975

(0.180) (0.510)U.S. real GDP growth × U.S. −5.399 −10.331 −1.623

(2.200)** (3.760)*** (0.190)Canadian real interest rate × −15.551 −7.299 −6.610

Canada (3.520)*** (1.820)* (0.400)French real interest rate × −0.204 −1.631 −11.591

France (0.130) (0.410) (1.650)*German real interest rate × 0.468 4.859 −7.393

Germany (0.170) (1.230) (0.400)Japanese real interest rate × −11.248 −11.576 −9.544

Japan (2.090)** (1.010) (0.860)Spanish real interest rate × −0.458 1.160 55.259

Spain (0.180) (0.460) (2.660)***U.K. real interest rate × U.K. 17.733 17.890

(1.070) (0.940)U.S. real interest rate × U.S. −7.122 −8.053 1.028

(2.740)*** (3.030)*** (0.110)%Δ(Private real claims on 0.146 0.130 0.347

other countries) (3.130)*** (2.200)** (0.680)


rate vis-à-vis the dollar, and the crisis dummy) and aggregate changes in claims (oncountries other than the host in question) are constrained to affect foreign bankclaims from all home countries in the same way. On the other hand, these tests sug-gest that banks from different home countries respond differently to shocks to theirown growth and interest rates. In other words, we cannot impose the restriction thatthe coefficients on home growth and real interest rates are the same across banksfrom different home countries (see Table A.3). Thus, in the estimations in Table 2,each home country variable is interacted with the corresponding home countrydummy.

The first column in Table 2 presents the selected restricted model for the wholesample period, 1985–2000. We find evidence that foreign banks responded tohome, host, and aggregate claims shocks. Focusing on the subset of home countryj variables, we find that banks from France, Germany, Spain, the United Kingdom,and the United States reduced claims in response to increased profit opportunitiesat home (that is, in response to higher home growth), but only the coefficient onU.S. growth is significant with a negative sign. Home growth has a positive and sig-nificant effect for Canadian bank claims. With the exception of banks fromGermany and the United Kingdom, the home real interest rate has the expected neg-ative impact. This variable is statistically significant for Canada, Japan, and theUnited States.34

34The finding that only some of the home variables are significant might arise from the fact that thesevariables tend to be significantly correlated within and across home countries.

Table 2. (Concluded)


Variable (t-Statistic) (t-Statistic) (t-Statistic)

Number of observations 804 426 378Adjusted R-squared 0.13 0.10 0.19F-test for significance of the 5.59*** 5.56*** 3.84***

overall regression

Source: Authors’ calculations.Notes: This table reports the estimates from a selected (restricted) version of equation (1),

according to the F-tests reported in Table A.3. The model assumes that banks’ reactions to host con-ditions is the same across home countries, but that their response to their home conditions is differ-ent across home country, hence each home country variable is interacted with a home countrydummy. The model estimated is equivalent to estimating a separate equation for each type of bank(that is Canadian, French, German, Japanese, Spanish, U.K. and U.S. banks) where the coefficientson the host variables are constrained to be the same across equations (that is, across home countries).The United Kingdom is omitted for the period 1985–1994, because data on private sector claims arenot available prior to 1993. Home dummies and home and host exchange rate changes are includedbut not shown. Robust t-statistics are in parentheses.

*, **, *** denote significance at 10, 5, and 1 percent, respectively.


445



446

Among the subset of host country variables, we find that the coefficient onhost growth is positive and significant, showing that foreign banks responded tohost country growth, increasing and decreasing claims over the cycle. However,we also find strong support for the notion that claims’ procyclicality (that is, sen-sitivity to host growth) falls as foreign banks’ exposure to the host rises. The co-efficient on host credit rating is positive and significant, while the interaction termof rating with exposure is negative but not significant. Controlling for host coun-try growth and risk ratings, the crisis variable is not significant. Therefore, it doesnot appear that crisis episodes cause any further decline in foreign bank claims.35

Finally, the coefficient on changes in private claims on other countries is positiveand significant, indicating that aggregate changes in foreign bank claims spill overto individual hosts.

Between 1995 and 2000, foreign bank penetration in Latin America—the par-ticipation of foreign banks in the local banking market—increased significantly. In1995, foreign bank loans (in local and foreign currency) represented approximately15 percent of total bank loans to the region. By 2000 this figure rose to 38 percent.36

Some have speculated that an increase in foreign banks’ brick-and-mortar opera-tions in developing countries could signal a stronger commitment to their hosts,since this type of presence makes it both physically and reputationally harder forforeign banks to “run” in the face of adversity at home, in the host country, or else-where in their portfolio (see Palmer, 2000; and Peek and Rosengren, 2000).

Ideally, to investigate whether the responsiveness of foreign banks to shockschanged as their brick-and-mortar investment (or local claims) in these countriesincreased, we would want to control for the ratio of local foreign bank claims tototal claims (that is, claims extended through brick-and-mortar operations as ashare of overall foreign bank claims). However, since such data are not consistentlyavailable, we are able to examine this issue only indirectly, by comparing the esti-mates of our model over two subsamples: 1985–94, when foreign participation wasin general low and in some instances prohibited, and 1995–2000, when foreignbank presence took off.37,38 Because this is a very indirect way of exploring this

35This result is robust including both the crisis dummy as well as other regressors contemporaneouslyrather than lagged. Also, the result holds if we separately control for banking, currency, and twin crises.

36These figures, which include the share of loans held by all foreign banks operating in all of LatinAmerica, come from Salomon Smith Barney (2000).

37BIS statistics lump together local claims in foreign currency with cross-border claims; therefore, itis not possible to calculate the importance of brick-and-mortar operations (that is, the sum of all localclaims, in foreign and domestic currency) as a share of total foreign bank claims. Another problem is thatlocal claims in local currency statistics are not broken down by sector (private, public, and so forth). Onecould in principle construct a series of foreign bank local claims on each Latin American host by aggre-gating balance sheet data for individual banks from Bankscope or from bank superintendencies in thehost countries. The problem with combining this data with the BIS data is that while the latter nets outinteroffice positions (the positions of foreign bank offices worldwide), the balance sheet data mentionedabove would not. Furthermore, banks’ balance sheet information for the period 1985–2000 is generallynot available.

38We separate the sample in the period before and after 1995, following Crystal, Dages, and Goldberg(2001) and García (2002), who argue that the increase in foreign bank presence in Latin America occurredafter this year.


issue, we need to be cautious in interpreting our results, especially given that manyother factors aside from the share of brick-and-mortar presence may have changedin the region over the period 1985–2000. Also, caution is warranted consideringthat the period of analysis is relatively short. Nevertheless, because the question isan interesting one, we present these estimations to illustrate some of the potentialissues that will have to be verified by future research.

For the period 1985–1994, we find that host real growth played a significantrole in explaining movements in real claims (see Table 2). The coefficient on thedummy variable capturing crises in the host countries is negative but not signifi-cant. Home factors appear to be significant only for Canada and the United States.Finally, the coefficient on the change in real claims on other countries is both pos-itive and significant, indicating that changes in claims on other countries affectedspecific hosts.

For the period 1995–2000, we find that in contrast to the results for the earlierperiod, there is no significant evidence that changes in claims elsewhere were trans-mitted to the host countries we focus on. Home factors continued to be significant,but only for Spain and Japan. As for the host factors, there is no evidence that for-eign banks retrenched at times of crises during this period either. On the other hand,changes in credit ratings had a positive and significant impact on foreign bank lend-ing. However, foreign banks’ responsiveness to this variable seems to decrease withexposure.

All in all, the findings from splitting the sample into the periods pre- and post-1995 suggest that in recent years foreign banks have become less inclined to cutand run when faced with adversity in the region or abroad. Whether this is the resultof the late 1990s increase in foreign bank brick-and-mortar presence will have tobe examined more thoroughly and directly in future research.

Not only is it possible that foreign bank behavior changes as their brick-and-mortar presence increases, but it is also feasible that their reaction to shocks dependsasymmetrically on the nature of these events. To examine this possibility, we dis-criminate between positive and negative changes in host GDP growth, host creditratings, and in other claims. Table 3 presents the results from this estimation for theoverall sample, 1985–2000.39 We find that while positive changes in host real GDPgrowth have a positive and significant sign, the coefficient on negative host GDPgrowth is negative but not significant. The same is true for credit ratings: claimsrespond to upgrades and not to downgrades in credit ratings. However, the higherthe exposure to a host country, the smaller the response of claims to upgrades incredit ratings, as indicated by the significant and negative coefficient on the inter-action term between upgrades and exposure. On the other hand, both positive andnegative changes in aggregate claims on other countries are statistically significant,but negative changes have a much stronger impact, and that difference is statisti-cally significant at standard significance levels.

39Note that we define negative changes in absolute terms so that we can interpret a negative coeffi-cient as indicating that larger drops in the variable in question lead to a decline in the growth of claims.


447



448

Table 3. The Impact of Positive and Negative Shocks on the Change in Foreign Bank Claims on Latin America

Variable Coefficient t-Statistic

Host country variablesPositive host real GDP growth 1.679 (1.68)*Negative host real GDP growth −2.179 (1.59)Positive host real GDP growth × exposure to host −0.016 (0.13)Negative host real GDP growth × exposure to host 0.358 (1.46)Host country rating upgrade 1.665 (1.88)*Host country rating downgrade −0.036 (0.06)Host country rating upgrade × exposure to host −0.245 (2.53)**Host country rating downgrade × exposure to host −0.117 (2.11)**Host crisis dummy −5.385 (0.92)Host crisis dummy × exposure to host 1.100 (0.87)

Home country variablesIncrease in private real claims on other countries 0.133 (2.79)***Decrease in private real claims on other countries −0.922 (1.99)**Canadian real GDP growth × Canada 8.894 (2.04)**French real GDP growth × France −2.291 (1.09)German real GDP growth × Germany −2.140 (0.86)Japanese real GDP growth × Japan 5.483 (1.69)*Spanish real GDP growth × Spain −5.889 (0.87)U.K. real GDP growth × U.K. −1.136 (0.07)U.S. real GDP growth × U.S. −4.618 (1.87)*Canadian real interest rate × Canada −14.433 (3.14)***French real interest rate × France −0.118 (0.07)German real interest rate × Germany −0.333 (0.12)Japanese real interest rate × Japan −13.031 (2.43)**Spanish real interest rate × Spain −0.110 (0.04)U.K. real interest rate × U.K. 21.010 (1.20)U.S. real interest rate × U.S. −5.372 (1.77)*

Number of observations 804Adjusted R-squared 0.13F-test for significance of the overall regression 6.15Prob > F 0.00

Source: Authors’ calculations.Notes: This table presents the results from a model where the impact of host GDP changes, rat-

ing changes, and changes in other claims is allowed to vary depending on the positive or negativenature of the shocks. The period of estimation is 1985–2000. Home dummies and host and homeexchange rate changes are included, but not shown. Growth rates and changes in ratings and claimsare expressed in terms of absolute values. *, **, ***denote significance at 10, 5, and 1 percent,respectively. t-statistics are obtained on the basis of robust standard errors.

A useful way of summarizing the relative importance of home, host, and aggre-gate claim shocks is provided in Table 4, which details the percentage of the vari-ance in private sector claims explained by each of these groups of variables. Inother words, for each group of variables, we compute the increase in the R-squared,


as a proportion of the total variance of the percentage change in claims explainedby all variables.40,41

We find that while changes in aggregate claims on other countries explain a sig-nificant share of the variance in the dependent variable (21 percent) in the1985–1994 period, this variable plays practically no role in explaining changes inprivate sector claims on host countries in Latin America during the later period.Similarly, home country conditions explain a large proportion (62 percent) of thevariance in private claims in the period 1985–1994, but their importance declinessignificantly in the late 1990s. On the other hand, host country conditions explainbetween 20 and 50 percent of the variance in claims in both periods, and, over-whelmingly, it is positive changes (positive growth and credit rating upgrades) thatplay the most significant role in explaining changes in international financialclaims. These patterns are observed comparing two periods over which the brick-and-mortar operations of foreign banks increased significantly in importance. Onceagain, whether the increase in brick-and-mortar presence is responsible for thesepatterns will have to be established more conclusively by future research.

40We rescale the percentage explained by each group of variables so that the sum of all three is 100.41The R-squared statistics for these regressions appear to be low, yet they are in line with those

obtained in other studies on foreign bank lending flows also using BIS data, as well as in studies that exam-ine capital flows in general. See, for example, Chuhan, Claessens, and Mamingi (1998); Goldberg (2002);and Van Rijckeghem and Weder (2003).

Table 4. Percentage of Variance of Change in Foreign Bank ClaimsExplained by Home, Host, and Portfolio Shocks

Home Country Host Country %ΔPrivate Real Claims Variables Variables on Other Countries

Entire sample 46.78 31.41 21.81Positive changes 59.451 28.43 6.39Negative changes 1.69 4.031985–1994 61.63 17.76 20.611995–2000 49.56 48.59 1.85

Source: Authors’ calculations.Notes: This table reports the percentage of the variance of the change in foreign bank claims that

can be explained by home, host, and portfolio shocks. The percent variance explained is calculatedas (R2

_full–R2_constrained)/R2

_full *100. The home country variables included are real GDP growth, realinterest rates, and the home/dollar exchange rate. Host country variables included are: real GDPgrowth, credit rating, the host/dollar exchange rate, and the crisis dummy. Positive changes refer tocredit rating upgrades, host positive real GDP growth, and increases in other claims. Negativechanges refer to credit rating downgrades, negative real host GDP growth, and decreases in otherclaims. We rescale the percent of the variance explained by each set of variables so that for a givenestimation the sum of all three groups adds to 100.

1We are unable to split home variables into positive and negative shocks, since between 1985–2000there are no periods when home variables take negative values. So, essentially the negative/positive estimation corresponds to one where host variables and other claims, but not the home vari-ables, are split into positive and negative changes (see Table 3).


449



450

Robustness Checks

We conduct a number of additional estimations to verify the robustness of our mainfindings. First, we include time dummies to control for the impact of events that cansimultaneously affect all home countries (for example, good and bad news aboutthe world economy and crises outside Latin America). Second, we explore whetherchanges in foreign bank claims to Latin America are driven by the recent wave ofmergers and acquisitions of domestic banks by foreign banks. Since the BIS defi-nition of claims includes not only loans but also equity participations, we are inter-ested in determining whether changes in foreign claims are driven by bankconsolidation, as opposed to new lending. Finally, we use alternative definitions ofexposure and of aggregate changes in claims that take into account the fact thatbanks do not only hold claims on the nonbank private sector but also extend claimsto the public sector and to other financial institutions in the host countries.

The first three columns of Table 5 repeat the baseline specifications in Table 2,including yearly time dummies. With time dummies, we find that host GDP growthcontinues to be strongly significant for the overall period and the period 1985–1994. On the other hand, the credit rating is significant at only 10 percent now. Asbefore, Canadian and U.S. home factors drive claims on Latin America. Also, wecontinue to find no support for the argument that foreign banks are likely to cut andrun during crises. Finally, changes in claims on other countries are less significantthan when the time dummies are excluded.

The last three columns of Table 5 include a dummy to control for recentepisodes of mergers and acquisitions between foreign and domestic financial insti-tutions in our sample of countries.42 This dummy variable takes the value of 1 foreach annual bilateral observation, where there was a purchase of a host countrybank by a foreign bank from one of the seven home countries (there are 89 suchcases out of our total 804 observations). We find that the coefficient on this dummyis positive but not significant. Most importantly, including this variable does notchange our main results.

Finally, Table 6 presents the results for our preferred specification, but with amodified definition of exposure and of aggregate changes in claims. So far, expo-sure was defined as claims from banks from home j to the nonbank private sectorin host i as a ratio of all nonbank private sector claims held by banks from home j.This definition of exposure was driven by the fact that we wanted to focus exclu-sively on the determinants of the behavior of private sector claims, since claims onthe public sector might not always respond purely to risk-return factors. However,this measure of exposure does not reflect the fact that in reality foreign banks lendto the public and banking sectors as well, and that these claims are part of the banks’overall exposure and may affect private sector claims. Therefore, we also report esti-mations considering private sector claims from banks in home j to host i as a ratioof all international financial claims from banks from home j, regardless of the

42Information on all acquisitions by banks from each of the home countries considered in our study inour 10 host countries (sale or merger of an entity in host country i to an entity from home country j in yeart) was obtained from Thomson Financials.


Table 5. The Determinants of the Change in Foreign Bank Claims Including Time Effects and Mergers and Acquisitions

(5.1) (5.2) (5.3) (5.4) (5.5) (5.6)

1985–2000 1985–1994 1995–2000 1985–2000 with 1985–1994 with 1995–2000 with Controlling Controlling Controlling Time Dummies Time Dummies Time Dummies for M&As for M&As for M&As

Coefficient Coefficient Coefficient Coefficient Coefficient Coefficient Variables (t-Statistic) (t-Statistic) (t-Statistic) (t-Statistic) (t-Statistic) (t-Statistic)

Mergers and acquisitions (M&As) dummy 11.065 10.996 2.691(1.37) (0.73) (0.26)

Host country variablesHost real GDP growth 2.495 2.432 0.871 2.199 2.294 1.158

(2.98)*** (3.02)*** (0.48) (2.98)*** (3.08)*** (0.73)Host real GDP growth × exposure to host −0.178 −0.228 0.070 −0.229 −0.244 0.037

(1.77)* (2.16)** (0.31) (2.31)** (2.35)** (0.16)%Δ(Host country rating) 0.799 0.117 1.950 1.113 0.041 2.275

(1.70)* (0.42) (1.77)* (2.39)** (0.15) (1.94)*%Δ(Host country rating) × exposure to host −0.080 0.015 −0.394 −0.093 0.028 −0.383

(1.26) (0.27) (2.17)** (1.51) (0.55) (2.18)**Host crisis dummy −7.615 −5.288 −5.619 −3.439 −6.924 −0.558

(1.31) (0.86) (0.48) (0.64) (1.21) (0.05)Host crisis dummy × exposure to host 0.101 −0.067 0.401 0.619 0.079 0.257

(0.09) (0.06) (0.19) (0.55) (0.08) (0.13)

Home country variablesCanadian real GDP growth × Canada 5.634 2.231 −0.071 8.719 2.122 −6.111

(1.28) (0.50) (0.00) (2.05)** (0.59) (0.28)French real GDP growth × France 2.169 −3.856 18.017 −1.489 −3.605 −1.693

(0.67) (1.01) (1.31) (0.72) (1.52) (0.32)German real GDP growth × Germany −2.200 −2.742 −2.208 −2.272 −2.841 −22.313

(0.60) (0.64) (0.10) (0.90) (1.06) (1.62)Japanese real GDP growth × Japan 2.804 −0.428 5.976 5.347 −0.864 11.877

(0.80) (0.09) (1.08) (1.66) (0.27) (2.46)**

BAN

KIN

G O

N FO

REIG

NER

S: THE BEH

AV

IOR

OF IN

TERN

ATIO

NA

L BAN

K C

LAIM

S

45

1


Ma

ria So

led

ad

Ma

rtine

z Pe

ria, A

nd

rew

Po

we

ll, an

d Iva

nn

a V

lad

kova

-Ho

llar

45

2


(5.1) (5.2) (5.3) (5.4) (5.5) (5.6)

1985–2000 1985–1994 1995–2000 1985–2000 with 1985–1994 with 1995–2000 with Controlling Controlling Controlling Time Dummies Time Dummies Time Dummies for M&As for M&As for M&As

Coefficient Coefficient Coefficient Coefficient Coefficient Coefficient Variables (t-Statistic) (t-Statistic) (t-Statistic) (t-Statistic) (t-Statistic) (t-Statistic)

Spanish real GDP growth × Spain −4.779 −2.764 −9.841 −4.826 −1.857 −5.012(0.66) (0.33) (0.45) (0.72) (0.25) (0.31)

U.K. real GDP growth × U.K. 3.949 19.264 −2.95 −8.963(0.20) (0.60) (0.18) (0.5)

U.S. real GDP growth × U.S. −9.686 −11.386 30.000 −5.624 −10.006 −1.577(2.52)** (2.80)*** (0.97) (2.30)** (3.62)*** (0.18)

Canadian real interest rate × Canada −11.646 −8.381 −8.606 −15.502 −7.247 −6.716(2.47)** (1.87)* (0.54) (3.50)*** (1.82)* (0.40)

French real interest rate × France 5.074 −3.289 −15.900 0.232 −1.965 −11.421(1.96)** (0.54) (1.44) (0.14) (0.48) (1.61)

German real interest rate × Germany 7.151 6.037 −13.863 0.661 4.905 −7.166(1.52) (0.98) (0.50) (0.24) (1.24) (0.39)

Japanese real interest rate × Japan −1.936 −6.638 −9.230 −11.234 −11.499 −9.484(0.31) (0.53) (0.68) (2.08)** (1.00) (0.85)

Spanish real interest rate × Spain 2.222 0.957 32.273 0.001 1.021 55.013(0.75) (0.32) (1.21) (0.00) (0.39) (2.69)***

U.K. real interest rate × U.K. 31.905 60.241 16.43 17.528(1.32) (1.37) (0.98) (0.91)

U.S. real interest rate × U.S. −5.746 −6.132 3.436 −7.396 −8.285 0.803(1.55) (1.35) (0.35) (2.86)*** (3.15)*** (0.08)

%Δ(Private real claims on other countries) 0.090 0.118 0.494 0.141 0.129 0.35(1.95)* (1.93)* (0.77) (2.96)*** (2.18)** (0.69)

Number of observations 804 426 378 804 426 378Adjusted R-squared 0.13 0.09 0.20 0.16 0.16 0.26

Source: Authors’ calculations.Notes: Columns (5.1)–(5.3) report the estimates from the selected (restricted) version of equation (1), according to the F-tests reported in Table A.3, including

time dummies (not shown). Columns (5.4)–(5.6) report the estimates from the selected version of equation (1), according to the F-tests reported in Table A.3, includ-ing a dummy to control for known episodes of mergers and acquisitions involving banks from the home countries with banks in the host countries. t-statistics are inparentheses (calculated on the basis of robust standard errors). *, **, *** denote significance at 10, 5, and 1 percent, respectively.


Table 6. The Determinants of the Change in Foreign Bank Claims with a Modified Definition of Total Claims


Variable (t-statistic) (t-statistic) (t-statistic)

Host country variablesHost real GDP growth 1.975 1.888 1.292

(2.73) (2.59)*** (0.79)Host real GDP growth × exposure to host −0.243 −0.205 −0.0001

(2.28)** (1.84)* (0.00)%Δ(Host country rating) 0.843 −0.301 2.247

(1.85)* (0.99) (1.97)**%Δ(Host country rating) × exposure to host −0.029 0.102 −0.402

(0.48) (2.53)** (2.07)**Host crisis dummy −3.124 −11.075 1.934

(0.58) (1.87)* (0.19)Host crisis dummy × exposure to host 0.127 0.789 −0.540

(0.16) (0.86) (0.40)

Home country variablesCanadian real GDP growth × Canada 8.003 1.837 0.596

(1.93)* (0.50) (0.02)French real GDP growth × France −1.626 −3.671 1.349

(0.75) (1.51) (0.22)German real GDP growth × Germany −1.488 −3.203 −18.246

(0.59) (1.22) (1.42)Japanese real GDP growth × Japan 4.492 −4.028 10.481

(1.42) (1.41) (2.06)**Spanish real GDP growth × Spain −2.827 0.736 −4.983

(0.43) (0.10) (0.30)U.K. real GDP growth × U.K. −5.024 −6.662

(0.29) (0.38)U.S. real GDP growth × U.S. −7.577 −11.256 −2.103

(2.82)*** (3.88)*** (0.23)Canadian real interest rate × Canada −14.799 −6.680 −12.221

(3.64)*** (1.61) (0.63)French real interest rate × France 0.190 −0.829 −10.463

(0.11) (0.21) (1.49)German real interest rate × Germany −1.560 4.865 −3.686

(0.58) (1.20) (0.20)Japanese real interest rate × Japan −9.170 −1.469 −8.697

(1.86)* (0.16) (0.80)Spanish real interest rate × Spain 0.069 1.007 50.504

(0.03) (0.39) (2.27)**U.K. real interest rate × U.K. −6.406 −11.914 18.746

(0.37) (1.58) (1.05)U.S. real interest rate × U.S. −5.378 −6.337 1.768

(1.77)* (2.01)** (0.18)%Δ(Total claims on other countries) 0.529 0.551 0.456

(1.83)* (1.28) (0.73)

Observations 810 435 375Adjusted R-squared 0.16 0.14 0.26F-test significance of the overall regression 5.20 5.00 3.61Prob > F 0.00*** 0.00*** 0.00***

Source: Authors’ calculations.Notes: This table reports the estimates from the selected model according to the F-tests reported

in Table A.3. Exposure and other claims are defined not in terms of claims to the private sector, butrather include total (private + public) claims. t-statistics are in parentheses (calculated on the basis ofrobust standard errors). *, **, *** denote significance at 10, 5, and 1 percent, respectively.



454

sector. Similarly, we replace our measure of aggregate claim shocks (%ΔPrivateClaims on Other Countries) for one that considers all international financial claimsrather than only those to the private sector (%ΔTotal Claims on Other Countries).As Table 6 shows, our main results are not driven by our definition of exposure orof aggregate changes in claims. Including the revised definition for these variablesthat considers all claims (public and private) does not alter our main findings inTable 2.

IV. Conclusions

As the trend toward greater international financial integration persists, the debateon the behavior of foreign banks is likely to continue. In this paper, we employeda comprehensive data set to explore the behavior of foreign bank claims to LatinAmerica, a region that witnessed a significant increase in foreign bank financingduring the 1990s. Our data set is rich in two dimensions. From a cross-sectionalperspective, we captured the behavior of banks from a number of home countriesthat differ both in their degree of exposure and in their importance as a source offinance to Latin America. From a time-series perspective, our data set allowed usto focus on periods of tranquility as well as on periods of crises, on periods of lowforeign bank penetration and on periods of strong brick-and-mortar presence.

Nevertheless, the data used in this paper have limitations. First, the data areaggregated at the country level. Lack of individual bank-level data precludes usfrom formally testing portfolio-balancing effects, among other things. Second, be-cause we lack separate data on local and cross-border claims, we cannot test theimpact of the increase in foreign bank brick-and-mortar presence on the behaviorof foreign bank claims directly. Finally, the time period considered, though longerthan in some of the previous studies, is still relatively short.

While this paper cannot answer all questions related to the behavior of foreignbanks, it contributes to this debate by highlighting some of the consequences ofrelying on foreign bank financing and allowing foreign bank entry in developingcountries. We found that “banking with foreigners” has the potential of makingdeveloping countries vulnerable to home country shocks. We also found that move-ments in claims to other countries may spill over to individual host countries. Thegood news, though, is that the importance of these external shocks appears to havedeclined in recent years.

Regarding the impact of host shocks, we found that while foreign banks respondto host conditions, they do not appear to pull out faster at times of crises or duringother periods of economic downturn. Furthermore, higher foreign bank exposureappears to be a stabilizing force, since we found that foreign banks’ responsivenessto host conditions becomes less procyclical as exposure increases.

Our results suggest a number of policy implications that will need to be con-firmed with further research. First, though further analysis would be required toestablish a clear causal relation, it appears that favoring brick-and-mortar presenceover cross-border lending might be a sensible policy to limit the scope for externalshocks that may result from relying on foreign bank financing. Second, regardingthe issue of claims’ procyclicality, our results suggest that countries dependent on


foreign bank financing might benefit more from dealing with a small number offoreign banks with high exposures than with a large number of banks with lowexposure and no commitment to the country. Finally, our finding that foreign bankclaims do not retrench during crises is consistent with the view that foreign bankfinancing and entry should be promoted on the basis that it can be have a stabiliz-ing influence on credit growth during crisis periods and in their aftermath.


455


Ma

ria So

led

ad

Ma

rtine

z Pe

ria, A

nd

rew

Po

we

ll, an

d Iva

nn

a V

lad

kova

-Ho

llar

45

6

Table A.1. Crises Classification: Banking and Currency Crises in the 10 Latin American Hosts, 1985–2000

Host Country 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000

Argentina CC BC, BC CC BC,CC CC

Brazil CC BC, BC BC BC CCCC

Chile

Colombia BC BC BC BC,CC

Costa Rica BC BC BC BC BC

Ecuador CC CC BC BC BC BC, BCCC

Mexico CC BC, BC, BCCC CC

Peru BC BC BC BC, BC, CCCC CC

Uruguay CC

Venezuela BC BC, CC BC, BC, BC BCCC CC CC

Note: BC denotes banking crises, while CC denotes currency crises.Banking crises: chronology follows Caprio and Klingebiel (1999), which documents episodes where much or all of bank capital was exhausted.Currency crises: includes episodes of forced changes in parity, abandonment of fixed exchange rate regimes, and those episodes identified by an index of

exchange market pressure. The index is a standard deviation weighted average of exchange rate changes, short-term interest rate changes, and reserve changes. Acurrency crisis is recorded when the index exceeds the critical threshold of 1.5 standard deviations above its mean.


Table A.2. Unrestricted Model of the Determinants of the Change in Foreign Bank Claims on Latin America

Variable Coeff. t-Statistic

Host country variablesHost real GDP growth × Canada 0.561 (0.14)Host real GDP growth × France 0.760 (0.84)Host real GDP growth × Germany 1.101 (0.72)Host real GDP growth × Japan 2.316 (2.00)**Host real GDP growth × Spain 4.203 (1.93)*Host real GDP growth × U.K. 3.435 (0.76)Host real GDP growth × U.S. 2.414 (2.61)***Host real GDP growth × exposure to host × Canada −0.011 (0.02)Host real GDP growth × exposure to host × France −0.199 (0.88)Host real GDP growth × exposure to host × Germany −0.112 (0.62)Host real GDP growth × exposure to host × Japan −4.349 (2.43)**Host real GDP growth × exposure to host × Spain −0.285 (1.22)Host real GDP growth × exposure to host × U.K. −0.599 (0.24)Host real GDP growth × exposure to host × U.S. −0.160 (1.16)%Δ(Host country rating) × Canada 2.852 (1.20)%Δ(Host country rating) × France −0.779 (2.39)**%Δ(Host country rating) × Germany 0.170 (0.35)%Δ(Host country rating) × Japan 1.537 (2.27)**%Δ(Host country rating) × Spain 1.207 (1.12)%Δ(Host country rating) × U.K. 0.919 (0.37)%Δ(Host country rating) × U.S. 1.735 (3.35)***%Δ(Host country rating) × exposure to host × Canada −0.313 (1.16)%Δ(Host country rating) × exposure to host × France 0.224 (2.54)**%Δ(Host country rating) × exposure to host × Germany −0.009 (0.13)%Δ(Host country rating) × exposure to host × Japan −0.127 (0.17)%Δ(Host country rating) × exposure to host × Spain −0.172 (1.18)%Δ(Host country rating) × exposure to host × U.K. −0.983 (0.61)%Δ(Host country rating) × exposure to host × U.S. −0.050 (0.82)Host crisis dummy × Canada 9.459 (0.48)Host crisis dummy × France −6.435 (0.92)Host crisis dummy × Germany −6.303 (0.72)Host crisis dummy × Japan 5.304 (0.26)Host crisis dummy × Spain 16.369 (0.79)Host crisis dummy × U.K. −38.074 (0.75)Host crisis dummy × U.S. −15.390 (1.53)Host crisis dummy × exposure to host × Canada 2.375 (0.53)Host crisis dummy × exposure to host × France −1.605 (0.73)Host crisis dummy × exposure to host × Germany −0.469 (0.36)Host crisis dummy × exposure to host × Japan −4.761 (0.19)Host crisis dummy × exposure to host × Spain 0.017 (0.00)Host crisis dummy × exposure to host × U.K. 12.441 (0.35)Host crisis dummy × exposure to host × U.S. 1.152 (0.82)

Home country variablesCanadian real GDP growth × Canada 3.358 (0.67)French real GDP growth × France −4.572 (2.43)**

(continued)


457



458

Table A.2. (Concluded)

Variable Coeff. t-Statistic

German real GDP growth × Germany −3.492 (1.31)Japanese real GDP growth × Japan 5.725 (1.87)*Spanish real GDP growth × Spain −3.618 (0.65)U.K. real GDP growth × U.K. −17.058 (1.32)U.S. real GDP growth × U.S. −5.564 (2.08)**Canadian real interest rate × Canada −0.167 (0.03)French real interest rate × France −0.188 (0.10)German real interest rate × Germany 2.640 (0.85)Japanese real interest rate × Japan −12.073 (2.16)**Spanish real interest rate × Spain −1.263 (0.44)U.K. real interest rate × U.K. −16.455 (0.64)U.S. real interest rate × U.S. −0.679 (0.19)%Δ(Private real claims on other countries) × Canada × Canada 1.632 (1.95)*%Δ(Private real claims on other countries) × France −0.011 (0.03)%Δ(Private real claims on other countries) × Germany 0.501 (1.46)%Δ(Private real claims on other countries) × Japan 0.132 (2.29)**%Δ(Private real claims on other countries) × Spain −0.214 (0.68)%Δ(Private real claims on other countries) × U.K. 2.048 (1.43)%Δ(Private real claims on other countries) × U.S. 0.583 (2.41)**

Number of observations 804Adjusted R-squared 0.11F-test significance of the overall regression 2.59Prob > F 0.00

Source: Authors’ calculations.Notes: This table presents the results of estimating equation (1) for the change in private sector

claims. The model estimated is equivalent to estimating a separate equation for each home/lendercountry where home and host countries are allowed to have a different coefficient for each home orlender country. Home and host exchange rate changes and home and host dummies are included butnot shown. *, **, *** denote significance at 10, 5, and 1 percent, respectively. t-statistics are calcu-lated on the basis of robust standard errors.


Table A.3. F - tests for Coefficient Restrictions Across Home Countries

F-test of Unrestricted Coefficient

Restricted Coefficients Coefficients Restrictions

Model A3.1Host real GDP growth Home real GDP growth F(96,658) = 1.14Host real GDP growth × exposure to host Home real interest rate Prob > F = 0.1785%Δ(Host local currency/US$ exchange rate) %Δ(Home local currency/%Δ(Host rating) US$ exchange rate)%Δ(Host rating) × exposure to host %Δ(Private real claims Host crisis dummy on all other countries)Host crisis dummy × exposure to host Home dummiesHost dummies

Model A3.2Host real GDP growth Home real GDP growth F(101,658) = 1.10Host real GDP growth × exposure to host Home real interest rate Prob > F = 0.2559%Δ(Host local currency/US$ exchange rate) %Δ(Private real claims %Δ(Host country rating) on all other countries)%Δ(Host country rating) × exposure to host Home dummiesHost crisis dummyHost crisis dummy × exposure to hostHost dummies%Δ(Home local currency/US$ exchange rate)

Model A3.3Host real GDP growth Home real GDP growth F(107,658) = 1.14Host real GDP growth × exposure to host Home real interest rate Prob > F = 0.1793%Δ(Host local currency/US$ exchange rate) Home dummies%Δ(Host country rating)%Δ(Host country rating) × exposure to hostHost crisis dummyHost crisis dummy × exposure to hostHost dummies%Δ(Home local currency/US$ exchange rate)%Δ(Private real claims on all other countries)

Model A3.4Host real GDP growth Home real GDP growth F(112,658) = 1.20Host real GDP growth × exposure to host Home dummies Prob > F = 0.0916*%Δ(Host local currency/US$ exchange rate)%Δ(Host country rating)%Δ(Host country rating) × exposure to hostHost crisis dummyHost crisis dummy × exposure to hostHost dummies%Δ(Home local currency/US$ exchange rate)%Δ(Private real claims on all other countries)Home real interest rate

Notes: This table presents the results from estimating a number of tests on equation (1), wherewe constrain coefficients across home countries. Starting from the specification reported in Table A.2we present a series of tests in which we try constraining several combinations of coefficients. * denotes significance at 10 percent. Source: Authors’ calculations.


459



460

REFERENCES

Bank for International Settlements, 2003, “Guide to the International Banking Statistics,” BISPapers No. 16 (Basel).

Barajas, Adolfo, Roberto Steiner, and Natalia Salazar, 2000, “The Impact of Liberalization andForeign Investment in Colombia’s Financial Sector,” Journal of Development Economics,Vol. 63, No. 1, pp. 157–96.

Bonin, John, Iftekhar Hasan, and Paul Wachtel, 2004, “Privatization Matters: Bank Efficiencyin Transition Countries,” William Davidson Institute Working Paper No. 2004-679 (AnnArbor, Michigan: University of Michigan Business School).

Bordo, Michael, Barry Eichengreen, Daniela Klingebiel, and Maria Soledad Martinez Peria,2001, “Is the Crisis Problem Growing More Severe?” Economic Policy, Vol. 32, pp. 53–75.

Buch, Claudia, 2000, “Why Do Banks Go Abroad? Evidence from German Banks,” FinancialMarkets, Institutions, and Instruments, Vol. 9, No. 1, pp. 33–67.

Calvo, Guillermo, Leonardo Leiderman, and Carmen Reinhart, 1993, “Capital Inflows andReal Exchange Rate Appreciation in Latin America,” Staff Papers, International MonetaryFund, Vol. 40, No. 1, pp. 108–51.

Calvo, Guillermo, and Enrique Mendoza, 2000, “Rational Contagion and the Globalization ofSecurities Markets,” Journal of International Economics, Vol. 51, No. 1, pp. 79–113.

Caprio, Gerald, and Daniela Klingebiel, 1999, “Episodes of Systemic and Borderline BankingCrises” (unpublished; Washington: World Bank).

Cho, Kang Rae, Suresh Krishnan, and Douglas Nigh, 1987, “The State of Foreign BankingPresence in the United States,” International Journal of Bank Marketing, Vol. 5, pp. 59–75.

Chuhan, Punam, Stijn Claessens, and Nlandu Mamingi, 1998, “Equity and Bond Flows to LatinAmerica and Asia: The Role of Global and Country Factors,” Journal of DevelopmentEconomics, Vol. 55, No. 2, pp. 439–63.

Claessens, Stijn, Asli Demirgüç-Kunt, and Harry Huizinga, 2001, “How Does Foreign EntryAffect Domestic Banking Markets?” Journal of Banking and Finance, Vol. 25, No. 5,pp. 891–911.

Crystal, Jennifer, B. Gerard Dages, and Linda Goldberg, 2001, “Does Foreign OwnershipContribute to Sounder Banks in Emerging Markets? The Latin American Experience,” inOpen Doors: Foreign Participation in Financial Systems in Developing Countries, ed. byRobert Litan, Paul Masson, and Michael Pomerleano (Washington: The BrookingsInstitution Press), pp. 217–66.

Dages, B. Gerard, Linda Goldberg, and Daniel Kinney, 2000, “Foreign and Domestic BankParticipation in Emerging Markets: Lessons from Mexico and Argentina,” Economic PolicyReview, Vol. 6 (September) (New York: Federal Reserve Bank of New York), pp. 17–36.

De Haas, Ralph, and Iman Van Lelyveld, 2003, “Foreign Banks in Central and Eastern Europe:Asset or Risk Factor?” Maandschrift Economie, Vol. 67, No. 3, pp. 188–209.

Denizer, Cevdet, 2000, “Foreign Entry in Turkey’s Banking Sector, 1980–1997,” in TheInternationalization of Financial Services: Issues and Lessons for Developing Countries, ed.by Stijn Claessens and Marion Jansen (The Hague: Kluwer Law International), pp. 389–406.

Fernandez-Arias, Eduardo, 1996, “The New Wave of Private Capital Inflows: Push or Pull?”Journal of Development Economics, Vol. 48, No. 2, pp. 389–418.

Fisher, A., and Philip Molyneux, 1996, “A Note on the Determinants of Foreign Bank Activity inLondon Between 1980 and 1989,” Applied Financial Economics, Vol. 6, No. 3, pp. 271–77.

García, J. A., 2002, “Análisis de los Efectos del Incremento de la Actividad de la BancaExtranjera en América Latina y el Caribe,” Monetaria, Vol. 3, pp. 275–314.


Goldberg, Lawrence, and Robert Grosse, 1994, “Location Choice of Foreign Banks in theUnited States,” Journal of Economics and Business, Vol. 46, No. 5, pp. 367–79.

Goldberg, Lawrence, and Denise Johnson, 1990, “The Determinants of U.S. Banking ActivityAbroad,” Journal of International Money and Finance, Vol. 9, No. 2, pp. 123–37.

Goldberg, Lawrence, and Anthony Saunders, 1980, “The Causes of U.S. Bank ExpansionOverseas: The Case of Great Britain,” Journal of Money, Credit, and Banking, Vol. 12,No. 4, pp. 630–43.

———, 1981a, “The Determinants of Foreign Banking Activity in the United States,” Journalof Banking and Finance, Vol. 5, pp. 17–32.

———, 1981b, “The Growth of Organizational Forms of Foreign Banks in the U.S.: A Note,”Journal of Money, Credit, and Banking, Vol. 13, No. 3, pp. 365–74.

Goldberg, Linda, 2002, “When Is U.S. Bank Lending to Emerging Markets Volatile?” inPreventing Currency Crises in Emerging Markets, ed. by Sebastian Edwards and JeffreyFrankel (Chicago: University of Chicago Press).

Hernandez, Leonardo, Pamela Mellado, and Rodrigo Valdes, 2001, “Determinants of PrivateCapital Flows in the 1970s and 1990s: Is There Evidence of Contagion?” IMF WorkingPaper 01/64 (Washington: International Monetary Fund).

Nigh, Douglas, Kang Rae Cho, and Suresh Krishnan, 1986, “The Role of Location-RelatedFactors in U.S. Banking Involvement Abroad: An Empirical Analysis,” Journal of Inter-national Business Studies, Vol. 17, pp. 59–72.

Palmer, David E., 2000, “U.S. Bank Exposure to Emerging-Market Countries During RecentFinancial Crises,” Federal Reserve Bulletin (February), pp. 81–96.

Peek, Joe, and Eric Rosengren, 2000, “Implications of the Globalization of the Banking Sector:The Latin American Experience,” New England Economic Review (September–October)(Boston: Federal Reserve Bank of Boston), pp. 45–62.

Salomon Smith Barney, 2000, Foreign Financial Institutions in Latin America (New York City).

Schinasi, Garry, and R. Todd Smith, 1999, “Portfolio Diversification, Leverage, and FinancialContagion,” IMF Working Paper 99/136 (Washington: International Monetary Fund).

Van Rijckeghem, Caroline, and Beatrice Weder, 2003, “Spillovers Through Banking Centers: APanel Data Analysis of Bank Flows,” Journal of International Money and Finance, Vol. 22,No. 4, pp. 483–509.


461


462


*Andrew Berg is a Division Chief in the Policy Development and Review Department of the IMF,Eduardo Borensztein from the Research Department of the IMF is currently on leave at the Inter-AmericanDevelopment Bank, and Catherine Pattillo is a Senior Economist in the African Department of the IMF.This paper was largely written in 2002–3, while all three authors were in the Research Department of theIMF. The authors would like to thank the editors of this journal, Francis Diebold, and many IMF staffmembers, including Paul Cashin and Robert Rennhack, for useful comments, and Manzoor Gill for superbresearch assistance.

Assessing Early Warning Systems: How Have They Worked in Practice?

ANDREW BERG, EDUARDO BORENSZTEIN, AND CATHERINE PATTILLO*

Since 1999, IMF staff have been tracking several early warning system (EWS) mod-els of currency crisis. The results have been mixed. One of the long-horizon modelshas performed well relative to pure guesswork and to available non-model-basedforecasts, such as agency ratings and private analysts’ currency crisis risk scores.The data do not speak clearly on the other long-horizon EWS model. The twoshort-horizon private sector models generally performed poorly. [JEL F31, F47]

Research on developing early warning system (EWS) models of currency cri-sis received a strong stimulus in recent years following the Mexican and

Asian crises. Both events took the international community somewhat by surpriseand thus focused attention on indicators and methods that could assist in the timelyidentification of highly vulnerable countries. Since the beginning of 1999, IMFstaff has been systematically tracking, on an ongoing basis, various models devel-oped in-house and by private institutions, as part of a broader forward-looking vulnerability assessment.

EWS models play a necessarily small part in vulnerability assessment rela-tive to more detailed country-specific analyses. We do not review the IMF’s expe-


ASSESSING EARLY WARNING SYSTEMS: HOW HAVE THEY WORKED IN PRACTICE?

463

rience with the broader vulnerability assessment process here.1 The question weare interested in, and to which we return in the conclusion, is whether EWS mod-els have any role to play at all.2

This paper looks in detail at the performance of these models in practice todate. We emphasize the distinction between in-sample and out-of-sample predic-tion. For an EWS model to be a useful tool in monitoring vulnerabilities, it musthold up in real time, after the model has been formulated. We thus focus on theactual forecasts made since 1999, though we also reexamine the run-up to the Asiacrisis.

A typical result from our earlier studies was that, while the model forecasts wereexamine here were statistically significant predictors of crises, whether they wereeconomically significant was harder to say. In other words, the forecasts were infor-mative compared with a benchmark of complete random guessing, but it was lessclear whether they were useful to an already informed observer. It is reasonable tosuppose that comprehensive, country-specific holistic assessments by informed ana-lysts, based on all available qualitative and quantitative information, must be betterthan inevitably simple EWS models. Indeed the ability to take all the informationinto account is clearly a huge potential advantage. But there have been no studieson whether such comprehensive assessments are in fact better. We gain perspectiveon this issue here by comparing EWS model forecasts to non-model-based indica-tors, such as bond spreads, agency ratings, and risk scores published by analysts.

I. EWS Models

Policy initiatives to monitor indicators of external vulnerability can be traced tothe Mexican peso crisis of December 1994. A seminal effort to use a systematicquantitative early warning system to predict currency crises was the “indicators”model of Kaminsky, Lizondo, and Reinhart (1998). The Asia crises of 1997/98provided further impetus for the effort. Evidence suggested that despite the daunt-ing challenges involved in this sort of exercise, this kind of model had some suc-cess in predicting these crises out of sample (Berg and Pattillo, 1999a). It alsosuggested that a variety of improvements could substantially improve model per-formance (Berg and Pattillo, 1999b).

In light of this research, IMF staff has implemented various models to predictcurrency and balance of payments crises since 1999, as described in Berg and oth-ers (2000). IMF staff has also tracked various private sector models, includingGoldman Sachs’ GS-WATCH and Credit Suisse First Boston’s Emerging MarketsRisk Indicator, and a more recent model, the Deutsche Bank Alarm Clock. Table 1summarizes the main features of the models under consideration here. It details

1Some discussion of the role of EWS models in broader vulnerability assessment can be found in IMF(2002) and in Berg and others (2000).

2Predicting currency crises is closely related to predicting exchange rate movements, so any successof EWS models is notable in the context of the extensive literature, starting with Meese and Rogoff (1983),which has shown how difficult it is to predict exchange rates out of sample. Berg and others (2000) dis-cuss this point further in the EWS context.


An

dre

w Be

rg, Ed

ua

rdo

Bore

nszte

in, a

nd

Ca

the

rine

Pa

ttillo

46

4

Table 1. Specification of Early Warning Models

DCSD1 EMRI4

(Berg, Borensztein, Crisis Signals2 GS-WATCH3 (Credit Suisse DB Alarm Clock5

Pattillo) (based on KLR) (Goldman Sachs) First Boston) (Deutsche Bank)

Crisis Definition

HorizonMethod

Variables

Weighted average of one-month changes inexchange rate andreserves more than 3(country-specific) stan-dard deviations abovecountry average2 yearsProbit regression with rhsvariables measured in(country-specific) per-centile terms

OvervaluationCurrent accountReserve lossesExport growthST debt/reserves

Same as DCSD

2 yearsWeighted (by frequencyof correct predictions)average of indicators.Variables measured as0/1 indicators accordingto threshold chosen tominimize noise/signalratioOvervaluationCurrent account6

Reserve lossesExport growth

Weighted average ofthree-month changes inexchange rate andreserves above country-specific threshold

3 monthsLogit regression with(most) rhs variables mea-sured as 0/1 indicatorsbased on thresholdsfound in autoregressionwith dummy (SETAR)

Overvaluation

Export growth

Depreciation > 5% and atleast double precedingmonth’s

1 monthLogit regression with rhsvariables measured inlogs, then deviation frommean and standardized

Overvaluation

Debt/exports

Various “trigger points”:Depreciation > 10% andInterest rate increase >25% typical

1 monthLogit two-equation simul-taneous systems onexchange rate and interestrate “events” of differentmagnitude

Overvaluation


ASSESSIN

G EA

RLY

WA

RN

ING

SYSTEM

S: HO

W H

AV

E THEY

WO

RK

ED IN

PR

AC

TICE?

46

5

Reserves/M2 (level)6

Reserves/M2 (growth)Domestic credit growth

Money multiplier changeReal interest rateExcess M1 balances

Reserves/M2 (level)

Financing requirement

Stock market

Political eventGlobal liquidityContagion

Growth of credit to private sector

Reserves/imports (level)Oil prices

Stock price growthGDP growth

Regional contagion

Industrial productionDomestic credit growth

Stock market

Devaluation contagionMarket pressure contagionRegional dummiesInterest rate “event”

Source: Berg and others (2000).Note: rhs = right-hand-side.1DCSD: Developing Country Studies Division of the IMF (Berg and others, 2000).2KLR: Kaminsky, Lizondo, and Reinhart (1998).3Goldman Sachs: Ades, Masih, and Tenengauzer (1998).4EMRI: Emerging Markets Risk Indicator, Credit Suisse First Boston: Roy and Tudela (2000).5Deutsche Bank: Garber, Lumsdaine, and van der Leij (2000).6Not included in the original KLR model.


Andrew Berg, Eduardo Borensztein, and Catherine Pattillo

466

the crisis definition employed, the prediction horizon, the method used to gener-ate predictions, and the predictor variables. Appendix I contains a more completedescription of the models. Table A.1 shows all the crisis dates for these modelssince January 1999.3

As Table 1 illustrates, the specification of EWS models involves a number ofdecisions that, while guided in some way by economic theory, are largely empiri-cal and judgmental in nature. Currency crises, for example, are not preciselydefined events, but the models must nonetheless define crisis dates in terms of thedata. (See Box 1 for a discussion of how the different models implement the cur-rency crisis concept.) The choice of a prediction horizon depends on the objectivesof the user. The in-house models adopt a relatively long horizon, which shouldallow time for policy changes that may prevent the crisis. The time horizon of pri-vate sector models is shorter and their criterion for evaluating the accuracy of pre-dictions (frequently, a trading rule) is sometimes different. Nevertheless, it is stillinformative to consider the predictions from private sector EWS models whenassessing vulnerability, especially because those predictions are disseminatedwidely within the investor community.

The predictive variables in the models are inspired by theories of balance ofpayments crises but constrained by data availability, but in the end reflect whatworks best in fitting the data. The choice of statistical method is an essentiallyempirical decision. Appendix I discusses considerations that apply to these speci-fication choices with special reference to the models tracked in the IMF.

II. The Value Added of EWS Models

Since early 1999 the IMF has been regularly producing forecasts from two EWSmodels, the Kaminsky-Lizondo-Reinhart (KLR) and the Developing CountryStudies Division (DCSD), and monitoring two private sector models, the GoldmanSachs (GS) and Credit Suisse First Boston (CSFB). This section examines the use-fulness of these models in providing early warnings of crises.

The evaluation of EWS models requires a benchmark. Typically the questionis whether a model provides a statistically and economically significant predictionof crisis. This might appear to be a low standard to meet. Given the difficultiesinvolved in crisis prediction, however, it is ambitious. To forecast crises reliablyimplies systematically outperforming the market at predicting sudden changes inthe exchange rate.

Assessments must focus on out-of-sample performance. Successful in-samplepredictions are much easier to achieve than out-of-sample predictions but muchless meaningful. First, the diligent analyst may have searched through enough trulyunrelated variables until finding some that, by coincidence, happen to be corre-

3More recently, other models have been developed at the IMF, such as Mulder, Perrelli, and Rocha(2002) and Abiad (2003). Other recent developments include models designed to predict other sorts ofcrises, such as Manasse, Roubini, and Schimmelpfennig (2003). For a review of recent developments inthis literature, see Abiad (2003).



467

Box 1. Crisis Definition

The early warning system (EWS) models considered in this paper attempt to predictcurrency crises, as distinct from other sorts of crises, such as debt and bankingcrises. Although opinions differ as to what constitutes a currency crisis and whenone is observed, the formulation of an EWS model requires a specific quantitativedefinition (Table 1 briefly describes the crisis definitions for the models discussed indetail in this paper. Table A.1 lists all the crisis dates for the various models for the1999–2001 period (1999–2002 for DCSD and KLR)).

Models that attempt to predict only successful speculative attacks, such as thatof Credit Suisse First Boston (CSFB), define crises solely by sufficiently largechanges in the exchange rate over a short period of time. For a private sector institu-tion, predicting sudden large changes in the exchange rate alone may be the mainobjective. EWS models implemented by the IMF, as well as the Goldman-Sachs(GS) model, attempt to predict both successful and unsuccessful speculative attacksby calculating an “exchange market pressure” index that combines exchange rateand reserve changes.

For the CSFB model, a crisis occurs when the exchange rate moves by morethan some threshold amount over a short period of time (see Table 1). For the in-house models, a crisis occurs when the exchange market pressure index is very highrelative to its historical average.1 The GS model also defines a crisis as an indexvalue that is high relative to a country-specific threshold, with the threshold definedso as to separate calm periods from those of unusual volatility. The GS definitiontends to produce more frequent crises, often occurring for several consecutive peri-ods, while the Developing Country Studies Division (DCSD) and the Kaminsky-Lizondo-Reinhart (KLR) crises are almost always isolated. Comparing over acommon sample of 16 countries from 1996 through mid-2001, the CSFB crisis defi-nition produces 34 crises, the DCSD/KLR definition produces 34 crises, and the GSdefinition 150 crises, with the latter grouped into 47 distinct episodes of one or moreconsecutive crisis months.

The omission of interest rates in the crisis definitions for most emerging marketEWSs, because of poor availability of historical data on market-determined interestrates, is increasingly recognized as a shortcoming in identifying crises. For example,the 1995 attack on the Argentine peso in the wake of the “tequila” effect was thekind of event the models should attempt to predict. However, this failed attack,which was evidenced mainly by a rapid increase in domestic interest rates, is notidentified as a crisis by many EWS models. The Deutsche Bank Alarm Clock(DBAC) defines substantial exchange rate depreciations and interest rate increasesas separate events but jointly estimates the probability of these two types of events.However, the model uses International Financial Statistics (IFS) data on moneymarket interest rates, which are deficient for many emerging economies.

1Taking into account whether the crisis index is high relative to its history in a particular coun-try has an advantage relative to defining a crisis by the same absolute cutoff depreciation for all coun-tries. For example, for a country that has had a pegged exchange rate regime and where the rate hasremained fixed for some time, a relatively small devaluation might be considered a crisis. The samesize depreciation might not constitute a crisis in a country where the exchange rate is flexible and hasbeen more volatile.



468

lated with crises in this particular sample. Such a spurious relationship is not likelyto persist in a new sample.4 Second, even if a true relationship is found in the sam-ple, the next set of crises may be fundamentally different from the last. A modelthat provides accurate out-of-sample forecasts has thus passed a much toughertest. Of course, only models passing this test are useful for actual crisis prediction.For these reasons, the focus here is on out-of-sample testing.

In this section, several approaches to testing these models are pursued. First,the stage is set by a review of the performance of a model designed prior to theAsia crises, KLR, in predicting those crises. Following earlier work, the model isimplemented as it might have been in early 1997 and forecasts are compared withactual outcomes. These predictions are compared with those implied by spreadson dollar-denominated sovereign bonds and sovereign credit ratings as well as theassessments of currency crisis risks produced by country experts from the Econo-mist Intelligence Unit (EIU). These results suggest that the EWS models showpromise. The KLR model decisively outperforms all the non-EWS based compara-tors in this period.

The second part of the section, and the core of the paper, looks at the how thevarious models that have been monitored at the IMF since early 1999 have per-formed in this period. First is a detailed analysis of the first set of forecasts officiallyproduced within the IMF, in May 1999. These are compared with the alternativeindicators, such as bond spreads, ratings, and analysts’ views described above,where possible. We follow with a more systematic examination of how well themodels have predicted crises over the full out-of-sample period.

4Similarly, models are likely to find an “overvaluation” of the exchange rate before currency criseswhen the long-run, or equilibrium, exchange rate is calculated, as is usual, as some form of average valueover the estimation period that includes the crisis events. Evaluating the out-of-sample performance of themodels also avoids any overstatement of the predictive value of the models through this channel.

Box 1. (Concluded)

It is unlikely that any simple formula, however well thought out, will always besuccessful in picking out crisis periods in the data. One possible improvement wouldbe to combine the results from the quantitative definition with country-specificknowledge about exchange market developments to make some adjustments to thedating of crisis periods. Events that are close calls according to the crisis formulacould receive particular scrutiny, and the analyst might judge whether to label someof these as crises. For example, Sri Lanka suffered a reserve loss of about 40 percentduring 2000, along with a currency depreciation of nearly 15 percent, culminating inthe abandonment of the crawling band exchange rate regime and further currencydepreciation in January 2001. Because no single month was sufficiently traumatic,however, the formula employed in the KLR and DCSD models registered a fairlyclose call, but not a crisis. This episode might have been called a crisis if it had beenassessed “by hand.”



469

EWS Models and Alternative Indicators in the Asia Crisis

Berg and others (2000) looked at various measures of performance of a variety ofEWS models, focusing in particular on their ability to predict the Asia crises of1997–98 out of sample.5 One main conclusion was that the original KLR model,which was designed prior to 1997 and hence without the benefit of hindsight, hadsubstantial predictive power over the Asian episodes. Column 1 of Table 2 showsthe ranking of countries according to the risk of currency crisis that KLR wouldhave produced in early 1997. These forecasts are fairly good, with many of themost vulnerable countries in fact being the hardest hit in terms of crisis severity.6For example, Korea and Thailand were among the top third of countries in termsof vulnerability, according to the KLR model. Although Brazil and the Philip-pines, which were not hit particularly hard over this period, were at the top of the vulnerability table, the forecasts are informative overall. Country rank in thepredicted vulnerability list is a statistically significant predictor of actual crisisincidence.

One lesson from Berg and others (2000) and other work is that there was clearscope for improvement of those earlier models. A variety of potentially importantcrisis indicators had not been tested, such as the current account deficit as a shareof GDP and the ratio of short-term external debt to GDP. Moreover, regression-based estimation techniques that more fully exploit the information in the dataseemed a promising alternative to the “indicator”-based method of the KLRmodel. A revamped KLR-based model and the DCSD model described in sectionI were the result of an effort to improve on the original KLR model. Not surpris-ingly, given the benefit of hindsight, these models perform substantially better inpredicting the Asia crises (Column 2 of Table 2 presents results for the DCSDmodel).

The predictions of EWS models were significantly better than random guessesin predicting the Asia crises, but they were not overwhelmingly accurate. How, incomparison, did the various non-model-based indicators fare over this period?Among these, sovereign spreads are a commonly watched indicator of countryrisk. While the spreads are important indicators of market access, and also marketsentiment, they do not fare particularly well as currency crisis predictors over thisperiod. The most affected countries had generally lower pre-crisis spreads as ofthe first quarter of 1997, as shown in the third column of Table 2. The spread aver-aged 90 basis points in the countries that subsequently suffered a crisis, while itaveraged 201 in the other countries.7

5See also Berg and Pattillo (1999c) on the implications of EWS models for the Asia crisis.6The measure of the severity of crisis for a particular country is the maximum value reached by the

exchange market pressure index in 1997, where the index itself is a weighted average of the depreciationof the exchange rate and the loss of international reserves.

7Although one could rationalize the low sovereign spreads in the Asian economies on the basis of theirrelatively low levels of external debt, spreads did increase after October 1997, suggesting that markets mayhave underestimated risks. The period before the Asian crises was characterized by unusually low spreadsfor almost all emerging market economies.



470

Table 2. Risk Assessments Prior to the Asia Crisis Based on KLR, DCSD,Bond Spread, Credit Rating, and the Economist Intelligence

Unit (EIU) Forecasts

Spread4 Rating5 EIU 1997 Q1 Country1 KLR2 DCSD3 1997 Q1 1997 Q1 Currency Risk6

Korea, Rep. of 22 24 50 18 22Thailand 20 40 51 25 42Indonesia 16 32 109 43 38Malaysia 14 39 37 20 36Zimbabwe 19 n.a. n.a. n.a. 58Philippines 34 14 165 55 36Taiwan Province of China 23 46 n.a. n.a. 12Colombia 15 41 129 45 35India 10 21 n.a. n.a. 35Brazil 31 15 233 65 51Turkey 16 18 416 66 56Venezuela 14 9 n.a. n.a. 53Pakistan 20 36 n.a. 68 49South Africa 19 26 85 48 39Jordan 14 15 n.a. n.a. 61Sri Lanka 12 17 n.a. n.a. 43Chile 11 14 n.a. n.a. 17Bolivia 10 5 n.a. n.a. 37Argentina 14 11 265 63 59Mexico 14 8 231 55 55Peru 20 26 n.a. 70 51Uruguay 10 14 135 50 37Israel 14 24 44 30 46

AverageCrisis countries 20 34 90 34 35Non-crisis countries 15 17 201 57 46

Rank correlation7 0.52 0.53 −0.31 −0.49 −0.33

Source: Authors’ calculations.1Countries that suffered a crisis in 1997 are in bold. The countries are ordered by severity of crisis.2Probabilities of currency crisis over a 24-month horizon, from average KLR model for 1996.3Probabilities of currency crisis over a 24-month horizon, from average of 1996 DCSD results.4The spread is expressed in basis points. It refers to the difference between the yield on U.S.

dollar–denominated foreign government eurobonds and the equivalent maturity U.S. treasury bonds.5Average of S&P and Moody’s ratings, each converted to a numerical rating ranging from 100

(S&P SD) to 0 (S&P AAA or Moody’s Aaa), following Ferri, Liu, and Stiglitz (1999). A lower numbermeans a better rating (unlike Ferri, Liu, and Stiglitz).

6Currency risk: “Scores and ratings assess the risk of a devaluation against the dollar of 20 percentor more in real terms over the two-year forecast period,” following EIU.

7Countries are ranked according to each indicator as well as according to crisis severity (in bothcases a lower number implies a worse actual or predicted crisis). The rank correlation relates thesetwo rankings.



471

Second, there is some evidence that sovereign ratings from agencies such asMoody’s and Standard and Poor’s (S&P’s) have also been poor predictors ofrecent currency crises.8 The fourth column of Table 2 shows the sovereign ratingsas of the first quarter of 1997, based on a quantitative conversion of Moody’s andS&P’s ratings, where higher numbers correspond to a better rating. Korea,Malaysia, and Thailand are the highest-rated countries, while Mexico is relativelypoorly ranked. Indeed, the average rating in the ten most affected countries wassubstantially better than in the ten least affected countries.

Other non-model-based predictions of currency crises are surveys of cur-rency market analysts, such as those prepared by the EIU. The EIU has regularlyproduced estimates of currency crisis risk, defined as the risk of a 20 percent realdepreciation of the currency over the two-year forecast horizon.9 These estimatesderive from the analysis of country experts who consider a broad set of quantita-tive and qualitative factors, ranging from macroeconomic and financial variablesto the strength of the banking system, the quality of economic decision making,and the stability of the political situation. The estimates are available for a largenumber of countries since 1996. As column 5 of Table 2 shows, these EIU fore-casts gave generally positive assessments to the Asian economies that were aboutto suffer from severe episodes. Indeed, countries with higher risk scores in the sec-ond quarter of 1997 were systematically less likely to have a crisis during the1997–98 period.10

Along similar lines, there exist surveys of estimates of future exchange ratechanges by foreign exchange traders and specialists in financial institutions andmultinational corporations. Goldfajn and Valdés (1998) examined the surveys bythe Financial Times Currency Forecaster and found that such market participants’expectations provided no useful early warning of currency crises in a large sampleof emerging markets, or in important cases such as Mexico in 1994 or Thailand in 1997.

To summarize, there is little evidence that “market views,” or analysts’ views,as expressed in spreads, ratings, and surveys, are reliable crisis predictors, impor-tant as they may be in determining market access. This conclusion is illustrated forthe case of Korea in Figure 1, which shows DCSD model predictions as well asvarious other indicators of crisis risk.

The results of this round of EWS testing were sufficiently promising to sug-gest the continued implementation of these models on an ongoing basis, alongwith further research and development. The in-sample results were fairly good.More remarkably, the out-of-sample results were also promising. Here, however,

8See Sy (2003) and Reinhart (2002). As with sovereign spreads, it could be argued that these ratingsare designed to predict default, not currency crisis. Against this, however, is the fact that currency crisesdo increase the risk of default and that, because of this, ratings have in fact been downgraded after mostcurrency crises. This suggests that the rating agencies would have likely downgraded the countries hadthey seen the currency crises coming.

9See Economist Intelligence Unit Currency Risk Handbook, June 2001.10The Economist Intelligence Unit forecasts are similarly unsuccessful when compared with crises as

defined by the EIU itself.



472

–1

0

1

2

3

4

5

6

7

8

9

Jan-9

6

Apr-96

Jul-9

6

Nov-96

Feb-97

Jun-9

7

Sep-97

Jan-9

8

Apr-98

Aug-98

Nov-98

Mar-99

Jun-9

9

Sep-99

Jan-0

0

Apr-00

Aug-00

Nov-00

Per

cent

age

poin

ts

0

200

400

600

800

1,000

1,200Forward premium (left scale)1

Rating (left scale)

KDB bond spread (right scale)2

Estimated Probability of Crisis by DCSD Model3

0

10

20

30

40

50

Jan-

96

May-9

6

Sep-9

6

Jan-

97

May-9

7

Sep-9

7

Jan-

98

May-9

8

Sep-9

8

Jan-

99

May-9

9

Sep-9

9

Jan-

00

May-0

0

Sep-0

0

Per

cent

0

10

20

30

40

50

Sources: Bloomberg, JPMorgan, and IMF staff estimates. Berg and others (2000). DCSD stands for Developing Country Studies Division of the IMF in which the model was originally formulated.

1The forward premium is defined as the log of the ratio of the 12-month forward rate to the spot exchange rate.

2Spread of Korean Development Bank Eurobond over comparable U.S. treasuries.3The probability of a crisis over the next 24 months estimated using the DCSD model.

Figure 1. Korea: Forward Exchange Rate Premium, Bond Spread andProbability of an Exchange Rate Crisis Estimated by the

DCSD Model, 1996–2000(Through November 20, 2000)



473

the results of the DCSD model results must be discounted, since this model bene-fited from hindsight in its formulation, though only pre–Asia crisis data were usedto produce the forecasts.11 Moreover, it may be that this good Asia-crisis perfor-mance was just lucky or, alternatively, that subsequent crisis episodes may havebeen sufficiently different that these models stopped working. This suggests anexamination of the recent performance of the EWS models.

How Well Have the EWS Models Done in Practice Since Their Implementation in January 1999?

We examine this question in two ways. First, we look at the results of the modelsthe first time the forecasts were produced “officially” in a forward-looking exer-cise, in July 1999.12 Second, more systemic measures of the “goodness of fit” ofthe models are examined, emphasizing the comparison of in-sample and out-of-sample model performance, and the trade-off between missing crises and generatingfalse alarms.

The July 1999 forecasts

Table 3 shows the predicted probabilities of crisis for the DCSD and KLR mod-els.13 Countries that suffered crises are in bold, with the dates of the crisis noted.The KLR, and particularly the DCSD, model did fairly well. Both countries withDCSD probabilities of crisis above 50 percent subsequently had crises, and no cri-sis country had a probability below 26 percent. Using the models’ own definition,there have been only three crises in the roughly two years since July 1999.

As before, it is useful to compare these forecasts with other indicators andestimates. Columns 3, 4, and 5 of Table 3 show spreads on dollar-denominatedbonds, sovereign ratings, and the EIU’s currency crisis risk scores, all as of thesecond quarter of 1999. The alternative predictors fared better than prior to theAsia crises but still not well. Spreads are moderately higher, at 549, for the threecrisis countries compared with 462 for the others. The average sovereign rating is56 for the three crisis countries, while it is slightly worse, at 55, for the rest of thecountries. The EIU estimates, in contrast, are substantially better here than theywere before the Asia crises. For example, the average risk score of the three crisiscountries was 59, compared with 42 for the others.14

11The main benefit from this hindsight was the inclusion of short-term debt and reserves as a predic-tive variable. The original KLR model had focused on M2/reserves instead. This latter variable also works,though not as well.

12Appendix II explains how in-sample and out-of-sample periods are determined for each of the mod-els considered.

13The two private sector models monitored at that time, GS and CSFB, forecast only over a one- tothree-month horizon. The snapshot of the first “official” July 1999 results is thus not informative. Theirperformance is examined in the discussion of overall goodness of fit, below.

14Using the EIU’s own crisis definition, the forecasts perform somewhat worse, with the average riskfor the crisis countries below the average for the non-crisis countries.



474

Table 3. Crisis Probabilities According to Different Models as of July 1999

EconomistCountry1 KLR2 DCSD3 Spread4 Ratings5 Intelligence Unit6

Colombia (Aug 99) 42 61 544 45 42Turkey (Feb 01) 45 50 554 68 58Zimbabwe (Aug 00) 24 26 n.a. n.a. 77

Bolivia n.a. 36 n.a. n.a. 42Chile 11 36 n.a. n.a. 29Venezuela 42 34 n.a. n.a. 58Argentina 20 31 471 58 62Peru 32 26 210 58 35Uruguay 32 23 216 45 36Brazil 24 21 451 68 47Mexico 14 19 296 55 42Pakistan n.a. 14 2,270 90 69Jordan 14 14 n.a. n.a. 34South Africa 32 9 141 48 40India 11 9 n.a. n.a. 38Sri Lanka n.a. 7 n.a. n.a. 51Israel 11 6 78 30 30Thailand 32 4 192 48 41Philippines 14 3 401 50 28Malaysia 11 3 174 45 36Indonesia 32 1 872 78 50Korea, Rep. of 24 1 238 45 30

AverageCrisis countries 37 46 549 56 59Non-crisis countries 22 16 462 55 42

Source: Authors’ calculations.Note: KLR: Kaminsky, Lizondo, and Reinhart (1998); DCSD: Developing Country Studies

Division of the IMF (Berg and others, 2000).1Countries with crises between June 1999 and June 2001 are in bold.2Probabilities of currency crisis over a 24-month horizon, from KLR model. Estimated using

data through March 1999, except for Brazil, Jordan, Korea, Mexico, Venezuela, and Zimbabwe(December 1998); Chile and Israel (May 1999); India and Indonesia (January 1999); Malaysia (April1999); South Africa (February 1999); and Turkey (November 1998).

3Probabilities of currency crisis over a 24-month horizon from CSFB model. Estimated throughMarch 1999, except for R.B. de Venezuela (December 1998), Malaysia (January 1999), and Mexico,Thailand, and Indonesia (April 1999).

4The spread is expressed in basis points. It refers to the difference between the yield on U.S. dollar–denominated foreign government eurobonds and the equivalent maturity U.S. treasury bonds.

5Average of Standard & Poor’s (S&P) and Moody’s ratings, each converted to a numerical rat-ing ranging from 100 (S&P SD) to 0 (S&P AAA or Moody’s Aaa), following Ferri, Liu, and Stiglitz(1999). A lower number means a better rating (unlike Ferri, Liu, and Stiglitz).

6Currency risk: “Sources and ratings assess the risk of a devaluation against the dollar of 20 per-cent or more in real terms over the two-year forecast period,” following Economist Intelligence Unit(2001).



475

The improved performance of the non-model-based indicators compared withbefore the Asia crisis, combined with the low incidence of crises over the out-of-sample period, suggests that the challenge for the models over this period wasmore to avoid a large number of false alarms than to call otherwise unforeseencrises.

The distinction between sovereign, or default, risk and currency crisis risksurely plays an important role in explaining the performance of ratings and spreadsin some important recent cases—a role that it did not play in the Asia crises.Colombia’s crisis in August 1999 involved a drop in the exchange rate as the coun-try abandoned a crawling band exchange rate regime, but there was little subse-quent concern about sovereign default. Thus, it is perhaps not surprising that theratings and spreads do not predict this incident. Conversely, Pakistan suffered adebt crisis but had no currency crisis over the period, and its exceedingly highspreads, which started to widen after the economic sanctions following the nucleartests in 1998, greatly increase the non-crisis-country average.15

Overall goodness of fit since January 1999

A look at the goodness of fit of the models over the entire out-of-sample periodprovides a more systematic assessment of the models (see Box 2 for details onthese measures). The computation of goodness-of-fit measures requires selectinga cutoff probability value, above which the prediction is classified as an “alarm,”implying that the model expects a crisis to ensue at some point along the predic-tion horizon. The threshold probability for an alarm can be chosen to minimize aloss function that weighs two types of errors: failing to predict a crisis and issuinga crisis alarm that does not materialize.

The specification of the loss function implies a decision on how much weightto give to both types of mistakes. The relative weights depend implicitly on thecost imputed to each type of error. From the point of view of an institution con-cerned with the stability of the international financial system and the well-beingof the individual economies that are part of the global system, it would seem thatthe highest priority would be never to fail to predict a crisis. After all, the very pur-pose of using EWS models is to prevent currency crises or at least lessen theirimpact by being able to respond early and in a well-planned fashion. This wouldargue for choosing a low cutoff probability. In such case, however, the EWS modelwould be prone to generate a high number of false alarms; namely, crisis predic-tions that do not materialize. This would impair the credibility of the model anddampen the inclination to take aggressive policy action to prevent a possible cri-sis. Throughout this chapter, equal weight is placed on the share of alarms that arefalse and the share of crises that are missed, although the issue could be exploredfurther both in the evaluation and the estimation of the models.16 From the point

15Excluding Pakistan, the average spread for non-crisis countries declines to 312 from 462.16Demirgüc-Kunt and Detragiache (1999), who employ a similar loss-function approach in looking at

banking crisis prediction, discuss the relative weights in terms of the costs of policies and regulations toincrease the resilience of the banking system versus the costs of rescuing failed institutions.



476

Box 2. Goodness-of-Fit Measures and Trade-Offs

Early warning systems (EWSs) typically produce a predicted probability of crisis.In evaluating their performance, it would be simple and informative to comparethese predicted probabilities with actual crisis probabilities. Because the latter are not directly observable, goodness-of-fit calculations measure how well the predicted probabilities compare with the subsequent incidence of crisis. The firststep is to convert predicted probabilities of crisis into alarms; that is, signals that a crisis will ensue within 24 months (assuming that is the model’s horizon). Analarm is defined as a predicted probability of crisis above some threshold level (the cutoff threshold). Then each observation (a particular country in a particularmonth) is categorized as to whether it is an alarm (i.e., whether the predicted prob-ability is above the cutoff threshold) and also according to whether it is an actualpre-crisis month.

The threshold probability for an alarm can be chosen to minimize a “loss func-tion” equal to the weighted sum of false alarms (as a share of total tranquil periods)and missed crises (as a share of total crisis periods). In this paper, equal weight isplaced on the share of alarms that are false and the share of crises that are missed.(The former might be thought of as Type 1 errors and the latter as Type 2 errors, ifthe null hypothesis is no crisis.) A higher weight on missed crises would imply alower cutoff threshold for calling a crisis, and the model would generate both fewermissed crises and more false alarms. Note that only in-sample information can beused to calculate a threshold for actual forecasting purposes. When using the modelout of sample to make predictions, however, there is no guarantee that anotherthreshold would not provide better goodness of fit.

The columns of Table 4 and Table 5 show how the signals from the variousmodels compare with actual outcomes, over various periods. Each number in thegoodness-of-fit table represents the number of observations that satisfy the criterialisted in the rows and columns. For example, for the Developing Country StudiesDivision (DCSD) model (Table 4, column 7) over the January 1999 to December2000 period, there were a total of 443 tranquil months, and for 90 of them the probability was above the cutoff threshold.

From this table various measures of accuracy can be calculated. For example,the percentage of crises correctly called is equal to the number of observations forwhich the alarm was sounded and a crisis in fact ensued divided by the total numberof actual crises. The footnotes to Table 4 define these various measures.

It is possible that one model might do better when great weight is placed onavoiding false alarms; that is, when cutoff thresholds are relatively high, whileanother might excel when the optimal cutoff threshold is low. This turns out notgenerally to be the case for the models examined here. To demonstrate this, figurescan be produced that show how the models perform for any cutoff and independentof the loss function chosen. For a given model over a given sample, each candidatecutoff threshold produces a certain percentage of crises correctly called and per-centage of false alarms. For example, a cutoff of 0 produces 100 percent crises cor-rectly called but also 100 percent false alarms (as a share of tranquil periods),because the model would predict a crisis every time, while a cutoff of 100 produces0 percent crises correctly called but 0 percent false alarms, because the model willnever predict a crisis. The upper panel of Figure 2 traces all these points for each



477

of view of a private investor, although the trade-off between missed crises andfalse alarms also exists, the evaluation of cost and benefits may be easier, as it isdirectly related to the returns on different asset positions. In this case, one can sus-pect a bias toward a loss function that predicts crises more often because, as hasbeen observed, interest differentials are never high enough to offset the impact ofa large depreciation that takes place over a very short period.

A first result is that the in-sample goodness of fit of the models has also beenreasonably stable as new data have come in, as Table 4 shows. For example, con-sider the results for the DCSD model when estimated for Berg and others (2000)in 1998 (column 1) with the same model when used to produce the first set of“official” internal forecasts, produced in July 1999 (column 5). The model’s accu-racy is actually slightly improved over the longer period. Moreover, the modelsthemselves have remained fairly stable as new data have come in and as the coun-try coverage has changed slightly. For the DCSD model, for example, the valuesand statistical significance of the coefficients have not changed much.

Box 2. (Concluded)

cutoff between 1 and 100, for the DCSD and the Kaminsky-Lizondo-Reinhart(KLR) models over the in-sample period. Points that are closer to the lower rightare unambiguously preferred for any loss function, in that the percentage of crisescorrectly called is higher while the percentage of false alarms is lower. As shown in the figure, the DCSD model dominates for all cutoff frequencies, in that theDCSD curve lies to the right and below the KLR curve. For any given percentageof crises correctly called, the DCSD model calls fewer false alarms. Figures 3through 6 show similar results, though they each show one model’s in-sample and out-of-sample results.

Another way to see how models compare that does not depend on the cutoff isthe loss function graph shown in the lower panel of Figure 3. This shows how themodels perform for various cutoff thresholds, for a given loss function. To read thisfigure, note that the loss function is the number of false alarms (in percent of totaltranquil periods) plus the number of missed crises (in percent of total pre-crisis peri-ods). A loss function value of 50, for example, implies 20 percentage points fewerfalse alarms and/or missed crises than a loss function of 70.

An alternative approach would be to apply quadratic probability scores (QPSs)to evaluate goodness-of-fit, as is standard in the literature on evaluation of businesscycle forecasts (Diebold and Rudebusch, 1989), which in turn derives from theweather forecast evaluation literature (see Diebold and Lopez, 1996, for a survey ofthe forecast evaluation literature). In the context of EWS forecast evaluation, theQPS measure, which is an analog of a mean-squared error, was applied by Berg andPattillo (1999b). A shortcoming of QPSs, however, is that the differences betweenmodels are hard to evaluate because there are no probability distributions availablefor them. It is not possible to tell if one model beats another by a statistically signifi-cant margin. Moreover, QPSs do not provide an intuitive interpretation of modelperformance as do the loss functions used in this paper.



478

Figure 2. DCSD1 and KLR2 in Sample Forecasts

The Trade-Off of False Alarms and Missed Crises

0

20

40

60

80

100

0 20 40 60 80 100

Percent of Crisis Months Correctly Called

Per

cent

Fal

se A

larm

s

0

20

40

60

80

100

DCSDKLR

A

B

50

60

70

80

90

100

0 20 40 60 80 100

Cutoff

Val

ue o

f Los

s F

unct

ion DCSD

KLR

A

The Loss Function

BOptimal in-sample cutoff

Notes: See Box 2 for an explanation. Point A corresponds to the cutoff that minimizes the loss function in-sample for the DCSD model. Point B indicates the same point for the KLR model.

1Berg and others (2000). DCSD stands for Developing Country Studies Division of the IMF in which the model was originally formulated.

2Kaminsky, Lizondo, and Reinhart (1998).


ASSESSIN

G EA

RLY

WA

RN

ING

SYSTEM

S: HO

W H

AV

E THEY

WO

RK

ED IN

PR

AC

TICE?

47

9

Table 4. Goodness of Fit: DCSD and KLR Models

Asia Crisis Recent Experience

In sample Out of sample In sample Out of sampleDec. 1985 to Apr. 1995 May 1995 to Dec. 1996 Dec. 1985 to May 1997 Jan. 1999 to Dec. 2000

DCSD KLR DCSD KLR DCSD KLR DCSD KLR

Cutoff1 18 15 18 15 23 15 23 15Value of loss function2 59 73 63 76 58 68 90 63Percent of observations 73 57 62 57 76 70 72 76

correctly calledPercent of crises in 24 months 65 73 84 75 63 60 31 58

correctly called3

Percent of tranquil periods in 75 54 53 49 79 72 80 7924 months correctly called4

False alarms as percent of 69 78 60 62 64 71 78 65total alarms5

Probability of crisis given 31 22 40 38 37 29 22 35signal6

Probability of crisis given 8 8 10 18 8 10 14 9no signal7

Statistical tests of forecasts8

Coefficient in regression 1.19 1.02 2.13 1.42 1.14 0.87 0.61 1.47of actual on predicted

Standard error 0.30 0.25 0.38 0.74 0.18 0.22 0.39 0.52p-value (coefficient = 0) 0.00 0.00 0.00 0.06 0.00 0.00 0.12 0.01


An

dre

w Be

rg, Ed

ua

rdo

Bore

nszte

in, a

nd

Ca

the

rine

Pa

ttillo

48

0


Predicted9 Predicted Predicted Predicted Predicted Predicted Predicted Predicted

Actual TranquilCrisis

Source: Authors’ calculations.Notes: KLR: Kaminsky, Lizondo, and Reinhart (1998); DCSD: Developing Country Studies Division of the IMF (Berg and others, 2000). See Box 2 for fur-

ther explanation of items in this table.1This is the cutoff probability above which a forecast is deemed to signal a crisis.2The loss function is equal to the sum of false alarms as a share of total tranquil periods and missed crises as a share of total pre-crisis periods.3This is the number of pre-crisis periods correctly called (observations for which the estimated probability of crisis is above the cutoff probability and a crisis

ensues within 24 months) as a share of total pre-crisis periods.4This is the number of tranquil periods correctly called (observations for which the estimated probability of crisis is below the cutoff probability and no crisis

ensues within 24 months) as a share of total tranquil periods.5A false alarm is an observation with an estimated probability of crisis above the cutoff (an alarm) not followed by a crisis within 24 months.6This is the number of pre-crisis periods correctly called as a share of total predicted pre-crisis periods (observations for which the estimated probability of

crisis is above the cutoff probability).7This is the number of periods in which tranquility is predicted and a crisis actually ensues as a share of total predicted tranquil periods (observations for which

the estimated probability of crisis is below the cutoff probability).8Crisis dummy (1 if pre-crisis month, 0 otherwise) is regressed on forecast probabilities, with HAC standard errors. See Box 2 for explanation.9The number in each cell represents the number of observations that are predicted to be (actually are) either tranquil (T) or crisis (C), depending on the col-

umn (row).

T C T C T C T C T C T C T C T C

1,571 533 1,196 1,009 170 151 159 166 1,970 531 1,838 711 353 90 351 92128 242 108 286 19 101 34 101 180 305 196 291 59 26 36 49



481

What about out-of-sample performance? As Appendix II explains, the out-of-sample period for the KLR and DCSD models extends from January 1999 throughDecember 2000, since it is too early to fully judge more recent forecasts. For theprivate sector GS and CSFB models, which have three-month and one-month hori-zons, respectively, it is possible to look at goodness of fit through April and August2001, respectively.

The out-of-sample results vary substantially by model. The KLR model per-forms better out of sample than in sample, calling 58 percent of pre-crisis monthscorrectly. The forecasts were highly informative: when the crisis probability wasbelow the cutoff, a crisis ensued only 9 percent of the time, compared with 35 per-cent of the time when the crisis probability was above the cutoff. The DCSDmodel’s performance deteriorated substantially over this sample period, with only31 percent of pre-crisis months correctly called. The model remained somewhatinformative, with crises following above-cutoff signals 22 percent of the time andbelow-cutoff signals only 14 percent of the time.

Figures 3 and 4 give a more complete picture of the models’ performance. Thetop panel of Figure 3 shows, for example, that the DCSD model has more falsealarms out of sample than in sample for all cutoff levels. At the in-sample optimalcutoff, the model has about the same fraction of false alarms but calls many fewercrises correctly. The bottom panel shows that a much higher cutoff (around 50 per-cent) would have been desirable. It would have avoided many false alarms with-out increasing the number of missed crises very much. This is reflected in the toppanel for the DCSD model, where the out-of-sample curve lies on the in-samplecurve in the neighborhood of point C, which corresponds to a cutoff probability of47 percent.

The out-of-sample period studied in this paper was comparatively calm.According to the specific definition applied in the models, there were two crisesper year during the recent period, while the average number of crises per year inthe estimation period is almost three. It is perhaps not surprising then that the mod-els tended to predict the crises well in the recent period, but that they also regis-tered a comparatively high number of false alarms; that is, crisis predictions thatdid not materialize.

Of course, it is impossible to know until it is too late what the best cutoff prob-ability to call crises is. However, the successful goodness-of-fit performance of themodels for some cutoffs does imply that the models were able to rank the obser-vations reasonably well according to crisis probability, with the higher probabili-ties assigned to observations that correspond to pre-crisis months.

How statistically good (in the case of KLR) or bad (in the case of DCSD) arethese results? Table 4 also shows the results of regressing the actual value of thecrisis variable on the model’s predicted probability of crisis for various models andsample periods. Thus, we run a regression of the form

where c24it = 1 if there is a crisis in the 24 months after period t (for country i) and0 otherwise. PredProbit is the predicted crisis probability for period t and country

c it it it24 = + +α β ε* ,PredProb



482


0

20

40

60

80

100

0 20 40 60 80 100


Per

cent

Fal

se A

larm

s

0

20

40

60

80

100In-sampleOut-of-sample

B

C

A

The Loss Function

50

60

70

80

90

100

110

0 20 40 60 80 100

Cutoff

Val

ue o

f Los

s F

unct

ion

In-sampleOut-of-sample

A

C

B

Optimal in-sample cutoff

Notes: See Box 2 for an explanation. Point A corresponds to the cutoff that minimizes the loss function in sample. Point B indicates the out-of-sample results corresponding to the same cutoff. Point C corresponds to a cutoff of 47 percent.

1Berg and others (2000). DCSD stands for Developing Country Studies Division of the IMF, the Division in which the model was originally formulated.

Figure 3. DCSD1 Forecasts



483


0

20

40

60

80

100

120

0 20 40 60 80 100


Per

cent

Fal

se A

larm

s

0

20

40

60

80

100

120


A

B

The Loss Function

50

60

70

80

90

100

110

0 10 20 30 40 50 60 70 80 90 100Cutoff

Val

ue o

f Los

s F

unct

ion


A

B Optimal in-sample cutoff

Notes: See Box 2 for an expanation. Point A corresponds to the cutoff that minimizes the loss function in-sample. Point B indicates the out-of-sample results corresponding to the same cutoff.

1Kaminsky, Lizondo, and Reinhart (1998).

Figure 4. KLR1 Forecasts



484

i. For informative forecasts, β should be significant; a coefficient of 1 implies thatthey are unbiased.17

The regression results confirm the strong out-of-sample KLR forecasts andsuggest that the data are not revealing for the DCSD model. First, the strong KLRand DCSD in-sample results are clear. The estimated β is always statistically dif-ferent from 0 and a value of 1 cannot be rejected. Turning to the recent out-of-sample period, the KLR model’s forecasts are highly significant, while the hypoth-esis that the true β is 1 cannot be rejected. The DCSD forecasts are not significantat the traditional confidence levels, with a p-value for β = 0 of 12 percent. However,the data are more consistent with the hypothesis that the forecasts are accurate thanthat they are useless: the p-value for β = 1 is 31 percent.

As suggested by the fact that neither the hypothesis that β = 0 nor that β = 1can be decisively rejected, the tests lack power. This in turn reflects the smallamount of information in the out-of-sample period. We illustrate this lack ofpower by carrying out the following simulation exercise. We suppose that, in fact,the crisis probabilities in the out-of-sample period result from a process that isexactly as described by the DCSD model, and use the DCSD model to generate“data” on which to test the value of the coefficient β. The “data” we create implythat β = 1 and any remaining errors are a result of noise inherent in the data-generating process—the forecasts are as good as they could be. We then simulatethe data-generating process implied by the estimated DCSD model 500 times,creating 500 sets of out-of-sample observations and associated model forecasts.We ask how often these ideal forecasts would look as bad as those actually pro-duced by the DCSD model using the true out-of-sample data. The answer is thatfor these ideal forecasts, the hypothesis that β = 0 would not be rejected 28 per-cent of the time.

Neither of the short-horizon private sector models performs well (Table 5 andFigures 5 and 6).18 Even though in-sample goodness of fit was adequate, the

17We estimate this regression using ordinary least squares (OLS) with heteroskedasticity- and autocorrelation-corrected (HAC) standard errors. This solves two sorts of problems. First, the c24 andPredProb variables are highly serially correlated, which causes the OLS standard errors to be incorrect.Monte Carlo exercises suggest that in our setup, the OLS standard errors are substantial underestimatesbut that a HAC correction largely solves this problem. Second, the c24 variable is qualitative, resultingin a heteroskedastic �, as is well known from the “linear probability” literature. The usual solution is torun a probit or logit regression. Here, though, the relationship between PredProb and c24 will be linearunder either the null (with β = 0) or the alternative (with β = 1). The heteroskedasticity is of known form,suggesting the use of feasible generalized least squares estimators (FGLS). However, some observationswill produce negative variances. The usual solution is to apply some ad hoc adjustment to these observa-tions, such as dropping them. Our own experience and some Monte Carlo exercises confirm much earlierconclusions that these procedures are unsatisfactory, and suggest that OLS with HAC standard errors pro-duces reasonable results with only a small loss of efficiency compared with Generalized Least Squares(GLS). (See Judge and others (1982) on the linear probability model.) Berg and Coke (2004) discuss sim-ilar problems in the estimation of EWS models themselves. Harding and Pagan (2003) address relatedissues in a different context.

18Each model’s own definition of crisis is used to evaluate its performance. See Box 1 on crisis definitions.


ASSESSIN

G EA

RLY

WA

RN

ING

SYSTEM

S: HO

W H

AV

E THEY

WO

RK

ED IN

PR

AC

TICE?

48

5

Table 5. Goodness of Fit: Short-Horizon Models

In Sample Out of Sample

GS CSFB GS CSFBJan. 1996 to Dec. 1998 Jan. 1994 to Jul. 2000 Jan. 1999 to Apr. 2001 Aug. 2000 to Aug. 2001

Cutoff1 10 35 10 35Value of loss function2 66 58 97 88Percent of observations correctly called 66 76 50 83Percent of crises correctly called3 66 65 54 27Percent of tranquil periods correctly called4 66 76 50 85False alarms as percent of total alarms5 74 92 87 96Probability of crisis given signal6 26 8 14 4Probability of crisis given no signal7 8 2 12 2Statistical tests of forecasts8

Coefficient in regression of actual on predicted 1.41 0.17 0.56 0.06Standard error 0.42 0.03 0.41 0.06p-value (coefficient<=0) 0.00 0.00 0.17 0.29


An

dre

w Be

rg, Ed

ua

rdo

Bore

nszte

in, a

nd

Ca

the

rine

Pa

ttillo

48

6


Predicted9 Predicted Predicted Predicted

Tranquil Crisis Tranquil Crisis Tranquil Crisis Tranquil Crisis

Actual Tranquil 540 279 1,951 618 325 332 361 65Crisis 50 98 29 54 45 52 8 3

Source: Authors’ calculations.Notes: GS: Goldman Sachs (Ades, Masih, and Tenengauzer, 1998); CSFB: Credit Suisse First Boston (Roy and Tudela, 2000). See Box 2 for further expla-

nation of items in this table.1This is the cutoff probability above which a forecast is deemed to signal a crisis.2The loss function is equal to the sum of false alarms as a share of total tranquil periods and missed crises as a share of total pre-crisis periods.3This is the number of pre-crisis periods correctly called (observations for which the estimated probability of crisis is above the cutoff probability and a crisis

ensues in 3 months (GS), or 1 month (CSFB)) as a share of total pre-crisis periods.4This is the number of tranquil periods correctly called (observations for which the estimated probability of crisis is below the cutoff probability and no crisis

ensues in 3 months (GS), or 1 month (CSFB)) as a share of total tranquil periods.5A false alarm is an observation with an estimated probability of crisis above the cutoff (an alarm) not followed by a crisis in 3 months (GS), or 1 month

(CSFB).6This is the number of pre-crisis periods correctly called as a share of total predicted pre-crisis periods (observations for which the estimated probability of

crisis is above the cutoff probability).7This is the number of periods in which tranquility is predicted and a crisis actually ensues as a share of total predicted tranquil periods (observations for which

the estimated probability of crisis is below the cutoff probability).8Crisis dummy (1 if pre-crisis month, 0 otherwise) is regressed on forecast probabilities, with HAC standard errors. See Box 2 for explanation.9The number in each cell represents the number of observations that are predicted to be (actually are) either tranquil or crisis, depending on the column (row).



487


0

20

40

60

80

100

0 20 40 60 80 100


Per

cent

Fal

se A

larm

s

0

20

40

60

80

100


B

A

The Loss Function

50

60

70

80

90

100

110

0 20 40 60 80 100

Cutoff

Val

ue o

f Los

s F

unct

ion

50

60

70

80

90

100

110


A

B


Notes: See Box 1 for an explanation. Point A corresponds to the cutoff that minimizes the loss function in-sample. Point B indicates the out-of-sample results corresponding to this same cutoff.

1Goldman-Sachs (Ades, Masih, and Tenengauzer, 1998) and authors’ calculations.

Figure 5. GS1 Model Forecasts



488

Figure 6. CSFB1 Model Forecasts


0

20

40

60

80

100

0 20 40 60 80 100


Per

cent

Fal

se A

larm

s

0

20

40

60

80

100

In-sample

Out-of-sample

BA

The Loss Function

50

60

70

80

90

100

110

120

0 20 40 60 80 100Cutoff

Val

ue o

f Los

s F

unct

ion

50

60

70

80

90

100

110

120



B

A

Notes: See Box 1 for an explanation. Point A corresponds to the cutoff that minimizes the loss function in-sample. Point B indicates the out-of-sample results corresponding to this same cutoff.

1Credit Suisse First Boston (Roy and Tudela, 2000) and authors’ calculations.



489

models’ out-of-sample forecasts deteriorate sharply.19 The statistical tests reflectthis, in that the forecasts are not significant predictors of actual crisis incidence outof sample. Again, the data are not completely definitive, particularly for the GSmodel. Here, the p-value for the hypothesis that the forecasts should be given noweight is 0.17, while the p-value that the forecasts are unbiased predictors of cri-sis risk is 0.28.

All these results should be interpreted cautiously. The number of crises actu-ally observed has been limited. This translates into a small effective sample size.20

In this context, small changes in sample can make a large difference in the goodness-of-fit indicators yielded by the models. A previous version of this paper found thatthe DCSD model forecasts performed as well in the January to June 1999 to out-of-sample period as they did in sample.

The private sector models set out to accomplish a distinctly different task thanDCSD or KLR did; namely, to predict the timing of a crisis with precision. Theadoption of a shorter horizon may make prediction easier, since signs of crisis mayemerge more clearly right before a crisis. On the other hand, the exact timing of acurrency crisis may be more difficult to predict than vulnerability over an intervalof time as wide as two years, in part because of the possibility of multiple equi-libria and a resulting difficulty in predicting the timing of speculative attacks. Inany case, the comparison of the short-horizon models and the long-horizon mod-els is not direct and must be treated with caution.21

A final and necessary qualification is that the terms “false alarm” and “missedcrisis” should not be taken too literally. An alarm is considered false if no crisis infact ensues. However, this signal may have been appropriate. First, the crisis defi-nitions employed may fail to classify some events as crises that we might well wanta model to warn about. Second, a warning may be followed by policy adjustmentor luck that causes the crisis to be avoided; the warning might nonetheless havebeen useful. Indeed, an examination of the 90 observations that generated false

19This assessment is based on this paper’s metric for evaluating performance. CSFB’s own methoduses a different loss function to choose a cutoff, putting more weight on missed crises. In effect, theirobjective is to minimize false alarms, subject to achieving a certain share of correctly called crises. Alsonote that CSFB uses the probabilities in a more complex way to generate various levels of risk warningsfor clients, based on changes in the probabilities over the most recent one to six months. We have not eval-uated how well this system does in predicting crises.

20To put this problem another way, the pre-crisis observations and the predicted probabilities arehighly serially correlated; adjusting for this factor greatly increases the standard errors in the model. Thisalso implies that adding observations through extending the time dimension of the out-of-sample period isnot as helpful as the increase in the number of total observations would suggest.

21In addition, there are some differences in the way the out-of-sample forecasts were generated. Thosefor the GS model come directly from contemporary monthly publications, so they necessarily reflect incom-plete data that had to be supplemented with estimates for various predictive variables. For example, theJanuary 1999 crisis probability (for April 1999) uses GS estimates of data as of January 1999. The CSFBestimates, in contrast, may use revised predictive variables, although it is not clear how substantial the revi-sions are. This issue is somewhat less serious for the DCSD and KLR models because they forecast overmuch longer horizons. The July 1999 forecasts, for example, made use of data available only through Aprilfor many series, but because the forecast horizon is so long, the use of such data did not make the forecastsobsolete.



490

alarms in the January 1999 to December 2000 period in the DCSD model suggeststhat half of them (46) can be readily classified in one or the other of these cases.22

III. Summary and Conclusions

Since the beginning of 1999, IMF staff has been systematically tracking, on anongoing basis, various models developed in-house and by private institutions, aspart of its broader forward-looking vulnerability assessment. This paper looks indetail at the performance of these models in practice.

We have monitored two long-horizon in-house models (DCSD and KLR) andtwo short-horizon private sector models (GS and CSFB) since 1999. This paperhas analyzed the forecasts made between January 1999 and December 2000 by the24-month-horizon DCSD and KLR models, between January 1999 and April 2001by the GS model, and between April 2000 and June 2001 by the CSFB model.These forecasts were “pure” out-of-sample forecasts in that no information aboutactual outcomes was used in the forecasts or, more generally, in the developmentor estimation of the models themselves.

The results are mixed. The forecasts of the KLR model are statistically andeconomically significant predictors of actual crises. The forecast accuracy in theout-of-sample period is only slightly inferior to the accuracy in the estimationperiod. The DCSD model performs substantially worse out of sample than in sam-ple. The forecasts are still somewhat informative, however, and the hypothesis thatthe forecasts are unbiased and informative is more likely (p-value = 0.31) than thehypothesis that the model’s forecasts were useless ( p-value = 0.12). The ambigu-ous statistical results reflect the fact that the out-of-sample period contains 528observations but only eight crises; the latter number is important in determiningthe amount of information in the data.23

On the whole, the short-horizon private sector models we examined performedpoorly out of sample, despite stellar in-sample performance. Both sets of crisispredictions were largely uninformative—the probability of a crisis was about thesame whether the forecast probability was above a cutoff threshold or not.

At least for the KLR model and the DCSD model, the forecasts were statisti-cally significant predictors of crisis (only at the 12 percent level for the latter).This means that they are likely better than could have been produced throwingdarts at a a suitable target. How do they compare, though, to the more challengingbenchmark of alternative forecasts? Here we have compared them with bondspreads, agency ratings, and, perhaps most relevant for IMF work, overall cur-rency crisis risk scores published by analysts. We find that during the Asia crisis,these alternative indicators fared very poorly, much worse than the DCSD and

22The countries involved in these 46 observations of technically false, but still useful, alarms areArgentina, Chile (before July 1999), Pakistan, Turkey, Uruguay, and República Bolivariana de Venezuela.

23The serial correlation in the data also reduces the effective amount of information, as discussed innote 17. An earlier version of this paper analyzed data through end-1999 and found that the DCSD modelperformed as well out of sample as in sample. This dependence of the results on the sample is captured bythe low power of the tests.



491

KLR EWS models. During the recent period, the non-model forecasts performedsomewhat better, though still generally not as well as the models.

Overall, these results reinforce the view that EWS models are not accurateenough to be used as the sole method to anticipate crises. However, they can con-tribute to the analysis of vulnerability in conjunction with more traditional surveil-lance methods and other indicators. It is worth underlining the relatively highstandard to which these models are being held. It is plausible to suppose that com-prehensive assessments by informed analysts, based on all available qualitativeand quantitative information, must be better than the inevitably simple EWS mod-els. But the evidence we have examined with respect to this question is not encour-aging concerning these more comprehensive assessments.

The advantage of EWS models lies in their objective, systematic nature. Themodels process data in a mechanical way and are not clouded by conventionalmisperceptions or biases based on past experiences. For example, as shown inSection II, Korea, a country with one of the most successful economic records inrecent years, was showing some serious signs of vulnerability to an external crisisin 1996–97, according to EWS models. However, possibly because of that suc-cessful record, analysts and markets did not signal any increase in risk prior to theDecember 1997 currency crisis.

Over the Asia crisis periods, the best EWS models did dramatically better thannon-model-based predictors, such as spreads, ratings, and assessments of informedanalysts. Over the more recent period, the performance of some of these alterna-tive predictors improved somewhat, so that the relative superiority of the modelsdeclined. This suggests that recent crises have simply not been the surprises that theAsia crises were, either because they were easier to predict or because analysts’sensitivity was heightened. In general, most analysts foresaw important risks in thecrisis countries in question. The expected strength of the EWS models is in iden-tifying important crisis risks that other forms of analysis do not expose, whileavoiding an excessive number of false alarms that would dilute the credibility ofthe crisis signals. The crises of 1999–2002 have not, fortunately, afforded thisopportunity.

Looking at events of the past few years, it is clear that several developmentsare under way that are changing the landscape for currency crisis prediction mod-els. First, we have observed a resurgence of crises in which sovereign and domes-tic debt dynamics play a central role. Debt and currency crises are related butdistinct: most debt crises are associated with currency crises, but the reverse is nottrue. Recent work has, appropriately, focused on predicting these sorts of crises.24

A second trend is the increased importance of floating exchange rates inemerging markets. Over the past decade, in large part in the aftermath of currencycrises, there has been a sharp increase in the number of emerging market countriesin which the capital account is broadly open, the de jure exchange rate regime isfloating, and there is substantial de facto flexibility as well. Ten years ago, perhaps

24Detragiache and Spilimbergo (2001) and Manasse, Roubini, and Schimmelpfennig (2003) look atdeterminants of debt crises. Hemming, Kell, and Schimmelpfennig (2003) look at fiscal vulnerabilities inemerging market economies. Sy (2003) emphasizes that debt and currency crises are distinct events.



492

only South Africa fit these characteristics among large developing countries. Nowmany, including Brazil, Chile, Colombia, Korea, Mexico, Poland, and Thailand,have joined the ranks. This sort of arrangement, often augmented by an inflation-targeting monetary policy, is indeed now more the rule than the exception for suchcountries. Going forward, what does this imply for currency crisis models?

In principle, a sharp depreciation of the exchange rate can happen under a float-ing exchange rate regime as much as under any other regime. Although floating rateregimes would help to avoid situations of extreme overvaluation, particularly thosedriven by policy inconsistencies, the economies could still be vulnerable to suddenchanges in market sentiment, unsustainable levels of debt, and financial sectorweaknesses, among other factors. Moreover, regimes that are broadly floating may,under speculative attack, evolve toward de facto pegs as policymakers resist thedownward pressure on the currency. In fact, according to the IMF’s de jure classi-fication, there is no evidence that floating rates have been more resistant to currencycrises.25 We draw no firm conclusions here. We suspect, though, that painful cur-rency crises will remain a feature of emerging markets for the foreseeable future.

APPENDIX I

Description of Early Warning System (EWS) Models and Specification Issues

Models Implemented at the IMF

Kaminsky-Lizondo-Reinhart (KLR) model

Perhaps the most prominent model for predicting currency crises proposed before the Asia cri-sis is the indicators approach of Kaminsky, Lizondo, and Reinhart (1998) (KLR), who moni-tor a large set of monthly indicators that signal a crisis whenever they cross a certain threshold.The model attempts to predict the probability of a crisis within the next 24 months, where acrisis occurs when there are extreme changes in a weighted average of the monthly exchangerate depreciation and reserve loss. A variable-by-variable approach is chosen so that a surveil-lance system based on the method would provide assessments of which variables are “out ofline.” In addition to overvaluation, the current account, reserve losses, and export growth, themodel also includes reserves to broad money as a measure of reserve adequacy and severalmonetary variables, such as domestic credit growth, real interest rates, and excess M1 bal-ances.26 The information from the separate variables is combined, using each variable’s fore-casting track record, to produce a composite measure of the probability of crisis (Kaminsky,1999). IMF staff has implemented a version of the KLR model, supplemented with severaladditional variables.

25A more complete analysis should correct for the influence of other factors that contribute to currencycrises and would consider de facto classifications such as those of Levy-Yeyati and Sturzenegger (2001)and Reinhart and Rogoff (2004).

26Goldstein, Kaminsky, and Reinhart (2000) add some new indicators and update the KLR model.They find that the best monthly indicators for predicting a currency crisis were real exchange rate appre-ciation, a banking crisis, a decline in equity prices, a fall in exports, a high ratio of broad money to reserves,and a recession, while the best annual indicators were a large current account deficit relative to both GDPand investment.



493

Developing Country Studies Division (DCSD) model

The current structure of the DCSD model has been influenced by the path of its development.Its origins were a project testing the out-of-sample performance of the KLR and other modelsin predicting the Asian crisis. Later work tested the usefulness of interpreting predictive vari-ables in terms of discrete thresholds, the crossing of which signals a crisis (Berg and Pattillo,1999a and 1999b). Using the same crisis definition and prediction horizon as KLR, but embed-ding the KLR approach into a multivariate probit regression, the authors found that a better sim-ple assumption is that the probability of crisis goes up linearly with changes in the predictivevariables. The variables are measured in percentile form; that is, relative to their own history.

The resulting “linear” probit model in that paper was composed of five variables: realexchange rate deviations from trend, the current account to GDP ratio, export growth, reservegrowth, and the level of M2 to reserves. This set of predictor variables was the result of start-ing with an extensive set of KLR variables, plus our additions, and selecting the most impor-tant variables through a specification search process.

Because the role of short-term debt in weak financial systems was brought to the forefrontby the Asian crises, a measure of short-term debt to reserves was added to the model (Berg andothers, 2000). It was found to be highly significant, while the ratio of M2 to reserves lost itssignificance and was dropped from the model, resulting in the current five-variable DCSDmodel.27

Policy Development and Review (PDR) model

A third EWS model recently developed by the IMF is the PDR model. This EWS adds balancesheet variables and proxies for standards to the DCSD model (Mulder, Perrelli, and Rocha,2002). The following variables have been deemed important in predicting the probability of acrisis: at the corporate level, leveraged financing and a high ratio of short-term debt to workingcapital; balance sheet indicators of bank and corporate debt to foreign banks as a share ofexports; and a legal regime variable proxying shareholder rights.

The corporate sector data are available only on an annual basis and with a significant lag.These variables are often slow moving, however, so they can still contribute to forecasting accu-racy. More up-to-date data, as well as a larger and more stable underlying sample of corpora-tions, would increase the analytical and forecasting usefulness of the model.

Private Sector Models

The interest of investment banks in developing EWSs as tools for advising their clients has fluc-tuated, based on what is “in fashion” and whether crises are in the daily headlines. Following theAsian crisis, most major banks developed in-house models attempting to predict currencycrashes. These models were designed either for explicit use in advising on foreign currency trad-ing strategies, or, more generally, to assess values and risks in emerging market currencies andsupplement economic forecasts provided to investors. Since that time, a number of these systemshave ceased operation: Lehman Brothers has abandoned its Currency Jump Probability model;Citicorp no longer implements its Early Warning System for anticipating balance-of-paymentscrises in Latin America; and JPMorgan has substituted a simple weighted vulnerability index forits Event Risk Indicator model. However, as volatility in emerging markets ratcheted up again inlate 2000 and in 2001, a number of new private models were brought out. For example, Deutsche

27The model uses mainly monthly data, but also some quarterly or, for some countries, annual data.These latter series are interpolated or extrapolated to generate monthly crisis predictions.



494

Bank introduced its Deutsche Bank Alarm Clock (DBAC) and Morgan Stanley Dean Witter hasrecently set up an Early Warning System for Currency Crises.

Goldman Sachs’s GS-WATCH

IMF staff regularly tracks Goldman Sachs’ GS-WATCH model and Credit Suisse First Boston’s(CSFB) Emerging Markets Risk Indicator (Roy and Tudela, 2000), both of which have been inoperation since 1998. GS-WATCH (Ades, Masih, and Tenengauzer, 1998) predicts the likeli-hood of a crisis in a three-month period, defined as a weighted average of three-monthexchange rate and reserve changes. The predictions are generated through a logit regression inwhich most explanatory variables are converted into zero/one signals. The predictor variablesinclude macro fundamentals such as measures of credit booms, real exchange rate misalign-ment, export growth, reserve growth, and external financing requirements, as well as changesin stock prices, political risk, contagion, and global liquidity. The latter two variables are mea-sured continuously, making the overall crisis probabilities follow a smoother path. While inclu-sion of political risk makes sense, the simple zero/one variable (one around the time ofelections, or when a revolution, coup, major riot, or strike takes place) only partially capturesthis type of risk. The model is estimated using monthly data, but predictions are updated weeklyfor inclusion in analysts’ reports. On a week-to-week basis, changes in the contagion variabledrive much of the movement in the crisis predictions. Contagion is measured for each countryas a weighted average of the changes in the exchange rate and reserve change index for theother countries in the sample, where the weights are the historical relationships between thoseindices across countries.

Credit Suisse First Boston’s Emerging Markets Risk Indicator (CSFB)

CSFB re-specified its model in September 2000, changing some of the predictor variables andreducing the number of variables (Roy and Tudela, 2000). A logit model predicts the one-month-ahead probability of a depreciation greater than 5 percent and at least double the pre-ceding month’s depreciation. The variables are standardized; that is, measured relative to thecountry-specific mean and variability for that variable. Many variables similar to those used inother models are included: real exchange rate deviations from trend; the ratio of debt to exports;growth in credit to the private sector; output changes; reserves to imports; changes in stockprices; oil prices; and a regional contagion dummy, measured simply as the number of coun-tries in the region recently experiencing a crisis.

Deutsche Bank Alarm Clock (DBAC)

The DBAC model defines separate exchange rate and interest rate “events” as depreciationsgreater than a certain size (estimated separately for levels ranging from 5 percent to 25 percent)and increases in money market interest rates of more than 25 percent in a month (Garber,Lumsdaine, and van der Leij, 2000). It uses a methodology to jointly estimate the probability ofthese two types of events, allowing the probability of a simultaneous interest rate event to influ-ence the likelihood of an exchange rate crisis and the probability of a depreciation event to affectthe prediction of an interest rate crisis. Relatively few predictors are included in the exchange rateevent model: changes in stock prices, domestic credit, industrial production, and real exchangerate deviations, as well as a contagion variable. All the investment bank models claim to demon-strate that an investor using trading strategies based on their models could earn substantial prof-its over a particular period. DBAC adds a twist to these calculations by proposing an “actiontrigger” to identify cutoff probability levels at which an alarm should be sounded and investors



495

should change their positions. The trigger is calculated to maximize profits, assuming a strategyin which the investor will be long in the local currency when the probability of a depreciation cri-sis is below the trigger and short whenever the probability crosses above the trigger.

Specification Issues

EWS models are econometric methods for generating predictions of currency crises, preciselydefined. Although there have long been empirical studies of currency crises, it was not untilafter the 1994–95 Mexican tequila crisis that the literature focused on finding methods for pre-dicting crises, rather than on explaining a particular set of historical crises or testing specifictheories. The largely unexpected Asia crisis, however, provided the real impetus for a new waveof papers and the development of systems for a continuous monitoring of crisis vulnerabilitiesat various institutions.

What is being predicted?

Most would agree that a sudden, large depreciation of the exchange rate constitutes a currencycrisis. Further, a situation of intense pressure on the foreign exchange market, resulting in largelosses of international reserves and/or a hike in domestic interest rates can also be considereda crisis, even if a step devaluation is avoided. In any event, one may be interested in forecast-ing both successful (those resulting in an exchange rate depreciation) and unsuccessful attackson the currency, so that both types of event would be considered a crisis for the purpose of aforecasting model. Box 1 discusses the difficulties involved in operationalizing the concept ofcurrency crisis and how they are addressed in the models considered in this paper. Table A.1lists crisis dates for the various models for the 1999–2001 period.

What variables should be included?

After identifying a set of crises, the next issue is the choice of a set of variables that may beuseful in predicting crises. Berg and others (2000) survey the literature on currency crises andlook for common symptoms of crises in past episodes. Drawing up a list of potential predictivevariables starts with theoretical models of currency crisis. “First-generation” models focus onmacroeconomic imbalances that lead to a depletion of foreign exchange reserves and make adevaluation inevitable. In second-generation models, the government weighs the cost and ben-efits of defending the currency. Because expectations affect the trade-off faced by the policy-maker, crises may be self-fulfilling, and thus much more difficult to predict. More recentmodels stress elements such as market failure in international capital markets and distortions indomestic financial markets. For example, information failures can lead to investor herdingbehavior and contagion, and implicit government guarantees of private sector liabilities cangenerate moral hazard and unsustainable implicit deficits.

The theoretical literature suggests classifying variables into three groups: first, macroeco-nomic fundamentals such as measures of real exchange rate overvaluation, the fiscal deficit,excess money growth, terms of trade, domestic credit, the current account deficit, and outputgrowth; and second, variables that gauge a country’s vulnerability to attacks, if, given relativelyweak fundamentals, an attack were to take place. These include measures of the adequacy ofinternational reserves relative to possible short-run liabilities of external and domestic origin,external financing needs, and soundness of the financial sector.

The third group of variables is composed of indicators of market expectations or sentiment,such as interest rate differentials, bond spreads, the forward exchange rate, the number of criseselsewhere or other contagion channels, and variables representing investors’ “risk appetite.”



496

Table A.1. Crisis Dates According to Different Models

DCSD/KLR Crisis Dates (Jan. 1999–Mar. 2003)

Brazil Jan 99Colombia Aug 99; Jul 02South Africa Dec 01Turkey Feb 01Uruguay Jul 02Venezuela Feb 02Zimbabwe Aug 00

GS Crisis Dates (Jul. 1998–Apr. 2001)

Brazil Jul 98–Jan 99; Jun–Jul 99; Mar–May 00; Sep 00; Jan–Apr 01Bulgaria Jan–Apr 99; Jan–Feb 00; Feb 01Chile Jun–Aug 99; Mar–Apr 01China Jul–Aug 98Colombia Jul–Aug 98; Mar–Jul 99; Mar–May 00Czech Republic Nov 98–Jan 99Ecuador Jul–Aug 98; Nov 98–Feb 99; Apr–May 99; Jul–Nov 99; Jan 01Egypt Jul 99; Sep–Oct 00Hong Kong SAR Jul–Sep 98; May–Jul 99Hungary Jul 98India May 00Indonesia Dec 98; Jun–Jul 99; Mar–May 00; Jan–Mar 01Israel Jul–Sep 98; Jul 99Korea, Rep. of Sep–Nov 00Malaysia Aug 00Mexico Jul 98Peru Jul–Nov 98; Jul 99; Nov 00Philippines Jul 98; Jun–Jul 99; May–Oct 00Poland Nov 98; Jul 99; Mar 00Russia Jul–Nov 98; Jul–Aug 99South Africa Sep–Oct 00Singapore Oct 98; Mar 01Taiwan Province of China Sep–Nov 00; Apr 01Thailand Jun–Jul 99; Feb–Aug 00Turkey Sep–Nov 98; Nov 00–Mar 01

CSFB Crisis Dates (Jul. 1998–Aug. 2001)

Brazil Mar 99Colombia Nov 98; Sep 99Croatia Apr 99Czech Republic Apr 99Ecuador Mar–Apr 99; Aug 99; Nov–Dec 99; Feb 00Indonesia Jul 98; Mar 99; Oct 99; Jan 00; Nov 00; Jun 01Israel Dec 98Korea, Rep. of Feb 01Mexico Oct 98Nigeria May 99; May 01Pakistan Nov 00Philippines Dec 00Poland Apr 99Russia Oct–Nov 98; Feb 99; Jun 99; Mar 00South Africa Aug 98Slovak Republic Oct 98; Jul 00



497

The task of specifying a model with variables that are useful predictors of crisis does notinvolve simply assembling all the a priori plausible variables. There is significant danger of“overfitting” a model by adding more and more variables through “data mining.” Typically, sucha model will explain a particular historical episode of crisis very well, but will have little powerin forecasting the next set of crises. Finding the best method to forecast crisis probabilities arguesfor a parsimonious model: a robust set of variables useful for predicting both past and futurecrises. There is a deeper problem associated with the statistical one. If the nature of criseschanges from one episode to the next, how can a model be robust to those changes? The answeris to focus on the symptoms that may be common to all external crisis episodes, even when theultimate causes of those crises are different.

It should also be kept in mind that the different indicators are interrelated, so that the inclu-sion of all of them is not necessary. The indicators may be covered indirectly, in that the vari-ables employed in the model may capture many of the important manifestations of these otherproblems. For example, a large fiscal deficit and high inflation may contribute to the risk of cri-sis, but may be already accounted for in a model that includes real exchange rate overvaluationand the current account deficit.

Finally, there are the issues of availability of consistent data over time and across coun-tries, and at a desirably high frequency. Data on the health of the financial sector, such as ratesof nonperforming loans, is an important example of factors that do not meet those standards.Political risk is another example of a factor that is intrinsically difficult to measure consistently.In addition, some variables may not fit well into the structure of a given model. A good exam-ple is the phenomenon of contagion. The transmission of crises from country to country, par-ticularly if the mechanism operates through financial channels, seems to occur quite rapidly.Thus, it is difficult to incorporate contagion in models attempting to predict the likelihood ofcrisis over a longer horizon, such as the next two years. Also, there are other idiosyncratic vari-ables (for example, oil prices) that, while particularly important for some countries, may haveinsignificant or contrary effects in other emerging markets.

How do you generate predictions?

Two conceptual questions underlie the choice of a methodology that uses the variables to pre-dict the crises. First, how should the information content of each explanatory variable beassessed? One option is the “signaling” approach, in which each indicator is said to issue a signal of impending crisis when its value exceeds a particular threshold. For example, if thecountry-specific threshold for the ratio of the current account deficit to GDP is 3 percent, a ratiobelow 3 percent would not contribute to the risk of crisis, while ratios above 3 percent wouldcontribute equally to the probability of a crisis. A second option is to introduce the variables

Table A.1. (Concluded)

Sri Lanka Aug 00; Mar 01Thailand Aug 98; Nov 99Turkey Jan 99; Apr–May 01; Aug 01Zimbabwe Jul 98; Oct 98; Mar 99; Oct 00

Source: Authors’ calculations.Note: KLR: Kaminsky, Lizondo, and Reinhart (1998); DCSD: Developing Country Studies

Division of the IMF (Berg and others, 2000); GS: Goldman Sachs (Ades, Masih, and Tenengauzer,1998); CSFB: Credit Suisse First Boston (Roy and Tudela, 2000).


28The Bank for International Settlements adopts a less common approach, in which, after each variableis converted to a score from a set scale, the scores are aggregated by summing, using judgmental weights.

29There are also a number of new approaches that are being explored in the literature. For example,Burkart and Coudert (2000) use linear discriminant analysis; Vlaar (2000) and Fratzscher (2003) developswitching regime models; and Osband and Van Rijckeghem (2000) use non-parametric methods to iden-tify safe zones.

30The Goldman Sachs GS-WATCH model also uses predictive indicators in zero/one form, but theseare used as regressors in a logit model. Therefore, the probabilities are less “jumpy” than in the KLR indi-cators model.


498

continuously so that, for example, any small increase in the current account/GDP ratio couldmarginally increase the crisis prediction.

It is also necessary to decide how the variables should be measured. Some models includethe variables in raw form, often in growth rates or ratios. Alternatively, the variables could bemeasured relative to their history for each country. For example, what matters in the DCSDmodel is not the level of the current account deficit per se but whether the deficit corresponds toa high percentile, relative to the history of the current account deficit in each country consideredindividually.

The second question is how to aggregate the information from the different variables intoa single prediction. A method associated with the signaling approach is the calculation of acomposite probability as the weighted sum of the number of indicators that are signaling, whereeach indicator is weighted by its reliability in predicting crises.28 An alternative is to use a probit (or logit) regression; that is, a regression in which the dependent variable takes the value of one when there is a crisis and zero otherwise.29

What are the relative advantages and disadvantages of each approach? The indicatorapproach is a popular one, because the framework of monitoring key variables for signs of“unusual” behavior accords well with the intuition of early warning. But, by evaluating eachvariable separately, the method does not consider how an interrelated set of conditions couldmake an economy more vulnerable to crisis. A practical difficulty with the indicator approachis that the crisis probabilities tend to be “jumpy,” as variables move in and out of the signal-ing territory, making interpretation difficult.30 A probit regression addresses many of the prob-lems with the indicator approach: it generates predictions taking into account the correlationamong all the predictive variables, and allows testing of the statistical significance of indi-vidual variables. However, because the probit is a nonlinear model, the contribution of a par-ticular variable depends on the magnitude of all the other variables. This means that therelationship between changes in the variables themselves and changes in their contribution tothe crisis prediction is not always transparent. In the final analysis, the relative merits of thetwo approaches are decided by one key factor: how successful is each method in predictingcrises?

Forecasting horizon

Another important design issue for models that attempt to predict both the cross-country inci-dence and timing of crises is how far in advance the prediction is to be made. Neither the KLRnor the DCSD model attempts to predict the exact timing of the crisis (which may be muchharder or impossible), but rather the likelihood that a crisis will occur sometime in the follow-ing 24 months. The relatively long prediction window could be useful for the IMF because itwould permit sufficient lead time for the authorities to make some policy adjustments. In fact,research on the DCSD model has shown relatively little difference in the estimated model usingany horizon between nine months and two years.



499

Private sector models tend to attempt to predict the probability of a crisis over a shorterhorizon, from one to three months. Some investment banks provide weekly updates of crisispredictions to their clients, although only a small subset of the variables changes at this fre-quency. This prediction horizon clearly relates to these firms’ objectives of providing advice toclients participating mainly in foreign exchange markets, who may use shifting short-term fore-casts to continually adjust their portfolios or hedge their positions. Different sets of variablesmay be important predictors at short horizons. For example, the three private sector modelstracked by the IMF staff all include a measure of contagion in the model, reflecting the fact thatcontagion can occur relatively rapidly in emerging markets. Changes in stock prices anddomestic credit to the private sector have also been found to be important predictive variablesin all three private sector models.

APPENDIX II

Meaning of In-Sample and Out-of-Sample Periods in Early Warning System (EWS) Models

The text emphasizes the distinction between in-sample and out-of-sample performance. Thisappendix defines these terms and explains their implementation in this paper. The designer of anEWS chooses the variables and estimates the parameters of the model in a way that best fits theobservations in a particular sample (the estimation sample). In-sample testing measures howwell the models fit the crises in a particular sample. Good in-sample testing is a sign of a usefulmodel but must be interpreted cautiously. Good in-sample performance may be a coincidence,perhaps resulting from a search through a large number of specifications until a good fit occursby chance. Moreover, the determinants of crises may vary over time.

In out-of-sample testing, the predictions of an existing model are compared with a new setof observations not belonging to the estimation sample. An unavoidable difficulty with out-of-sample testing is that a forecast can be properly judged only after the entire forecast windowhas closed. This paper examines the forecasts through June 1999 of the Kaminsky-Lizondo-Reinhart (KLR) and Developing Country Studies Division (DCSD) models, since it is too earlyto fully judge more recent forecasts. A prediction of risks as of August 1999, for example, can-not be fully judged until September 2001. Before then, it cannot be known whether August1999 was in fact a pre-crisis or a tranquil month, since it would not yet be known whether a cri-sis followed within 24 months. Given the two-year model horizon, these forecasts apply to thetwo-and-a-half-year period through July 2001. For the private sector Goldman Sachs (GS) andCredit Suisse First Boston (CSFB) models, which have three-month and one-month horizons,respectively, it is possible to look at goodness of fit through April and August 2001.31

Out-of-sample testing should mimic the way a forecasting model would be used in practice.In the strictest and most interesting form of out-of-sample testing, the modeler has no knowledgeof the out-of-sample observations when generating the forecasts to be tested. Sometimes, in con-trast, the modeler may withhold the most recent observations from the estimation sample, usingthem for subsequent out-of-sample testing. The modeler may nonetheless use information fromthese observations to create the model. For example, the DCSD model was estimated over thepre–Asia crises period and used to predict the Asia crises out of sample in Berg and others(2000). However, the authors created the model in 1998, after the Asia crises, and they added the

31Similarly, in-sample estimation periods for KLR and DCSD must end some 24 months before themodel is estimated. For example, the in-sample period for the DCSD model in Berg and others (2000) endedin May 1995, so that the estimation did not reflect knowledge of the Asia crises that began in July 1997.



500

short-term debt and reserves variable because they knew it was likely to be important in explain-ing the Asia crises.

The start dates for the out-of-sample periods examined in this paper were chosen becausethey followed the dates at which they could have informed the estimation of the models. TheKLR and DCSD forecasts examined here, for the period of January 1999 to December 2000,correspond to the versions used for the “official” internal July 1999 forecasts and subsequentinternal forecasts. The model specifications were finalized in late 1998. The GS out-of-sampleforecasts come directly from contemporary monthly publications over the period of January1999 to April 2001, so they could not have reflected out-of-sample information. The CSFB out-of-sample estimates for the April 2000 to June 2001 period were produced in August 2001using the model as it had been estimated a year earlier, so in principle they should not have beeninfluenced by out-of-sample events.

REFERENCES

Abiad, Abdul, 2003, “Early Warning Systems: A Survey and a Regime-Switching Approach,”Working Paper 03/32 (Washington: International Monetary Fund).

Ades, Alberto, Rumi Masih, and Daniel Tenengauzer, 1998, “GS-Watch: A New Framework forPredicting Financial Crisis in Emerging Markets,” (New York: Goldman Sachs).

Berg, Andrew, Eduardo Borensztein, Gian Maria Milesi-Ferretti, and Catherine Pattillo, 2000,Anticipating Balance of Payments Crises: The Role of Early Warning Systems, IMFOccasional Paper 186 (Washington: International Monetary Fund).

Berg, Andrew, and Rebecca Coke, 2004, “Autocorrelation-Corrected Standard Errors in PanelProbits: An Application to Currency Crisis Prediction,” IMF Working Paper 04/39(Washington: International Monetary Fund).

Table A.2. Model Samples

Sample

Asia crisisIn sample

KLR/DCSD Dec. 1985 to Apr. 1995Out of sample

KLR/DCSD May 1995 to Dec. 1996Recent experience

In sampleKLR/DCSD Dec. 1985 to May 1997GS Jan. 1996 to Dec. 1998CSFB Jan. 1994 to Jul. 2000

Out of sampleKLR/DCSD Jan. 1999 to Dec. 2000GS Jan. 1999 to Apr. 2001CSFB Aug. 2000 to Aug. 2001

Source: Authors’ calculations based on models.Note: KLR: Kaminsky, Lizondo, and Reinhart (1998); DCSD: Developing Country Studies

Division of the IMF (Berg and others, 2000); GS: Goldman Sachs (Ades, Masih, and Tenengauzer,1998); CSFB: Credit Suisse First Boston (Roy and Tudela, 2000).



501

Berg, Andrew, and Catherine Pattillo, 1999a, “Are Currency Crises Predictable? A Test,” IMFStaff Papers, Vol. 46, No. 2 (June), pp. 107–38 (also issued as IMF Working Paper 98/154and published in popularized form in 2000, “The Challenge of Predicting EconomicCrises,” Economic Issues, No. 22, Washington: International Monetary Fund).

———, 1999b, “Predicting Currency Crises: The Indicators Approach and an Alternative,”Journal of International Money and Finance, Vol. 18, No. 4 (August), pp. 561–86.

———, 1999c, “What Caused the Asian Crises: An Early Warning System Approach,”Economic Notes, Vol. 28, No. 3 (November), pp. 285–334.

Burkart, Olivier, and Virginie Coudert, 2000, “Leading Indicators of Currency Crises inEmerging Economies,” Notes d’Etudes et de Recherche #74 (Paris: Banque de France),also 2002, Emerging Markets Review, Vol. 3, No. 2, pp. 107–33.

Demirgüc-Kunt, Asli, and Enrica Detragiache, 1999, “Monitoring Banking Sector Fragility:A Multivariate Logit Approach,” IMF Working Paper 99/147 (Washington: InternationalMonetary Fund).

Detragiache, Enrica, and Antonio Spilimbergo, 2001, “Crises and Liquidity: Evidence andInterpretation,” IMF Working Paper 01/02 (Washington: International Monetary Fund).

Diebold, Francis, and José Lopez, 1996, “Forecast Evaluation and Combination,” in Handbookof Statistics 14: Statistical Methods in Finance, ed. by G. S. Maddala and C. R. Rao(Amsterdam: North-Holland), pp. 863–83.

Diebold, Francis, and Glenn Rudebusch, 1989, “Scoring the Leading Indicators,” Journal ofBusiness, Vol. 62, No. 3, pp. 369–91.

Economist Intelligence Unit, 2001, “EIU Country Risk Service June Handbook, 2001”(London: Economist Intelligence Unit).

Fratzscher, Marcel, 2003, “On Currency Crises and Contagion,” International Journal ofFinance and Economics, Vol. 8, No. 2, pp. 109–29.

Ferri, G., L-G. Liu, and J. E. Stiglitz, 1999, “The Procyclical Role of Rating Agencies:Evidence from the East Asian Crisis,” Economic Notes, Vol. 28, No. 3 (November),pp. 335–55.

Garber, Peter M., Robin L. Lumsdaine, and Marco van der Leij, 2000, “Deutsche Bank AlarmClock: Forecasting Exchange Rate and Interest Rate Events in Emerging Markets” (NewYork: Deutsche Bank).

Goldfajn, Ilan, and Rodrigo O. Valdés, 1998, “Are Currency Crises Predictable?” EuropeanEconomic Review, Vol. 42, No. 3–5 (May), pp. 873–85.

Goldstein, Morris, Graciela L. Kaminsky, and Carmen M. Reinhart, 2000, Assessing FinancialVulnerability: An Early Warning System for Emerging Markets (Washington: Institute forInternational Economics).

Harding, Don, and Adrian Pagan, 2003, “Synchronization of Cycles” (manuscript; Victoria,Australia: University of Melbourne).

Hemming, Richard, Michael S. Kell, and Axel Schimmelpfennig, 2003, Fiscal Vulnerabilityand Financial Crises in Emerging Market Economies, IMF Occasional Paper No. 218(Washington: International Monetary Fund).

International Monetary Fund, 2002, “Early Warning System Models: The Next Steps Forward,”in Global Financial Stability Report (March) (Washington).

Judge, George G., William E. Griffiths, R. Carter Hill, Helmut Lütkepohl, and Tsoung-ChaoLee, 1982, The Theory and Practice of Econometrics (New York: John Wiley and Sons).

Kamin, Steven B., John W. Schindler, and Shawna L. Samuel, 2001, “The Contribution ofDomestic and External Factors to Emerging Market Devaluation Crises: An Early Warning



502

Systems Approach,” International Finance Discussion Paper No. 711 (Washington: Boardof Governors of the Federal Reserve System).

Kaminsky, Graciela L., 1999, “Currency and Banking Crises: The Early Warnings of Distress,”Working Paper 99/178 (Washington: International Monetary Fund).

Kaminsky, Graciela, Saul Lizondo, and Carmen M. Reinhart, 1998, “Leading Indicators ofCurrency Crises,” IMF Staff Papers, Vol. 45, No. 1 (March), pp. 1–48.

Levy-Yeyati, Eduardo, and Federico Sturzenegger, 2001, “Exchange Rate Regimes andEconomic Performance,” IMF Staff Papers, Vol. 47 (Special Issue), pp. 62–98.

Manasse, Paolo, Nouriel Roubini, and Axel Schimmelpfennig, 2003, “Predicting SovereignDebt Crises,” IMF Working Paper 03/221 (Washington: International Monetary Fund).

Meese, Richard A., and Kenneth Rogoff, 1983, “Empirical Exchange Rate Models of theSeventies: Do They Fit out of Sample?” Journal of International Economics, Vol. 14,No. 1–2 (February), pp. 3–24.

Mulder, Christian, Roberto Perrelli, and Manuel Rocha, 2002, “The Role of Corporate, Legal,and Macroeconomic Balance Sheet Indicators in Crisis Detection and Prevention,” IMFWorking Paper 02/59 (Washington: International Monetary Fund).

Osband, Kent and Caroline Van Rijckeghem, 2000, “Safety from Currency Crashes,” IMF StaffPapers, Vol. 47, No. 2, pp. 238–58.

Reinhart, Carmen M., 2002, “Default, Currency Crises, and Sovereign Credit Ratings,” WorldBank Economic Review, Vol. 16, No. 2, pp. 151–70.

Reinhart, Carmen, and Kenneth Rogoff, 2004, “The Modern History of Exchange RateArrangements: A Reinterpretation,” Quarterly Journal of Economics, Vol. 119, No. 1,pp. 1–48.

Roy, Amlan, and Maria M. Tudela, 2000, “Emerging Market Risk Indicator (EMRI): Re-Estimated Sept 00” (New York: Credit Suisse First Boston).

Sy, Amadou, 2003, “Rating the Ratings Agencies: Anticipating Currency Crises or DebtCrises,” IMF Working Paper 03/122 (Washington: International Monetary Fund).

Vlaar, Peter J. G., 2000, “Currency Crisis Models for Emerging Markets,” DNB Staff ReportsNo. 45/2000 (Amsterdam: De Nederlandsche Bank).


503



Does SDDS Subscription Reduce Borrowing Costs for Emerging Market Economies?

JOHN CADY*

Does macroeconomic data transparency—as signaled by subscription to the IMF’sSpecial Data Dissemination Standard (SDDS)—help reduce borrowing costs ininternational capital markets? This question is examined using data on new issuesof sovereign foreign-currency-denominated (U.S. dollar, yen, and euro) bonds forseveral emerging market economies. Panel econometric estimates indicate thatspreads on new bond issues declined on average by close to 20 percent, or by an average of about 55 basis points for sample countries, following SDDS sub-scription. [JEL C22, F33, F34]

In 1996, the International Monetary Fund introduced the Special Data Dissem-ination Standard (SDDS). Development of this international macroeconomic

data standard was prompted by the widely held view that the emerging marketcrises of the mid-1990s were partially attributable to a lack of market informationand transparency, particularly with respect to macroeconomic and financial statis-tics. The SDDS is intended to guide countries that have, or seek to have, access tointernational capital markets in their provision of economic and financial data tothe public. An important aspect of access is the cost at which it is provided; thus,it is natural to inquire whether subscription to the SDDS has reduced borrowingcosts, particularly for emerging market participants.

Subscription to the SDDS is voluntary and involves no direct monetary costs,but it does require subscribers to observe the standard and provide information ondata and dissemination practices (the metadata) to the IMF for redissemination,

*John Cady is a Senior Economist in the IMF’s Statistics Department. He would like to thank Carol S.Carson, Robert Flood, A. Pellechio, J. R. Rosales, and colleagues in the IMF’s Statistics, InternationalCapital Markets, and Policy and Development Review Departments for helpful comments and suggestions.


John Cady

504

which could entail costs in upgrading a country’s statistical reporting and compi-lation systems. The standard identifies four dimensions of data dissemination, pre-scribing monitorable elements in the areas of access, integrity, quality, and the datathemselves. In particular, the data dimension lists 18 data categories providingcoverage for four (real, fiscal, financial, and external) sectors of the economy andprescribes minimum timeliness and frequency standards, summarized in Table 1.1

Several recent studies have examined the impact of the SDDS on emergingmarket economies’ (EMEs’) access and borrowing costs in international capitalmarkets. The Institute for International Finance (IIF, 2002) found that SDDS sub-scription led to a 200–300 basis point decline in U.S. dollar Eurobond spreads fora sample of emerging market economies. Subsequently, Christofides, Mulder, andTiffin (2003), in a study of the impact of international standards and codes onspreads and credit ratings, found that adherence to international standards, includ-ing SDDS subscription, contributed significantly to explaining changes in sov-ereign credit costs and ratings. Spreads, measured by JPMorgan’s Emerging MarketBond Index (EMBI), were found to be reduced by about 15 percent after SDDSsubscription. Glennerster and Shin (2003) provide econometric evidence that im-plementation of transparency measures reduces emerging market spreads; in thecase of the SDDS, EMBI spreads declined by 4–12 percent (equivalent to 20–60basis points in their sample) in the period following SDDS subscription.

All of these papers investigate the impact of the SDDS on secondary-marketyield spreads, with two utilizing the JPMorgan Emerging Market Bond Index thattracks the value of country-specific portfolios of U.S. dollar–denominated sov-ereign or quasi-sovereign debt instruments trading in secondary markets.2 The sec-ondary market for existing emerging market debt instruments is predominantly aglobal over-the-counter market composed of brokers, dealers, and investors world-wide linked daily through broker computer and telecommunications networks,offering the advantages of counterpart anonymity, efficiency, and transparency inprice determination.

In the primary market for sovereign debt, new issues are generally marketedby investment banks acting as “managers of the transaction,” first advising theissuer on the terms of the bond and the size of the issue, then typically organizing“road shows” to publicize the client country’s debt issue to potential investors.Subsequent public or private offerings are not conducted through a formal auctionprocess but rather by investment bankers taking orders from clients (“building thebook”). At this stage, it is possible, but not routine, for new price guidance to beprovided or the size of the potential issue to be altered in line with the degree ofclient interest. Once the order book has firmed, the managers of the transactionunderwrite the bonds, set a final price, and allocate bonds to clients. Clearly, theprimary and secondary markets for emerging market debt differ in terms of struc-ture, operation, and efficiency, with the secondary market more closely resembling

1Further information on the SDDS is available on the IMF’s Dissemination Standard Bulletin Board(DSBB): http://dsbb.imf.org/Applications/web/sddshome/.

2The spread for a particular country is defined as its EMBI portfolio yield over a theoretical U.S. zero-coupon curve, where the sovereign yield is set to equate the net present value of the sovereign cash flowsto zero.


http://dsbb.imf.org/Applications/web/sddshome/

DOES SDDS SUBSCRIPTION REDUCE BORROWING COSTS?

505

a tâtonnement process. Eichengreen and Mody (1998) provide evidence of a ten-dency for primary market spreads to follow secondary market spreads with athree- to four-quarter lag, but they note that launch spreads can, and do, move dif-ferently from secondary market spreads over the shorter run. Consequently, theempirical support for an SDDS discount found in secondary market studies doesnot necessarily provide evidence of a similar discount operating in primary mar-kets, justifying separate analysis.

This paper contributes to filling the gap in primary market evidence by directlyexamining the influence of SDDS subscription on the cost of issuance in primarysovereign bond markets. This is important since issuers in primary markets areunambiguously the beneficiaries of any cost reductions, or discounts, associatedwith SDDS participation—a significant consideration for many emerging marketgovernments, since international borrowing costs, ultimately borne by taxpayers,can play a pivotal role in public finances. The paper also contributes to the existingliterature in two other areas. The empirical work analyzes launch spreads for bondsdenominated in the three principal currencies used in private international bond

Table 1. SDDS Data Categories and Related Periodicity and Timeliness Standards

SDDS Data Category Periodicity Minimum Timeliness

Real SectorNational accounts Quarterly 1 quarterProduction indices Monthly 6 weeksEmployment, unemployment, wages/earnings Quarterly 1 quarterConsumer price index Monthly 1 month

Fiscal SectorGeneral government operations Annual 2 quartersCentral government operations Monthly 1 monthCentral government debt Quarterly 1 quarter

Financial SectorAnalytical accounts of the banking sector Monthly 1 monthAnalytical accounts of the central bank Monthly 2 weeksInterest rates Daily *Stock market Daily *

External SectorBalance of payments Quarterly 1 quarterInternational reserves Monthly 1 weekMerchandise trade Monthly 8 weeksInternational investment position Annual 2 quartersExternal debt Quarterly 1 quarterExchange rates Daily *

Addendum: Population Annual . . .

Source: IMF Statistics Department.Note: * indicates no timeliness standards set given that data are widely available from private

sources; dissemination by official data producers may be less time sensitive.


John Cady

506

markets—the U.S. dollar, yen, and euro—and finds evidence of an SDDS discountfor all three. Some empirical evidence is also found that mature market interestrates play a role in the determination of EME yield spreads.3

I. Data and Estimation Methodology

The influence of SDDS subscription on external borrowing costs is investigated for17 emerging market countries using information on foreign-currency-denominatedsovereign bond issuance and quarterly data on the principal macroeconomic deter-minants of interest rates and yield spreads. As subscription to the SDDS wasopened in April 1996, and a dozen countries subscribed in that year, the estimationstrategy seeks to consider a maximum of launch spreads for those subscribingcountries both before and after their subscription, and consider data for countriessubscribing to the initiative at a later stage to alleviate the bunching problem in1996. To properly gauge the impact of the SDDS it is also important to includedata on bonds issued by nonsubscribing countries. On the basis of these criteria, apanel of 17 countries was composed, including the following SDDS subscribers:Argentina, Brazil, Colombia, Croatia, Hungary, Korea, Lithuania, Malaysia,Mexico, the Philippines, Poland, South Africa, Tunisia, and Turkey, along with thenonsubscribing countries: China, Uruguay, and Venezuela (Table 2). Largely deter-mined by the availability of macroeconomic data, the time dimension of the panelis composed of a ragged sample ranging from 1990:3 to 2002:4 (country-specificsample periods are also reported in Table 2).4 The empirical analysis thus spans atime frame approximately six years prior to and following the opening of subscrip-tion to the SDDS in April 1996.

Data were drawn from three principal sources: sovereign bond characteristicsof new issues from the Bonds, Equities, and Loans (BEL) database of the IMF (assourced from Capital Data); macroeconomic data from the IMF’s InternationalFinancial Statistics (IFS) and World Economic Outlook (WEO); and external debtindicators from the World Bank’s Global Development Finance (GDF). Informationon Paris Club debt reschedulings and IMF arrangements and SDDS subscriptionswere drawn from the respective external websites of these organizations.

While the BEL database provides information on bonds issued in several currencies, spread data is available for only fixed interest rate bonds denominatedin U.S. dollars, yen, and euros.5 The panel includes observations on some 240

3Recently, Ferrucci (2003) and Arora and Cerisola (2001) report positive correlations between U.S.interest rates and EME spreads, consistent with their theoretical role. Earlier empirical investigations intothe determinants of EME yield spreads found either no role for mature market interest rates (Min, 1998)or unexpected negative correlations (Eichengreen and Mody, 1998; and Kamin and von Kleist, 1999).

4The countries included in the panel were chosen to include large emerging market countries sub-scribing to the SDDS that had launched a significant number of foreign-currency-denominated bonds dur-ing the period under consideration, and for which adequate quarterly macroeconomic data are available toconduct empirical analysis. Certain other large EMEs, including India and Singapore, did not issue anysovereign foreign-currency-denominated bonds between 1990 and 2002. Similarly, Korean sovereignissues were quite limited during this period, and data on bonds issued by the Korean Development Bankhave been used to extend the panel database.

5And prior to the introduction of the euro in 1999, both deutsche mark– and ECU-denominated bonds.



507

sovereign bond issues denominated in these currencies issued by the 17 countriesincluded in the panel.6 On average over the period 1999–2002, these 17 countriesaccounted for more than 65 percent of the value of all new emerging market debtissues.7 Bonds of various maturities, ranging from 1 to 30 years (the sample periodmean maturity is about 71⁄2 years), are represented in the panel data set.

Launch spreads reported in the BEL are defined as the annual yield to matu-rity at the time of issue on the emerging market instrument less a “risk-free”benchmark yield. The risk-free yield is approximated by the annualized yield onan industrial country government bond of the same currency and maturity as theemerging market instrument. More formally:

where SP is the yield spread, y represents the annualized yield on the emergingmarket debt instrument, and Y is the annualized yield on the industrial countrybenchmark bond of the same currency and maturity. In this paper all interest rates,yields, and launch spreads are measured in basis points.

SP y Y= − , ( )1

6Most foreign-currency-bond issues are denominated in these three currencies. Additionally, the BELdatabase does not calculate spreads for variable interest rate bond issues, regardless of the currency ofdenomination.

7International Monetary Fund (2004).

Table 2. SDDS Subscription Dates, Sample Periods,and Number of Bonds in Sample

Date of SDDS Number of Bonds Country Subscription Sample Period in the Sample

Argentina 16 August 1996 1994:2 to 2002:4 24Brazil 14 March 2001 1995:3 to 2002:4 16China* Nonsubscriber 1994:1 to 2000:4 10Colombia 31 May 1996 1995:2 to 2002:4 19Croatia 20 May 1996 1997:2 to 2001:4 8Hungary 24 May 1996 1996:1 to 2001:2 7Korea 20 September 1996 1990:3 to 2002:4 27Lithuania 30 May 1996 1996:1 to 2001:4 9Malaysia 21 August 1996 2001:2 to 2002:4 2Mexico 13 August 1996 1991:2 to 2002:4 24Philippines 5 August 1996 1993:3 to 2002:4 8Poland 17 April 1996 1996:2 to 2002:4 7South Africa 2 August 1996 1990:3 to 2002:4 13Tunisia 20 June 2001 1995:2 to 2002:4 6Turkey 8 August 1996 1990:3 to 2002:4 34Uruguay* 12 February 2004 1992:4 to 2001:4 12Venezuela* Nonsubscriber 1990:3 to 2001:4 15

Sources: IMF Statistics Department and BEL database (sourced from Capital Data).Note: * indicates that the country did not subscribe to the SDDS during the sample period.


John Cady

508

Spreads are related to a range of issue and issuer characteristics and funda-mental macroeconomic variables, in a standard log-linear specification:

where the dependent variable is the logarithm of the spread for country i in periodt, Xi,t is a vector of issue and issuer characteristics and macroeconomic fundamen-tals, and ui,t is a random error term. The vector Xi,t is composed of bond and issuercharacteristics and global economic conditions. The objective is to determine ifSDDS subscription plays a role in the determination of launch spreads after theinfluences of the economic fundamentals and bond and issuer characteristics havebeen taken into consideration. The choice of variables to represent economic fun-damentals has been guided by the existing long literature on the determinants ofspreads for emerging market economies (see Edwards, 1984; and Kamin and vonKleist, 1999).8 The influence of other country-specific factors, such as official debtrescheduling with Paris Club creditors and IMF financial support, are to beaccounted for with dummy variables. Other specific bond characteristics to be con-sidered include the maturity of the bond and its currency of denomination, also tobe accounted for with dummy variables indicating denomination in yen or euros.9

Following Eichengreen and Mody (1998) and Kamin and von Kleist (1999),the maturity of the bond is specified as an exogenous variable. However, cognizantthat launch yields and maturities might be simultaneously determined, Grangercausality tests were conducted on theses variables to test statistically if maturitycan be considered as exogenous (Table 3). For all of the countries considered,these tests permitted acceptance of the hypothesis of exogeneity, save for Mexicoand the Philippines, where the results were mixed and inconclusive. To fore-shadow the discussions of the estimation results somewhat, the panel econometricestimates proved robust to the inclusion or exclusion of the data for these twocountries, obviating the simultaneity question.

The factor of principal interest, the date of subscription to the SDDS, is repre-sented with a dummy variable, with values of zero prior to subscription and onethereafter for SDDS subscribers, and only zero in the case of nonsubscribers. Thesubscription date is considered as exogenous, an assumption particularly appropri-ate for the countries subscribing in 1996, which benefited from a transition period.10

Pooled time-series cross-section estimation was carried out using cross-sectionSeemingly Unrelated Regression, a procedure that corrects for equation-specific

log , ( )SP f X ui,t i,t i,t( ) = ( ) + 2

8Previous studies in this literature have investigated variables such as real GDP growth, the rate ofinflation, short- and long-term mature market interest rates, the fiscal and external current account bal-ances in relation to GDP, the external debt stock- and debt service-to-exports ratios, and short-term exter-nal debt in relation to international reserves.

9U.S. dollar–denominated bonds serve as the excluded category to preclude perfect multicollinearity.10The SDDS incorporated a formal transition period, beginning with the opening of subscription in

early April 1996 and ending December 31, 1998. During this period member countries could subscribe tothe SDDS even if their dissemination practices were not fully in line with the SDDS, permitting them tobring their data and dissemination practices into line with the standard according to a transition plan agreedwith the IMF.



509

Table 3. Pairwise Granger Causality Tests: Launch Spreads and Maturity

Null Rejected at 1 Percent

Country Hypothesis Observations F-Statistic Probability Level

Argentina A 39 2.536 0.061 NoB 0.618 0.653 No

Brazil A 27 0.670 0.621 NoB 1.294 0.310 No

China A 33 2.046 0.120 NoB 0.182 0.945 No

Colombia A 28 1.445 0.258 NoB 0.542 0.707 No

Croatia A 16 0.053 0.994 NoB 2.341 0.154 No

Hungary A 43 1.017 0.413 NoB 0.712 0.589 No

Korea A 43 1.619 0.192 NoB 1.733 0.165 No

Lithuania A 25 0.164 0.954 NoB 0.814 0.535 No

Malaysia* A 8 0.274 0.777 NoB 0.059 0.943 No

Mexico A 43 0.751 0.564 NoB 4.104 0.008 Yes

Philippines A 35 4.652 0.006 YesB 2.133 0.105 No

Poland A 27 0.422 0.791 NoB 1.527 0.237 No

South Africa A 43 0.817 0.524 NoB 0.530 0.715 No

Tunisia A 28 1.196 0.345 NoB 0.906 0.480 No

Turkey A 43 3.022 0.031 NoB 2.881 0.037 No

Uruguay A 39 0.726 0.581 NoB 1.015 0.415 No

Venezuela A 43 1.569 0.205 NoB 1.248 0.309 No

Source: Model estimates.Notes: Null hypotheses: A: Maturity (MAT) does not Granger cause spread (SP); B: Spread (SP)

does not Granger cause maturity (MAT); lags = 4; * based on 2 rather than 4 lags due to limiteddegrees of freedom.

serial correlation and cross-section heteroskedasticity.11 However, prior to estima-tion, the time-series properties of all the variables in the panel were investigatedusing panel unit root tests. For the most part, these panel unit root tests (Table 4)

11Tests with other panel estimation methods yielded broadly similar results.


John Cady

510

point to the absence of unit roots at conventional levels of significance.12 This isnot, however, the case with U.S. interest rates, and particularly the federal fundsrate, which is widely considered to be nonstationary time series. Similar conclu-sions about the nonstationarity of the U.S. interest rates are largely borne out by thestandard unit root tests performed over the sample period under consideration (bot-tom of Table 4).

To foreshadow the results again, regressions including and excluding short-and long-term U.S. interest rates had no appreciable impact on the estimated co-efficients, and particularly that attached to the SDDS dummy variable, permittingthe conclusion that the nonstationarity of the interest rate variables is benign in thiscontext and does not require the resort to panel cointegration techniques. Theinteresting results from estimations including U.S. interests rates are reported sincethey are consistent with others’ findings that U.S. interest rate increases are lessthan proportionately reflected in emerging market spreads.13

II. Estimation Results

In the estimating equation, spreads (SP) are modeled as a function of several fun-damental macroeconomic variables, including real GDP growth (YDOT), the rateof consumer price inflation relative to that in the United States (PDOT), and theexternal public debt stock relative to exports (DXR).14 The U.S. federal funds rate(USFED) and the yield on the 10-year U.S. treasury bond (USLONG), both mea-sured in basis points, have been included to proxy global monetary and liquidityconditions that could possibly influence emerging market yields spreads, inde-pendent of country-specific fundamentals.15 Other variables account for specificbond characteristics, including the term to maturity (MAT), measured in years, andthe currency of denomination (EURO and YEN). Dummy variables account for therespective official debt rescheduling (PARIS) and program status (IMF) history ofthe country with the Paris Club and the IMF, as well as the issuing country’s dateof SDDS subscription (SDDS) when applicable. A time trend (TIME) is includedin some of the estimated equations. The estimating equation is specified as:

log logSP YDOT PDOT MATi,t i,t i,t i( ) = + + +β β β β0 1 2 3 ,,t i,t

t

DXR

USFED USLO

( ) + ( )+ ( ) +

ββ β

4

5 6

log

log log NNG EURO YEN

IMF PARIt i,t i,t

i,t

( ) + ++ +

β ββ β

7 8

9 101 SS SDDS TIME ui,t i,t i,t+ + +β β11 12 3. ( )

12Annual data for external debt (public and publicly guaranteed) stock-to-exports ratios, drawn fromthe World Bank’s GDF database, were converted to a quarterly frequency (same value for all quarters) thensmoothed with the Hodrick-Prescott filter with standard quarterly parameters prior to testing the order ofintegration.

13See Eichengreen and Mody (1998, page 8).14Several other macro variables (including fiscal and current account balance measures and the exter-

nal short-term debt-to-international reserves ratio) proved insignificant or of the wrong sign and were sub-sequently omitted.

15The federal funds rate enters the regressions lagged one quarter, while USLONG is smoothed witha four-quarter moving average. Using U.S. interest rates to represent global liquidity conditions is consis-tent with the fact that the U.S. dollar–denominated bonds represent the omitted category for other dummyvariables.


DO

ES SDD

S SUBSC

RIP

TION

RED

UC

E BOR

RO

WIN

G C

OSTS?

51

1

Table 4. Panel and Single Variable Unit Root Tests

Accept Hypothesis Test Value/ Test Value/ of Unit Root at

Variable Test (Probability)* (Probability)** 5 Percent Level

SP (yield spread) Levin, Lin, and Chu T-statistic −3.324 (0.004) −5.889 (0.000) NoBreitung T-statistic −4.775 (0.000) −2.295 (0.011) NoIm, Pesarin, and Shin W-statistic −3.412 (0.000) −4.618 (0.000) NoADF-Fisher Chi-square 69.707 (0.003) 99.452 (0.000) NoPP-Fisher Chi-square 67.758 (0.000) 96.562 (0.000) NoHadri Z-statistic 12.353 (0.005) 4.085 (0.000) No

YDOT (real GDP growth) Levin, Lin, and Chu T-statistic −2.979 (0.001) −1.449 (0.074) MixedBreitung T-statistic −4.100 (0.000) −4.694 (0.000) NoIm, Pesarin, and Shin W-statistic −6.882 (0.000) −6.161 (0.000) NoADF-Fisher Chi-square 114.29 (0.000) 99.670 (0.000) NoPP-Fisher Chi-square 86.310 (0.000) 68.043 (0.000) NoHadri Z-statistic 4.490 (0.000) 2.086 (0.019) No

PDOT (inflation differential) Levin, Lin, and Chu T-statistic −14.629 (0.000) −23.835 (0.000) NoBreitung T-statistic −0.120 (0.452) 0.461 (0.678) YesIm, Pesarin, and Shin W-statistic −27.815 (0.000) −32.199 (0.000) NoADF-Fisher Chi-square 124.48 (0.000) 596.94 (0.000) NoPP-Fisher Chi-square 100.23 (0.000) 337.01 (0.000) NoHadri Z-statistic 3.478 (0.000) −0.247 (0.597) No


Joh

n C

ad

y

51

2


Accept Hypothesis Test Value/ Test Value/ of Unit Root at

Variable Test (Probability)* (Probability)** 5 Percent Level

MAT (maturity) Levin, Lin, and Chu T-statistic −4.505 (0.000) −6.383 (0.000) NoBreitung T-statistic −3.814 (0.000) −4.299 (0.000) NoIm, Pesarin, and Shin W-statistic −5.889 (0.000) −5.904 (0.000) NoADF-Fisher Chi-square 111.31 (0.000) 97.756 (0.000) NoPP-Fisher Chi-square 117.19 (0.000) 103.07 (0.000) NoHadri Z-statistic 4.590 (0.000) 4.553 (0.000) No

DXR (debt-exports ratio) Levin, Lin, and Chu T-statistic −1.616 (0.053) −0.797 (0.213) YesBreitung T-statistic −0.834 (0.202) −2.751 (0.003) MixedIm, Pesarin, and Shin W-statistic −1.729 (0.040) −5.366 (0.000) NoADF-Fisher Chi-square 82.363 (0.000) 144.02 (0.000) NoPP-Fisher Chi-square 41.741 (0.170) 48.857 (0.048) NoHadri Z-statistic 16.928 (0.000) 14.719 (0.000) No

USFED (U.S. federal funds rate) Augmented Dickey-Fuller −2.710 (0.079)+ −2.806 (0.202) MixedPhillips-Perron −1.913 (0.324) −1.973 (0.602) Yes

USLONG (U.S. 10-year treasury Augmented Dickey-Fuller −1.286 (0.627) −3.80 (0.0252)++ Mixedbond yield) Phillips-Perron −1.286 (0.629) −2.477 (0.338) Yes

USYC (slope U.S. yield curve) Augmented Dickey-Fuller −2.447 (0.135) −2.600 (0.282) YesPhillips-Perron −1.946 (0.309) −2.091 (0.538) Yes

Source: Model estimatesNotes: Panel unit root tests based on individual effects (*) and individual effects and linear trends (**); automatic lag length selected using the Schwarz infor-

mation criterion. For standard unit root tests, plus signs indicate rejection of the null hypothesis that the series has a unit root at the 10 percent level (+) and at the5 percent level (++). The critical values of the Augmented Dickey-Fuller test statistics are from MacKinnon (1996).



513

Coefficient estimates and key summary statistics for the basic specification arereported in column 1 of Table 5. There is a relatively high degree of fit, with allestimated coefficients of the expected sign and statistically significant at conven-tional confidence levels.

Of principal interest is the estimated coefficient attached to the SDDS dummyvariable. It is negative, strongly significant, and indicates that spreads decline byabout 0.186, or close to 20 percent, following SDDS subscription.16 An “SDDS dis-count” in this range is equivalent to a reduction of some 40 basis points when eval-uated at the sample mean of 214 basis points and a reduction of 55 basis pointswhen evaluated at 300 basis points, a level more representative of spreads prevail-ing since the establishment of the SDDS. The estimated impact of SDDS subscrip-tion is very much in line with those from the secondary market studies noted above.

Interesting results and interpretations can be given to the dummy variablesrepresenting financial and rescheduling arrangements with the IMF and ParisClub. The equation was found to fit best when these two variables were specifiedin change form, implying changes in launch spreads at both the outset and termi-nation of such special arrangements. In addition, although correctly signed, initialcoefficient estimates for the Paris Club dummy variable were not statistically sig-nificant. Only in the cases of rescheduling of larger Organization for EconomicCooperation and Development (OECD) member countries (Poland and Turkey)was this variable statistically significant. The coefficient attached to the OECDParis Club dummy variable indicates that spreads for rescheduling OECD coun-tries widen by about 30 basis points at the beginning of the consolidation period,presumably to compensate for increased default risk, then narrow at the end ofthe consolidation period. For countries with an IMF program,17 spreads declineby an estimated 10 basis points at the outside of a program, perhaps reflectingmarket expectations that Fund-supported programs help to restore macroeconomicstability, then increase by a similar magnitude following the expiration of thearrangement.

The estimated coefficients attached to the dummy variables for yen- and euro-denominated bonds are both negative, reflecting systematically lower yield spreadsthan for U.S. dollar-denominated instruments through the sample period. Theimportance of macroeconomic performance, as proxied by real GDP growth, isreflected by its negative coefficient estimate, indicating that favorable performancetends to narrow spreads. Higher inflation relative to that in the United States tendsto increase spreads. The length of maturity of the bond is estimated to exert astrong, positive influence on spreads, with longer maturities generally exhibitinghigher spreads. The U.S. federal funds rate exerts a positive impact on spreads,while long-term U.S. rates are negatively correlated with spreads. Consideredtogether, these estimates imply that a steeper U.S. yield curve is associated with

16A 95 percent confidence interval for the −0.186 point estimate ranges from −0.08 to −0.29, repre-senting an SDDS discount ranging between −24 and −87 basis points, when evaluated at the sample aver-age spread of about 300 basis points for the SDDS period 1996:2 to 2002:4.

17All countries in the panel except South Africa have had financial arrangements with the IMF duringsome part of the sample period.


John Cady

514

lower EME spreads.18 The coefficient attached to the external debt stock-to-exportsratio is positive, providing evidence that higher premiums are demanded as exter-nal indebtedness increases.

As previously suggested, the basic model has been reestimated in variousways to investigate several questions and potential econometric estimation issues.

18Ferrucci (2003) reports similar results and suggests that this might be attributable to carry trades,with leveraged investors borrowing in the industrial countries at short-term rates to purchase longer matu-rity EME bonds, pushing up their prices, depressing their yields and, ultimately, yield spreads.

Table 5. Panel Estimation Results

Estimation Period: 1990:3–2002:4* (1) (2) (3) (4) (5)

Constant 5.736 9.311 0.810 4.016 1.067(3.42) (6.16) (2.39) (2.46) (3.00)

Real GDP growth (YDOT) −0.532 −0.559 −0.580 −0.509 −0.600(−2.22) (−2.29) (−2.40) (−1.98) (−2.42)

Inflation differential (PDOT) 0.210 0.173 0.155 0.177 0.162(1.46) (1.04) (1.06) (1.17) (1.09)

Maturity (log MAT) 0.056 0.062 0.054 0.051 0.052(2.80) (3.08) (2.73) (1.87) (2.50)

Debt-to-exports ratio (log DXR) 0.777 0.815 0.772 0.775 0.760(13.66) (12.86) (13.71) (13.60) (13.12)

Federal funds rate (log USFED) 0.134 0.086 — 0.153 —(2.03) (1.22) (1.94)

U.S. 10-year bond yield (log USLONG) −0.868 −1.318 — −0.628 —(−3.42) (−5.40) (−1.92)

Slope U.S. yield curve (log USYC) — — — — −0.036(−2.88)

Euro-denominated issue (EURO) −0.301 −0.292 −0.283 −0.272 −0.306(−11.42) (−11.13) (−10.39) (−9.16) (−10.73)

Yen-denominated issue (YEN) −0.445 −0.426 −0.428 −0.393 −0.442(−13.43) (−12.86) (−12.96) (−8.89) (−12.65)

IMF arrangement (ΔIMF) −0.096 −0.096 −0.091 −0.111 −0.101(−3.26) (−3.23) (−3.04) (−2.57) (−3.22)

Paris Club rescheduling (ΔPARIS) 0.288 0.289 0.275 0.290 0.285(1.71) (1.68) (1.62) (1.70) (1.75)

SDDS subscription (SDDS) −0.186 −0.106 −0.172 −0.156 −0.192(−3.50) (−1.93) (−3.17) (−2.44) (−3.44)

Time trend (TIME) 0.018 — 0.024 0.019 (0.024)(4.03) (6.47) (3.83) (6.22)

Adjusted R2 0.995 0.994 0.993 0.991 0.995Durbin-Watson statistic 2.166 2.193 2.140 2.115 2.048Total pool observations 572 572 572 487 523Mean of the dependentVariable (in basis points) 213.5 213.5 213.5 203.6 210.4

Source: Model estimates.Notes: * = global estimation period; see Table 2 for country-specific sample periods; t-statistics

reported in parentheses.



515

First, as the basic equation includes a highly significant time trend variable, theequation was reestimated without the time trend to evaluate its impact. Reportedin column 2 of Table 5, estimates of this equation are broadly similar to the firstestimates. The estimated SDDS coefficient remains statistically significant butdoes move to −0.11. It would seem that the time trend in the first is picking upunspecified time-related developments in EME launch spreads that apparentlyimpart a downward bias to the SDDS coefficient in the second equation.

Second, the inclusion of maturity as an explanatory variable raised the ques-tions of its exogeneity and the possibility of simultaneous equation bias. TheGranger causality tests discussed earlier indicate that maturity can be considered asexogenous in all cases, except for Mexico and the Philippines. Therefore, the basicequation was reestimated excluding the data for Mexico and the Philippines to helpevaluate their impact on the estimates. The results of this regression are reported incolumn 4 of Table 5. All of the estimated coefficients are quite similar in size, sign,and statistical significance to those of the basic model, and particularly the matu-rity variable itself, indicating that simultaneity bias, if any, is minimal. The esti-mated SDDS coefficient is only marginally different than the full panel estimateand remains statistically significant.

Finally, the inclusion of nonstationary U.S. interest rates in the regressionsraises concerns over the appropriateness of the estimation techniques employed.This question was investigated in two ways. First, the basic model was reestimatedwith all U.S. interest rates excluded. Reported in Table 5 under column 3, thesecoefficient estimates are similar in sign, magnitude, and statistical significance tothe initial estimates, and in particular the estimated coefficient of the SDDS vari-able remains very close to the initial estimate. On this basis it is concluded that thenonstationarity of the interest rate variables is benign and that the application ofpanel cointegration techniques would be of marginal benefit.

A second approach was to replace the U.S. interest rates with a single variable,the slope of the yield curve, USYC19 (lagged three quarters). In the event, thisapproach does not prove useful since the slope of the yield curve also appears tobe a nonstationary series (Table 4). The estimation results are, nonetheless, inter-esting in their own right and illustrate the general parameter constancy of the esti-mated equation, and particularly the estimated impact of SDDS subscription onlaunch spreads. Reported in Table 5 under column 5, all of the estimated coeffi-cients are broadly similar in terms of sign, magnitude, and statistical significance,and most notably the SDDS coefficient, at −0.192.

Establishing whether the magnitude of the SDDS discount changes over timeor remains stable is also important. To investigate this aspect the basic equationwas reestimated recursively. Figure 1 presents a plot of recursive SDDS dummyvariable coefficient estimates, starting with the estimates from the shortest feasi-ble sample period (1992:2 to 2000:1), then successively adding two additionalobservations at the beginning and end of the estimation period until the full sam-

19Measured as the 10-year treasury bond rate (USLONG) minus the three-month treasury bill rate,both measured in basis points.


John Cady

516

ple period (1990:3 to 1992:4) is utilized. The plot indicates that the recursive esti-mates of the SDDS coefficient are indeed quite stable (as are all of the otherparameter estimates). The signs, magnitudes, and statistical significance of all theestimated coefficients, along with the degree of fit, of the recursively estimatedequations are broadly similar to those of equation (1), indicative of a relativelystable empirical relationship over time.

III. Conclusions

This paper presents econometric evidence of an SDDS discount in primary interna-tional capital markets for sovereign countries issuing foreign-currency-denominatedbonds, very much in line with the findings of secondary-market-based research.Specific bond characteristics—the currency denomination, global liquidity condi-tions, and a country’s fundamental macroeconomic and debt situation—remain theprimary determinants of sovereign borrowing costs, but subscription to the SDDSalso appears to lead to significant savings. Based on sovereign foreign currencybond issues in primary capital markets over the period 1990 to 2002, the SDDSspread discount for a group of 17 emerging market economies borrowing in thethree principal global currencies is estimated to be close to 20 percent.

The policy implications for sovereign borrowers are clear: macroeconomicand public debt fundamentals are of primary importance in the determination ofinternational borrowing costs, but subscription to the SDDS can provide importantcost savings to the sovereign borrower and ultimately its taxpayers.

Figure 1. Recursive Estimates of the SDDS Coefficient(Solid and dotted bands denote one and two standard error bands)

–0.35

–0.30

–0.25

–0.20

–0.15

–0.10

–0.05

0

36 38 40 42 44 46 48 50

Number of time series observations

Poi

nt a

nd in

terv

al e

stim

ates

Source: Model estimates. Note: SDDS = Special Data Dissemination Standard



517

REFERENCES

Arora, Vivek, and Martin Cerisola, 2001, “How Does U.S. Monetary Policy Influence SovereignSpreads in Emerging Markets?” IMF Staff Papers, Vol. 48 (November), pp. 474–98.

Christofides, Charis, Christian Mulder, and Andrew Tiffin, 2003, “The Link Between Adherenceto International Standards of Good Practice, Foreign Exchange Spreads, and Ratings,” IMFWorking Paper 03/74 (Washington: International Monetary Fund).

Edwards, Sebastian, 1984, “LDC Foreign Borrowing and Default Risk: An Empirical Investiga-tion,” American Economic Review, Vol. 74, No. 4, pp. 726–34.

Eichengreen, Barry, and Ashoka Mody, 1998, “What Explains Changing Spreads on Emerging-Market Debt: Fundamentals or Market Sentiment?” NBER Working Paper No. 6408(Cambridge, Massachusetts: National Bureau of Economic Research).

Ferrucci, Gianluigi, 2003, “Empirical Determinants of Emerging Market Economies’ SovereignBond Spreads,” Bank of England Working Paper No. 205 (London: Bank of England).

Glennerster, Rachel, and Yongseok Shin, 2003, “Is Transparency Good for You, and Can theIMF Help?” IMF Working Paper 03/132 (Washington: International Monetary Fund).

Institute for International Finance, 2002, IIF Action Plan Proposals and Dialogue with thePrivate Sector, Appendix D, “Does Subscription to the IMF’s Special Data DisseminationStandard Lower a Country’s Credit Spread?” (Washington).

International Monetary Fund, 2004, Global Financial Stability Report, World Economic andFinancial Surveys (Washington, September).

Kamin, Steven B., and Karsten von Kleist, 1999, “The Evolution and Determinants of EmergingMarket Credit Spreads in the 1990s,” BIS Working Paper No. 68 (Basel: Bank for Inter-national Settlements).

MacKinnon, James G., 1996, “Numerical Distribution Functions for Unit Root and Cointegra-tion Tests,” Journal of Applied Econometrics, Vol. 11, No. 6, pp. 601–18.

Min, Hong G., 1998, “Determinants of Emerging Market Bond Spread: Do Economic Funda-mentals Matter?” World Bank Policy Research Paper No. 1899 (Washington: World Bank).


518


Domestic Debt Markets in Sub-Saharan Africa

JAKOB CHRISTENSEN*

This study discusses the role of domestic debt markets in sub-Saharan Africa (SSA)based on a new data set covering 27 SSA countries during the 20-year period 1980–2000. The study finds that domestic debt markets in these countries are generallysmall, highly short term, and often have a narrow investor base. Domestic interestpayments present a significant burden to the budget, despite much smaller domesticthan foreign indebtedness. The use of domestic debt is also found to have signifi-cantly crowded out private sector lending. Finally, the study identifies significantdifferences among the size, cost, and maturity structure of domestic debt markets inheavily indebted poor countries (HIPCs) and non-HIPCs. [JEL E43, E44, H63,O23, O55]

In the past decades, the external debt burden and its impact on fiscal sustain-ability and economic growth in low-income countries have been extensively

debated. This debate has culminated in various debt reduction plans, such as the recent Heavily Indebted Poor Countries (HIPC) Initiative, which sought toreduce the external debt stocks in these countries and free up resources for pro-growth government spending. However, at least until recently, much less atten-tion has been given to the issue of domestic debt in low-income countries, despiteits potentially significant impact on government budgets, macroeconomic stabil-ity, private sector lending, and, ultimately, growth performance. Existing studieshave been limited mostly to individual country assessments in the context of thejoint World Bank and IMF Financial Sector Assessment Programs or theoretical

*Jakob Christensen is an Economist in the African Department of the IMF. The author is indebted toScott Rogers for his continuous advice and support. In addition, he would like to thank seminar partici-pants at the IMF’s African Department for discussion and comments and desk economists for their gener-ous data provision.


DOMESTIC DEBT MARKETS IN SUB-SAHARAN AFRICA

519

analyses of domestic debt. Moreover, data on domestic debt are scarce and lim-ited to a few African countries.

The main objective of this study is to discuss long-term developments andidentify key characteristics of African domestic debt markets based on a newlycollected database for 27 non-CFA1 sub-Saharan African (SSA) countries for theperiod 1980–2000. The discussion will focus on the following issues: (1) thedevelopment in domestic debt markets over the period relative to various indica-tors, such as GDP, foreign debt, and broad money; (2) the investor base of the debtmarkets, especially the degree to which the bank and nonbank sectors hold domes-tic debt; (3) the development in real treasury bill interest rates in view of the sig-nificant financial sector reforms that have taken place over the period; (4) thematurity structure of the domestic debt portfolio in these countries, including acomparison with debt markets in more developed countries; and (5) the impact ofdomestic borrowing on government budgets and private sector credit.

Of particular interest is the difference between domestic debt markets inHIPCs and non-HIPCs. First, the paper examines to what extent HIPCs have reliedon foreign versus domestic debt. Given that HIPCs have access to highly conces-sional foreign resources, it would be surprising if they had accumulated a signifi-cant amount of domestic debt, since domestic interest rates are higher than foreignones. Finally, the relatively underdeveloped financial systems in many HIPCs mayhave been an obstacle to developing sound and well-functioning domestic debtmarkets compared with those of non-HIPCs.

I. Description of Database

This study is the first attempt to compile a comprehensive database on domesticdebt for sub-Saharan African countries. While other databases exist, they are scarceand limited to only a few African countries. The most comprehensive database todate is the government financial statistics in the IMF’s International FinancialStatistics. However, it contains data for only 19 of 38 non-CFA countries, and thedata for many of these countries are incomplete. Another source is the WorldBank’s World Development Indicators, but the problem of missing data is evenlarger in this case.

The database for this study contains information about the characteristics ofdomestic debt markets for the period 1980–2000. The study focuses on gross secu-ritized domestic government debt, composed of treasury bills, development stocks,and bonds. Hence, the data set excludes domestic debt arising from domestic arrearsaccumulation and direct advances from the central bank and commercial banks.2While an attempt was made to collect information on outstanding stocks of centralbank debt (since it represents a quasi-fiscal cost to the government), it was not pos-sible to obtain information for all countries. In addition to the stock of domestic

1CFA denotes Franc de la Coopération Financière en Afrique.2Most countries do not have significant stocks of direct advances, because governments generally

clear these at the end of the year through issuance of treasury bills or transfers from other governmentaccounts. However, a few of the countries in the database have accumulated significant stocks of direct lia-bilities to the banking system.


Jakob Christensen

520

debt, the database also contains information on the maturity structure, holdings ofdebt by sectors, real treasury bill interest rates, and domestic interest burden on thebudget.

The choice of countries was limited to non-CFA countries, since CFA countriesuntil very recently did not have any domestic debt markets. Among the non-CFAcountries, Angola, Botswana, the Democratic Republic of the Congo, Mozambique,and São Tomé and Príncipe did not have domestic government debt markets at thetime of collection. Apart from these countries, it was possible to obtain completeseries for all countries except Guinea. Information on domestic debt data wasobtained primarily from individual IMF country reports, such as recent economicdevelopment reports and country desk databases. In cases where these were insuffi-cient, central bank reports or IMF country desk economists helped to fill the gaps.

II. Important Aspects of Domestic Debt Management

The Need for Domestic Debt Issuance

The need to issue domestic debt can arise both from government deficits that arenot fully foreign financed and from implementation of monetary policy. Generally,a deficit leads to a change in government net assets. Hence, a budget deficit can befinanced by either drawing down assets or incurring new liabilities, either domes-tic or foreign. The use of assets entails selling property or reducing deposits. Thistype of financing, however, is constrained by the stock and attractiveness of assets(the feasibility of privatization), and governments, therefore, normally resort todomestic or foreign borrowing to finance large parts of fiscal deficits. The choicebetween foreign and domestic borrowing, in turn, depends on cost (interest rates),maturity structure, and risks. Most of the SSA countries have access to foreignfinancing at very low interest rates (well below market interest rates) and at verylong maturity from international aid agencies or on grant terms. These terms areoften more favorable than for domestic borrowing, since domestic debt instrumentscarry much higher interest rates and have shorter maturities. Another advantage offoreign borrowing is that it increases the supply of foreign exchange, which is crit-ical to meet import requirements. One drawback to foreign borrowing is currencyrisk, which may increase along with foreign indebtedness, given that a growing for-eign debt service increases the demand for foreign exchange. However, Beaugrand,Loko, and Mlachila (2002) found that highly concessional foreign loans—whenavailable—are still the most attractive way to finance budget deficits, even if thereare significant devaluation risks, given the high levels of domestic interest rates.

Despite the attractiveness of foreign borrowing, governments may still con-sider domestic borrowing for a number of reasons. First, the supply of foreign(concessional) financing may be determined by the aid agencies’ budgets and theirassessment of the economic performance of the recipient country. Second, inter-national aid is very often linked to project financing and therefore cannot financea government’s recurrent expenditures or capital projects not supported by donors.Hence, governments with large recurrent budget deficits may be forced to tapdomestic savings, including through issuance of domestic debt, to close theirbudget gaps.



521

Domestic debt can also be used to achieve monetary policy targets. This is par-ticularly the case in countries with large balance of payment surpluses, created bylarge aid inflows or oil exports, for example. In those situations, the inflows of for-eign exchange increase liquidity, which can undermine macroeconomic stability,and the central banks often decide to intervene by selling government or centralbank bills to stem inflationary pressures from excess liquidity.

Macroeconomic Risks Related to Domestic Debt Financing

Extensive use of domestic borrowing can have severe repercussions on the econ-omy. Domestic debt service can consume a significant part of government rev-enues, especially given that domestic interest rates are higher than foreign ones.The interest cost of domestic borrowing can rise quickly along with increases inthe outstanding stock of debt, especially in shallow financial markets. In such mar-kets, given that financial resources are limited, expansions in domestic debt willmore easily lead to higher domestic interest rates. The increase in interest ratesmay be even more pronounced if the investor base is relatively narrow, since thegovernment may be held hostage by a particular group of investors (World Bankand IMF, 2001). A diverse investor base reduces the monopoly power of a partic-ular group of investors, bringing down not only costs but also rollover risks.Hence, an important aspect of debt management is broadening the investor base.This can be achieved through a combination of efforts, including promoting invest-ment by retail investors and developing and reforming pension and retirement fundsto encourage their investment in government bonds.

Another risk concerns the crowding out of private investment. When issuingdomestic debt, governments tap domestic private savings that would otherwise beavailable to the private sector. This is normally followed by an increase in domes-tic interest rates, if these are flexible, adversely affecting private investment.However, even when interest rates are controlled, domestic borrowing can lead tocredit rationing and crowding out of private sector investment (Fischer andEasterly, 1990). The impact of government borrowing will, to some extent, beaggravated if there are capital account restrictions, since banks cannot as easilycircumvent higher domestic interest rates through foreign borrowing. Last, but notleast, an investor base that is dominated by commercial banks may exacerbate theabove-mentioned effect. The crowding-out effect may, therefore, be more pro-nounced in the absence of nonbank investors, such as pension funds and retire-ment funds, to which the government could sell its debt without necessarilycrowding out private sector credit. Hence, a diverse investor base prevents exces-sive reliance on commercial bank funds and thereby reduces the risk of crowdingout (World Bank and IMF, 2001).

Maturity Structure

The government debt portfolio should adequately comprise short- and long-termpaper. If the debt portfolio consists mainly of short-term debt, the government mayface considerable risks. First, with more frequent rollovers, the government is highly


Jakob Christensen

522

vulnerable to a sudden increase in interest rates, which can raise debt service sig-nificantly. This can lead to further deterioration in the market’s confidence in gov-ernment bonds, prompting even higher interest rates on government debt. Second,administrative costs tend to be higher with a short maturity structure, because thegovernment must frequently roll over large parts of its debt, notably in countrieswithout an automated book-entry system. Third, the maturity structure is importantfor investors as they seek to diversify their asset portfolios. In many African coun-tries, government debt is the only investment opportunity besides lending to the pri-vate sector, since stock markets are either absent or highly illiquid (Gelbard andPereira Leite, 1999). The provision of government longer-term paper is thereforehighly important for investors to balance their long-term liabilities with long-termassets and for banks to increase profitability by taking on interest rate risk.

However, the government may experience several obstacles in pursuing alonger-term debt portfolio. First, the market may not be willing to hold long-termpaper in view of significant inflation and default risks. Second, it may not be suffi-ciently advanced to demand long-term paper, especially in the absence of institu-tional investors (Impavido, Musalem, and Tressel, 2003). Finally, the governmentmay hesitate to extend the maturity, since longer-term bonds can entail higher inter-est rates, in view of a rising yield curve, which would increase financing costs.

III. Characteristics of Domestic Debt Markets in Non-CFA SSA Countries

Developments in Domestic Debt: 1980–2000

Table 1 shows developments in domestic and external debt for 27 non-CFA SSAcountries for the period 1980–2000. It is apparent that domestic debt is not a recentphenomenon in African countries; most of the countries have relied on domesticborrowing since the beginning of the observation period. However, the averageratio of domestic debt increased from 11 percent of GDP in the 1980s to 15 percentin the late 1990s, with the median increasing from 4 percent to 10 percent over thesame period. An increasing number of countries became heavily domestically in-debted, and the number of countries with debt-to-GDP ratios exceeding 20 percentrose from three at the beginning of 1980 to nine by 2000.

There are wide differences across the countries with respect to the size ofgovernment securities markets. One group of countries has relied extensively ondomestic debt since the beginning of the period. This group includes Ethiopia,Kenya, Mauritius, Nigeria, South Africa, Tanzania, Zambia, and Zimbabwe. In con-trast, countries such as Angola, Botswana, the Democratic Republic of the Congo,Mozambique, and São Tomé and Príncipe have not used or have only recentlydeveloped government securities markets.3 Between these extremes, there is a vastgroup of countries that either have fairly small debt markets or have recently expe-rienced a considerable increase in their domestic debt burden, including TheGambia, Ghana, Namibia, and Seychelles.

3Botswana and Mozambique have fairly developed markets for central bank notes.


Table 1. Domestic and External Debt, 1980–2000(In percent of GDP, unless otherwise indicated)

1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000

Type of Domestic/total debt Country Domestic debt1 Domestic debt External debt Total debt (In percent)

Angola . . . 0 0 0 158 113 81 158 113 81 0 0 0Botswana . . . 0 0 0 5 4 10 5 4 10 0 0 0Burundi TB, TC 3 2 6 40 96 138 44 98 144 8 2 4Cape Verde TB 0 11 34 42 42 40 42 53 74 0 20 46Congo, Dem. Rep. of . . . 0 0 0 50 126 254 50 126 254 0 0 0Ethiopia TB, B 16 19 10 31 115 109 47 134 120 34 14 9Gambia, The TB, DN, S 3 13 23 80 84 104 83 96 127 3 13 18Ghana TB 12 8 24 19 55 83 32 64 106 38 13 22Guinea TB . . . . . . . . . 0 0 91 . . . . . . . . . . . . . . . . . .Kenya TB, B, S 21 23 22 61 77 52 81 100 74 25 23 29Lesotho TB, B 8 8 5 40 49 58 48 58 62 17 15 8Madagascar TB 3 3 3 71 120 110 74 123 113 4 2 2Malawi TB, S 13 8 9 65 100 126 78 109 135 16 7 7Mauritius TB, S 27 29 33 39 21 15 66 50 48 41 57 69Mozambique . . . 0 0 0 75 207 121 75 207 122 0 0 0Namibia TB 0 8 19 0 4 2 0 12 21 . . . 69 89Nigeria TB, B, TC, S 28 29 16 49 93 80 77 122 97 37 24 17



1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000

Type of Domestic/total debt Country Domestic debt1 Domestic debt External debt Total debt (In percent)

Rwanda TB, B 8 9 5 17 55 70 25 65 75 31 14 7São Tomé and Príncipe . . . 0 0 0 155 422 643 155 422 643 0 0 0Seychelles TB, B, S 14 45 68 29 24 20 43 69 88 33 65 77Sierra Leone TB, B, S 13 5 7 34 94 143 47 99 150 28 5 5South Africa TB, B 30 37 45 0 0 0 30 37 45 100 100 100Swaziland TB, B, S 4 1 1 20 21 16 24 22 17 16 5 7Tanzania TB, S 26 6 12 71 131 100 96 137 112 27 5 11Uganda TB, S 2 1 2 0 73 57 2 74 59 100 1 4Zambia TB, B 25 9 6 134 178 196 159 186 202 16 5 3Zimbabwe TB, B, S 35 29 37 27 34 48 62 63 86 56 45 44Average 11 12 15 49 87 103 62 102 118 25 19 22HIPC 9 6 8 56 124 156 69 138 169 22 6 6

Decision point reached2 10 7 8 58 126 150 73 143 164 25 7 7Eligible3 2 1 3 45 111 196 47 112 199 4 1 2

Non-HIPC4 14 18 23 39 40 35 53 59 59 30 35 40

Sources: IMF staff reports; and selected central bank statistics.1TB=Treasury bills; TC=Treasury certificates; B=Bonds; S=Government stocks; DN=Discount note series; HIPC=Heavily Indebted Poor Countries.2Includes Ethiopia, The Gambia, Ghana, Guinea, Madagascar, Malawi, Mozambique, Rwanda, São Tomé and Príncipe, Sierra Leone, Tanzania, and Uganda.3Includes Burundi and the Democratic Republic of the Congo.4Includes Angola, Botswana, Cape Verde, Kenya, Lesotho, Mauritius, Namibia, Nigeria, Seychelles, South Africa, Swaziland, and Zimbabwe.



525

Table 1 also shows that the domestic debt burden in HIPCs is much smallerthan in non-HIPCs. HIPCs, which, almost by definition, have relied heavily on for-eign financing, have not developed their domestic debt markets to the same degreeas non-HIPCs. On average, domestic debt in HIPCs amounted to about 8 percentof GDP, although this ratio increased slightly in the latter half of the 1990s, mostlybecause of large increases in outstanding domestic debt in Ghana and The Gambia.However, other HIPCs have managed to obtain significant reductions in the ratio ofdomestic debt to GDP over the same period, notably Ethiopia and Zambia. In con-trast, domestic debt markets have grown steadily in non-HIPCs, as the average ratioof domestic debt to GDP increased from 14 percent in the 1980s to 23 percent bythe end of the 1990s.

Finally, Table 1 shows that while domestic debt stocks have grown in recentyears relative to GDP, their size is still negligible compared with the size of foreignindebtedness. Domestic debt accounted for just over one-fifth of total debt in thelatter half of the 1990s, slightly lower than in the 1980s. However, there are markeddifferences between HIPCs and non-HIPCs. While domestic debt financing hasgrown relative to foreign borrowing in non-HIPCs, domestic borrowing in HIPCshas been dominated by a huge accumulation of external debt in the 1990s. As aresult, the ratio of domestic debt to total debt between the two groups diverged sig-nificantly in the 1990s, from relatively similar levels in the 1980s as the proportionof domestic debt fell to less than 10 percent of total debt in HIPCs while increas-ing to almost 40 percent in non-HIPCs by the end of the 1990s.

As mentioned above, the potential for expanding domestic debt depends on thedepth of the financial sector. A useful indicator in that regard is the ratio of broadmoney to GDP. Table 2 shows that African financial sectors generally appear to berelatively small and, on average, they tend to be much smaller in HIPCs than innon-HIPCs. The “deepest” financial sectors were found in Cape Verde, Kenya,Mauritius, Seychelles, and South Africa, where broad money amounted to morethan 50 percent of GDP in the late 1990s.

The small financial sectors in most countries limit the potential for expandingdomestic debt. The ratio of domestic debt to broad money is shown in Table 2. Anumber of countries had very large ratios of domestic debt to broad money at theend of the 1990s, including The Gambia, Ghana, Nigeria, Seychelles, South Africa,and Zimbabwe. The ratio for Ghana was even larger than 100 percent. Interestingly,the average ratio is almost the same in HIPCs as in non-HIPCs, even though the for-mer group has much less domestic debt. In other words, the potential for expandingdomestic debt in HIPCs appears to be more limited, particularly in The Gambia,Ghana, Malawi, Sierra Leone, and Tanzania, compared with non-HIPCs, becausefurther expansions of domestic debt in HIPCs would decrease the availability ofcommercial bank resources and, thereby, curb credit to the private sector.

The Investor Base

As mentioned above, a diverse investor base is crucial to lowering the cost of gov-ernment debt and the volatility of market yields. Furthermore, a narrow investorbase, consisting mainly of commercial banks, increases the risk of crowding out


Jakob Christensen

526

private investment, especially in SSA countries where private companies have torely on bank financing, given the absence of corporate debt markets. As such, animportant component of debt management is stimulating a diverse investor baseand developing instruments, trading facilities, and distribution networks that bestsuit the needs of investors (World Bank and IMF, 2001). In most developed mar-ket economies, there are traditionally four general categories of potential investors

Table 2. Financial Sector Depth and Domestic Debt, 1980–2000

M2 (In percent of GDP) Domestic Debt (In percent of M2)

Country 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000

Angola 107 72 18 0 0 0Botswana 19 20 21 0 0 0Burundi 18 18 19 19 11 30Cape Verde 47 64 64 0 17 53Congo, Dem. Rep. of 8 14 7 0 0 0Ethiopia 28 41 41 57 47 25Gambia, The 21 22 29 13 57 80Ghana 15 16 22 83 47 106Guinea 10 9 10 . . . . . . . . .Kenya 29 38 50 71 63 44Lesotho 49 34 31 18 25 16Madagascar 21 22 21 15 13 12Malawi 22 22 16 59 38 57Mauritius 47 67 77 57 44 43Mozambique 37 22 21 0 0 1Namibia 12 30 42 0 25 44Nigeria 27 21 17 106 137 95Rwanda 13 16 17 62 59 30São Tomé and Príncipe 59 31 32 0 0 0Seychelles 32 42 78 43 107 86Sierra Leone 19 12 14 71 38 50South Africa 56 53 56 53 71 81Swaziland 33 32 26 12 3 4Tanzania 27 17 16 93 38 74Uganda 9 8 13 24 7 16Zambia 17 20 19 145 44 30Zimbabwe 27 22 42 129 130 91Average 31 30 32 39 39 42HIPC 23 20 21 38 27 37Decision point reached1 24 21 22 43 31 41Eligible2 13 16 13 9 6 15

Non-HIPC3 41 41 43 41 52 46

Sources: IMF staff reports; and selected central bank statistics.Notes: HIPC=Heavily Indebted Poor Countries; M2=M2 money supply.1Includes Ethiopia, The Gambia, Ghana, Guinea, Madagascar, Malawi, Mozambique, Rwanda,

São Tomé and Príncipe, Sierra Leone, Tanzania, and Uganda.2Includes Burundi and the Democratic Republic of the Congo.3Includes Angola, Botswana, Cape Verde, Kenya, Lesotho, Mauritius, Namibia, Nigeria,

Seychelles, South Africa, Swaziland, and Zimbabwe.



527

in government securities instruments: domestic and foreign and, for each of thesecategories, the banking sector (composed of commercial banks as well as centralbanks) and the nonbank sector, consisting of the contractual savings sector (pen-sion funds), collective investment funds, and nonfinancial entities such as non-financial corporations and individual investors. The presence of foreign investorsin African securities markets is generally limited. To date, only a few countrieshave had active participation of foreign investors in their debt markets, whichmay be a result of underdeveloped trading facilities, high country risk, and capitalaccount restrictions.

Commercial banks are the main holders of government debt in the Africandebt market, holding half of all outstanding domestic debt (Table 3). While theyenjoy a relatively high income from government debt, their large holdings ofdomestic debt may reflect some fundamental shortcomings in their commercialbanking operations (World Bank and IMF, 2001). These shortcomings includeinstitutional weaknesses that undermine lending to the private sector, given inef-fective screening and monitoring capabilities of loans, little reliable informationon creditworthy borrowers, and weak legal systems (such as a lack of commercialcourts to settle payment disputes).4

The nonbank sector was found to be the second biggest holder, accounting forone-third of outstanding debt. The limited role played by the nonbank sector com-pared with commercial banks may be attributed to the absence of large-scale insti-tutional investors in the nonbank sector. However, the nonbank sector has playedan important role in Kenya, Madagascar, Mauritius, Rwanda, and South Africa.Insurance companies and pension funds were in these cases the most commoninvestors, but building societies, post office savings banks, public enterprises, andthe general public also played a role.

Finally, central banks accounted for a modest share of government debt, withthe exception of Burundi, Nigeria, and Tanzania. While such holdings can be uti-lized for monetary policy purposes, central bank purchases of government debt arebasically identical to monetizing budget deficits.

Real Treasury Bill Rates and Financial Sector Reforms

Financial systems in most African countries were highly controlled in the 1980s.However, many countries embarked on a series of financial sector reforms in the late1980s aimed at liberalizing their financial sectors to improve financial inter-mediation. In many cases, these reforms included a move toward more liberal gov-ernment debt markets based on flexible and market-determined interest rates,subject to the level of inflation, the amount of outstanding debt, and the risk ofdefault. This replaced a system in which the government often had forced the state-controlled financial system to hold government debt despite minimal returns.

4A good measure of these shortcomings is the number of nonperforming loans (NPLs). Mehran andothers (1998) found that the ratio of NPLs to total loans averaged 16 percent in 16 non-CFA countries. Theratio was significantly higher for HIPCs than for non-HIPCs, with almost one-fourth of total loansrecorded as NPLs in the former group.


Jakob Christensen

528

Whereas real interest rates on bonds were often negative in the prereform period,in the early 1990s they needed to increase to more positive realms to make bondsattractive.

The positive impact of these reforms on financial development and liberal-ization is evident from the index numbers in the first two columns of Table 4.These numbers are drawn from Gelbard and Pereira Leite (1999) and comprisesix subcategories, each based on the following additional subindicators: (1) mar-ket structure, (2) financial products, (3) financial liberalization, (4) institutionalenvironment, (5) financial openness, and (6) monetary policy instruments. Ahigher index number corresponds to a more developed financial system. It can be

Table 3. Holdings of Government Debt Across Sectors(In percent)

Banking Sector

Country Total Central bank Commercial banks Nonbank Sector

Burundi 77 55 22 23Cape Verde 78 30 48 22Ethiopia 81 24 57 19Gambia, The 52 0 52 48Ghana 66 27 39 34Kenya 50 11 39 50Lesotho 81 1 80 19Madagascar . . . . . . . . . . . .Malawi 100 0 100 0Mauritius 45 5 40 55Nigeria 96 66 30 4Rwanda 21 0 21 79Seychelles 86 0 86 14Sierra Leone 63 4 60 37South Africa . . . . . . . . . . . .Swaziland 66 0 66 34Tanzania 86 44 42 14Uganda 90 17 73 10Zambia 78 0 77 22Zimbabwe 53 19 35 47Average 70 17 54 30HIPC 71 17 54 29Decision point reached1 71 13 58 29Eligible2 77 55 22 23

Non-HIPC3 69 16 53 31

Note: HIPC=Heavily Indebted Poor Countries.1Includes Ethiopia, The Gambia, Ghana, Guinea, Madagascar, Malawi, Mozambique, Rwanda,





529

Table 4. Financial Development, Real Treasury Bill Rates,and Private Sector Lending, 1980–2000

Financial Development Real Treasury Bill Rates Credit to Private Sector

(Index)1 (In percent) (In percent of broad money)

Country 1987 1997 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000

Angola 9 21 . . . . . . . . . . . . . . . 14Botswana 47 62 . . . . . . . . . 40 54 51Burundi . . . . . . . . . . . . . . . 42 63 72Cape Verde 34 54 . . . . . . 6 . . . . . . 44Congo, Dem. 20 52 . . . . . . . . . 23 7 . . .

Rep. ofEthiopia 9 23 (2) (5) 2 11 9 46Gambia, The 43 60 (4) 9 11 77 44 36Ghana 31 75 (32) 5 5 19 26 38Guinea 30 50 . . . 14 8 . . . . . . . . .Kenya 44 75 5 8 15 68 57 60Lesotho 20 44 (1) (0) 5 26 48 53Madagascar 38 63 . . . (11) (1) 103 79 48Malawi 24 47 (6) (2) (4) 61 51 29Mauritius 62 85 (2) 0 3 53 56 68Mozambique 24 53 . . . . . . 9 . . . 43 58Namibia 42 72 . . . 3 8 . . . 90 96Nigeria 27 61 (12) (19) (10) 52 47 65Rwanda . . . . . . . . . 49 43 53São Tomé and 22 30 . . . . . . . . . . . . . . . 25

PríncipeSeychelles 8 11 6 36 18 19Sierra Leone (49) (15) (2) 22 25 19South Africa 77 87 (2) 1 7 93 110 121Swaziland 43 60 (4) 0 4 65 75 64Tanzania 30 65 . . . 17 2 14 54 22Uganda 36 64 (101) 2 4 29 40 38Zambia 47 75 (27) (71) 9 40 36 41Zimbabwe 38 65 (5) (4) 2 43 84 89Average 34 58 (16) (3) 4 46 50 51HIPC 28 53 (32) (6) 4 41 40 40Decision point 29 53 (32) (6) 4 42 41 38

reached2

Eligible3 20 52 . . . . . . . . . 32 35 72Non-HIPC4 40 62 (1) (0) 4 53 64 62

Sources: IMF, International Financial Statistics; and Gelbard and Pereira Leite (1999).Notes: Numbers in parentheses are negative; HIPC=Heavily Indebted Poor Countries.1Numbers based on Gelbard and Leite (1999).2Includes Ethiopia, The Gambia, Ghana, Guinea, Madagascar, Malawi, Mozambique, Rwanda,




Jakob Christensen

530

seen that the financial systems improved in virtually all countries, with the largestimprovements observed in HIPCs (although they came from the lowest base).

The liberalization of the financial system appears to have been accompanied bya sharp rise in real interest rates (Table 4). At the end of the 1990s, all countries hadpositive real treasury bill rates except for Madagascar, Malawi, Nigeria, and SierraLeone. In comparison, 13 of 15 countries for which data were available had nega-tive real treasury bill rates during the 1980s. In contrast, real treasury bill rates in TheGambia and Kenya exceeded 10 percent at the end of the 1990s. As noted above,financial sectors in HIPCs developed more rapidly than in non-HIPCs, and this mayalso explain why they witnessed the largest increase in real treasury bill rates, froman average of −32 percent in the 1980s to 4 percent by the end of the 1990s. In con-trast, securities markets in non-HIPCs were already relatively liberal in the 1980s,and only small increases were needed to achieve positive real treasury bill rates.

Maturity Structure

The maturity structure of government debt can affect both the costs and risks ofusing domestic debt instruments. In general, the government should attempt toissue debt whose maturity mirrors the maturity structure of short-term current andlong-term capital expenditures. However, governments may be tempted to issuemainly short-term debt if the yield curve slopes sufficiently upward. Furthermore,while there are obvious benefits to extending the maturity structure, including areduction in market and rollover risks, the market may not be ready to absorb long-term paper, especially if there is considerable macroeconomic instability. In addi-tion, the absence of a contractual savings sector and mutual funds with sufficientlylong investment horizons may also limit the ability of the government to extendthe maturity structure. To some extent, the length of the maturity structure can beviewed as a measure of the degree of market development.

Short-term paper dominates debt markets in Africa (Figure 1). Three-monthbills are the most frequently used, accounting for almost 50 percent of outstand-ing debt stocks (implying that African governments, on average, must roll overhalf of their debt portfolio four times a year). The second most common maturityis 12 months, accounting for about one-fifth of the bonds, while one-tenth of allbonds have a 6-month maturity.

The average maturity for the African countries is 231 days, or about 10 months(Table 5). Domestic debt markets in HIPCs appear to have the shortest maturitystructure: 177 days. Burundi has the shortest average maturity, with 77 days,closely followed by Uganda, with 93 days. In contrast, non-HIPCs benefiting frommore sophisticated markets have longer maturities: South Africa and Swaziland topthe list with an average maturity length of 1,748 and 1,145 days, respectively.

As mentioned above, the dominance of short-term paper in African securitiesmarkets greatly increases rollover and market risk, especially in countries withlarge outstanding debt stocks. Financial liberalization has led to more interest rateflexibility and made countries with a large amount of short-term debt vulnerableto changes in market conditions. Some governments must roll over debt amount-ing to one-fourth of GDP three or four times a year (on average).


DO

MESTIC

DEBT M

AR

KETS IN

SUB-SA

HA

RA

N A

FRIC

A

53

1

Figure 1. Average Maturity Structure of Domestic Debt in 14 Non-CFA SSA Countries

0

5

10

15

20

25

30

35

40

45

50

1 month 3 months 4 months 6 months 9 months 12 months 18 months 2 years 3 years 3.5 years 4 years 5 years 6 years 7 years 10 years

Source: Author’s calculations.Note: SSA = Sub-Saharan Africa


Jakob Christensen

532

The short-term nature of African debt markets is even more evident comparedwith debt markets in more developed countries. In six developed and emergingmarket countries for which data were available, the average maturity was about fiveand half years, or seven times longer than in African countries. Roughly speaking,maturity length seems to be more closely related to general economic development(in terms of per capita income) than to the size of debt markets (relative to GDP).As a country gains wealth (with the exception of India) and the demand for moresophisticated economic arrangements expands, the need for longer-term savingsinstruments increases. This implies that as African countries continue to develop,their debt markets may be expected to become more advanced and long term. Thiswould help reduce the significant risks of portfolios that are dominated by short-term debt. A word of caution may be necessary in that some fairly poor countries

Table 5. Average Maturity of Domestic Debt for Selected African Countries and Emerging Market Countries

Domestic Debt/GDP GDP Per Capita Maturity Country (In percent)1 (In U.S. dollars)2 (In days)

Burundi 9 141 77Uganda 2 348 93Gambia, The 31 371 112Ghana 29 413 122Malawi 11 169 177Sierra Leone 10 147 190Lesotho 11 551 203Nigeria 21 254 228Cape Verde 26 1,519 256Zambia 5 392 296Rwanda 6 242 351Kenya 22 328 382Namibia 19 2,408 859Swaziland 1 1,476 1,145South Africa 41 3,985 1,748Average 15 850 231HIPC 13 293 177Non-HIPC 17 1,089 512

Memorandum items:Mexico 23 3,819 720Brazil . . . 4,624 1,085Italy 105 20,885 1,376Lithuania . . . 2,056 1,715India . . . 459 3,050New Zealand 35 17,548 3,720

Average 54 8,232 1,945

Sources: Selected country staff reports; World Bank and IMF (2001).Note: HIPC=Heavily Indebted Poor Countries.1Data for 2000.2GDP per capita (at constant 1995 U.S. dollars) in 2000.



533

have expanded their debt markets significantly in relatively undeveloped financialmarket settings—thereby ending up with sizable amounts of short-term debt, caus-ing a significant burden on, and risk to, the budget.

IV. The Impact of Domestic Debt on the Budget and Private Sector Credit

Budget Implications

A key concern regarding domestic debt management is the cost, in terms of amor-tization and interest payments, to the budget. This section will focus on interest,since most African governments have been net borrowers in domestic markets,rolling over existing debt. Two issues deserve attention: (1) the interest burden onthe budget and (2) the relative cost of domestic versus foreign borrowing.

Domestic interest payments are sizable compared with revenues and GDP(Table 6). Average interest payments, as a percentage of revenues, increased in bothHIPCs and non-HIPCs over the period 1980–2000. However, there are large varia-tions among countries—Ethiopia, Lesotho, Madagascar, Nigeria, and Rwanda havecut their interest payments significantly as their debt stocks have fallen. In contrast,The Gambia, Ghana, Malawi, Sierra Leone, and Zimbabwe have witnessed a sharpincrease in domestic interest payments, to more than 15 percent of their revenues.Relative to GDP, domestic interest payments, on average, account for more than2 percent of GDP in these countries. The ratio is slightly higher in non-HIPCs thanin HIPCs.

Surprisingly, domestic interest payments are as large as foreign ones, despitemuch lower levels of domestic than foreign debt. In fact, domestic interest pay-ments exceeded foreign interest payments in 10 (half of which were HIPCs) of the22 countries for which data were available. Despite the drastic decline in domestic-to-total-debt ratio in HIPCs, domestic interest payments as a percentage of totalinterest payments remained relatively constant at about 40 percent throughout theperiod. Furthermore, domestic interest payments in HIPCs account for almost thesame share of total interest payments as in non-HIPCs, despite much smallerdomestic debt relative to foreign debt in HIPCs. Hence, in addition to the large for-eign interest burden, recently highlighted by the HIPC Initiative, African govern-ments have to pay a significant part of their revenues to service domestic debt.

The significant domestic interest burden is a result of relatively high domesticinterest rates. Various comparisons of the cost of domestic versus foreign borrow-ing suggest that domestic interest rates are much higher than foreign ones (Table 7).To measure the cost of borrowing, the average implicit interest rates for bothdomestic and foreign borrowing were calculated by dividing the interest paymentsin the budget by the actual debt stock.5 At the end of the 1990s, the implicit domes-

5A more common approach is to look at the uncovered interest rate parity. However, using foreignmarket interest rates for African countries may overstate the cost, since most of their borrowing is onhighly concessional terms with interest rates well below market rates. The implicit interest rate calculatesthe average interest rate, which takes into account (ex post) exchange rate depreciation.


Jakob Christensen

534

Table 6. Domestic Interest Payments, 1980–2000

In Percent of In Percent of Total Debt Service Revenues1 In Percent of GDP

1980– 1990–94 1995– 1980– 1990–94 1995– 1980– 1990–94 1995–Country 1989 2000 1989 2000 1989 2000

Angola . . . . . . . . . . . . . . . . . . . . . . . . . . .Botswana . . . . . . . . . . . . . . . . . . . . . . . . . . .Burundi 54.8 28.0 34.5 2.9 2.9 5.1 0.5 0.5 0.9Cape Verde 20.4 38.8 69.2 1.8 2.4 11.0 0.3 0.5 2.2Congo, Dem. . . . . . . . . . . . . . . . . . . . . . . . . . . .

Rep. ofEthiopia 74.6 68.8 64.9 4.1 11.6 8.1 0.8 1.5 1.5Gambia, The 33.4 53.9 72.8 6.7 9.1 18.1 1.4 2.0 3.3Ghana . . . 66.6 73.6 . . . 11.9 24.4 . . . 1.8 4.3Guinea . . . 4.0 14.4 . . . 0.6 2.0 . . . 0.1 0.2Kenya 64.1 71.7 74.5 13.4 22.2 15.9 3.1 5.7 4.2Lesotho . . . 70.2 40.6 5.0 6.3 1.7 1.8 2.6 0.7Madagascar 21.0 32.5 20.7 2.5 9.4 5.9 0.3 0.9 0.6Malawi 92.1 54.8 64.4 24.1 10.5 17.4 4.6 2.0 2.9Mauritius 62.8 80.6 85.8 14.7 13.2 14.9 3.3 3.0 2.9Mozambique . . . . . . . . . . . . . . . . . . . . . . . . . . .Namibia . . . . . . . . . . . . . . . . . . . . . . . . . . .Nigeria 54.7 38.1 37.4 20.6 25.4 9.7 2.2 3.3 1.6Rwanda 63.9 71.4 41.0 4.1 25.1 7.4 0.6 1.6 0.6São Tomé and . . . . . . 2.5 . . . . . . 1.3 . . . . . . 0.2

PríncipeSeychelles . . . . . . 85.7 . . . . . . 16.8 . . . . . . 7.5Sierra Leone 42.9 31.2 48.7 17.4 11.7 24.9 1.3 1.3 2.0South Africa 95.5 97.9 96.1 12.4 19.1 22.1 2.9 4.4 5.3Swaziland 32.5 16.8 25.1 1.8 0.5 0.5 0.5 0.1 0.1Tanzania 78.3 51.9 54.5 11.9 8.8 10.6 1.6 1.1 1.4Uganda . . . 14.4 29.4 . . . 3.2 2.6 . . . 0.2 0.3Zambia 33.4 29.7 31.5 11.0 11.3 10.2 2.4 3.0 2.0Zimbabwe 72.1 72.3 73.7 9.9 12.9 21.7 2.7 3.4 6.0Average 56.0 49.7 51.9 9.7 10.9 11.5 1.8 2.0 2.3HIPC 54.9 42.3 42.5 9.4 9.7 10.6 1.5 1.3 1.6Decision point 54.9 43.6 43.2 10.2 10.3 11.1 1.6 1.4 1.6

reached2

Eligible3 54.8 28.0 34.5 2.9 2.9 5.1 0.5 0.5 0.9Non-HIPC4 57.5 60.8 65.3 9.9 12.7 12.7 2.1 2.9 3.4

Sources: IMF, staff reports and International Financial Statistics.Note: HIPC=Heavily Indebted Poor Countries.1Excluding grants.2Includes Ethiopia, The Gambia, Ghana, Guinea, Madagascar, Malawi, Mozambique, Rwanda,





535

Table 7. The Cost of Domestic and Foreign Borrowing

Nominal Treasury Bill Rate1 Implicit Domestic2 Implicit Foreign2

Country 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000 1980–89 1990–94 1995–2000

Angola . . . . . . . . . . . . . . . . . . . . . . . . . . .Botswana . . . . . . . . . . . . . . . . . . . . . . . . . . .Burundi . . . . . . 12 16 25 16 2 1 1Cape Verde . . . . . . 8 . . . 2 6 2 2 2Congo, Dem. . . . . . . . . . . . . . . . . . . . . . . . . . . .

Rep. ofEthiopia 3 7 6 5 9 14 1 1 1Gambia, The 13 17 14 32 16 15 3 2 1Ghana 17 28 37 . . . 31 19 . . . 2 2Guinea . . . 18 12 . . . . . . . . . . . . . . . 1Kenya 12 33 22 13 24 20 3 3 3Lesotho 13 13 13 11 29 18 2 2 2Madagascar . . . 14 16 9 31 26 2 2 2Malawi 12 18 35 36 24 38 2 2 1Mauritius 10 9 10 13 10 9 5 3 3Mozambique . . . . . . 17 . . . . . . . . . . . . . . . . . .Namibia . . . 15 15 . . . . . . . . . . . . . . . . . .Nigeria 9 17 13 7 12 10 5 6 4Rwanda 8 . . . . . . 8 17 12 1 1 1São Tomé and . . . . . . . . . . . . . . . . . . . . . 2 2

PríncipeSeychelles 12 13 9 . . . . . . 8 . . . . . . 6Sierra Leone 14 50 20 17 31 27 3 3 1South Africa 13 14 14 10 12 12 . . . . . . . . .Swaziland 11 11 12 13 13 13 2 3 3Tanzania . . . 47 17 8 17 12 1 1 1Uganda 21 27 10 . . . 35 16 . . . 2 1Zimbabwe 8 22 37 8 12 18 4 4 3Average 12 22 18 14 21 17 3 2 2HIPC 12 28 20 16 25 21 2 2 1Decision point 12 28 20 16 25 21 2 2 1

reached3

Eligible4 . . . . . . 8 . . . 2 6 2 2 2Non-HIPC5 11 16 15 11 14 12 3 3 3

Sources: IMF, International Financial Statistics and country reports; and author’s calculations.Note: HIPC=Heavily Indebted Poor Countries.1Nominal treasury bill rates.2The implicit interest rate was calculated by dividing the interest payments in the budget by the actual debt stock

and multiplying by 100.3Includes Ethiopia, The Gambia, Ghana, Guinea, Madagascar, Malawi, Mozambique, Rwanda, São Tomé and

Príncipe, Sierra Leone, Tanzania, and Uganda.4Includes Burundi and the Democratic Republic of the Congo.5Includes Angola, Botswana, Cape Verde, Kenya, Lesotho, Mauritius, Namibia, Nigeria, Seychelles, South

Africa, Swaziland, and Zimbabwe.


Jakob Christensen

536

tic interest rate was found to average about 17 percent, compared with 2 percent forforeign borrowing. At the same time, implicit domestic borrowing costs were foundto be higher in HIPCs than in non-HIPCs. Implicit domestic interest rates are sim-ilar to nominal treasury bill rates.

What makes a government borrow domestically when the interest rates aremuch higher? First, amortization on external borrowings requires foreign exchange.Hence, external vulnerability may increase dramatically if external indebtednessrises significantly. In contrast, the authorities can, at least in the short run, roll overdomestic debt without major macroeconomic implications. Second, to limit exter-nal vulnerability, many Fund-supported programs in poor countries include a capon nonconcessional borrowing. Thus, if the governments in these countries cannotobtain sufficient concessional foreign assistance to meet their financing require-ments, they must resort to relatively expensive domestic borrowing rather than fill-ing the financing gap by more favorable nonconcessional foreign borrowing.

Impact on Private Sector Credit

As mentioned above, domestic debt can crowd out private sector credit with ad-verse consequences for private investment. To examine this effect, a simple paneldata model was estimated, regressing private sector lending on domestic debt (bothvariables were in percentage of broad money) for the 27 countries over the period1980–2000. The results from this regression (shown in Table 8) yielded significantsupport for the crowding-out hypothesis: on average across countries, an expansionin domestic debt of 1 percent relative to broad money causes the ratio of private sec-tor lending to broad money to decline by 0.15 percent.

The Gambia showed one of the strongest decreases in the ratio of private sec-tor lending to broad money, dropping from about three-fourths to about one-thirdduring the 20-year period from 1980 to 2000. This coincided with a strong expan-sion in domestic borrowing as the ratio of domestic debt to broad money rose to106 percent by the end of the 1990s from an average of 13 percent in the 1980s.Another interesting case is Malawi, which witnessed a sharp reduction in privatesector lending in the latter half of the 1990s. Despite a relatively small ratio ofdomestic debt to GDP, domestic debt assumed a relatively large proportion ofbroad money, given the relatively underdeveloped financial sector (see Table 4).One exception is South Africa, where the ratio of credit to the private sectorincreased despite expansion in domestic debt. This can be attributed to the small

Table 8. Regression Results

Domestic Debt Constant

Coef. Std. error Coef. Std. error Observations R2

Private Sector Credit −0.15 (0.03) 52.7 (1.54) 492 0.0007

Note: Both variables in percent of M2.



537

commercial bank holdings of government debt, which helped reduce the negativeimpact of debt expansion on private sector lending.

V. Conclusion, Policy Implications, and Need for Reforms

This study has examined different features of domestic debt markets in non-CFAsub-Saharan African countries. Overall, the use of domestic debt instruments isnot a recent phenomenon; 19 out of 27 countries had domestic debt markets in1980, a number that had increased to 21 by the year 2000. The ratio of domesticdebt increased from 11 percent in the 1980s to 15 percent of GDP by the end ofthe 1990s. However, the domestic debt burden is still small compared with foreignindebtedness.

Even though the ratio of domestic debt to GDP is modest, domestic borrowingstill assumes a large part of financial resources, given the thin and shallow financialmarkets in the countries. The ratio of domestic debt to broad money was constantat about 40 percent throughout the period, although some countries had ratios ofalmost 100 percent. Because commercial banks hold more than half of the out-standing domestic debt, expansion in domestic debt has had a significant negativeimpact on private sector lending. The nonbank sector plays a limited role, given arelatively underdeveloped institutional investment sector in many of the countries.In addition, domestic markets were mainly short term, with the most commonmaturity being three months; the average maturity for 15 SSA countries for whichdata were available was only 231 days, far shorter than in selected emerging mar-ket countries.

Domestic debt financing was found to be much more expensive than foreignborrowing. This may be explained by the ongoing financial liberalization, whichhas resulted in sharply rising real treasury bill rates, but also by the fact that mostcountries borrow externally on highly concessional terms. Consequently, domesticinterest payments present the same burden to the budget as the foreign debt does,even though the domestic debt burden comprises only a fraction of the total debtburden. While domestic interest payments, on average, assumed about one-tenthof total revenue, some countries, such as The Gambia, Ghana, Kenya, Malawi,Seychelles, Sierra Leone, South Africa, and Zimbabwe, have to set aside more than15 percent of their revenues to pay interest on domestic debt.

The study also identified marked differences in the size, cost, and maturity ofdomestic debt markets between HIPCs and non-HIPCs. Given the significantreliance on external financing, HIPCs have accumulated less domestic debt,although some face a significant domestic debt burden in addition to their largestock of foreign debt. Despite the lower domestic-debt-to-GDP ratio, HIPCs havean almost identical ratio of domestic debt to broad money, given the smallerdegree of financial intermediation than in non-HIPCs. Thus, further expansions indomestic debt are more likely to crowd out private investments in HIPCs. HIPCsembarked on comprehensive financial liberalization in the first part of the 1990s,and real interest rates surged considerably as a result. This, combined with a highdegree of concessional foreign borrowing, explains why domestic interest pay-ments almost equal foreign ones in these countries, despite smaller amounts of


Jakob Christensen

538

domestic debt. Further, governments in HIPCs face a much higher market risk asa result of the shorter maturity structure of domestic debt than do governments innon-HIPCs.

The significant debt problems in many countries, both domestic and foreign,raise considerable concern about fiscal sustainability. In the worst case, these debtproblems may call for reforms. One option would be to pursue debt reductionschemes for domestic debt similar to the HIPC Initiative. However, an outrightreduction in domestic debt would increase liquidity in the system and therebyendanger macroeconomic stability. Instead, one might consider a debt reductionscheme similar to that enacted in Cape Verde, whereby a donor-financed trust fundwas established. The foreign exchange from this fund was used to retire domesticdebt without injecting liquidity into the system, because the foreign exchangetransaction essentially absorbed the liquidity.

Another consideration is that countries could benefit from extending the matu-rity structure of domestic debt, since Africa’s debt markets tend to be of anextremely short duration. While this might entail greater debt-service costs to gov-ernments, since bonds with longer terms may carry higher interest rates, it wouldlower the significant market and rollover risks that they currently face. In light ofthese countries’ nascent capital markets, such reforms should be accompanied bybroader reforms that promote long-term paper and strengthen and expand theinsurance and pension sectors as well as corporate governance and institutions.

Finally, domestic debt markets would greatly benefit from improved foreignaccess to holdings of domestic debt. In addition to strengthening competition,which would reduce financing costs, a strong foreign investor presence would con-tribute to the introduction of financial technology and innovation, thereby leadingto higher market efficiency.

REFERENCES

Beaugrand, Philippe, Bolieau Loko, and Montfort Mlachila, 2002, “The Choice BetweenExternal and Domestic Debt in Financing Budget Deficits: The Case of Central and WestAfrican Countries,” IMF Working Paper 02/79 (Washington: International Monetary Fund).

Fischer, Stanley, and William Easterly, 1990, “The Economics of the Government BudgetConstraint,” The World Bank Research Observer, Vol. 5, No. 2, pp. 127–42.

Gelbard, Enrique, and Sérgio Pereira Leite, 1999, “Measuring Financial Development in Sub-Saharan Africa,” IMF Working Paper 99/105 (Washington: International Monetary Fund).

Impavido, Gregorio, Alberto R. Musalem, and Thierry Tressel, 2003, “The Impact of Contrac-tual Savings Institutions on Securities Markets,” World Bank Policy Research WorkingPaper No. 2948 (Washington: The World Bank).

Mehran, Hassanali, and others, 1998, Financial Sector Development in Sub-Saharan AfricanCountries, IMF Occasional Paper No. 169 (Washington: International Monetary Fund).

World Bank and International Monetary Fund, 2001, Developing Government Bond Markets, aHandbook (Washington: World Bank).


539

IMF Staff PapersVol. 52, No. 3


Volume 52 Index

Volume 52 (2005) comprises three regular issues and one special issue, as follows:

No. 1, 1–148

No. 2, 149–366

No. 3, 369–551

Special Issue, 1–180

Authors

Aiyar, Shekhar, and Carl-Johan Dalgaard. Total Factor Productivity Revisited: A DualApproach to Development Accounting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .82

Aliber, Robert Z. The 35 Most Tumultuous Years in Monetary History: Shocks, the TransferProblem, and Financial Trauma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-142

Arora, Vivek, and Athanasios Vamvakidis. How Much Do Trading Partners Matter forEconomic Growth? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24

Baqir, Reza, Rodney Ramcharan, and Ratna Sahay. IMF Programs and Growth: Is OptimismDefensible? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .260

Barron, Greg, and Felix Várdy. The Internal Job Market of the IMF’s Economist Program 410

Berg, Andrew, Eduardo Borensztein, and Catherine Pattillo. Assessing Early Warning Systems:How Have They Worked in Practice? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .462

Berger, Wolfram, and Helmut Wagner. Interdependent Expectations and the Spread of CurrencyCrises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .41

Borensztein, Eduardo, Andrew Berg, and Catherine Pattillo. Assessing Early Warning Systems:How Have They Worked in Practice? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .462


Volume 52 Index

540

Bown, Chad P., and Rachel McCulloch. U.S. Trade Policy and the Adjustment Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-107

Boyer, Russell S., and Warren Young. Mundell’s International Economics: Adaptations andDebates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-160

Brandner, Peter, and Harald Grech. Why Did Central Banks Intervene in ERM I? The Post-1993 Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .120

Cady, John. Does SDDS Subscription Reduce Borrowing Costs for Emerging MarketEconomies? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .503

Cargill, Thomas F. Is the Bank of Japan’s Financial Structure an Obstacle to Policy? . . . .311

Cerra, Valerie, and Sweta Chaman Saxena. Did Output Recover from the Asian Crisis? . . . .1

Chami, Ralph, Connel Fullenkamp, and Samir Jahjah. Are Immigrant Remittance Flows aSource of Capital for Development? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .55

Choudhri, Ehsan U., and Mohsin S. Khan. Real Exchange Rates in Developing Countries:Are Balassa-Samuelson Effects Present? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .387

Christensen, Jakob. Domestic Debt Markets in Sub-Saharan Africa . . . . . . . . . . . . . . . . . .518

Cordella, Tito, and Eduardo Levy Yeyati. Country Insurance . . . . . . . . . . . . . . . . . . . . .SI-85

Dalgaard, Carl-Johan, and Shekhar Aiyar. Total Factor Productivity Revisited: A DualApproach to Development Accounting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .82

Engerman, Stanley. On the Accuracy of Some Past and Present Forecasts . . . . . . . . . . .SI-15

Fogel, Robert W. Reconsidering Expectations of Economic Growth After World War II fromthe Perspective of 2004 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-6

Frankel, Jeffrey. Mundell-Fleming Lecture: Contractionary Currency Crashes in DevelopingCountries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .149

Fullenkamp, Connel, Ralph Chami, and Samir Jahjah. Are Immigrant Remittance Flows aSource of Capital for Development? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .55

Genberg, Hans, and Alexander K. Swoboda. Exchange Rate Regimes: Does What CountriesSay Matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-129

Grech, Harald, and Peter Brandner. Why Did Central Banks Intervene in ERM I? The Post-1993 Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .120

Ize, Alain. Capitalizing Central Banks: A Net Worth Approach . . . . . . . . . . . . . . . . . . . . .289

Jahjah, Samir, Ralph Chami, and Connel Fullenkamp. Are Immigrant Remittance Flows aSource of Capital for Development? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .55

Jeanne, Olivier, and Jeromin Zettelmeyer. The Mussa Theorem (and Other Results on IMF-Induced Moral Hazard) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-64

Khan, Mohsin S., and Ehsan U. Choudhri. Real Exchange Rates in Developing Countries: AreBalassa-Samuelson Effects Present? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .387


Kose, M. Ayhan, Eswar S. Prasad, and Marco E. Terrones. Growth and Volatility in an Era ofGlobalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-31

Levchenko, Andrei A. Financial Liberalization and Consumption Volatility in DevelopingCountries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .237

Levy Yeyati, Eduardo, and Tito Cordella. Country Insurance . . . . . . . . . . . . . . . . . . . . .SI-85

Martinez Peria, Maria Soledad, Andrew Powell, and Ivanna Vladkova-Hollar. Banking on Foreigners: The Behavior of International Bank Claims on Latin America,1985–2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .430

McCulloch, Rachel, and Chad P. Bown. U.S. Trade Policy and the Adjustment Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-107

Pattillo, Catherine, Andrew Berg, and Eduardo Borensztein. Assessing Early Warning Systems:How Have They Worked in Practice? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .462

Powell, Andrew, Maria Soledad Martinez Peria, and Ivanna Vladkova-Hollar. Banking on Foreigners: The Behavior of International Bank Claims on Latin America,1985–2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .430

Prasad, Eswar S., M. Ayhan Kose, and Marco E. Terrones. Growth and Volatility in an Era ofGlobalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-31

Ramcharan, Rodney, Reza Baqir, and Ratna Sahay. IMF Programs and Growth: Is OptimismDefensible? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .260

Rodrik, Dani, and Arvind Subramanian. From “Hindu Growth” to Productivity Surge: TheMystery of the Indian Growth Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .193

Romeu, Rafael B. Why Are Asset Markets Modeled Successfully, But Not Their Dealers? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .369

Sahay, Ratna, Reza Baqir, and Rodney Ramcharan. IMF Programs and Growth: Is OptimismDefensible? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .260

Saxena, Sweta Chaman, and Valerie Cerra. Did Output Recover from the Asian Crisis? . . . .1

Shiells, Clinton R. VAT Design and Energy Trade: The Case of Russia and Ukraine . . . . .103

Stella, Peter. Central Bank Financial Strength, Transparency, and Policy Credibility . . . . .335

Subramanian, Arvind, and Dani Rodrik. From “Hindu Growth” to Productivity Surge:The Mystery of the Indian Growth Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .193

Swoboda, Alexander K., and Hans Genberg. Exchange Rate Regimes: Does What CountriesSay Matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-129

Terrones, Marco E., M. Ayhan Kose, and Eswar S. Prasad. Growth and Volatility in an Era ofGlobalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-31

Vamvakidis, Athanasios, and Vivek Arora. How Much Do Trading Partners Matter forEconomic Growth? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24

Várdy, Felix, and Greg Barron. The Internal Job Market of the IMF’s Economist Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .410

VOLUME 52 INDEX

541


Volume 52 Index

542

Vladkova-Hollar, Ivanna, Maria Soledad Martinez Peria, and Andrew Powell. Banking onForeigners: The Behavior of International Bank Claims on Latin America,1985–2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .430

Wagner, Helmut, and Wolfram Berger. Interdependent Expectations and the Spread of CurrencyCrises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .41

Young, Warren, and Russell S. Boyer. Mundell’s International Economics: Adaptations andDebates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-160

Zettelmeyer, Jeromin, and Olivier Jeanne. The Mussa Theorem (and Other Results on IMF-Induced Moral Hazard) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-64

Titles

On the Accuracy of Some Past and Present Forecasts. Stanley Engerman . . . . . . . . . . . .SI-15

Are Immigrant Remittance Flows a Source of Capital for Development? Ralph Chami, ConnelFullenkamp, and Samir Jahjah . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .55

Assessing Early Warning Systems: How Have They Worked in Practice? Andrew Berg,Eduardo Borensztein, and Catherine Pattillo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .462

Banking on Foreigners: The Behavior of International Bank Claims on Latin America,1985–2000. Maria Soledad Martinez Peria, Andrew Powell, and Ivanna Vladkova-Hollar. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .430

Capitalizing Central Banks: A Net Worth Approach. Alain Ize . . . . . . . . . . . . . . . . . . . . .289

Central Bank Financial Strength, Transparency, and Policy Credibility. Peter Stella . . . . .335

Country Insurance. Tito Cordella and Eduardo Levy Yeyati . . . . . . . . . . . . . . . . . . . . . .SI-85

Did Output Recover from the Asian Crisis? Valerie Cerra and Sweta Chaman Saxena. . . . . .1

Does SDDS Subscription Reduce Borrowing Costs for Emerging Market Economies? JohnCady . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .503

Domestic Debt Markets in Sub-Saharan Africa. Jakob Christensen . . . . . . . . . . . . . . . . . .518

Exchange Rate Regimes: Does What Countries Say Matter? Hans Genberg and Alexander K.Swoboda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-129

Financial Liberalization and Consumption Volatility in Developing Countries. Andrei A.Levchenko . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .237

From “Hindu Growth” to Productivity Surge: The Mystery of the Indian Growth Transition.Dani Rodrik and Arvind Subramanian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .193

Growth and Volatility in an Era of Globalization. M. Ayhan Kose, Eswar S. Prasad, and MarcoE. Terrones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-31

How Much Do Trading Partners Matter for Economic Growth? Vivek Arora and AthanasiosVamvakidis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24


VOLUME 52 INDEX

543

IMF Programs and Growth: Is Optimism Defensible? Reza Baqir, Rodney Ramcharan, andRatna Sahay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .260

Interdependent Expectations and the Spread of Currency Crises. Wolfram Berger and HelmutWagner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .41

The Internal Job Market of the IMF’s Economist Program. Greg Barron and Felix Várdy. .410

Is the Bank of Japan’s Financial Structure an Obstacle to Policy? Thomas F. Cargill . . . . .311

Mundell-Fleming Lecture: Contractionary Currency Crashes in Developing Countries. JeffreyFrankel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .149

Mundell’s International Economics: Adaptations and Debates. Russell S. Boyer and WarrenYoung . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-160

The Mussa Theorem (and Other Results on IMF-Induced Moral Hazard). Olivier Jeanne andJeromin Zettelmeyer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-64

Real Exchange Rates in Developing Countries: Are Balassa-Samuelson Effects Present? EhsanU. Choudhri and Mohsin S. Khan. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .387

Reconsidering Expectations of Economic Growth After World War II from the Perspective of2004. Robert W. Fogel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-6

The 35 Most Tumultuous Years in Monetary History: Shocks, the Transfer Problem, andFinancial Trauma. Robert Z. Aliber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-142

Total Factor Productivity Revisited: A Dual Approach to Development Accounting. ShekharAiyar and Carl-Johan Dalgaard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .82

U.S. Trade Policy and the Adjustment Process. Chad P. Bown and Rachel McCulloch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-107

VAT Design and Energy Trade: The Case of Russia and Ukraine. Clinton R. Shiells . . . . .103

Why Are Asset Markets Modeled Successfully, But Not Their Dealers? Rafael B. Romeu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .369

Why Did Central Banks Intervene in ERM I? The Post-1993 Experience. Peter Brandner andHarald Grech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .120

Subjects

To facilitate electronic storage and retrieval of bibliographic data, IMF Staff Papers has adopted thesubject classification scheme of the Journal of Economic Literature (Nashville, Tennessee).

B METHODOLOGY AND HISTORY OF ECONOMIC THOUGHT

B1 History of Economic Thought through 1925

B10 General

On the Accuracy of Some Past and Present Forecasts. Stanley Engerman . . . . . . .SI-15


Volume 52 Index

544

B2 History of Economic Thought since 1925

B20 General

On the Accuracy of Some Past and Present Forecasts. Stanley Engerman . . . . . . .SI-15

B29 Other

Mundell’s International Economics: Adaptations and Debates. Russell S. Boyer andWarren Young . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-160

B3 History of Thought: Individuals

B31 Individuals


C MATHEMATICAL AND QUANTITATIVE METHODS

C2 Econometric Methods: Single Equation Models

C22 Time-Series Models

Does SDDS Subscription Reduce Borrowing Costs for Emerging Market Economies?John Cady . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .503

C3 Econometric Methods: Multiple/Simultaneous Equation Models

C32 Time-Series Models

Did Output Recover from the Asian Crisis? Valerie Cerra and Sweta Chaman Saxena . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1

C7 Game Theory and Bargaining Theory

C78 Bargaining Theory; Matching Theory

The Internal Job Market of the IMF’s Economist Program. Greg Barron and Felix Várdy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .410

D MICROECONOMICS

D6 Economic Welfare

D64 Altruism

Are Immigrant Remittance Flows a Source of Capital for Development? Ralph Chami,Connel Fullenkamp, and Samir Jahjah . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .55

D7 Analysis of Collective Decision-Making

D73 Bureaucracy; Administrative Processes in Public Organizations

The Internal Job Market of the IMF’s Economist Program. Greg Barron and Felix Várdy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .410

D8 Information and Uncertainty


D82 Asymmetric and Private Information


E MACROECONOMICS AND MONETARY ECONOMICS

E1 General Aggregative Models

E10 General


E3 Prices, Business Fluctuations, and Cycles

E31 Price Level; Inflation; Deflation

Is the Bank of Japan’s Financial Structure an Obstacle to Policy? Thomas F. Cargill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .311

E32 Business Fluctuations; Cycles

Growth and Volatility in an Era of Globalization. M. Ayhan Kose, Eswar S. Prasad, andMarco E. Terrones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-31

E4 Money and Interest Rates

E42 Monetary Standards and Regimes; Government and the Monetary System

Central Bank Financial Strength, Transparency, and Policy Credibility. Peter Stella . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .335

Exchange Rate Regimes: Does What Countries Say Matter? Hans Genberg andAlexander K. Swoboda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-129

E43 Determination of Interest Rates; Term Structure of Interest Rates

Domestic Debt Markets in Sub-Saharan Africa. Jakob Christensen . . . . . . . . . . . . .518

E44 Financial Markets and the Macroeconomy


E5 Monetary Policy, Central Banking, and the Supply of Money and Credit

E58 Central Banks and Their Policies

Capitalizing Central Banks: A Net Worth Approach. Alain Ize . . . . . . . . . . . . . . . . .289


Interdependent Expectations and the Spread of Currency Crises. Wolfram Berger andHelmut Wagner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .41


VOLUME 52 INDEX

545


Why Did Central Banks Intervene in ERM I? The Post-1993 Experience. Peter Brandnerand Harald Grech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .120

E6 Macroeconomic Aspects of Public Finance, Macroeconomic Policy, and GeneralOutlook

E61 Policy Objectives; Policy Designs and Consistency; Policy Coordination


E63 Comparative or Joint Analysis of Fiscal and Monetary or Stabilization Policy



F INTERNATIONAL ECONOMICS

F02 International Economic Order; Economic Integration: General

Financial Liberalization and Consumption Volatility in Developing Countries. Andrei A.Levchenko . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .237

F12 Models of Trade with Imperfect Competition and Scale Economies

VAT Design and Energy Trade: The Case of Russia and Ukraine. Clinton R. Shiells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .103

F13 Commercial Policy; Protection; Promotion; Trade Negotiations

U.S. Trade Policy and the Adjustment Process. Chad P. Bown and Rachel McCulloch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-107

F14 Country and Industry Studies of Trade


F15 Economic Integrations

How Much Do Trading Partners Matter for Economic Growth? Vivek Arora andAthanasios Vamvakidis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24


F2 International Factor Movements and International Business

F22 International Migration


Volume 52 Index

546


F3 International Finance

Why Are Asset Markets Modeled Successfully, But Not Their Dealers? Rafael B. Romeu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .369

F30 General

Country Insurance. Tito Cordella and Eduardo Levy Yeyati . . . . . . . . . . . . . . . . .SI-85

F31 Foreign Exchange

Assessing Early Warning Systems: How Have They Worked in Practice? Andrew Berg,Eduardo Borensztein, and Catherine Pattillo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .462

Real Exchange Rates in Developing Countries: Are Balassa-Samuelson Effects Present?Ehsan U. Choudhri and Mohsin S. Khan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .387


F32 Current Account Adjustment; Short-term Capital Movements

Mundell-Fleming Lecture: Contractionary Currency Crashes in Developing Countries.Jeffrey Frankel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .149

The Mussa Theorem (and Other Results on IMF-Induced Moral Hazard). Olivier Jeanneand Jeromin Zettelmeyer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-64

F33 International Monetary Arrangements and Institutions


Exchange Rate Regimes: Does What Countries Say Matter? Hans Genberg andAlexander K. Swoboda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-129

IMF Programs and Growth: Is Optimism Defensible? Reza Baqir, Rodney Ramcharan,and Ratna Sahay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .260



The Mussa Theorem (and Other Results on IMF-Induced Moral Hazard). Olivier Jeanneand Jeromin Zettelmeyer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-64


F34 International Lending and Debt Problems



VOLUME 52 INDEX

547


F35 Foreign Aid

IMF Programs and Growth: Is Optimism Defensible? Reza Baqir, Rodney Ramcharan,and Ratna Sahay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .260

F36 Financial Aspects of Economic Integration

Financial Liberalization and Consumption Volatility in Developing Countries. Andrei A.Levchenko . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .237


F39 Other


F4 Macroeconomic Aspects of International Trade and Finance


F41 Open Economy Macroeconomics




Real Exchange Rates in Developing Countries: Are Balassa-Samuelson Effects Present?Ehsan U. Choudhri and Mohsin S. Khan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .387

F42 International Policy Coordination


F43 Economic Growth of Open Economies


How Much Do Trading Partners Matter for Economic Growth? Vivek Arora andAthanasios Vamvakidis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24

F47 Forecasting and Simulation

Assessing Early Warning Systems: How Have They Worked in Practice? Andrew Berg,Eduardo Borensztein, and Catherine Pattillo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .462

Volume 52 Index

548


F49 Other


G FINANCIAL ECONOMICS

G1 General Financial Markets


G15 International Financial Markets


G2 Financial Institutions and Services

G21 Banks; Other Depository Institutions; Mortgages

Banking on Foreigners: The Behavior of International Bank Claims on Latin America,1985–2000. Maria Soledad Martinez Peria, Andrew Powell, and Ivanna Vladkova-Hollar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .430

G22 Insurance; Insurance Companies


H PUBLIC ECONOMICS

H2 Taxation and Subsidies

H21 Efficiency; Optimal Taxation


H5 National Government Expenditures and Related Policies

H50 General


H6 National Budget, Deficit, and Debt

H63 Debt; Debt Management



H8 Miscellaneous Issues

VOLUME 52 INDEX

549


H87 International Fiscal Issues

The 35 Most Tumultuous Years in Monetary History: Shocks, the Transfer Problem, andFinancial Trauma. Robert Z. Aliber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-142

M BUSINESS ADMINISTRATION AND BUSINESS ECONOMICS; MARKETING;ACCOUNTING

M4 Accounting

M40 General


N ECONOMIC HISTORY

N1 Macroeconomics and Monetary Economics; Growth and Fluctuations

N10 General, International, or Comparative


N2 Financial Markets and Institutions

N20 General, International, or Comparative


N26 Latin America; Caribbean

Banking on Foreigners: The Behavior of International Bank Claims on Latin America,1985–2000. Maria Soledad Martinez Peria, Andrew Powell, and Ivanna Vladkova-Hollar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .430

O ECONOMIC DEVELOPMENT, TECHNOLOGICAL CHANGE, AND GROWTH

O1 Economic Development

O10 General

Reconsidering Expectations of Economic Growth After World War II from thePerspective of 2004. Robert W. Fogel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .SI-6

O11 Macroeconomic Analyses of Economic Development

From “Hindu Growth” to Productivity Surge: The Mystery of the Indian GrowthTransition. Dani Rodrik and Arvind Subramanian . . . . . . . . . . . . . . . . . . . . . . . . . .193

O2 Development Planning and Policy

O23 Fiscal and Monetary Policy in Development


Volume 52 Index

550


O4 Economic Growth and Aggregate Productivity

O47 Measurement of Economic Growth; Aggregate Productivity


Total Factor Productivity Revisited: A Dual Approach to Development Accounting.Shekhar Aiyar and Carl-Johan Dalgaard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .82

O5 Economywide Country Studies

O53 Asia Including Middle East


O55 Africa


O57 Comparative Studies of Countries

Total Factor Productivity Revisited: A Dual Approach to Development Accounting.Shekhar Aiyar and Carl-Johan Dalgaard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .82

Q AGRICULTURAL AND NATURAL RESOURCE ECONOMICS

Q4 Energy

Q43 Energy and the Macroeconomy


VOLUME 52 INDEX

551


This page intentionally left blank


In statistical matter throughout this issue,

dots (. . . ) indicate that the data are not available;

a dash (—) indicates that the figure is zero or less than half the final digit shown, or that the item does not exist;

a single dot (.) indicates decimals;

a comma (,) separates thousands and millions;

“billion” means a thousand million; and “trillion” means a thousand billion;

a short dash (–) is used between years or months (for example, 1998–99 or January–June) to indicate a total of the years or months inclusive of the beginning and ending years or months;

a slash (/) is used between years (for example, 1998/99) to indicate a fiscal year or a cropyear; and

components of tables may not add to totals shown because of rounding.

The term “country,” as used in this publication, may not refer to a territorial entity that is a state as understood by international law and practice; the term may also cover some territorialentities that are not states but for which statistical data are maintained and provided internationallyon a separate and independent basis.

Design: Luisa Menjivar-Macdonald and Sanaa Elaroussi


Documents

IMFstaffpapers - IMF eLibrary€¦ · ttt t (6) Empirical Dealer-Level Tests—Lyons (1995): ()* eIID it it it t=− − +μ α γ (1) Open Macroeconomics Theory—Parity conditions,