51
A Semi-Nonparametric Model of the Pricing Kernel and Interest Rates Yuriy Kitsul University of North Carolina - Chapel Hill [email protected] November 29, 2004

2004-11-29

Embed Size (px)

Citation preview

Page 1: 2004-11-29

A Semi-Nonparametric Model of the Pricing Kernel andInterest Rates

Yuriy KitsulUniversity of North Carolina - Chapel Hill

[email protected]

November 29, 2004

Page 2: 2004-11-29

Introduction and Motivation

• Models of the short-term (risk-free) interest rate describe the time-seriesdynamics of the yield of the bond of instantaneous maturity (in the continuoustime set-up), which serves as a discount rate of risk-adjusted cash-flows innumerous financial applications

• Dynamic models of the term structure of interest rates describe:

- how bond yields depend on time to maturity at each moment of time

- how the form of this dependence changes with time

• Need a model, which would produce a good fit of bond yields

Yuriy Kitsul 1

Page 3: 2004-11-29

Objectives

• To model the pricing kernel and bond yields in a flexible pure-diffusionframework that allows

- to introduce various levels of model complexity ⇐⇒ semi-nonparametrics

- to have closed-form solutions for bond yields ⇐⇒ eigenfunctions

• Within the offered framework to let the data decide on the level of modelcomplexity/number on semi-nonparametric terms/elements

Yuriy Kitsul 2

Page 4: 2004-11-29

Two Major Themes

• Semi-nonparametrics:

- Flexibility (as in nonparametrics) ⇒ 1) information about unknown relationships2) identification of mispriced securities

- Closed form expressions of the stochastic processes of interest (as inparametric models)

• Empirical sufficiency of pure diffusions (motivated by problematic hedgingproperties of jumps, e.g. review and discussion by Jones (2003) in the contextequity options)

Yuriy Kitsul 3

Page 5: 2004-11-29

In this presentation

• Methodology:

- Model the unknown functional relationship between the pricing kernel andthe state variables

- Explore the implications for the bond yields, the short risk-free interest rateand the market price of risk.

- Offer two methods of implementing the ”no-arbitrage” condition.

• Empirics:

- Estimate the model with Gallant&Tauchen’s (2004) EMM-MCMC

- Find that one factor model with a sufficient, but not too high, number ofsemi-nonparametric terms can not be rejected by the univariate yields data.

Yuriy Kitsul 4

Page 6: 2004-11-29

Current Issues and Some Recent Literature - 1, Modeling 1

• Bt(T ), price at time t of a zero-coupon bond maturing and paying 1 dollarat time T > t:

MtBt(T ) = EPt {MT · 1} (1)

where P is the physical, or real-world, probability measure,Mt is the instantaneous stochastic discount factor, or pricing kernel, whichassigns prices at time t to payoffs at time t + dt.

• The pricing kernel, Mt,T , which assigns prices at time t to payoffs at time

T > t is Mt,T = MTMt

and the price of bond becomes:

Bt(T ) = EPt {Mt,T · 1} (2)

Yuriy Kitsul 5

Page 7: 2004-11-29

Current Issues and Some Recent Literature - 1, Modeling 2

• Under risk-neutral probability measure, Q, which exists and is unique under”absence of arbitrage” and completeness:

Bt(T ) = EQt {e−

R Tt rsds · 1} (3)

where rs is instantaneous (short-term) risk-free interest rate, or money marketspot interest rate.

• To switch back to physical measure, P , Girsanov theorem is used:

Bt(T ) = EPt {e−

R Tt rsds · ξt,T · 1} (4)

where Radon-Nikodym derivative ξt,T = e(R Tt λsdWs−1

2

R Tt λ2

sds) depends onλs, known as the market price of risk.

Yuriy Kitsul 6

Page 8: 2004-11-29

Current Issues and Review of Some Recent Literature - 2, TermStructure Literature 1.

• Depending on how rs is modeled, the literature can be categorized into thefollowing types of models:

– affine (eg. Dai/Singleton(00))– non-linear (eg. quadratic: Ahn/Dittmar/Gallant(02), Ahn/Dittmar/Gallant/Gao(03))– regime-shift1(eg. Bansal/Zhou(02),Bansal/Tauchen/Zhou(03))– jump-diffusion (eg. Duffie/Pan/Singleton(00))

• Relatively simple specification for underlying factors to obtain closed-formbond pricing formulas. Complexity is usually achieved by adding extra factors,jumps or regime shifts.

1Often in a discrete time framework

Yuriy Kitsul 7

Page 9: 2004-11-29

Current Issues and Review of Some Recent Literature - 2, TermStructure Literature 2

Some empirical features of the existing term-structure models:

• A test of empirical model is a joint test of specifications for the short-termrisk-free interest rate and the market price of risk, i.e. test of a model forthe pricing kernel.

• The debate on which approach is empirically preferable is unresolved yet (eg.constant sign of risk premium and trade-off between a structure of volatilityand negative correlation in affine models).

Yuriy Kitsul 8

Page 10: 2004-11-29

Current Issues and Some Recent Literature - 3, Time Series Dynamicsof the Short Interest Rate - 1

• Use a proxy for the short interest rate, rt, e.g. three-month Treasury bill rate,which may be a problem according to Chapman, Long and Pearson (1997)

• No need for closed-form bond price solutions

• More sophisticated specifications for rt, which are tested separately, notjointly with specification for the market price of risk

Yuriy Kitsul 9

Page 11: 2004-11-29

Current Issues and Some Recent Literature - 3, Time Series Dynamicsof the Short Interest Rate -2

• Can categorized into:

- one-factor non-linear diffusions (e.g. Aıt-Sahalia (96))

- stochastic volatility (e.g. Andersen and Lund (97))

- regime-shifts (e.g Ang and Bekaert (00))

- jump-diffusions (e.g. Johannes (04))

• A variety of estimation methods: Aıt-Sahalia (96), Conley, Hansen, Luttmerand Scheinkman (97), Gallant and Tauchen (98), Eraker(01), Durham(03)and others

Yuriy Kitsul 10

Page 12: 2004-11-29

In This Work - 1, Questions

• Is a flexible pure diffusion framework empirically sufficient?

• How many factors/state variables are needed?

• Do we really need any proxies for instantaneous risk-free interest rate to studynon-linearities in its dynamics? What do bond yields of longer maturities tellus about its behavior?

• What is the functional form that describes how investor’s risk preferencesreact to the underlying shocks?

Yuriy Kitsul 11

Page 13: 2004-11-29

Essence of Approach - 1

• A1: ∃ a factor/state variable xt, which drives the economy and is governedby:

dxt = µ(xt)dt + σ(xt)dWt, l < x < u (5)

• A2: ∃ a pricing kernel, Mt and it is some real-valued and positive function,M(xt) ∈ L2(q):

Mt = M(xt) such that

∫ u

l

M(x)2q(x)dx < ∞ (6)

Yuriy Kitsul 12

Page 14: 2004-11-29

Essence of Approach - 2

• ∀ such function, M(x), can be expanded on an infinite number of orthogonalterms, φi(x), such that

∫ u

lφi(x)φj(x)q(x)dx =

{h2

i , if i = j, where hi is called norm0 otherwise

Therefore,

Mt =∞∑

i=0

aiφi(xt) (7)

where ai = 1h2

i

∫ u

lφi(x)M(x)q(x)dx - an inner product w/r to the weighting

function q(x).

Yuriy Kitsul 13

Page 15: 2004-11-29

Essence of Approach - 3

• φi(x) are eigenfunctions =⇒ closed form bond-prices

• Truncate expansion to a finite number of terms, assume some specific formfor xt, and, thus, for eigenfunctions, and estimate ai from the data

• H0: Given parametric specification for underlying factors and a number ofterms in the expansion, are there parameter values, for which the modeledpricing kernel is an acceptable representation of the true pricing kernel?

Yuriy Kitsul 14

Page 16: 2004-11-29

Another Brief Digression into the Literature

• Non-linear empirical pricing kernel literature: Bansal/Viswanathan(1993),Bansal/Hseih/Viswanathan(1993), Chapman(1997), and others.

• What is different in this work:

- Latent state variables, interest rates application and closed form prices

• Papers that discuss eigenfunctions2:

Hansen/Scheinkman/Touzi(1998), Chen/Hansen/Scheinkman(2000), Meddahi(2001-a,b), Florens/Renault/Touzi(1998), Gorovoi/Linetsky(2003) and others

• Rogers (1997) uses eigenfunctions to form the potential and presents bondpricing formulas, but implementation issues are not obvious

2Some of the following discussion will be based on the first three sources

Yuriy Kitsul 15

Page 17: 2004-11-29

Eigenfunctions Framework - 1

• A diffusion process that governs xt can also be described by an infinitesimalgenerator, A:

Aφ(x) = µ(x)φ(x)′ + 0.5σ(x)φ(x)′′ (8)

• Define scale function, S(x), and speed density, m(x), (Karlin/Taylor(81))

S(x) =∫ x

s(ξ)dξ (9)

where s(x) = exp(− ∫ x 2µ(ξ)σ2(ξ)

dξ)

m(x) =1

s(x)σ2(x)(10)

Yuriy Kitsul 16

Page 18: 2004-11-29

Eigenfunctions Framework - 2

• Density of stationary distribution, defined as q(x) = lims→∞ p(x, s|y, t):

q(x) =m(x)∫ r

lm(ξ)dξ

(11)

• Consider the following equation (l < x < u):

Aφ(x) = −δφ(x) (12)

which is the same as

µ(x)φ(x)′ +12σ2(x)φ(x)′′ + δφ(x) = 0 (13)

Yuriy Kitsul 17

Page 19: 2004-11-29

Eigenfunctions Framework - 3

• Under some appropriate boundary protocol there will be a countable numberof functions, φi(x) and constants δi, which solve this equation and areorthogonal w/r to a stationary density q(x).

φi(x) - i’th eigenfunction of A and δi - corresponding i’th eigenvalue.

• Crucial property:E(φi(xT )|Ft) = e−δi(T−t)φi(xt) (14)

Yuriy Kitsul 18

Page 20: 2004-11-29

Eigenfunctions Framework - 4, Some Examples

• O-U process of a form below has q(x) = e−x2

2 with eigenfunctions of itsgenerator equal to Hermite polynomials and eigenvalues δi = κi

dxt = −κxtdt +√

2κdWt

H0(xt) = 1,H1(xt) = xt,Hi(xt) = 1√i(xtHi−1 −

√i− 1Hi−2(xt))

• CIR process of a form below has q(x) = xαe−x with eigenfunctions of itsgenerator equal to Laguerre polynomials and eigenvalues δi = κi

dxt = κ(α + 1− xt)dt +√

2κ√

xtdWt

Lα0 (xt) = 1, L1(xt) = 1+α−xt√

1+α

Yuriy Kitsul 19

Page 21: 2004-11-29

Bond Pricing - 1

• Recall that we represent a stochastic discount factor, Mt, as

Mt =∑n

i=0 aiφi(xt)

and Bt(T ), price at time t of a zero-coupon bond maturing at T > t:

MtBt(T ) = Et(MT · 1)

• Substituting an expression for the pricing kernel, we obtain:

(n∑

i=0

aiφi(xt))Bt(T ) = Et(n∑

i=0

aiφi(xT )) (15)

Yuriy Kitsul 20

Page 22: 2004-11-29

Bond Pricing - 2

• Using the property of eigenfunctions, Et(φi(xT )) = e−δi(T−t)φi(xt),

Bt(T ) =∑n

i=0 aie−δi(T−t)φi(xt)∑n

i=0 aiφi(xt)(16)

=n∑

i=0

aiφi(xt)∑nj=0 ajφj(xt)

e−δi(T−t)

• Instantaneous risk-free interest rate:

rt = lim(T−t)→0

−ln(Bt(T ))T − t

=∑n

i=0 aiδiφi(xt)∑ni=0 aiφi(xt)

(17)

Yuriy Kitsul 21

Page 23: 2004-11-29

Market Price of Risk - 1

• Duffie (2001): assume dMt = µM(t)dt + σM(t)dWt

and consider some arbitrary security, with cumulative-return process,

dRt = dStSt

= µS(t)St

dt + σS(t)St

dWt

Then, the Sharpe ratio, or the market price of risk, −λt:

−λt = µR(t)−rtσR(t) = −σM(t)

Mt

Yuriy Kitsul 22

Page 24: 2004-11-29

Market Price of Risk - 2

Using Ito lemma:

dMt = {A(n∑

i=0

aiφi(xt))}dt + σ(xt)(n∑

i=0

aiφi(xt))′dWt (18)

Therefore,

λt =σ(xt)(

∑ni=0 aiφi(xt)′)∑n

i=0 aiφi(xt)(19)

Risk premium is very general and can change sign.

Yuriy Kitsul 23

Page 25: 2004-11-29

Positivity of Pricing Kernel

• Hansen/Richard(87) and Cochrane(2001): M > 0 ⇐⇒ no-arbitrage.

• Example of how to impose positivity numerically:

- Consider: Mt = a0 + a1H1(xt) + a2H2(xt)

- Find a map between the coefficients (a0, a1, a2) of this combination ofHermite polynomials and coefficients (b, c) of a regular polynomial of thesecond order that has only complex roots and, thus, never crosses zero:

a0 + a1H1(x) + a2H2(x) = (x− b− ic)(x− b + ic)

- Mt is either positive or negative on the whole state space of xt. If negative,multiply it by −1.

• Impose a prior in the process of MCMC estimation.

Yuriy Kitsul 24

Page 26: 2004-11-29

Multi-Factor Extension

• Independent factors, x1t and x2t, with eigenfunctions φ1i and φ2j,respectively.

Mt =n1∑

i=0

n2∑

j=0

aijφ1i(x1t)φ2j(x2t) (20)

• Define φij(xt) = φ1i(x1t)φ2j(x2t), where xt = [x1t, x2t] The followingproperty3 and all the previous results hold:

Et{φij(xT )} = e−δij(T−t)φij(xt) (21)

where δij = δ1i + δ2j

3See Meddahi(2001 a) for details

Yuriy Kitsul 25

Page 27: 2004-11-29

Data and Estimation Method

• Original and updated versions of a data set of Ahn/Dittmar/Gallant/Gao(03)who combine data of McCulloch and Kwon(93) and data provided by DanielWagooner and obtained by methods described in Bliss(97)

• Range: 01/52 - 12/99 (extended data is available till 12/02). Maturity: 6months.

• Method of estimation: Efficient Method of Moments proposed by Gallant andTauchen(96) and extended in Gallant and Tauchen(04) with Markov ChainMonte Carlo methodology along the lines of Chernozhukov and Hong(2003).

Yuriy Kitsul 26

Page 28: 2004-11-29

Overview of Efficient Method of Moments (EMM) - 1

• Advantages:

- Used when likelihood methods are not practical.

- Avoids an ad hoc selection of moment conditions when compared to othermethods of moments.

- Perfect for unobservable/latent factor structures and continuous-timemodels.

• Simulation-based method (related to Duffie and Singleton(93) andGourieroux, Monfort and Renault(93)):

Choose the parameters, ρ, of a model of interest (let us call it the mainmodel) in such a manner that the data simulated from the model is as closeto the observed data as possible.

Yuriy Kitsul 27

Page 29: 2004-11-29

Overview of Efficient Method of Moments (EMM) - 2

• Step 1: Projection: Summarize the data using the so-called auxiliary model,which is not the true model, but which approximates the data sufficientlywell and has a readily computable likelihood function f(·) in a closed form.

• Estimate the parameters of the auxiliary model, θ, by maximizing its likelihoodand and using the observed data, yt, xt−1:

θn = arg maxθ∈Θ1n

∑nt=1 log[f(yt|xt−1, θ)]

where n is the number of observations. Therefore, by construction

∂∂θ

1n

∑nt=1 log[f(yt|xt−1, θ)] = 0

Yuriy Kitsul 28

Page 30: 2004-11-29

Overview of Efficient Method of Moments (EMM) - 3

• Step 2: Estimation: The obtained score vector is used to generate momentconditions by simulating {yt, xt−1} from the main model:

m(ρ, θ) = 1N

∑Nt=1

∂∂θlog[f(yt|xt−1, θ)]

• We choose the parameters of the main model, ρ so that the generatedmoment conditions are as close to zero as possible:

ρn = arg minρ m′(ρ, θn)(In)−1m(ρ, θn)

where In is a quasimaximum likelihood information matrix and

nsn(ρ) = nm′(ρ, θn)(In)−1m(ρ, θn) ∼ χ2dim(ρ)−dim(θ)

Yuriy Kitsul 29

Page 31: 2004-11-29

Overview of Efficient Method of Moments (EMM) - 4, MCMCExtention

• Construct the analog to the likelihood in the Bayesian Markov Chain MonteCarlo (MCMC) methods:

L(ρ) = e−nsn(ρ)

• Use the standard Metropolis-Hastings algorithm to simulate a chain (and useits mode as the estimate):

{ρ(1), ...., ρ(Nch)}.

• 1) Easy to impose prior restrictions on non-linear functionals ψ(ρ)

2) Much better optimizer than the traditional methods (eg. a hill-climber)

3) Can use an analog of Bayesian posterior for econometric inference.

Yuriy Kitsul 30

Page 32: 2004-11-29

Metropolis-Hastings algorithm

• A candidate for the next value in the chain, ρnew is drawn from a proposaldensity q(ρnew|ρold), which, among other things, is easy to simulate from.

• Using the candidate value, ρnew the data simulation of length N is obtainedand the objective function, sn(ρnew), the functional of the parameters, ψnew,the prior, π(ρnew, ψnew), and the likelihood, L(ρnew) = e−nsn(ρnew) arecomputed.

• The chain moves from ρold to ρnew with probability

min{L(ρnew)π(ρnew,ψnew)q(ρnew|ρold)L(ρold)π(ρold,ψold)q(ρold|ρnew) , 1}

Yuriy Kitsul 31

Page 33: 2004-11-29

Data and Projected Conditional Mean

53 58 63 68 73 78 83 88 93 980

2

4

6

8

10

12

14

16

18Data and projected conditional mean, Six−month Treasury Bill, 03/53 − 12/99

time

percen

t

projecteddata

Figure 1: 6-month Treasure Bill: Data, 03/53-12/99 and conditional mean obtained with

11s1s0s1s1400000 SNP score .

Yuriy Kitsul 32

Page 34: 2004-11-29

Gaussian Factor Case: Empirical Model and Observation Equation

• Factor/state variable dynamics:

dxt = κ(θ − xt)dt + σdWt

• Pricing kernel within a model with n Hermite polynomials, H(n)-model:

Mt =∑n

i=0 aiHi(xt)

• Yield:

yldt(T ) = − 1T−t{ln(

∑ni=0 aie

−κi(T−t)Hi(xt))− ln(∑n

i=0 aiHi(xt))}

• Estimated parameters: {κ, θ, σ, a1, ...an}.

Yuriy Kitsul 33

Page 35: 2004-11-29

Estimation Results: 1 Gaussian factor, 2, 3 and 4 Hermite polynomials,Score: 11s1s0s1s1400000

Coefficient/Model CIR H(2) H(3) H(4)

κ 0.12294 0.40527 0.73877 0.604( 0.072865) (0.048579) (0.051739) (0.015484)

θ 5.1059 0.081055 0.19678 -0.14795(0.092615) (0.036123) (0.049139) (0.026694)

σ 0.19717 0.16699 0.2915 0.17139(0.0081534) (0.021523) (0.021654) (0.010689)

a0 1, fixed 1, fixed 1, fixeda1 1, fixed 1, fixed 1, fixeda2 -0.2373 -0.10693 -0.23877

(0.042031) (0.011367) (0.0021196)a3 0.31494 0.36616

(0.016185) (0.0017625)a4 -0.12256

(0.0016904)

χ2 21.323 21.716 9.9722 7.308p− value 0.0033 0.0014 0.0760 0.1205

df 7 6 5 4

Yuriy Kitsul 34

Page 36: 2004-11-29

0 10000 20000 30000 40000 50000

0.20.4

0.60.8

1.0

0 10000 20000 30000 40000 50000

4.85.0

5.25.4

0 10000 20000 30000 40000 50000

0.180.20

0.220.24

0.26−32

−30−28

−26−24

−22

Figure 2: CIR model. Chain of each of the parameters, {κ, θ, σ}. Every 50th point is plotted.

.

Yuriy Kitsul 35

Page 37: 2004-11-29

0 20 40 60 80 100

0.00.2

0.40.6

0.81.0

0 20 40 60 80 100

0.00.2

0.40.6

0.81.0

0.00.2

0.40.6

0.81.0

Figure 3: CIR model. Auto-correlation function for each of the parameters, {κ, θ, σ}..

Yuriy Kitsul 36

Page 38: 2004-11-29

0.0 0.2 0.4 0.6 0.8 1.0 1.2

4.5 5.0 5.5

Figure 4: CIR model. Kernel density for each of the parameters, {κ, θ, σ}..

Yuriy Kitsul 37

Page 39: 2004-11-29

V1

V2

V3

Figure 5: CIR model. Scatter plots of pairs of parameters {κ, θ, σ}. Every 50th point is

plotted. .

Yuriy Kitsul 38

Page 40: 2004-11-29

0 10000 20000 30000 40000 50000

0.20.5

0.8

0 10000 20000 30000 40000 50000

−0.5−0.2

0.1

0 10000 20000 30000 40000 50000

0.050.15

0 10000 20000 30000 40000 50000

0.61.0

1.4

0 10000 20000 30000 40000 50000

0.61.0

1.4

0 10000 20000 30000 40000 50000

−0.8−0.4

0 10000 20000 30000 40000 50000

−30−26

−22

Figure 6: H(2) model. Chain of each of the parameters, {κ, θ, σ, a0, ..., a2}. Note: a0 and

a1 are fixed for the identification purposes. Every 50th point is plotted..

Yuriy Kitsul 39

Page 41: 2004-11-29

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

Figure 7: H(2) model. Auto-correlation function for each of the parameters,

{κ, θ, σ, a0, ..., a2}. Note: a0 and a1 are fixed for the identification purposes..

Yuriy Kitsul 40

Page 42: 2004-11-29

0.2 0.4 0.6 0.8 1.0

−0.5 0.0 0.5

−0.1 0.0 0.1 0.2 0.3

−1 0 1 2 3

−1 0 1 2 3

−1.0 −0.8 −0.6 −0.4 −0.2 0.0 0.2

Figure 8: H(2) model. Kernel density for each of the parameters, {κ, θ, σ, a0, ..., a2}. Note:

a0 and a1 are fixed for the identification purposes..

Yuriy Kitsul 41

Page 43: 2004-11-29

V1

V2

V3

V4

V5

V6

Figure 9: H(2) model. Scatter plots of pairs of parameters {κ, θ, σ, a0, ..., a2}. Note: a0

and a1 are fixed for the identification purposes. Every 50th point is plotted..

Yuriy Kitsul 42

Page 44: 2004-11-29

0 10000 20000 30000 40000 50000

0.500.60

0.70

0 10000 20000 30000 40000 50000

−0.20

−0.05

0 10000 20000 30000 40000 50000

0.140.18

0.22

0 10000 20000 30000 40000 50000

0.61.0

1.4

0 10000 20000 30000 40000 50000

0.61.0

1.4

0 10000 20000 30000 40000 50000

−0.245

−0.230

0 10000 20000 30000 40000 50000

0.360

0.370

0 10000 20000 30000 40000 50000

−0.130

−0.115

−20−14

−8

Figure 10: H(4) model. Chain of each of the parameters, {κ, θ, σ, a0, ..., a4}. Note: a0 and

a1 are fixed for the identification purposes. Every 50th point is plotted..

Yuriy Kitsul 43

Page 45: 2004-11-29

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.8

0 20 40 60 80 100

0.00.4

0.80.0

0.40.8

Figure 11: H(4) model. Auto-correlation function for each of the parameters,

{κ, θ, σ, a0, ..., a4}. Note: a0 and a1 are fixed for the identification purposes..

Yuriy Kitsul 44

Page 46: 2004-11-29

0.4 0.5 0.6 0.7

−0.3 −0.2 −0.1 0.0 0.1

0.10 0.15 0.20 0.25

−1 0 1 2 3

−1 0 1 2 3

−0.25 −0.24 −0.23 −0.22

0.355 0.360 0.365 0.370 0.375 0.380

Figure 12: H(4) model. Kernel density for each of the parameters, {κ, θ, σ, a0, ..., a4}. Note:

a0 and a1 are fixed for the identification purposes..

Yuriy Kitsul 45

Page 47: 2004-11-29

V1

V2

V3

V4

V5

V6

V7

V8

Figure 13: H(4) model. Scatter plots of pairs of parameters {κ, θ, σ, a0, ..., a4}. Note: a0

and a1 are fixed for the identification purposes. Every 50th point is plotted..

Yuriy Kitsul 46

Page 48: 2004-11-29

53 58 63 68 73 78 83 88 93 980

2

4

6

8

10

12

14

16

18Projected and reprojected conditional means, Six−month Treasury Bill, 03/53 − 12/99

time

percent

projectedreprojected

53 58 63 68 73 78 83 88 93 98−4

−3

−2

−1

0

1

2

3Relative difference between projected and reprojected conditional mean, Six−month Treasury Bill, 03/53 − 12/99

time

percent

Figure 14: Conditional projected and reprojected first moments and their difference; ”1

Gausssian factor - 4 Hermite polynomials” model, estimated using 11s1s0s1s1400000 SNP

score.Yuriy Kitsul 47

Page 49: 2004-11-29

53 58 63 68 73 78 83 88 93 980

1

2

3

4

5

6

7Projected and reprojected conditional volatilities, Six−month Treasury Bill, 03/53 − 12/99

time

projectedreprojected

53 58 63 68 73 78 83 88 93 98−40

−20

0

20

40

60

80Relative difference between projected and reprojected conditional volatility, Six−month Treasury Bill, 03/53 − 12/99

time

percent

Figure 15: Conditional projected and reprojected second moments and their difference; ”1

Gausssian factor - 4 Hermite polynomials” model, estimated using 11s1s0s1s1400000 SNP

score.Yuriy Kitsul 48

Page 50: 2004-11-29

Conclusions

• It is possible to have a flexible diffusion framework with ”no-arbitrage” andclosed-form bond prices.

• EMM-MCMC is very well suited to handle our framework, which is highlynon-linear, with prior restrictions on non-linear functionals of parameteres

• A one Gaussian factor model with a sufficient, but not overly excessivenumber of semi-nonparametric terms can not be rejected by the time-seriesof 6-month to maturity yields data.

• Further work: fit the model to the joint dynamics of several yields of differentmaturities.

Yuriy Kitsul 49

Page 51: 2004-11-29

Extensions

• Incorporate macroeconomic variables.

• Apply to a multi-country case.

• Time-varying functional forms.

• Derivatives.

50