Upload
zelda
View
23
Download
0
Tags:
Embed Size (px)
DESCRIPTION
Lecture 4: Section 1. Generic Currency Signals. Momentum or Trend: bet on continuation of directional moves. Many ways to calculate the autoregressive nature of currency returns BofA: use past month and Deutsche Bank: use last 12-month returns and buy top-3 and sell bottom-3 - PowerPoint PPT Presentation
Citation preview
Lecture 4: Section 1
Generic Currency Signals
Momentum or Trend: bet on continuation of directional moves
Many ways to calculate the autoregressive nature of currency returns
• BofA: use past month and
• Deutsche Bank: use last 12-month returns and buy top-3 and sell bottom-3
• Could use MA cross-over rules (like 20-day & 200 day)
Why (and when) should momentum strategies work?
1
1
ln / 0
ln / 0
t t
t t
long foreign currency if S S
short foreign currency if S S
Carry: long high interest rate, short low interest rate
What interest rate differential?
• BofA: use 1 month LIBOR
• Deutsche Bank uses 3 month LIBOR & buy top-3 & sell bottom 3 – Rebalance every 3 months
How can you manage the risk of carry unwinds?
for usd
for usd
buy foreign currency if r r
sell foreign currency if r r
Fundamental Value: bet on reversion to long-term equilibrium value
Purchasing Power Parity (PPP) is most common measure
• BofA: compare PPP, adjusted by return to carry, to current spot rate
• Deutsche Bank: buy top-3 undervalued currencies and sell top-3 overvalued currencies
– Valuation determined by OECD PPP versus current spot rate
When will PPP be most likely to add value?
for usd
for usd
r r
r r
Buy foreign currency if PPP e S
Sell foreign currency if PPP e S
Lecture 4: Section 2
Liquidity Considerations
How do currencies rank in terms of liquidity?
0
20
40
60
80
100
EURUSD GBPUSD USDJPY EURCHF USDCHF AUDUSD EURSEK USDCAD EURNOK NZDUSD
What do you do about liquidity differences?
BofA has a “most liquid” portfolio:
• USD, EUR, JPY, GBP, CHF
And a less liquid portfolio:
• AUD, CAD, NOK, NZD, SEK
Can impose position caps that reflect liquidity tiers
• Depends upon size of portfolio
Could use IC weighting to alter alphas
*alpha IC volatility score
Modeling transaction costs
Liquidity will determine market impact of trades
• If you have large portfolio will have “backlog” where actual positions differ from desired positions
• Tcosts affected by inventory risk to broker (G&K, Ch. 16)– Estimated time for opposing trades to match a trade of size V:
• Trade 1 day’s volume, expect it to take a day for counterparties to clear risk
– Inventory risk facing liquidity provider depends on volatility:
• Convert annual vol to vol over relevant horizon for inventory management
• Measure in units of days, with 250 trading days in year
clear
Vtrade
Vdaily
250clear
inventory
Modeling transaction costs
• ADV varies across currencies– Use BIS survey data
• Vol varies across currencies– Calculate trailing vol or use implied vols
• Smooth some to avoid vol spikes distorting relative costs
• Beyond market impact costs, have fixed costs of commissions and spreads
i i i
tradesizetcosts a
ADV
Modeling transaction costs
Objective function:
• Tcost amortization factor– Higher to trade patiently– Lower to trade faster– Could simulate model to find “optimal” TCAF
• Experience with backlog will guide
• How often to rebalance? – Could use calendar interval or backlog risk trade trigger rule
• Only trade when risk >threshold
' ' | |th h Vh TC h
'
backlog opt act
backlog backlog backlog
h h h
h Vh
Lecture 4: Section 3
Risk Modeling
Risk modeling: drawdown risk
Clients often want to think about extreme value risk
• What is maximum drawdown? – Duration and magnitude
– Experience for fund with history
– Simulation for start-up
– Could make part of every proposed signal diagnostic
• Investment officers have different horizon than fund might suggest– Insurance companies or pension funds have very long-run horizon?
• Investment officer bonus paid annually• A major drawdown can get you fired
Risk modeling: covariance model
Fast or slow, historic or forward-looking?• Speed of COV should match rate of model alpha decay• Do you want “normal” correlations across assets?• Do you want to capture short-term shifts in correlations?• Do you want slowly-evolving correlation matrix?
– Exponentially smoothed with long HL
– Just update once-a-month?
– Impose priors in a “shrinkage” matrix to help capture important relations• Look at “hedge bundles” implied by construction: AUD & NZD; NOK & SEK;
EUR & CHF; ………• Regional funding; commodity currencies; “flight to quality” currencies;…..
• A fast model would require a faster COV matrix– Update daily
– Use implied vol only or in a blend with historic
Realized Volatility: the most accurate predictor
• Compute daily vol from the sum of squared intradaily returns– Motivated by approaching limit of Brownian motion
– Consider a continuous-time log process for the quote q:
– The corresponding discretely-sampled returns are:
– The variance of the h-period returns is:
– While that variance is unobservable, a measure of realized vol that is consistent (in m) for the integrated vol is summing the high-frequency squared returns:
Realized Volatility: the most accurate predictor
• Don’t sample at very highest frequency as get “microstructure noise”– Non-synchronous prices
– Data holes
• Can treat volatility as observable rather than latent – See Melvin & Melvin (Review of Economics & Statistics, 2003) for FX
example
• Use daily vol estimates to construct cov matrix– Better for fast models than implied vol, which contains risk premium
Lecture 4: Section 4
Model Building Tips
Follow SPCA Method for signal development
Sensible– First step is to convince yourself (and others) that the idea makes sense– No data mining allowed! – If it conflicts with finance theory, why?
Predictive– The proposed signal is useful for predicting returns
Consistent– Performance occurs through time (not in 1 episode)– May like signal that is not very consistent, but in certain market
conditions adds good diversification• For instance, a value signal that only has positive performance when carry
unwinds
Additive– Does it add value beyond existing model?
Useful Diagnostics for Signal Development
• IR for whole backtest sample and subsamples
• Cumulative return plot
• Drop one asset at a time to determine dependence across assets
• Lead-lag plots of IR around implementation day– Determine speed of alpha decay – Inform trading process: what do you lose by trading patiently?
• Tilt-timing analysis– Tilt: what if we held a constant portfolio– Timing: residual of actual returns - tilt
'
' '
tilt hr
timing hr hr
Lecture 4: Section 5
Pojarliev & Levich, Do Professional Currency Managers Beat the Benchmark?
Currency Benchmark
• There is no accepted practice– There is no “market portfolio”
• All positions long/short
– Zero assumption from random-walk findings of academic literature
– Cash: if you can’t beat that, why invest?
• P&L create currency indexes off of generic signals– Are these benchmarks?
– More like style betas
Estimate factor model
• Returns from Barclay Currency Trader Index (BCTI)
• 1 month LIBID for risk-free returns
t i it ti
t
R F
R excess return total return riskfree rate
skill
factor loading
F beta factor
Style Factors
• Carry– Citi G10 Carry Index: equally-weighted basket of 13 pairs
• Trend– AFX index: equally-weighted portfolio of 3 MA rules
• 32, 61, 117 days over 7 pairs weighted by turnover
• Value– Citi Beta1 G10 PPP Index: long currencies <20% undervalued & short
currencies >20% overvalued• OECD PPP for 13 pairs compared to spot
• Volatility– Avg of implied vol for EURUSD & USDJPY
• This is relevant for vol strategies rather than directional bets• But will capture carry unwinds so would want to be long JPY & short high-
yielders
All manager index results
• Trend is useful factor; carry & PPP not useful
• Alpha insignificant when factors are accounted for
• Short sample results– Carry sigificant 2001-2006 but Trend dominates
– Vol significant in both periods, but more significant in latter period• They say currency derivatives trading increased after 2001• In the low vol period, selling vol was a successful strategy
– Vol quite low until 2007, I kept saying “just wait”
Individual Manager Results
• 34 have data over 2001-2006 period (survivors)• Alpha positive on avg with big range
– Reflects different risk levels of funds
• IR puts on equal footing– Mean=0.47; median=0.45
• Estimate IR from 4-factor model– Only 8 firms have sig & positive alpha
– So returns are combination of beta & alpha
– With similar returns, find some managers generative beta while others generate alpha
– Trend sig for 15 firms; Carry for 8; Vol for 7; Value for 5• Firms have different approaches
Timing Ability, Alpha & R-square
• A passive portfolio has asset weights uncorrelated with returns– Regress returns on factors and factors squared
– Some evidence of timing ability for about ½ of firms
• Finally, the more a firm loads on 4 factors, the smaller alpha
Critique
• Don’t believe that it is cheap & easy to get active manager performance by just betting on generic strategies– Active managers adjust portfolio to changing market conditions
• Putting a generic strategy on cruise control is very dangerous– 2007 was good lesson for carry indexers
• Paper has small sample of active managers & conclusions are too sweeping from this limited universe