MS&E448: Statistical Arbitrage

Group 5: Carolyn Soo, Zhengyi Lian, Jiayu Lou, Hang Yang

Statistical Arbitrage (Recap)
How to systematically
identify stationary
- Stock returns - systematic and idiosyncratic components spreads X(t) and profit
from their mean
reversion?

- Form beta-neutral long-short positions using “group (n ≥ 1)” trades:

- long: for every $1 long in asset P, short $ j of ETF(j) for j = 1 to n
- short: for every $1 short in asset P, long $ j of ETF(j) for j = 1 to n

- Profit from mean-reversion of spread X

- test for stationarity
- model as OU process (mean-reverting) long-run
X mean
- check for fast mean reversion ( < 30 days)
- compute trading signals Bounded
- execute deviations
time
Stationarity Test & Fitting an OU-Model
- Augmented Dickney Fuller (ADF) Test for Stationarity of X
H0: = 0 (Xt is non-stationary)
test-statistic:
reject if p-value < 0.05

- X(t) for each stock modelled as an Ornstein-Uhlembeck (OU) process

Wt : Wiener process

- Assume κ, m, σ for each stock stays constant over 60-day trailing window,
accept only if
mean reversion
time-constant
Compute Trading Signals
- In equilibrium (i.e. as Δt → ∞), Model Adjustments
60 day trailing window to estimate
X and parameters (κ, m, σ)
- reject if mean reversion time
- Signal: constant = 1/κ > 30 days
- adjust for bias in m

Open short
+Sso
+Ssc Close short - volume adjusted returns to
S reduce signal on high
-Sbc Close long
-Sbo trading volumes.

Open long

4
Reference: Avellaneda & Lee, 2008
Midterm Recap (2003-2012)
$100k, static choice of 2003’s top 100 largest-cap
Net return: 115.3%, Sharpe 0.99

2003 2004 2005 2006 2007 2008 2009 2010 2011 2012

We remained beta neutral as desired over the entire period, however

our returns showed stagnation after 2008...
Post-2008: Why Stagnate?
● Avellaneda & Lee’s paper published → strategy exploited, no
more arbitrage! Greater competition for the same opportunities
● Avellaneda & Lee didn’t even cover the impact or aftermath of
crisis… perhaps the strategy is too simple to succeed in complex
environments
● 2008 financial crisis → Changed market environment
● Small capital base, small universe → Not enough opportunities
● Struggling to take the other side of the spread (illiquid stocks)
Post-2008: Delving Down For Improvements
● Period: 2008-2012 focused training ||| 2013-2016 testing
● Capital base: $1 Million
● Costless transactions
○ Costs were obfuscating the predictive power of our signals
○ In practice, funds with large capital pools can trade via
intermediaries to avoid driving up the prices for themselves
Post-Midterm Investigation
We have:

● Tested spreads for stationarity before applying our model (prev. discussed)
● Expanded the trading universe
● Addressed incompletely-filled orders (3 ways)
● Optimized signal cutoffs
● Optimized daily trading weights
● Experimented with mapping stocks to ETFs (pre-selecting regressors for
stocks)
Expanded Universe (Training 2008-2012)
● Previously: fixed universe with top 100 market cap stocks in 2003
● Now: every year, update the universe to the top 500 market cap stocks
○ Max limit of 100 long and 100 short stocks at any one time
● Problem with larger universe (1000)
○ Illiquidity of small-cap stocks leads to frequent unfilled orders and bias in pair trading

Universe Open Close Sharpe Annual

Max Limit Net returns
Size Cutoff Cutoff Ratio Return

100 20 1.2 0.6 0.57 21.4% 3.9%

500 100 1.2 0.6 1.23 42.1% 7.1%

1000 200 1.2 0.6 0.58 16.4% 3.0%

Unfilled Orders: Can’t fill our orders on time...
● Persist incompletely-filled stock orders on day 2, 3, etc. until they are
filled
○ But violates our beta neutrality in the long term
○ Large position of ETF, small incremental positions in stocks
● Delay the purchase of ETFs to day 2
○ Based on how many shares were successfully filled on the 1st day,
we will purchase corresponding ETFs on the 2nd day
○ Overcorrection, results in large-scale 1-day lag
● Sell overbought or buy back oversold ETFs on day 2
○ Make sure the consistency is restored on day 2
Cutoff Optimization (Training 2008 - 2012)
● Different combination of static open and close cutoffs
● Dynamic cutoff
○ Incorporates market volatility
○ Higher threshold in volatile market environment

● Conclusion: difficulty to find universally applicable parameters for drastically

different pre- and post-crisis market landscapes; all combinations faced issues
with lack of return after 2009
Cutoffs: Sensitivity Testing (Training 2008 - 2012)
Open Cutoff Close Cutoff Sharpe Ratio Return Beta

1.2 0.4 1.20 45.1% 0.00

1.2 0.5 1.18 42.2% 0.00

1.2 0.6 1.23 42.1% 0.00

1.3 0.4 0.86 27.5% 0.01

1.3 0.5 0.83 25.4% 0.02

1.3 0.6 1.01 30.1% 0.01

1.2 ＋ ⅕ 0.55 － ⅕
1.16 40.5% 0.0
＊sqrt(σ) ＊sqrt(σ)
Weights Optimization (Training 2008 - 2012)
● Assign weights proportional to the signal magnitude
○ Pairs with stronger signals will receive higher weights, and vice versa
○ Modest improvement

● Magnify the signal

○ Square the signal, etc.
○ Best Sharpe Ratio: 1.25

● Conclusion: lack of variation between signals above the threshold; magnifying

the signal too much means lack of diversification
Variant: Use Sector ETFs as regressors?
● Previously: Forecast expected returns by regressing on market…

● What if we try regressing the stock on a set of logical ETFs?

○ “Regress out” the market from each stock’s returns
○ Regressions and F-testing/ANOVA to select the ETFs which have
predictive power for that stock
Variant: Use Sector ETFs as regressors?
● 3 different approaches to map stocks to ETFs:

1. Map each stock to its sector ETF

○ i.e., Microsoft to HHH
2. Map each stock to the most significant sector ETF
○ Regress stock returns against ETF returns and find the ETF
with positive beta and smallest p-value
3. Map each stock to all significantly correlated ETFs
○ Identify all ETFs with regression p-values < 0.05
Risk Management

● Do not "reuse" signals

○ Change our position only when we see the appropriate signal

● In a crisis: starved of shorting / profit-taking opportunities

○ Scale by VWAP

● Leg risk when trading illiquid tickers

○ Next-day adjustment to regain beta neutrality
Risk Management 2
● Capital constraints
○ Restrict # of longs and shorts that we could hold at any instant
○

● Metrics:
○ Overall beta (systematic risk)
○ Rolling beta (systematic risk)
○ Volatility
○ Max drawdown (worst-case downside risk)
○ VAR (tail risk)
○ Sharpe ratio (risk-to-reward ratio)
○ Sortino ratio (downside risk-to-reward ratio)
Best Models (Training 2008-2012)
2008 - 2012 (5 yr): No Transaction Costs
Performance Metrics Risk Metrics
Max
Universe Strategy Annual Max
Stocks Sharpe Beta Volatility
Return Drawdown
Baseline (SPY) 9.5% 0.78 0.01 0.13 -14.4%
Baseline (SPY) + stationary 4.0% 0.57 -0.01 0.07 -7.7%
100 20
SPY + multiple ETF + stationary 4.1% 0.65 -0.01 0.07 -9.0%
SPY + best sector ETF + stationary 5.7% 0.86 0 0.07 -10.2%
Baseline (SPY) 6.3% 0.67 0.01 0.1 -11.3%
Baseline (SPY) + stationary 7.3% 1.23 0 0.06 -8.4%
500 100
SPY + multiple ETF + stationary 2.6% 0.54 0.01 0.05 -12.2%
SPY + best sector ETF + stationary 4.2% 0.84 0.01 0.05 -6.6%
Baseline (SPY) 8.1% 0.92 -0.01 0.09 -14.6%
Baseline (SPY) + stationary 3.1% 0.58 -0.01 0.06 -10.4%
1000 200
SPY + multiple ETF + stationary 3.6% 0.81 -0.01 0.04 -5.9%
SPY + best sector ETF + stationary 3.4% 0.77 -0.01 0.04 -5.8%
 Training (2008-2012)
Baseline (SPY) SPY
Baseline (SPY) + Multiple Sector ETF(s)
+ Stationary
+ Stationary
Cumulative
Return

2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012
Rolling Sharpe
Ratio (6 mths)

2008 2009 2010 2011 2012

2008 2009 2010 2011 2012
2008 2009 2010 2011 2012
 Training (2008-2012) SPY + Multiple Sector ETF(s)
Baseline (SPY) + Stationary + Stationary
Baseline (SPY)
Sharpe Ratio
(6 mths)

2008 2009 2010 2011 2012

2008 2009 2010 2011 2012
2008 2009 2010 2011 2012
(to SPY)
Beta

2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012
Drawdown

2008 2009 2010 2011 2012 2008 2009 2010 2011 2012
2008 2009 2010 2011 2012
Out-of-Sample Testing (2013-2016)
2013 - 2016 (4 yr): No Transaction Costs
Performance Metrics Risk Metrics
Max
Universe Strategy Annual Max
Stocks Sharpe Sortino Beta Volatility VaR*
Return Drawdown
Baseline (SPY) 2.2% 0.31 0.44 0.07 0.08 -18.7% -0.9%
Baseline (SPY) +
2.2% 0.54 0.75 0.01 0.04 -6.6% -0.47%
stationary
SPY +
500 100 Multiple ETF + 2.4% 0.59 0.84 0.01 0.04 -6.9% -0.53%
Stationary
SPY +
Best Sector ETF 2.5% 0.59 0.84 0.01 0.04 -6.9% -0.53%
+ stationary

* VaR = bottom 5th

percentile of daily returns
Out-of-Sample Testing (2013-2016)
Baseline (SPY) Baseline (SPY) + Stationary
Cumulative
Return

2013 2014 2015 2016 2013 2014 2015 2016

Sharpe Ratio
(6 mths)

2013 2014 2015 2016

2013 2014 2015 2016

 Out-of-Sample Testing (2013-2016)
Baseline (SPY) Baseline (SPY) + Stationary
Sharpe Ratio
(6 mths)

2013 2014 2015 2016

2013 2014 2015 2016
(to SPY)
Beta

2013 2014 2015 2016 2013 2014 2015 2016

Drawdown

2013 2014 2015 2016

2013 2014 2015 2016

 Mediocre Performance --- Seeking Reasons Why

● Avellaneda & Lee ended their study before the crisis hit…
○ Maybe overfitted to pre-crisis period?
○ Or core change in how markets move / how traders think

● 2012 Fed quantitative easing →Liquidity pushes up markets → Kill

off short positions and opportunities
Sector ETFs as regressors? Slight improvement...
SPY + Sector ETF(s) + Stationary Baseline (SPY) + Stationary
Cumulative
Return

2013 2014 2015 2016 2013 2014 2015 2016

Rolling Sharpe
Ratio (6 mths)

2013 2014 2015 2016 2013 2014 2015 2016

 Sector ETFs as regressors? Slight improvement...
SPY + Sector ETF(s) + Stationary Baseline (SPY) + Stationary
Sharpe Ratio
(6 mths)

2013 2014 2015 2016 2013 2014 2015 2016

Sharpe: 0.54
(to SPY)

Sharpe: 0.59
Beta

2013 2014 2015 2016 2013 2014 2015 2016

Drawdown

2013 2014 2015 2016 2013 2014 2015 2016

Sector ETFs as regressors? Slight improvement...
Probably because the predictive power of all sector ETFs with respect to
the market-independent component of returns is very poor…

Adjusted R^2 values for various tickers regressed on all sector ETFs (2008-2012)
Risk Metrics

Performance Metrics Risk Metrics

Max
Universe Strategy Annual Max
Stocks Sharpe Sortino Beta Volatility VaR*
Return Drawdown
Baseline (SPY) 2.2% 0.31 0.44 0.07 0.08 -18.7% -0.9%
Baseline (SPY) +
2.2% 0.54 0.75 0.01 0.04 -6.6% -0.47%
stationary
SPY +
500 100 Multiple ETF + 2.4% 0.59 0.84 0.01 0.04 -6.9% -0.53%
Stationary
SPY +
Best Sector ETF 2.5% 0.59 0.84 0.01 0.04 -6.9% -0.53%
+ stationary
Conclusion
● With stationary check, great pre-2008 before crisis hit
● Stationarity check controls max drawdown, preserves profit in
spite of unpredictable markets
● Not very robust when economic environment fundamentally
changes - wildly unpredictable when we vary starting year
● Too slow, not competitive enough? Rise of HFT
● Need better factors, to know environment
Avenues for Exploration
● Alternative Factor Selection
○ A model which includes sector ETFs performs only slightly better
○ Simple linear regression: Too crude?
● More Robust Parameter Optimization Procedures
● Conditionally-Heteroskedastic / Non-Stationary Models
Q+A