Use of Intraday Alphas in Algorithmic Trading

QWAFAFEW Boston Chapter June 21, 2011

Thorsten Schmidt, CFA Thor Advisors LLC

Execution Algorithms: Traditional Area of Focus

Minimize market impact & volatility of the result Vs a price benchmark (e.g. VWAP, arrival price, open, last close) Techniques commonly used to execute efficiently: - Forecast liquidity distribution across time and venues - Forecast adverse selection by venue (esp. dark pools) - Estimated price impact & high frequency volatility, optimize the trade-off between the two - Tactics to minimize signaling & information leakage (e.g. latency-adjusted Smart Router spraying) - Judicious spread-crossing Not so widely exploited (yet): - Generic alphas (Micro Alpha Models) - Portfolio implicit alphas (Alpha Profiling)

Generic Alpha Example: Periodicity

Intraday Patterns in the Cross-section of Stock Returns (Heston, Korajczyk, Sadka 2010) Intra-day segment (excess) return continuation at daily intervals Documented for U.S. stocks for 8-year time period 2001 2009 Heston, Korajczyk, Sadka robustness checks: - Robust after correcting for bid-ask bounce - Not an artifact of thinly traded stocks - Not an artifact of intra-day spread changes - Magnitude of effect inversely proportional to market cap, but not limited to small cap stocks - Periodicity in volume doesnt explain return periodicity Lets explore if it is relevant in a more recent time period and how useful in algorithmic execution trading context

Honing in on the Periodicity Effect

ACF Stock Returns
0.15 Autocorrelation Coefficient -0.15 -0.10 -0.05 0.00 0.05 0.10

13

26

39

52

65 30min Bin #

78

91

104

117

130

Returns normalized by stock trailing 20 day volatility

The first intraday lag shows high negative autocorrelation due to temporary liquidity imbalances and bid-ask bounce.

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. The trading day has 390 minutes, i.e. there are 13 30min bins per day.

Improved Visibility of Periodicity by Removing 1st Lag

ACF Stock Returns
0.15 Autocorrelation Coefficient -0.15 -0.10 -0.05 0.00 0.05 0.10

13

26

39

52

65 30min Bin #

78

91

104

117

130

Returns normalized by stock trailing 20 day volatility

ACF Stock Returns (ex 1st Lag)

0.02 Autocorrelation Coefficient -0.02 -0.01 0.00 0.01

13

26

39

52

65 30min Bin #

78

91

104

117

130

Returns normalized by stock trailing 20 day volatility

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. The trading day has 390 minutes, i.e. there are 13 30min bins per day.

Enhanced Periodicity Signal via Removal of Market and Sector Noise

1) ACF Stock Returns
0.02 0.02 Autocorrelation Coefficient
13 26 39 52 65 78 91 104 117 130

2) ACF CAPM Excess Returns, normalized by Tracking Error

Autocorrelation Coefficient

0.01

0.00

All ACF charts are without the 1st Lag

-0.01

-0.02

-0.02

-0.01

0.00

0.01

13

26

39

52

65

78

91

104

117

130

30min Bin #

30min Bin #

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. The trading day has 390 minutes, i.e. there are 13 30min bins per day.

Enhanced Periodicity Signal via Removal of Market and Sector Noise

1) ACF Stock Returns
0.02 0.02 Autocorrelation Coefficient
13 26 39 52 65 78 91 104 117 130

2) ACF CAPM Excess Returns, normalized by Tracking Error

Autocorrelation Coefficient

0.01

0.00

All ACF charts are without the 1st Lag

-0.01

-0.02

-0.02

-0.01

0.00

0.01

13

26

39

52

65

78

91

104

117

130

30min Bin #

30min Bin #

3) ACF 2-Factor Model Market/GICS Sector Residuals, normalized by Residual Volatililty

0.02 Autocorrelation Coefficient -0.01 0.00 0.01

-0.02

T -1 day
13

T -2 days
26

T -3 days
39

T -4 days
52

T -5 days
65 30min Bin #

etc
78 91 104 117 130

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. The trading day has 390 minutes, i.e. there are 13 30min bins per day.

Decile Spread High Frequency Trading Portfolios by Day Part

Best risk-adjusted returns with sector & market noise removed. Strongest effect can be seen end-of-day, but that might be artifact of recent momentum regime. Not tradable for investors except HFTs as expect return is less than typical spread.
Black: Return in Bps, Red: Sharpe w 0% RF Rate

Decile Spread Trading Portfolios by Day Part (CAPM Excess Returns)

8 0
10:00

10:30

11:00

11:30

12:00

12:30

13:00 Day Part Bin

13:30

14:00

14:30

15:00

15:30

16:00

Black: Return in Bps, Red: Sharpe w 0% RF Rate

Decile Spread Trading Portfolios by Day Part (2-Factor Model Market/GICS Sector Residual)
8 6

e.g. ~ 1 Bps Return, ~ 2.5 Sharpe

0
10:00

10:30

11:00

11:30

12:00

12:30

13:00 Day Part Bin

13:30

14:00

14:30

15:00

15:30

16:00

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. Using T-1 day lag to build trading portfolios. Holding period is 30mins.

Decile Spread Portfolios are Constantly Recomposed throughout the Trading Day

Correlation Matrix of All-day Returns* Sector-Market Model Periodicity Decile Spread Trading Portfolios on same Trading Day

Low and negative correlations of allday returns of adjacent decile spread portfolios shows that their make-up changes throughout the day. Does this mean that a winning portfolio in one bin becomes a losing one in the next?

10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30 16:00 10:00 1.00 10:30 0.00 11:00 -0.04 11:30 0.08 12:00 0.00 12:30 0.01 0.00 1.00 -0.12 0.12 0.02 0.06 -0.04 -0.12 1.00 -0.01 -0.19 -0.03 0.08 0.12 -0.01 1.00 -0.08 0.04 0.00 0.02 -0.19 -0.08 1.00 0.07 0.01 0.06 -0.03 0.04 0.07 1.00 -0.02 0.07 0.07 0.12 -0.15 0.03 -0.07 -0.08 0.03 0.00 -0.07 -0.01 0.06 0.01 -0.01 0.02 0.06 0.10 -0.08 0.05 -0.07 -0.04 0.03 0.02 -0.04 0.00 0.02 0.00 -0.02 0.05 -0.04 0.10 0.13 0.12 -0.13 0.08 -0.02 0.03 -0.06 -0.03 0.05 0.12

13:00 -0.02
13:30 -0.07 14:00 0.06

0.07
-0.08 0.01

0.07
0.03 -0.01

0.12
0.00 0.02

-0.15
-0.07 0.06

0.03
-0.01 0.10

1.00
0.01 0.04

0.01
1.00 0.03

0.04
0.03 1.00

-0.04
-0.01 0.03

0.01
-0.02 0.07

0.04
0.07 0.00

0.03
0.02 -0.06

14:30 -0.08
15:00 -0.04 15:30 -0.04

0.05
0.00 0.10

-0.07
0.02 0.13

-0.04
0.00 0.12

0.03
-0.02 -0.13

0.02
0.05 0.08

-0.04
0.01 0.04

-0.01
-0.02 0.07

0.03
0.07 0.00

1.00
0.05 -0.02

0.05
1.00 0.01

-0.02
0.01 1.00

0.07
0.11 0.13

16:00 -0.02

0.03

-0.06

-0.03

0.05

0.12

0.03

0.02

-0.06

0.07

0.11

0.13

1.00

* As proxy for similarity of Decile Spread Portfolio Composition

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. Using T-1 day lag to build trading portfolios.

however, Returns do not fully Decay, but show some Staying Power.

Decile Spread Trading Portfolios by Day Part (Cumulative Returns for Remainder of Day)

The effect is not so fleeting as to subsequently fully erase any returns within the same day.
Return Bps

This means there should be value in this Alpha factor for longer duration trades.

10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30 16:00 Day Part Bin

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. Using T-1 day lag to build trading portfolios.

Can the Periodicity Alpha be Utilized to Improve Typical All-day Institutional Orders?
Our experiment: Create hypothetical VWAP orders comprising the entire S&P 500 ($5+) universe for every trading day of our 168 day sample (79,557 hypothetical orders) Long/short mix: Approximately same number of buys and sells every day, randomly assigned each day using binomial distribution Volume forecast: We use trailing 20-day stock-specific volume distribution, smoothed by a 5th order polynomial fit, to build our VWAP curve forecast. This is for both operational (order placement) and economical (market impact) reasons. Our Alpha signal:

Z = (RstockT-1bin - RmarketT-1bin * market - RsectorT-1bin * sector ) / residuals

volatility of residuals is calculated for the entire rolling factor model estimation window of trailing 60 days

Transforming our Alpha Z Score into reweighed VWAP Curves

TradeDummy is 1 for buys, -1 for sells Z = Z * TradeDummy * -1 E.g. we are a VWAP buyer and Z is positive, Zsided will be negative because we want to underweight this expensive bin We want to overweight positive Zsided bins and underweight negative Zsided bins, proportional to the size of Zsided We super-impose our 13 coarse grid Zsided per stock/day on our refined grid of 390 1-min VWAP bins. reweighted = trailingavg * ( (Z Z ) * (1/390) + 1 )) ^ scalefactor reweighted is then rescaled to sum to 1 for every stock/day The VWAP curves are then refitted based on reweighted using a 5th order polynomial
sided sided sided

Example in Visuals: Creating reweighted VWAP Curve. Step 1

Residual Return (=Z Score), original (black) and sided (red)
Residual Return Normalized

Hypothetical all-day VWAP buy order for AAPL on 12/06/2010

-2

-1

30

60

90

120

150

180

210

240

270

300

330

360

390

Buy order & positive residual: Positive Z score flipped to negative Zsided for this bin

Trading Day Minute

Example in Visuals: Creating reweighted VWAP Curve. Step 2

Residual Return (=Z Score), original (black) and sided (red)
2

Hypothetical all-day VWAP buy order for AAPL on 12/06/2010

Residual Return Normalized

-2

-1

30

60

90

120

150

180

210

240

270

300

330

360

390

Trading Day Minute

Trailing Volume Fraction, original (black) and reweighted (red)

0.020

Overweight cheap bins

Volume Fraction

0.005

0.010

0.015

30

60

90

120

150

180

210

240

270

300

330

360

390

Trading Day Minute

Example in Visuals: Creating reweighted VWAP Curve. Step 3

Residual Return (=Z Score), original (black) and sided (red)
2

Hypothetical all-day VWAP buy order for AAPL on 12/06/2010

Residual Return Normalized

-2

-1

30

60

90

120

150

180

210

240

270

300

330

360

390

Trading Day Minute

Trailing Volume Fraction, original (black) and reweighted (red)

0.020

Overweight cheap bins

Volume Fraction

0.005

0.010

0.015

30

60

90

120

150

180

210

240

270

300

330

360

390

Trading Day Minute

Trailing VolumeFractionSmoothed, original (black) and reweighted (red)

0.007

Volume Fraction

Overweight cheap bins (after polynomial fit smoothing)

0.001

0.003

0.005

30

60

90

120

150

180

210

240

270

300

330

360

390

Trading Day Minute

Result: Periodicity Alpha Signal can Enhance S&P 500 VWAP Performance
VWAP Outperformance vs VWAP Volatility

Mean VWAP outperformance Bps

Enhanced mean VWAP outperformance can be traded off for increased volatility, as function of scalefactor, i.e. the extent to which we introduce the signal in building of VWAP curves. Achieves up to 0.4 bps of mean VWAP outperformance in our S&P 500 universe simulation.

0.4

0.5

2000

0.3

1000

0.2

500

0.1

350 200 100

0.0 -0.1
Original VWAP performance not using signal

15

20

25

30

35

40

Stdev Bps vs VWAP

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. Each simulation run is based on 79,557 hypothetical VWAP orders randomly sided.

Additional Simulation Runs: Performance Varies, but Positive Relationship is Robust

VWAP Outperformance vs VWAP Volatility
0.5 Mean VWAP outperformance Bps -0.1 0.0
Original VWAP performance not using signal

A total of 954,684 hypothetical orders, randomly sided. Shows performance gains are not result of some lucky buy/sell distribution of orders.

0.1

0.2

0.3

0.4

15

20

25

30

35

40

Stdev Bps vs VWAP

Universe: S&P 500 stocks $5+, July 2010 Feb 2011. Each simulation run is based on ca 79,557 hypothetical VWAP orders, randomly sided.

Important Disclaimers

Other factors also affect VWAP performance: The study does not factor in any other performance contributors, such as spread costs/savings, market impact, more sophisticated volume forecasting. These vary based on the algorithm provider and can both enhance or deteriorate results. Polynomial fit smoothing will avoid any radical overweighting of specific time bins. This is important as the intent of VWAP is to trade with volume to minimize market impact. Any extreme overweighting of a specific day part could more than offset the positive effect of the alpha signal, especially with larger % ADV orders.

Important Disclaimers (contd)

Suitability: In practice, one would likely allow the user to control the scalefactor, i.e. the extent to which the signal reweights VWAP curves. Users with many orders would be in a better position to stomach the increased volatility to harvest the expected performance gain. Exploiting the signal is not limited to VWAP orders, can be used for Implementation Shortfall orders, tactical child orders, etc. Results from Heston, Korajczyk, Sadka paper indicate performance gains should be greater on smaller caps, but not tested as part of this study.

Areas for Improvement and Further Research

Consider inter-day signal memory: The Periodicity effect has an memory of multiple days. There are some small benefits of incorporating multiple days in the signal via an EWMA process. Consider intra-day spill over: The process works well for 30min bins, but autocorrelation shows up in surrounding bins. Explore effect of spill-over from adjacent bins. Bin structure: Explore equal volume time bins vs. clock time bins. Effect of noise reduction model: Market-Sector works well, but could use PCA model, proprietary risk model. Volume distribution forecasting: Could help reduce volatility of result through more sophisticated volume forecasting (volume PCA decomposition, exchange-specific curves.) Other Alphas: There are other generic alphas as well as those implicit in the portfolio you are trading. This is a subject for another presentation. But I will touch on it briefly

Example other Intraday Alphas Decay Considerations: 10-day Decay Profile 12M Price Momentum vs 5-day CAPM Excess Return.
Return
20

Different Alpha factors have different decay speeds.

12M Price Momentum Factor

Mean P/L Bps

10

15

5 day CAPM Excess Return Factor

Sharpe Ratio

Understanding the Alpha decay profile of your portfolio (i.e. its underlying factor returns) will allow you to trade MUCH more intelligently.

6 # of Lag Days

10

Sharpe
1.2

12M Price Momentum Factor

0.8

5 day CAPM Excess Return Factor

0.4

6 # of Lag Days

10

Universe: S&P 500 stocks $5+, April 2007 May 2011

Thank you

Thor Advisors LLC [email protected]