Quantitative Stock Selection

1
Quantitative Stock Selection

• James F. Page III, CFA
• May 2005

2
Project Summary

1. Why Quant Selection is Attractive
2. Methodology
3. Historical Back Testing
4. Model Results
5. Dynamic Weights / Regime Change
6. Benchmarks
7. Next Generation Models
8. Concluding Thoughts

3
I. Quantitative Stock Selection
4
Quant Stock Selection

• Premise
• In aggregate, certain fundamental, expectational,
  and macro variables may contain valuable
  information in predicting stock returns
• Not unlike traditional fundamental analysis, just
  more systematic

5
Quant Stock Selection

• Pros
• Anecdotal evidence suggests 80 of stock picking
  is done by hand (individuals making calls on
  fundamentals)
• Relies heavily on talent (or luck) of individual
  analyst
• Individuals can only process so much information
  (sector focus)
• Human nature suggests cognitive biases likely
• Market structure may perpetuate mis-pricings
  (Street incentives, value weighted benchmarks,
  short sale restrictions)
• Little academic research on subject (trade rather
  than publish)
• Evidence suggests that investors systematically
  over pay for growth
• Quantitative selection is scaleable

6
Quant Stock Selection

• Cons
• Black box nature of model
• Explain approach without revealing too much
  information
• Attribution analysis must be able to explain
  performance
• Protecting against common modeling errors
• Credibility of simulated results
• Adapting to individual client restraints

7
Quant Stock Selection

• Market Neutral
• Generate returns from both undervalued and
  overvalued stocks
• At present, high market valuation low future
  returns
• Market exposure is commodity but good stock
  selection is valued (higher management fees)
• Low return expectations combined with
  geo-political environment suggests absolute
  return approach prudent

8
II. Methodology
9
Methodology

• Hypothesize
• Develop candidate list of potential factors that
  may assist in predicting stock returns
  (valuation, growth, etc.)
• Priors reduce data mining
• Back Test
• Decide on universe for testing (capitalization,
  index, sector, etc)
• Use sorting or regressions to test individual
  candidate variables
• FactSets AlphaTester currently available to Duke
  students
• Rebalance
• Periodically rebalance portfolios (monthly,
  annually, etc.)

10
Methodology

• Analyze Results
• Consider factor performance and consistency (both
  long and short candidates) in predicting returns
  balanced against turnover
• Select most promising factors for inclusion in
  the model
• Weight
• Once individual factors selected must decide on
  weights for final model by either
• Eye balling best factors and assigning weights
  for a scoring model
• Pushing individual factor portfolios into a
  mean-variance optimizer

11
III. Historical Back Testing
12
Historical Back Testing

• Access to reasonably accurate historical data is
  costly
• FactSets AlphaTester is currently available to
  Duke students
• Two approaches common in practice
• Regression of factors on security returns (Panel,
  etc.)
• Sorting universe into fractiles based on factor
  characteristics (AlphaTester)
• Must protect against common modeling errors
• Survivorship bias
• Information / reporting lags
• Data mining
• Inaccuracies in data
• Credibility of simulated returns is critical

13
Historical Back Testing

• Term 3 Model Discredited
• Errors in Historical Returns
• Scrub Example.xls
• Survivorship Bias
• Difficult to rule out unless you spend a lot of
  time examining results
• Fractile Misspecification
• MSFT grouped in F1 Div Yield for 85-04 because of
  Special Dividend
• Betas not believable
• Subject to similar errors as returns information
• Makes market neutral simulation difficult
• Combing factors into comprehensive model
  increases complexity

14
Historical Back Testing

• To Mitigate Potential Errors
• Universe Selection is critical component
• Market Cap weighted
• Adds to turnover (98-00)
• Unstable sector allocations
• Less undervalued firms to buy
• Revenue weighted
• Sector bias
• Less overvalued firms to sell
• Actual Indices (Preferred method)
• Limit universe to actual benchmark
• Limit survivorship bias
• Historical indices available (but option not
  turned on for Duke)
• Greatly enhance credibility look to acquire for
  next years class

15
Historical Back Testing

• To Mitigate Potential Errors
• Factor Syntax
• If you do not get this right data is worthless
  (lots of opportunities to get it wrong)
• Consider consolidating our approved syntax for
  future students as starting point
• Expectational (instead of accounting/fundamental)
  produced significantly fewer errors
• Survivorship Bias
• Selecting Research Companies does not protect
  without
• Appropriate Syntax on Factors
• Correct specification of Universe
• Sanity checks on early period companies
• of NA companies can be signal
• Errors
• You must clean historical data
• Consider median returns as back of envelope
  option

16
Historical Back Testing

• Recommendations
• Use historical indices as universe
• SP 500
• Barra 1000
• Start with approved list of factor syntax
• Clean historical results (particularly returns)
• Do not rely on betas to construct market neutral
  portfolio
• Research ways to limit reliance on AlphaTester
• Look for other data providers ask managers what
  they use
• Interface with CompuStat/IBES directly?
• Once comfortable with model, begin sorting real
  time ASAP

17
IV. Model Results
18
Model Results

• Desired Universe SP 500
• Why
• Considered to be highly efficient
• Value weighted index suggests low hanging fruit
• Historical data for testing is plentiful and
  reasonably accurate
• Highly liquid (market impact costs and borrow)
• Very scaleable because of market capitalizations
• Actual Universe
• First choose US Companies with highest sales (
  500)
• Had to switch to Market Cap because of data
  limitations

19
Model Results

• Universe Comments
• Unstable during bubble period (1998-2000)
• Less undervalued firms to buy (but more
  overvalued firms to sell)
• Sector allocations float with market sentiment
• Other
• Rebalanced official results annually due to
  time consuming nature of cleaning returns
• Equal number of companies in each bucket
• Equal weight returns
• Did not impose sector constraints
• Included two groups of Factors Fundamental and
  Expectational
• Actively looking for Quality factor to add to
  the model
• Assume beta exposure is equal is both
  portfolios probably conservative
• Results seem too good further cleaning
  necessary

20
Model Results
Individual Factor Performance Monthly Statistics
1989 2004 Long Factors correlated with Value
and visa versa View Portfolios
21
Model Results
Fixed Weighting Scheme
22
Model Results
Scoring Model Heat Map
23
Model Results
Summary Statistics
24
V. Dynamic Weights / Regime Change
25
Dynamic Weights / Regime Change

• A factors effectiveness may vary in different
  states of nature (PE ratios impacted by
  inflation)
• Certain market / macro conditions may favor
  growth or value (value was dog in late 1990s)
• Dynamic factor weights allow model to capitalize
  on conditional information
• Few managers currently employ dynamic weighting
  schemes
• This area is the Holy Grail of Quant Strategies

26
Dynamic Weights / Regime Change

• Forecasting Regime Change
• Inflection point for style (growth or value)
  relative performance
• Used SP 500 Barra Value and Growth Indices as
  Proxies
• Examined macro economic variables that might
  assist in forecasting these inflection points
• Two variables demonstrated promise in
  forecasting style relative performances over the
  following year

27
Dynamic Weights / Regime Change

• Regime Change Factor 1

28
Dynamic Weights / Regime Change

• Regime Change Factor 2

29
Dynamic Weights / Regime Change

• The Same Can Be Applied to View Portfolios
• Expectational Factor 2 and Regime Change Factor
  1
• Prediction of Long outperforming Short

30
Dynamic Weights / Regime Change

• The Same Can Be Applied to View Portfolios
• Expectational Factor 2 and Regime Change Factor
  2
• Prediction of Long Outperforming Short

31
VI. Benchmarks
32
Benchmarks

• Value or Equal Weight?
• Since 1990, EWI has outperformed by 177 basis
  points
• Turnover for EWI is 6x which begs the question
• Can we separate turnover between model signals
  and weighting scheme?

33
Benchmarks

• Value or Equal Weight?
• Significant Implications for Sector Weights /
  Tracking Error

34
Benchmarks

• Value or Equal Weight?
• Correlations drift through time implications
  for tracking error

35
Benchmarks

• Value or Equal Weight?
• EWI had positive loading on the size premium
• EWI has significant exposure to the value premium
• Fama-French Risk Factor Exposures

Source http//mba.tuck.dartmouth.edu/pages/facul
ty/ken.french/data_library.html
36
Benchmarks

• Value or Equal Weight?
• EWI has 82 correlation with 500 / Barra Growth
• EWI has 96 correlation with 500 / Barra Value
• Further proof of value tilt

37
Benchmarks

• Value or Equal Weight?
• Obvious Pros and Cons to both
• EWI benchmark will make returns look less
  impressive, but help explain turnover
• EWI may be a better match for style
• Provide more stable weighting for sector
  allocations
• Equal weight is newer idea historical data is
  limited
• If possible, choice should match weighting scheme
  of portfolio

38
VII. Next Generation Models
39
Next Generation Models

• Refining Dynamic Factor Weights
• Preferably done outside of FactSet
• Migration Tracking
• May contain information to enhance returns or
  limit turnover

40
Next Generation Models

• Modified Versions of SP 500 Model
• Separate Models for Sector and Stock Selection
• More Conservative
• More positions
• Limited tracking error
• More Aggressive
• Directional
• Less positions
• Leverage
• Other Domestic Models
• SP Mid-Cap 400 / Russell 2000
• International Models
• Developed / Emerging markets

41
VIII. Concluding Thoughts
42
Concluding Thoughts

• Theoretical
• How long will excess returns exist
• How to stay ahead of the curve
• Implementation
• Cost of data
• Credibility of simulation
• Returns during first 12 24 months
• Balance between turnover and model signals

43
Concluding Thoughts

• Overall
• Quantitative Stock Selection Appealing
• Outperformance Seems Possible
• Long/Short Consistent with Absolute Return
  Approach

44
Bio

• James F. Page III
• Jimmy became interested in quantitative stock
  selection during Campbell Harveys Global Asset
  Allocation and Stock Selection class and a
  follow-up course dedicated to quantitative stock
  selection. He received his Bachelor of Science
  degree from the University of Florida and will
  receive his MBA from Duke Universitys Fuqua
  School of Business in May 2005. Prior to
  enrolling at Duke, he spent four years in the
  Equity Research Department of Raymond James
  Associates in St. Petersburg, FL. He is also a
  CFA charter holder.