The Investment Value of Mutual Fund Portfolio Disclosure

1
The Investment Value of Mutual Fund Portfolio
Disclosure

• Russ Wermers
• University of Maryland
• (Joint with Tong Yao and Jane Zhao)
• Presentation at Colorado CFA
• March 17, 2009

2
Prior Studies Find Little Persistence

• Brown and Goetzmann (1995)
• Persistence is mainly due to persistence of poor
  performers
• Carhart (1997) Only Losers Strongly Repeat
• Hendricks, Patel, and Zeckhauser (1993)
• Fund managers have hot hands in year-over-year
  results
• Carhart shows that this is due to momentum
• Bollen and Busse (2002)
• Find persistence at short (quarterly) horizons

3
Stronger Persistence in Mutual Fund Portfolio
Returns (Rather than Net Returns)

• Daniel, Grinblatt, Titman, and Wermers (1997)
• Top quintile fund portfolios beat size, btm, and
  momentum benchmarks by 1 during following year
• Wermers (2003)
• Flow-related buying by funds outperforms by 2-3
  per year for up to 4 years

4
We Should Buy Portfolio, Not Mutual Fund

• Fund expenses too high
• Trading costs probably too high
• Loads and taxes also create hurdles
• Short-term trade fees for some funds
• Disposition effect (Frazzini 2006)
• Funds hold losers too long, sell winners too
  quickly
• Cannot short-sell funds (except ETFs)
• Limits ability to exploit mispricing of funds

5
Dont Buy Entire Portfolio of Winning Managers

• Some stockholdings underperform
• Disposition effect
• Limits in talents
• Tax, momentum, or other reasons for
  buying/selling
• Losing managers should be considered, too
• Adds information about what not to buy

6
Cross-Sectional Aggregation of Portfolio
Ownership Data Adds Power

• Buy only stocks that are overweighted by funds
  with high past alpha
• Underweighted by funds with low past alpha
• Sell stocks underweighted by high alpha funds
• Likely to be less successful, due to binding
  short-sale constraint of most mutual funds

7
Approach of Our Study

• We aggregate ownership to select our long-short
  portfolios
• Many stocks, fewer mutual funds complicates the
  aggregation
• Several aggregation strategies are used
• Simple alpha-weighted strategy
• Stock alpha is weighted-average past fund alphas
• Generalized inverse approach
• Bayesian approach

8
Summary of Results

• Weighted-average alpha (WAA) model performs well
• Long-short portfolio 3 during quarter 1 (above
  size, btm, and momentum benchmarks)
• 2.5 during quarter 2
• 8 compounded return during year 1
• Generalized inverse and Bayesian perform
  similarly
• Using portfolio-weight changes (rather than
  portfolio weights) shows fund buys (not sells)
  are informative
• These profits are not subsumed by 12 quantitative
  strategies documented in past academic studies
  (including momentum, accruals, analyst forecasts,
  etc.)
• WAA predicts operating profitability and works
  better during earnings announcement months

9
A Model of Stock Selection Based on Fund
Portfolio Weights

• Fund alpha is weighted-average stock alpha
• Funds are assumed to have (constant level of)
  persistent performance (all funds have same lt
  1)

?
10

• Past fund alpha is observed with noise
• Combining these three equations

11

• Rearranging,
• where

12

• Across all funds (stacked equations)
• where

13

• And
• Error vector, e, is assumed white noise with
  covariance matrix O

14
Solutions

• OLS/GLS
• However, W matrix is not full rank, since number
  of rows (M funds) is lower than number of columns
  (N stocks)

15
Solution 1 Weighted-Average Alpha

• Assume diagonal matrix W
• Stock alpha is weighted-average fund alpha
  weight measures size of bet, alpha measures
  information precision

16
Solution 2 Generalized-Inverse Alpha

• Let V be the NxN matrix of eigenvectors for
• Let D be the NxN (diagonal) matrix of eigenvalues
• Then,
• Define diagonal matrix, such that

17

• Generalized inverse estimator is then

18
Solution 3 Bayesian Alpha

• Let the prior distribution for stock alpha be
• where
• Combined with our equation,
• Results in

19
Stock Signals Based on Trades

• We substitute weight changes to modify our alpha
  signals
• First, decompose portfolio weights

20

• Weighted-average alpha is then decomposed into
• Generalized inverse and Bayesian estimators are
  decomposed in a similar manner

21
Datasets and Alpha Measurement Methodology
22
U.S. Domestic-Equity, Actively Managed Open-End
Mutual Funds, 1980-2002

• Aggressive Growth, Growth, and Growth Income
  funds
• Semi-annual portfolio holdings (quarterly, in
  many cases) from Thomson Financial
• Net returns, expenses, etc., from CRSP dataset
  (originally created by Mark Carhart)
• Thomson and CRSP linked through MFLINKS of
  Wharton, originally created in Wermers (2000)
• Stock returns from CRSP, accounting data from
  Compustat, analyst earnings forecasts from IBES

23
Measuring Performance at the Net Return Level

• Measure 1 Alpha from Jensen model (MKT)
• Measure 2 Alpha from Fama-French model (MKT,
  SMB, HML)
• Measure 3 Alpha from Carhart model (MKT, SMB,
  HML, UMD)
• Regress time-series of monthly mutual fund excess
  returns on portfolio returns accruing to four
  zero-investment factor-mimicking portfolios
• CRSP value-weighted index less 30-day T-bills
  (MKT)
• Small size minus big size (SMB)
• High book-to-market minus low book-to-market
  (HML)
• High prior-year return less low prior-year return
  (UMD)

24
Summary Statistics (Table 1)
25
Results
26
Test of Performance of Signals

• Compute alpha signal (forecast) for each stock at
  the end of each calendar quarter
• Using 12-month lagged fund alphas
• Form equal-weighted long and short portfolios,
  rebalanced quarterly for Quarters 1 to 4
• Compute DGTW-adjusted returns
• Three versions of signal (1) all fund holdings,
  (2) fund buys only, and (3) fund sells only

27
Holdings-Based Signals
28
Signal 1 Weighted-Average Alpha (EW Decile
Portfolios)
29
Signal 2 Generalized Inverse Alphas (EW
Decile Portfolios)
30
Signal 3 Bayesian Alphas (EW Decile
Portfolios)
31
Results of Three Signals

• The signals successfully pick stocks
• Long-short portfolio outperforms DGTW benchmarks
  by 5-7 during Year 1
• Simple weighted-average alpha model works best
• Underperformance by short portfolio roughly equal
  to outperformance by long portfolio
• Performance is significant during Quarters
  2, 3, and 4 (after portfolio disclosure
  occurs)over 4

32
Trade-Based Signals
33
Weighted-Average Alpha (Fund Buys) (EW Decile
Portfolios)
34
Weighted-Average Alpha (Fund Sells) (EW
Decile Portfolios)
35
Results of Signal Strength Among Lagged Fund Buys
vs. Sells (Since Last Disclosure)

• Buy signal is effective (gt4 during Year 1)
• Sell signal is ineffective (and goes wrong way)
• Skilled managers appear to sell too early, or to
  sell a positive alpha stock to fund an even
  bigger alpha stock also, short-sale constraint
  limits this signal
• Remainder of performance reflects that even
  lagged holdings (3 or 6 months) have investment
  value

36
Further Results (Unreported in Paper)

• Extension 1 Weight stocks by their alpha signal
  (rather than EW)
• Result Year 1 performance gt 8
• Extension 2 Use only funds that disclose
  semi-annually (vs. funds that disclose quarterly)
• Result No discernable difference in Year 1
  returns
• Extension 3 Look at subgroups of stocks
• Result A Slightly better performance for small
  stocks vs. large stocks
• Result B Equal performance for growth stocks vs.
  value stocks
• Result C Much better performance for high vs.
  low breadth-of-ownership stocks
• Result D Equal performance for higher turnover
  stocks vs. low turnover stocks

37

• Extension 4 Look at subgroups of fundsbetter
  stock signal is obtained from
• Smaller funds
• Older funds
• Higher turnover funds
• Lower expense funds
• Higher industry concentration funds
• These results indicate more persistent skills
  among these groups, not necessarily better
  average skills!

38

• Extension 5 use weighted-average t-statistic
  (WAT)
• Slightly lower alpha, but higher t-statistic for
  differenced portfolio

39
A Higher Frequency Look at Disclosure

• Look at weekly (rather than quarterly) returns
  around disclosure date
• Result Holdings signal exhibits interesting
  return pattern, indicating that market already
  reacts to disclosure, but only partially

40
Comparison With Aggregate Trading and Breadth of
Ownership Signals

• Approach
• Regress future quarterly characteristic-adjusted
  stock returns, cross-sectionally, on WAA signal
  Trade signal Breadth signal

41
Stock Returns Regressed on WAA and 12
Quantitative Signals of Past Literature
42
Price-Pressure or Private Information?

• Flows and/or herding do not fully explain the
  success of WAA (it remains significant in
  predicting stock returns)
• WAA predicts changes in accounting ROE

43
(No Transcript)
44
Weighted Average Alpha when Persistence Varies
Across Funds and/or Stocks

• Recall our alpha persistence assumption, where
  all funds have same assumed persistence strength
  (same lt 1) across all stocks
• First, we relax this assumption across funds with
  the following model

?
45

• We model fund j persistence as a function of P
  characteristics at time t,
• Substituting into our weighted-average alpha
  model results in

46

• where
• Therefore, the new signal is
• Unconditional weighted-average alpha, plus
• Fund characteristic-scaled weighted-average alpha
• Actual stock (DGTW) alpha is regressed on these
  two signals each quarter to estimate coefficients

47
Weighted-Average Alphas vs. Fund Characteristics
48
Results for Fund Characteristics

• Total net assets (TNA) negative coefficient
  (lower persistence for large funds)
• Turnover positive coefficient
• Expense ratio negative coefficient
• Fund Age positive coefficient
• Industry concentration positive coefficient

49
Next, Model is Extended to Allow Different
Persistence Across Stocks

• Construct stock alphas, scaled by stock
  characteristics
• Thus, conditional signal is weighted-average
  alpha plus (stock) characteristic-adjusted
  weighted-average alpha

50
Weighted-Average Alphas vs. Stock Characteristics
51
Results for Stock Characteristics

• Size coefficient is negative (funds show
  persistence more strongly in small stocks)
• BTM coefficient insignificant
• Volume-of-trade coefficient is insignificant
• Breadth-of-ownership coefficient is positive
  (i.e., widely held stockscertain large
  stockshave higher persistence)
• Volatility coefficient is negative

52
Conditional Weighted-Average AlphaOut-of-Sample
Performance
53
Summary of Findings

• Mutual fund portfolio holdings present a strong
  stock selection signal
• Important to weight holdings by past fund alphas
• Weighted-average alpha signal outperforms size,
  btm, and momentum by 8 percent during following
  year
• Fund buys are informative, funds sells are not
• Due to binding short-sale constraint
• Signal outperforms 12 quantitative signals by
  more than 2 during Year 1
• Residual potentially represents pure fundamental
  analysis signal