anomalies and news
play

ANOMALIES AND NEWS JOEY ENGELBERG (UCSD) R. DAVID MCLEAN - PowerPoint PPT Presentation

ANOMALIES AND NEWS JOEY ENGELBERG (UCSD) R. DAVID MCLEAN (GEORGETOWN) JEFFREY PONTIFF (BOSTON COLLEGE) 3 RD ANNUAL NEWS & FINANCE CONFERENCE COLUMBIA UNIVERSITY MARCH 8, 2018 Background and Motivation 2 Academic research has uncovered


  1. ANOMALIES AND NEWS JOEY ENGELBERG (UCSD) R. DAVID MCLEAN (GEORGETOWN) JEFFREY PONTIFF (BOSTON COLLEGE) 3 RD ANNUAL NEWS & FINANCE CONFERENCE COLUMBIA UNIVERSITY MARCH 8, 2018

  2. Background and Motivation 2  Academic research has uncovered many predictors of cross-sectional stock returns  E.g., long-term reversal, size, momentum, book-to-market, accruals, and post-earnings drift.  This “anomalies” research goes back to at least Blume and Husick (1973)  Yet 43 years later, academics still cannot agree on what causes this return predictability  Important Question: What explains cross-sectional return predictability?

  3. Theories of Stock Return Predictability 3  Three popular explanations for cross-sectional predictability  Differences in discount rates, e.g., Fama (1991, 1998)  Mispricing, e.g., Barberis and Thaler (2003)  Data-mining, e.g., Fama (1998)  This Paper:  Uses 97 anomalies along with firm-specific news and earnings announcements to differentiate between the three explanations

  4. The Discount Rate Story 4  Cross-sectional return predictability is expected  The predictability may be surprising to academics, but it is not to other market participants  Ex-post return differences reflect ex-ante differences in discount rates  There are no surprises here  Ex-post returns were completely expected by rational investors ex-ante  E.g., Fama and French (1992, 1996)

  5. Discount Rates and News 5 Anomaly Returns around an Earnings Announcement 0.015 0.01 Long Short 0.005 0 -5 -4 -3 -2 -1 0 1 2 3 4 5 -0.005 -0.01 -0.015

  6. Mispricing – Biased Expectations 6  Investors have systematically biased expectations of cash flows and cash flow growth  Expectations are too high for some stocks, too low for others  The anomaly variables are correlated with such expectations  New information causes investors to update their beliefs, which corrects prices, and creates the return-predictability.  Goes to back to at least (Basu, 1977)

  7. Mispricing and News 7 Anomaly Returns around an Earnings Announcement 0.06 Long 0.04 Short 0.02 0 -5 -4 -3 -2 -1 0 1 2 3 4 5 -0.02 -0.04 -0.06

  8. Data Mining 8  As Fama (1991) suggests, academics have likely tested thousands of variables  It’s not surprising to find that some predict returns in -sample  Realization of a “multiple testing bias” in empirical research dates at least back to Bonferroni (1935)  This is stressed more recently in the finance literature by Harvey, Lin, and Zhu (2015).

  9. Mispricing vs. Data Mining 9  Most anomalies focus on monthly returns  Stocks with high (low) monthly returns likely had good (bad) news during the month  A spurious anomaly would therefore likely perform better in- sample on earnings days and news days  Do anomaly strategies still have high returns on news and earnings days after controlling for this?

  10. Our Findings 10  Anomaly returns are higher by  7x on earnings announcement days  2x on corporate news days

  11. Returns in Event Time (3-day window) 11

  12. Financial Analysts 12  We also examine financial analysts’ forecasts errors  For stocks in long portfolios, forecasts are too low  For stocks in the short portfolios, forecasts are too high

  13. Interpretation – Difficult to Reconcile with Risk 13  Hard to tie stock-price reactions to firm-specific news to systematic risk  Anomalies do worse on days when macroeconomic news is announced  Anomalies do worse when market returns are higher, i.e., anomalies have a negative market beta  Risk cannot explain the analyst forecast error results

  14. Interpretation – Not (just) Data Mining 14  A spurious anomaly would likely perform better in- sample on earnings days and news days  However, controlling for contemporaneous monthly return, anomalies still perform better on news days  Out-of-sample anomalies perform better on news days and have the forecast error results  The relation between anomalies and news is stronger in small stocks

  15. Interpretation – Consistent with Mispricing 15  The results are easy to explain with a simple behavioral theory of biased expectations  Expectations are too high for some stocks, too low for others  The anomaly variables are correlated with such expectations  New information causes investors to update their beliefs, which corrects prices, and creates the return-predictability.  The analyst forecast error results fit this framework too

  16. Our Place in the Literature 16 We build on previous studies showing anomalies predict returns on earnings  announcement days  E.g., Chopra Lakonishok and Ritter (1992), La Porta et al. (1994), and Sloan (1996)  Edelen, Kadlec, and Ince (2015) – anomalies and institutions Our paper:   Investigates 6 million news days that are not earnings announcements  Uses 97 anomalies – compare across anomaly types  Relates a large sample of anomalies to analyst forecast errors  Develops new data-mining tests

  17. The Anomalies 17  Choosing the Anomalies  The list is from McLean and Pontiff (2016)  The anomaly has to be documented in an academic study  Primarily top 3 finance journals  Can be constructed with COMPUSTAT, CRSP, and IBES data  Cross-sectional predictors only

  18. The Anomalies 18  97 in Anomalies in Total  Oldest: Blume and Husic (1973)  Stocks sorted each month into long and short quintiles  16 of the 97 variables are binary  Can be replicated with CRSP, COMPUSTAT and I/B/E/S  Average pairwise correlation of anomaly returns is low (.05)

  19. The Sample 19  Earnings announcements from COMPUSTAT  Corporate news from the Dow Jones Archive  Used in Tetlock (2010)  Sample period is 1979-2013  40,220,437 firm-day observations in total

  20. The Sample 20

  21. Aggregate Anomaly Variables 21  We construct 3 aggregate anomaly variables  The variables are the sum of the number of stock i ’s anomaly portfolio memberships in month t  Long, Short, and Net  Net = Long - Short

  22. Aggregate Anomaly Variables 22 Variable Mean Std. Min Max Dev. Long 8.61 5.07 0 35 Short 9.21 5.93 0 45 Net -0.61 6.10 -36 32

  23. The Main Specification 23

  24. Main Specification 24

  25. Economic Magnitudes 25 Annualized Buy and Net = 10 Daily Basis Points Hold Return No Earnings Day 2.59 6.7% Earnings Day 22.39 75.7%

  26. Long and Short Separately 26

  27. Economic Magnitudes 27 Annualized Buy and Long = 10 Daily Basis Points Hold Return No Earnings Day 3.69 9.7% Earnings Day 25.61 90.5% Annualized Buy and Short = 10 Daily Basis Points Hold Return No Earnings Day -1.93 -5% Earnings Day -21.55 -72%

  28. Robustness 28 Are the results related to a day of the week effect (Birru, 2016)?   Controlling for day-of-week does not alter our findings Macroeconomic news (Savor and Wilson, 2016)?   Perhaps firm-specific news reflects systematic risk?  No, anomalies do worse on macro announcement days Endogeneity of news?   Stock return volatility causes news?  We control for daily volatility and nothing changes

  29. Anomaly Types 29  The effects are robust across anomaly types Event – Corporate events, changes in performance, 1. downgrades Fundamental – constructed only with accounting data 2. Market – Constructed only with market data and no 3. accounting data Valuation – Ratios of market values to fundamentals 4.

  30. Analyst Forecast Errors 30  Biased expectations suggests biases in analysts’ earnings forecasts, risk does not  Forecasts should be too low for stocks on the long side of the anomaly portfolios.  Forecasts should be too high for stocks on the short side of the predictor portfolios.

  31. Analysts’ Forecast Error 31

  32. Data Mining Tests 32  A spurious anomaly would likely perform better in-sample on earnings days and news days  Stocks with high (low) monthly returns likely had good (bad) news during the month  Do anomaly strategies still have high returns on news and earnings days after controlling for this?

  33. Data Mining Tests 33

  34. Data Mining Tests – Analyst Forecast Errors 34

  35. Conclusions 35  Evidence of cross-sectional return-predictability goes back at least 43 years to Blume and Husick (1973) – still disagreement over why  In this paper we provide evidence that the cross-section of stock returns is best explained by a cross-section of biased expectations.  Anomaly returns 9x on info days  Anomaly signal predicts analyst forecast errors  Difficult to explain the results with risk  Harder to rule out data mining, but it does not seem to explain the full effects

Recommend


More recommend