Transcript David Mclean

ANOMALIES AND NEWS
JOEY ENGELBERG (UCSD)
R. DAVID MCLEAN (GEORGETOWN)
JEFFREY PONTIFF (BOSTON COLLEGE)
11TH ANNUAL HEDGE FUND CONFERENCE
DECEMBER 8, 2016
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

See the 2013 Nobel Prize
Important Question: What explains cross-sectional return predictability?
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
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)
Discount Rates and News
5
Anomaly Returns around an Earnings Announcement
0.015
0.01
Long
Short
0.005
0
-5
-0.005
-0.01
-0.015
-4
-3
-2
-1
0
1
2
3
4
5
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)
Mispricing and News
7
Anomaly Returns around an Earnings Announcement
0.06
Long
0.04
Short
0.02
0
-5
-0.02
-0.04
-0.06
-4
-3
-2
-1
0
1
2
3
4
5
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).
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 insample on earnings days and news days

Do anomaly strategies still have high returns on news and
earnings days after controlling for this?
Our Findings
10

Anomaly returns are higher by
 7x
on earnings announcement days
 2x
on corporate news days
Returns in Event Time (3-day window)
11
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
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
Interpretation – Not (just) Data Mining
14

A spurious anomaly would likely perform better insample 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
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
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
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
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)
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
The Sample
20
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,
 Net
Short, and Net
= Long - Short
Aggregate Anomaly Variables
22
Variable
Mean
Std. Dev.
Min
Max
Long
8.61
5.07
0
35
Short
9.21
5.93
0
45
Net
-0.61
6.10
-36
32
The Main Specification
23
Main Specification
24
Economic Magnitudes
25
Net = 10
No Earnings Day
Earnings Day
Annualized Buy and
Daily Basis Points
Hold Return
2.59
6.7%
22.39
75.7%
Long and Short Separately
26
Economic Magnitudes
27
Long = 10
No Earnings Day
Earnings Day
Annualized Buy and
Daily Basis Points
Hold Return
3.69
9.7%
2.56
90.5%
Short = 10
No Earnings Day
Earnings Day
Annualized Buy and
Daily Basis Points
Hold Return
-1.93
-5%
-19.62
-72%
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
Anomaly Types
29

The effects are robust across anomaly types
1.
Event – Corporate events, changes in performance,
downgrades
2.
Fundamental – constructed only with accounting data
3.
Market – Constructed only with market data and no
accounting data
4.
Valuation – Ratios of market values to fundamentals
Robust Across Anomaly Types
30
Analyst Forecast Errors
31

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.
Analysts’ Forecast Error
32
Data Mining Tests
33

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?
Data Mining Tests
34
Data Mining Tests – Analyst Forecast Errors
35
Conclusions
36

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