Documente Academic
Documente Profesional
Documente Cultură
Paul Gompers
pgompers@hbs.edu
hb d
Robin Greenwood
R bi G d
rgreenwood@hbs.edu
Josh Lerner
jlerner@hbs edu
jlerner@hbs.edu
Spring 2009
Spring 2009
Gompers/Greenwood/Lerner 2009
Good Empirical Research
1. Answers a focused and well specified question
2. Rely on the simplest techniques possible required to
establish support for your hypothesis
3. Is well argued and well written
• This course aims to help you with 1 and 2 but mainly 2.
• It takes time to build up the ability to ask good questions,
but it is easier when you have an expanded toolkit.
Gompers/Greenwood/Lerner 2009
Administration
• Papers are listed on the syllabus
• We’ll print and bring to class at least one week in
advance.
• If you would like to schedule an appointment to meet
with any of us, please email Peggy
h f l l
(pmoreland@hbs.edu).
Gompers/Greenwood/Lerner 2009
Course Requirements
• Four referee reports over course of semester
• 5‐15 minute paper proposal and presentation at the end
of the semester
• If you audit the course, we hope you still consider doing
the paper proposal and presentation
h l d
• While not required, we recommend that you try to
replicate some of the main results in papers we cover in
class particularly if in a topic that interests you
class, particularly if in a topic that interests you.
Gompers/Greenwood/Lerner 2009
Schedule
Date Topic (s) Instructor
5 F b Long-run
5-Feb L returns R bi
Robin
Gompers/Greenwood/Lerner 2009
Other Useful Resources
Gompers/Greenwood/Lerner 2009
Other Useful Resources
Gompers/Greenwood/Lerner 2009
Lecture 1: Event Studies
Lecture 1: Event Studies
Empirical Methods in Corporate Finance
Empirical Methods in Corporate Finance
Robin Greenwood
January 2009
Gompers/Greenwood/Lerner 2009
Event Studies
• “Analysis of whether there was a statistically significant
reaction in financial markets to past occurrences of a
given type of event that is hypothesized to affect firms’
i t f t th t i h th i d t ff t fi ’
market values.”
• Depending on the application, can be used as a measure
Depending on the application can be used as a measure
of the information content of an event (under market
efficiency assumption), or as evidence of irrationality (if
efficiency assumption), or as evidence of irrationality (if
no news, or statistically significant drift post‐
announcement).
• Many famous results in finance derived from event
studies…
Gompers/Greenwood/Lerner 2009
Entrance into the S&P500 Index
Shleifer (1986); Harris and Gurel (1986)
3.5
2.5
%)
Percent Return (%
1.5
0.5
0
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6
-0.5
05
Days from Event
Gompers/Greenwood/Lerner 2009
Bernard and Thomas 1989
(post‐earnings announcement drift)
Stocks that have positive earnings surprises continue to outperform afterwards…
Gompers/Greenwood/Lerner 2009
Mikkelson and Ruback (1991)
Targeted Repurchases and Common Stock Returns
Gompers/Greenwood/Lerner 2009
Fama, Fisher, Jensen, Roll (1969)
Stock Splits
40 Cumulative
abnormal
35
return %
30
25
20
15
10
5
0
Month relative to split
p
Gompers/Greenwood/Lerner 2009
Huberman and Regev (2001)
Gompers/Greenwood/Lerner 2009
A rose.com…
Gompers/Greenwood/Lerner 2009
Valuable tool
• If market rationality, impact of news on stock price
should allow us to assess impact on firm value of:
– Mergers and acquisitions.
– Financing decisions.
– Public policy choices.
Public policy choices
• Extensive use in litigation and policy discussions.
Gompers/Greenwood/Lerner 2009
Market Efficiency: Reaction to news
Early Reaction
Delayed Reaction
Days relative to announcement day
‐t 0 +t
The timing for a positive news event
Gompers/Greenwood/Lerner 2009
Long history…
• First study: Dolley [1933] on stock splits.
• Refinement of technique by late 1960s.
• Formal testing of alternative methods by Brown and
Warner [1980, 1985].
• Gradual extension into industrial organization, other
economics literature.
Gompers/Greenwood/Lerner 2009
Lecture Plan
• Event Study Design
• Simple example
• Standard significance tests (see CLM, Chapter 4)
• Tips on arranging data for effective use in event studies
• Measures of Abnormal Returns
• Cross‐sectional models
– Simple example
• Econometric concerns and more examples
– Brown and Warner 1 and 2
– Bittlingmayer and Hazlett
– Greenwood – Returns are correlated for a single event
– Simulating Standard Errors – Barberis, Shleifer and Wurgler
Gompers/Greenwood/Lerner 2009
Next Class‐ Long Horizon Returns
• Benchmark adjustment becomes more important
• Overlap and clustering of windows becomes important
– Calendar time regressions
• Statistical properties of the estimators becomes
important
– BHARs versus Wealth Ratios versus CARs
Gompers/Greenwood/Lerner 2009
Crucial elements of any event study
• Event dates.
• Size of “event window”:
– Will reflect possibility of leakage, precision of event dating.
• Actual return.
• Expected return.
• Measure of variance.
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Step 1: Arrange your data in event‐time
Stock 1: IBM Stock 2: MSFT
Date Event day
Event day Return (%)
Return (%) Date Event day
Event day Return (%)
Return (%)
12/26/2008 ‐5 4.95 8/20/2008 ‐5 ‐4.63
12/29/2008 ‐4 1.13 8/23/2008 ‐4 4.78
12/30/2008 ‐3 1.38 8/24/2008 ‐3 ‐0.39
12/31/2008 ‐2 2.34 8/25/2008 ‐2 2.93
1/1/2009 ‐1 1.92 8/26/2008 ‐1 3.82
1/2/2009 0 7.79 8/27/2008 0 3.72
1/5/2009 1 ‐1.75 8/30/2008 1 0.03
1/6/2009 2 2.24 8/31/2008 2 1.31
1/7/2009 3 2.40 9/1/2008 3 3.21
1/8/2009 4 ‐1.48 9/2/2008 4 4.42
1/9/2009
/ / 5 ‐0.88 9/3/2008
/ / 5 ‐2.16
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Step 2: Make risk adjustment
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Step 2: Make risk adjustment
After this, sometimes helpful to plot the data
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Step 3: Cumulate over desired window
Stock 1: IBM
Stock 1: IBM Stock 2: MSFT
Stock 2: MSFT
Date Event day Return (%) Date Event day Return (%)
12/26/2008 ‐5 2.67 8/20/2008 ‐5 4.60
12/29/2008 ‐4 1.11 8/23/2008 ‐4 13.01
12/30/2008 ‐3 3.18 8/24/2008 ‐3 3.26 Stock 1: IBM Stock 2: MSFT
12/31/2008 ‐2 11.29 8/25/2008 ‐2 ‐0.01
1/1/2009 ‐1
1 ‐5
5.55
55 8/26/2008 ‐1
1 ‐0
0.41
41
1/2/2009 0 10.77 8/27/2008 0 5.55 CAR [‐1,0] 5.22 5.14
1/5/2009 1 ‐10.58 8/30/2008 1 ‐3.43
1/6/2009 2 1.39 8/31/2008 2 ‐6.62
1/7/2009 3 ‐4.92 9/1/2008 3 ‐0.82
1/8/2009 4 0.31 9/2/2008 4 ‐2.79
/ /
1/9/2009 5 0.09 9/3/2008
/ / 5 ‐4.86
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Step 4: Can test for significance of Average CAR, or individual CARs
Stock 1: IBM
Stock 1: IBM Stock 2: MSFT
Stock 2: MSFT
Date Event day Return (%) Date Event day Return (%)
12/26/2008 ‐5 2.67 8/20/2008 ‐5 4.60
12/29/2008 ‐4 1.11 8/23/2008 ‐4 13.01
12/30/2008 ‐3 3.18 8/24/2008 ‐3 3.26 Stock 1: IBM Stock 2: MSFT
12/31/2008 ‐2 11.29 8/25/2008 ‐2 ‐0.01
1/1/2009 ‐1
1 ‐5
5.55
55 8/26/2008 ‐1
1 ‐0
0.41
41
1/2/2009 0 10.77 8/27/2008 0 5.55 CAR [‐1,0] 5.22 5.14
1/5/2009 1 ‐10.58 8/30/2008 1 ‐3.43
1/6/2009
1/7/2009
2
3
1.39
‐4.92
8/31/2008
9/1/2008
2
3
‐6.62
‐0.82
Take Average
1/8/2009 4 0.31 9/2/2008 4 ‐2.79
/ /
1/9/2009 5 0.09 9/3/2008
/ / 5 ‐4.86
5.18%
Gompers/Greenwood/Lerner 2009
Testing for Statistical Significance
Pre‐event window Event Post‐event window
t‐k t t+j
• Objective is to reject the null hypothesis that the average
return is zero.
• Reminder:
– Power = probability that test will reject a false null hypothesis =
Probability that it will not make a type II error.
b bl h ll k
– Test size = Significance Level = Type 1 error = Probability of
rejecting a true null hypothesis
rejecting a true null hypothesis
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Test for statistical significance for a single event
1. Calculate variance of residual returns on pre
Calculate variance of residual returns on pre‐event
event and event
and event‐window
window
2. T‐test for one‐day event = Abnormal Return / sqrt(estimated variance)
3. T‐test for k‐day event = CAR / sqrt(k*estimated residual variance of daily returns)
Application: Litigation, case studies, assessing impact of individual news announcements
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Test for statistical significance for a single event
1. Calculate variance of residual returns on pre
Calculate variance of residual returns on pre‐event
event and event
and event‐window
window
2. T‐test for one‐day event = Abnormal Return / sqrt(estimated variance)
3. T‐test for k‐day event = CAR / sqrt(k*estimated residual variance of daily returns)
Related technique: Estimate a regression of R = a + bRm + cEventDummy
Less useful for long window events.
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Test for statistical significance for multiple events
CAR (τ 1 ,τ 2 )
t= 1
⎡σˆ 2 (τ 1 ,τ 2 ) ⎤ 2
⎣ ⎦
1
σˆ 2 (τ 1 ,τ 2 ) = 2 ∑ i =1σˆ i2 (τ 1 ,τ 2 )
N
N
⇒
N ⋅ CAR (τ 1 ,τ 2 )
t=
∑ σ (τ 1 ,τ 2 )
N 2
i =1
ˆ i
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Test for statistical significance for multiple events
CAR (τ 1 ,τ 2 )
t= 1
⎡σˆ 2 (τ 1 ,τ 2 ) ⎤ 2
⎣ ⎦
1
σˆ 2 (τ 1 ,τ 2 ) = 2 ∑ i =1σˆ i2 (τ 1 ,τ 2 )
N
N
⇒
N ⋅ CAR (τ 1 ,τ 2 ) Remember to adjust variance estimate
t=
f k h l
for k, the length of
h f
∑ σ (τ 1 ,τ 2 )
N 2
ˆ
i =1 i the event window!
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Test for statistical significance for multiple events
CAR (τ 1 ,τ 2 )
t= 1
Note that this is not the same
Note that this is not the same
⎡σˆ 2 (τ 1 ,τ 2 ) ⎤ 2 As the cross‐sectional
⎣ ⎦ Standard error of the CARs
1
σˆ 2 (τ 1 ,τ 2 ) = 2 ∑ i =1σˆ i2 (τ 1 ,τ 2 ) because we use pre‐event
N
N d t t i f
data to inform our estimate
ti t
⇒ of σ!!!
N ⋅ CAR (τ 1 ,τ 2 ) But cross-sectional standard
t=
∑ σ (τ 1 ,τ 2 ) error won’t be too far off in
N 2
i =1
ˆ i
most cases.
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Test for statistical significance for multiple events
CAR (τ 1 ,τ 2 )
t= 1
…And should be exactly the
…And should be exactly the
⎡σˆ 2 (τ 1 ,τ 2 ) ⎤ 2 same if the residual standard
⎣ ⎦ deviation is the same across
1
σˆ 2 (τ 1 ,τ 2 ) = 2 ∑ i =1σˆ i2 (τ 1 ,τ 2 ) securities
N
N
⇒
N ⋅ CAR (τ 1 ,τ 2 )
t=
∑ σ (τ 1 ,τ 2 )
N 2
i =1
ˆ i
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Test for statistical significance for multiple events
CAR (τ 1 ,τ 2 )
t= 1
…One common criticism of
…One common criticism of
⎡σˆ 2 (τ 1 ,τ 2 ) ⎤ 2 event studies has been that
⎣ ⎦ the variance of returns
1
σˆ 2 (τ 1 ,τ 2 ) = 2 ∑ i =1σˆ i2 (τ 1 ,τ 2 ) around the event is higher, in
N
N which case the t‐stat will be
hi h th t t t ill b
⇒ overstated.
N ⋅ CAR (τ 1 ,τ 2 )
t= Cross‐sectional standard
∑ σ (τ 1 ,τ 2 )
N 2
ˆ error partially helps here.
i =1 i
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
Cross‐sectional estimator of the variance (Boehmer et al 1991, Page 167 CLM):
∑ ( )
1 2
Var[CAR(τ 1 ,τ 2 )] = CARi (τ 1 ,τ 2 ) − CAR(τ 1 ,τ 2 )
N
N2 i =1
…Consistent if abnormal returns are uncorrelated in the cross‐section
Cross‐sectional homoskedasticity not required!
Cross‐sectional homoskedasticity not required!
Benefit of this approach: Very easy to implement, even in Excel !!
Gompers/Greenwood/Lerner 2009
Fairly Typical Event Study Design:
N non‐overlapping events (e.g., Stock Splits, Index adds)
Pre‐event window Event Post‐event window
t‐k t t+j
One more technique is to aggregate the individual scaled returns:
CARi
SCARi =
σi
1/2
⎛ N ( L1 − 4) ⎞
J2 = ⎜ ⎟ SCARi ~ N (0,1).
⎝ 2L − 4 ⎠
…CLM 4.4.24
Relevant if CAR variance is higher for high CAR events
In practice, I don’t see this very often, stick with simpler method
Gompers/Greenwood/Lerner 2009
Coding tips for doing event studies
• SAS
– Steps to compute event‐time
• Start with panel
Start with panel‐>Merge
>Merge event dates
event dates‐>Compute
>Compute Event time
Event time
• Transpose data in event time and dump to STATA for easy use
• Matlab
– Arrange your data in an event‐time matrix
• Stata/Eventus
– Eventus can produce a file that already identifies your returns in
can produce a file that already identifies your returns in
event time
• Excel
– SSuitable only for studies with small samples and for getting to know
i bl l f di ih ll l df i k
your data
• WARNINGS
– Watch out for events with duplicate stocks (ie, acquirer buys twice)
Gompers/Greenwood/Lerner 2009
Measures of Abnormal Returns
• Classics
– Mean Return model : AR(I,t)= R(I,t)
– Market adjusted model: AR(I,t)= R(I,t)‐RM(t)
– Market and risk adjusted model: AR(I,t)=R(I,t)‐Beta*RM(t)
• For this model: pre‐estimate using pre‐event data to get the
For this model: pre estimate using pre event data to get the
factor exposure
• Then on event, adjust using estimated beta
j g
– Fama‐French risk adjusted model:
AR(I,t)=R(I,t)‐β1*RM(t)‐β2*HML(t) –β3*SMB(t)
– Matched Portfolio Approach: AR(I,t)= R(I,t)‐RMatch(I,t)
– If
If your results using daily returns are sensitive to choice of
l i d il ii h i f
model, you should rethink your topic!
Gompers/Greenwood/Lerner 2009
Cross‐sectional models
• Often we are interested in whether there is an
association between abnormal returns and
characteristics of the event in question
characteristics of the event in question
– Earnings announcement returns and the type of surprise (i.e,
positive, negative)
– Acquirer announcement returns and the size of the premium
paid to target firm shareholders
– Asquith and Mullins (1986): Two‐day cumulative returns for the
q ( ) y
announcement of equity offerings is regressed on the size of
the offering as a percentage of the total equity of the firm.
• As long as events do not overlap, can usually estimate
OLS with White (1980) standard errors to
heteroskedasticity (e.g., “,Robust” in stata)
– CAR(i) = a + bX+ epsilon
Gompers/Greenwood/Lerner 2009
Cross‐sectional model example…
Earnings
Surprise: ‐1.00 ‐0.75 ‐0.50 ‐0.25 0.00 0.25 0.50 0.75 1.00
Really Really Really Really
bad bad Bad Bad Good Good Good Good
Description news news News News No news
No news News News News News
Event Date:
‐5 0.29 ‐1.52 ‐3.93 ‐3.40 ‐4.55 ‐4.58 ‐0.73 2.90 ‐3.30
‐4 2.86 4.34 ‐4.07 3.77 2.33 1.25 4.79 4.35 2.82
‐3 3.41 ‐0.54 2.03 2.27 ‐1.35 0.99 ‐3.25 4.26 1.87
‐2 0.75 ‐2.50 ‐1.98 2.69 3.09 4.68 4.98 1.83 ‐4.32
‐1 ‐2.50 0.39 ‐2.59 ‐0.48 4.85 3.04 3.19 ‐2.37 ‐4.52
0 ‐4.40
4.40 ‐6.65
6.65 ‐4.81
4.81 ‐0.05
0.05 ‐0.93
0.93 1.15 1.18 7.06 ‐1.94
1.94
1 2.09 3.85 0.46 3.69 0.78 ‐4.53 3.25 0.28 ‐4.10
2 ‐4.44 ‐3.51 1.29 0.55 ‐1.54 ‐3.66 3.84 1.74 0.50
3 ‐1.48 ‐2.52 4.58 ‐2.76 2.29 1.41 4.18 ‐3.25 1.36
4 0 80
0.80 2 69
2.69 4 69
4.69 ‐2.27
2 27 4 97
4.97 3 23
3.23 ‐2.58
2 58 4 16
4.16 3 84
3.84
5 3.41 ‐3.97 ‐3.74 1.34 1.55 ‐4.80 ‐3.09 0.42 ‐3.94
Gompers/Greenwood/Lerner 2009
Cross‐sectional model example…
Earnings
Surprise: ‐1.00 ‐0.75 ‐0.50 ‐0.25 0.00 0.25 0.50 0.75 1.00
Really Really Really Really
bad bad Bad Bad Good Good Good Good
Description news news News News No news
No news News News News News
Event Date:
‐5 0.29 ‐1.52 ‐3.93 ‐3.40 ‐4.55 ‐4.58 ‐0.73 2.90 ‐3.30
‐4 2.86 4.34 ‐4.07 3.77 2.33 1.25 4.79 4.35 2.82
‐3 3.41 ‐0.54 2.03 2.27 ‐1.35 0.99 ‐3.25 4.26 1.87
‐2 0.75 ‐2.50 ‐1.98 2.69 3.09 4.68 4.98 1.83 ‐4.32
‐1 ‐2.50 0.39 ‐2.59 ‐0.48 4.85 3.04 3.19 ‐2.37 ‐4.52
0 ‐4.40
4.40 ‐6.65
6.65 ‐4.81
4.81 ‐0.05
0.05 ‐0.93
0.93 1.15 1.18 7.06 ‐1.94
1.94
1 2.09 3.85 0.46 3.69 0.78 ‐4.53 3.25 0.28 ‐4.10
2 ‐4.44 ‐3.51 1.29 0.55 ‐1.54 ‐3.66 3.84 1.74 0.50
3 ‐1.48 ‐2.52 4.58 ‐2.76 2.29 1.41 4.18 ‐3.25 1.36
4 0 80
0.80 2 69
2.69 4 69
4.69 ‐2.27
2 27 4 97
4.97 3 23
3.23 ‐2.58
2 58 4 16
4.16 3 84
3.84
5 3.41 ‐3.97 ‐3.74 1.34 1.55 ‐4.80 ‐3.09 0.42 ‐3.94
Gompers/Greenwood/Lerner 2009
Plot the data
8.00
6.00
4.00
2.00
0.00
urns
Retu
-1.50
1 50 -1.00
1 00 -0.50
0 50 0 00
0.00 0 50
0.50 1 00
1.00 1 50
1.50
-2.00
-4.00
-6.00
-8.00
Earnings Surprise
Gompers/Greenwood/Lerner 2009
Estimate Regression
8.00
6.00
R = 4.28 EARNSURPRISE - 1.04
4.00
2.00
0.00
0 00
Returnss
-4.00
4 00
-6.00
-8
8.00
00
Earnings Surprise
Gompers/Greenwood/Lerner 2009
Interpretation
• Often the intention is to explain an average effect, ie,
stocks that go into the S&P 500 go up in price because
people forecast higher earnings.
l f t hi h i
• Here, the initial puzzle is the average abnormal return,
and the cross sectional relation can help explain it
and the cross‐sectional relation can help explain it
• Careful, because oftentimes the constant term still
remains significant in the regression suggesting that you
remains significant in the regression, suggesting that you
haven’t explained away the effect
Gompers/Greenwood/Lerner 2009
Brown and Warner 1980
Brown and Warner 1980
“Measuring Security Price
Performance” [JFE, 1980].
Gompers/Greenwood/Lerner 2009
BW1980: Looks at monthly return data
• Construct 250 samples of 50 “events”:
– Random monthly observations of different securities between
1944 d 1971
1944 and 1971.
– Add amount of abnormal performance (from 0% to 5%) in event
month.
• Subtract 1% of abnormal performance from months in (‐
89,+10) interval to insure zero overall abnormal
performance.
f
Gompers/Greenwood/Lerner 2009
BW 1980: Mean adjusted return
• Compute normalized return in month 0:
1 −11
Ki = ∑
79 t = −89
Rit
1/ 2
⎡1 −11
2⎤
σ ( Ri ) = ⎢
78
∑
⎣ t − −89
( Rit − K i ) ⎥
⎦
Ait = ( Rit − K i ) / σ ( Ri )
Gompers/Greenwood/Lerner 2009
BW 1980: Mean adjusted return (2)
• t‐test looks at whether A (the ratio of returns to variance)
is anomalous:
– Compare to the average A ratio for portfolio over (‐49,‐11)
window:
• “Crude
Crude dependence adjustment.
dependence adjustment ”
– Could also compute the average of the variances:
• “No dependence adjustment.”
Gompers/Greenwood/Lerner 2009
BW 1980: Mean adjusted return (3)
• Test statistic for N observations is
1 N
N
∑A
i =1
i0
1/ 2
⎡ 1 ⎛ −11 ⎡⎛ 1 N
⎞ ⎤
2
⎞⎤
⎢ ⎜ ∑ ⎢⎜ ∑ A ⎟ − A *⎥
⎟⎥
⎢ 38 ⎜⎝ t = −49 ⎣⎝ N ⎟⎥
it
⎣ i =1 ⎠ ⎦ ⎠⎦
⎡ −11 N ⎤ 1
A* ≡ ⎢ ∑ ∑ Ait ⎥
⎣t = −49 i =1 ⎦ 39 N
Gompers/Greenwood/Lerner 2009
BW 1980: Market adjusted return
• In this case, assume beta of one:
– Somewhat simpler because no normalization for variance:
• Heteroscedasticity presumed to be less of a problem.
– Generally uses equal‐weighted index.
– Otherwise identical.
Otherwise identical
Ait = Rit − Rmt
Gompers/Greenwood/Lerner 2009
BW 1980: Market model adjusted return
• Estimate coefficients for market model using months in (‐
89,‐11) interval.
• Define residual as Ait.
• Test statistic resembles market adjusted return:
1 N
N
∑A
i =1
i0
1/ 2
1 ⎛⎜ N ⎡ 1 −11 ⎛ ⎤⎞
2
⎛ −11
A ⎞⎞
∑ ⎢ ∑ ⎜⎜ Ait − ⎜ ∑ it ⎟ ⎟⎟ ⎥⎟
⎜
N i =1 ⎢ 77 t = −89 ⎝ ⎝ t = −89 79 ⎠ ⎠ ⎥⎦ ⎟⎠
⎝ ⎣
Gompers/Greenwood/Lerner 2009
BW 1980: Basic test
• Examine probability of:
– False conclusion of significant difference from zero returns
(T
(Type I error).
I )
– Failure to identify abnormal returns (Type II error).
– Generally favorable performance by mean adjusted return.
Generally favorable performance by mean adjusted return
• Objective is to reject the null hypothesis that the average
return is zero when it is actually not zero!
etu s e o e t s actua y ot e o
• Reminder!
– Power = probability that test will reject a false null hypothesis =
p y j yp
Probability that it will not make a type II error.
– Test size = Significance Level = Type 1 error = Probability of
rejecting a true null hypothesis
j ti t ll h th i
Gompers/Greenwood/Lerner 2009
Table 1: Main Results
Gompers/Greenwood/Lerner 2009
Table 1: Main Results
Gompers/Greenwood/Lerner 2009
Table 1: Main Results
Gompers/Greenwood/Lerner 2009
Table 1: Main Results
Gompers/Greenwood/Lerner 2009
BW 1980: Other results
• Dramatic fall‐off in power when uncertainty in event
dates.
• When shifts in overall risk, market model adjustments
may actually make worse off.
• Value‐weighted index adds little power.
l h d d dd l l
Gompers/Greenwood/Lerner 2009
BW 1980: An important issue
• Many events are clustered in time, e.g.:
– Merger waves.
– Regulatory shifts.
• Reduces degree of independence across observations.
• To test, induce event in same calendar month for each
sample of 50 firms.
• Weak performance of mean adjusted without
W k f f dj d ih
dependence adjustment.
Gompers/Greenwood/Lerner 2009
BW 1980: Take‐aways
• In general, simplest methods work as well as more
complex:
– In some cases even better!
• Poor performance of mean‐adjusted returns when
clustering.
clustering
• Getting event dates right is most critical!
Gompers/Greenwood/Lerner 2009
Brown and Warner
Brown and Warner
“Using Daily Stock Returns: The Case
of Event Studies” [JFE, 1995].
Gompers/Greenwood/Lerner 2009
BW 1985: Repeats many of same analyses
• Rationale for new analysis:
– Daily data has greater problems with non‐normality.
– Daily data has greater problems with non‐synchronous trading.
– Daily variance displays more non‐stationarity, serial
dependence.
dependence
Gompers/Greenwood/Lerner 2009
BW 1985: Same methodology
• 250 portfolios of 50 firms between 1962 and 1979.
• Gather data in (‐244,+5) window.
• Induce excess performance in event window.
• Compute the excess return measures (Ait) using three
methods above.
Gompers/Greenwood/Lerner 2009
BW 1985: Test statistics
• For an event day t, compute:
At / Sˆ ( At ), where
1 Nt
At ≡
N
∑A
i =1
i ,t
∑( )
−6
~2
Sˆ ( At ) ≡ At − A / 238
t = −244
~ 1 −6
A≡ ∑
239 t = −244
At
Gompers/Greenwood/Lerner 2009
BW 1985: Basic tests
• Little difference between three methodologies.
• Considerable greater power than in monthly analysis (not
surprising).
• Deterioration with longer event windows, smaller sample
sizes.
Gompers/Greenwood/Lerner 2009
BW 1985: Other key analyses
• Clustering of events:
– As before, deterioration in mean‐adjusted returns; little impact
otherwise.
th i
• Non‐synchronous trading:
– Scholes
Scholes‐Williams correction to market model makes little
Williams correction to market model makes little
difference:
• While estimate of ββ is biased downward, estimate of α is
biased upward!
Gompers/Greenwood/Lerner 2009
BW 1985: Take‐aways
• By and large, use of daily data is straightforward:
– Little impact of non‐normality, controls for non‐synchronous
t di
trading.
– Mean adjusted returns again perform worst.
– With longer windows, may need to control for autocorrelation:
With longer windows may need to control for autocorrelation:
• But substantial loss of power.
Gompers/Greenwood/Lerner 2009
Bittlingmayer and Hazlett
and Hazlett
“DOS Kapital” [JFE, 2000].
Gompers/Greenwood/Lerner 2009
Looks at market reaction to antitrust case
• “Stylized fact”: Microsoft harms competitors.
• Alternative explanations:
– “Good competitor” that benefits other firms:
• E.g., network effects.
– Antitrust may thus impose social costs, end efficient
i h i i l d ffi i
behavior.
– Private harm but social benefit:
Private harm but social benefit:
• Will not be able to discern this argument.
– Regulatory capture theory
Gompers/Greenwood/Lerner 2009
Explores effect on several classes of firms
• Microsoft itself.
• Firms that use Microsoft products:
– Input prices should fall after suit.
• Firms that produce complementary products:
– Increased demand.
• Rivals:
– End of predatory behavior.
Gompers/Greenwood/Lerner 2009
The sample
• 54 events in Microsoft antitrust investigation between
1991 and 1997:
– 29 pro‐enforcement events.
– 8 anti‐enforcement events.
– 17 ambiguous ones.
17 ambiguous ones
• Determined from Wall Street Journal Index.
Gompers/Greenwood/Lerner 2009
The sample (2)
• 159 firms in seven computer industry segments.
• Determined through Hoover’s Guide to Computer
Companies 1995.
• Concerns:
– Missing small firms (most hurt by predation?)
– Entering or exiting firms?
Gompers/Greenwood/Lerner 2009
Computing excess returns
• Multivariate regression model:
5
Rit = ai + bi M t + ∑ ci ,t + k Di ,k + ε it
k =−5
• Uses returns over entire period:
– Regresses daily return on:
• Constant.
• Market.
M k t
• Dummies for each day of event period.
– Uses indexes for each segment and overall:
Uses indexes for each segment and overall:
Does this make sense?
Gompers/Greenwood/Lerner 2009
Computing excess returns
Estimate market beta
Event
“Multivariate” approach
Effectively aggregates all of
Effectively aggregates all of
These estimates
Gompers/Greenwood/Lerner 2009
Computing excess returns ‐ Limitations
Estimate market beta
Event
Some data gets underweighted
Some data
gets overweighted
Problematic only if we think that the
P bl ti l if thi k th t th
beta is moving around a lot
Gompers/Greenwood/Lerner 2009
Computing excess returns (2)
• For each category:
– Report sum of coefficients on dummy variables.
• Eg.
• Reg ret mret dum_2 dum_1 dum0 dum1 dum2
• Test dum_2+dum_1+dum0+dum1+dum2=0
Test dum 2+dum 1+dum0+dum1+dum2=0
– Reports joint test statistic.
• Not really explained in paper
Gompers/Greenwood/Lerner 2009
Overall effect
• Show positive gains on anti‐enforcement event
announcements.
• Weaker, but significant, negative drop on pro‐
enforcement event announcements.
• Even stronger when restrict to events where Microsoft
h h f
stock rose or fell.
Gompers/Greenwood/Lerner 2009
Looking for social welfare harm
• Seek events where 1.5σ reactions of Microsoft, index, in
opposite directions:
– Is there any evidence of change in social welfare here.
• Only a few events‐‐19 in all!
• Almost all firm‐specific events (earnings announcements,
(
other litigation, etc.)
• Provides little support for claims.
P id littl tf l i
Gompers/Greenwood/Lerner 2009
Common Problem – Clustering of events
• Ideal Event Study
d l d
time
1 2 3 4 5 6 7 8 9
• More typical studies
yp
1 2 5 8 9
time
3 6
4 7
• Assuming event independence is likely to lead to over rejection of
the null hypothesis because treatment firms are similar
the null hypothesis, because treatment firms are similar
• Relax! Not usually a big deal for daily return event studies
• Quick solution: Aggregate events that happen on the same day or
overlap into a single observation
l i t i l b ti
• Alternative estimation: Panel estimation clustering standard errors
by period
• This is too conservative (because true ρ<1), but a good starting
h (b ) b d
point. We’ll return to this issue next class.
Gompers/Greenwood/Lerner 2009
Greenwood
“Short‐ and Long‐term demand curves
for stocks: Theory and Evidence on the
dynamics of arbitrage” [JFE, 2005].
Gompers/Greenwood/Lerner 2009
Event Summary
• Reweighting of Nikkei 225 Index
• Nikkei index equal to sum of prices of constituents divided by index
divisor
• 30 Additions experience positive demand shock, 30 deletions
negative demand shock, 195 remainders negative demand shock
• Under assumption of 2,430 billion yen of index linked assets,
Under assumption of 2 430 billion yen of index linked assets
rebalancing caused 2,000 billion yen of trading in one week (~ $US
20 billion).
• Substantial variation in demand shocks between stocks:
• By Yen size of shock
• Max 127 billion Yen, Min 221 million Yen
Max 127 billion Yen Min 221 million Yen
• As a fraction of market capitalization
• One deletion experienced a shock equivalent to 17.59% of its
market capitalization
k l
• One additions experienced a shock equivalent to 9.61% of its
market capitalization
Gompers/Greenwood/Lerner 2009
A Single Event
At close,
30 index
stocks
replaced
Gompers/Greenwood/Lerner 2009
Unusual Features of Event
• 30 Adds, 30 Deletes, 195 Remainders
• Adds = 30% of new index, Deletes = 3% of old index
• Implication: Remainders get significantly downweighted
Implication: Remainders get significantly downweighted
• Can compute cumulative returns for each of these portfolios:
30
20 Additions
Cumulative Percentage Reeturn
10
TOPIX
0
April 7, 2000 April 17, 2000 April 27, 2000 May 7, 2000
-10
10 R
Remainders
i d
-20
-30
-40
Gompers/Greenwood/Lerner 2009
Econometric Concerns
• Usual concern with this type of event is that security
returns are correlated before the event, and therefore
there are not 255 distinct events here
there are not 255 distinct events here.
– Intuition: Stocks A and B get added and go up, but A and B
exhibit prior positive correlation. Unfair to treat these as two
separate datapoints.
t d t i t
• Tricky in this case because additions go up while
deletions go down, even though additions and deletions
deletions go down, even though additions and deletions
had positive correlation in the past.
– This logic suggests that we might have more than 255
d t
datapoints.
i t
• My main struggle writing this paper was how to deal with
this issue correctly.
this issue correctly.
Gompers/Greenwood/Lerner 2009
Theory 1
• Capital Market includes N risky securities in fixed supply
given by supply vector Q.
t
Di ,t = Di ,0 + ∑ ε i , s
s =1
• Consider a mean variance world, where risk averse
arbitrageurs have mean variance preferences over next
b h f
period’s returns.
max N Et [− exp(−γ Wt +1 )]
Wt +1 = Wt + Nt' [ Pt +1 − Pt ]
• Yields Demand
ld d
1
Nt = [Vart ( Pt +1 )]−1 ( Et ( Pt +1 ) − Pt ).
γ
• With supply Q
Pt = Et ( Pt +1 ) − γ Vart ( Pt +1 )Q.
Gompers/Greenwood/Lerner 2009
Theory 2
• Consider a shock to supply QÆ to Q‐u
Pt* − Pt* −1 = ε t* + γΣ((T − t*)u + Q)
Gompers/Greenwood/Lerner 2009
Econometric Concerns
• Usual concern with this type of event is that security
returns are correlated before the event, and therefore
there are not 255 distinct events here
there are not 255 distinct events here.
– Intuition: Stocks A and B get added and go up, but A and B
exhibit prior positive correlation. Unfair to treat these as two
separate datapoints.
t d t i t
• Tricky in this case because additions go up while
deletions go down, even though additions and deletions
deletions go down, even though additions and deletions
had positive correlation in the past.
– This logic suggests that we might have more than 255
d t
datapoints.
i t
• My main struggle writing this paper was how to deal with
this issue correctly.
this issue correctly.
Gompers/Greenwood/Lerner 2009
Solution: GLS? GLS standard errors?
• Recall the OLS Estimator minimizes
( y − X β ) 'σ 2 ( y − X β )
• Yielding estimator I went with OLS estimator
β = ( X ' X ) ( X ' y)
−1
With GLS standard errors
Guiding principle: OLS
Guiding principle: OLS
Estimates are not sensitive
• But if instead we think that To covariance matrix.
Var (ε ) = Σ However, we want to be
C
Conservative in our
ti i
• Then we want the GLS estimator Standard errors.
Gompers/Greenwood/Lerner 2009
Solution: GLS? GLS standard errors?
• Recall the OLS Estimator minimizes
( y − X β ) 'σ 2 ( y − X β )
• Yielding estimator Roughly speaking, t‐stats
β = ( X ' X ) ( X ' y)
−1
Go down from about 10
To range of 3 or 4
To range of 3 or 4
• But if instead we think that
Var (ε ) = Σ
• Then we want the GLS estimator
β = ( X ' Σ −1 X )−1 ( X ' Σ −1 y )
• With standard errors given by
Var ( β ) = ( X ' Σ −1 X )
Gompers/Greenwood/Lerner 2009
Figure 5
Gompers/Greenwood/Lerner 2009
Figure 5
Looks like a t‐stat of 30…
But GLS adjustment takes it down considerably
Gompers/Greenwood/Lerner 2009
Gompers/Greenwood/Lerner 2009
“Comovement”
Nicholas Barberis – Chicago and NBER
Andrei Shleifer – Harvard and NBER
Andrei Shleifer Harvard and NBER
Jeffrey Wurgler – NYU Stern
JFE 2004
Gompers/Greenwood/Lerner 2009
Comovement: Motivation
• There are many strong patterns of return
comovement
– There are “common factors” in returns of:
• Small stocks
• Value stocks
Value stocks
• Closed‐end funds
• Stocks in same industry
• Bonds of same rating/maturity
/
• ...
• Why do certain assets comove, others do not?
– What determines “betas” or “loadings” on common factors?
Gompers/Greenwood/Lerner 2009
Comovement: Motivation 2
• Traditional view: Fundamentals‐based comovement
– Assets comove because their “fundamental values” comove
– A very useful paradigm… but the whole story?
• Some puzzling evidence
S li id
– Siamese‐twin stocks
– Closed‐end country funds
– Small stocks, value stocks
– Commodities
• Alternative view: Trading‐driven comovement
– In addition to fundamentals
Gompers/Greenwood/Lerner 2009
Comovement: Motivation 3
• Trading‐driven comovement idea: perhaps investor
trading patterns
gp also affect comovement
• Above and beyond fundamentals
• This
This builds on evidence that trading/demand per se
builds on evidence that trading/demand per se
affects prices (given limits on arbitrage)
• Additions to S&P 500 increase price
• Exchange holidays reduce volatility
• Mispricing in tech‐stock carveouts (Palm/3com)
• Common factors in returns match common factors in order flow
• If trading affects returns, then correlated trading leads
to correlated returns
to correlated returns
• Even without changes in fundamentals
Gompers/Greenwood/Lerner 2009
Comovement: Conjectures
• Trading‐induced comovement models make 3 predictions
• H1: Suppose asset j, previously a member of Y, gets
reclassified into X. Then j’s univariate beta w.r.t. the
X index will increase, as will its R
, 2 w.r.t. X index.
• H2: Suppose asset j, previously a member of Y, gets
reclassified into X. Then j’s bivariate beta w.r.t. X will
increase, and its bivariate beta w.r.t. Y will decrease.
• H3: More and more noise traders (who use X and Y as
categories/habitats) mean less and less correlation
between X and Y.
Gompers/Greenwood/Lerner 2009
Tests: Event Study
• Setting: S&P 500 additions and deletions
– S&P 500 is a category/habitat/classification used by many investors
– Membership changes don’t have obvious info. about fundamentals
– X = S&P 500
– Y = non‐S&P 500 (complement)
• H1: Does univariate S&P beta goes up following addition?
• Yes (in daily & weekly returns, but not monthly): Table 1
• H2: Following addition, does bivariate S&P beta go up, non‐
S&P beta go down?
• Yes (at every horizon)
Yes (at every horizon) – Figure 1
Figure 1
Gompers/Greenwood/Lerner 2009
Table 1
Gompers/Greenwood/Lerner 2009
Computing Standard Errors for Δβ….
Gompers/Greenwood/Lerner 2009
Computing Standard Errors for Δβ….
• Suppose that Yahoo and Google enter the SP500 around the same time
h h d l h d h
period:
Yahoo
G
Google
l
Before
After
• Because these stock returns are correlated (missing factor F), if the beta
goes up for both stocks, hard to say that this is really two independent
observations.
• Tough problem!!! Key: Need to gauge (a) importance of common factor
Tough problem!!! Key: Need to gauge (a) importance of common factor
and (b) degree of overlap. BSW‐> Use ρ(Δβij) as summary statistic.
Gompers/Greenwood/Lerner 2009
T=3000;
Replicating the simulation
N=100;
Niterations=500;
fac stdev=1;
eventdates=1010:10:1000+N*10;
eventdates=eventdates';
teststat=zeros(Niterations,1)+NaN;
for k=1:Niterations;
mrt=randn(T,1);
sprt=mrt+0.2*randn(T,1);
nonsprt=mrt+0.2*randn(T,1);
fac=(fac_stdev^2)*randn(T,1);
w=randn(T,N);
ret = kron(nonsprt,ones(1,N)) + kron(fac, ones(1,N)) + w;
deltab=zeros(N,1)+NaN;
window=100;
for i=1:length(eventdates);
range=eventdates(i)-window+1:eventdates(i);
y=ret(range,i);
x=[ones(window,1), sprt(range)];
b pre=inv(x'*x)*(x'*y);
p y
range=eventdates(i)+1:eventdates(i)+window;
y=ret(range,i);
x=[ones(window,1), nonsprt(range)];
b post=inv(x'*x)*(x'*y);
p ( ) ( y);
deltab(i)=b_post(2)-b_pre(2);
end;
teststat(k)=mean(deltab);
end;
Gompers/Greenwood/Lerner 2009
Now use simulated distribution…
Simulated Test Statistics for SigmaF = 1
60
50
40
30
20
10
0
-0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1 0.12
Gompers/Greenwood/Lerner 2009
Locate your test statistic on distribution…
Simulated Test Statistics for SigmaF = 1
60
50
40
30
20
10
0
-0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1 0.12
Significance level
Gompers/Greenwood/Lerner 2009
Event Studies – Summary and Extensions
Gompers/Greenwood/Lerner 2009
Methodological refinements
• Tests above make assumptions about distribution that
may be unappealing.
• Non‐parametric tests address this concern by looking at
sign, ranking.
• Campbell and Wasley
b ll d l [1993] argue that non‐parametric
[ ] h
tests provide more robust conclusions than parametric
ones.
ones
• When you are worried about the distribution, usually not
that difficult to simulate
that difficult to simulate
Gompers/Greenwood/Lerner 2009
Other methodological refinements (2)
• Often event day is uncertain, e.g.:
– Wall Street Journal Index entries.
– SEC filings.
• Can approach in two ways:
– Expand event window.
d i d
– Model uncertainty in maximum likelihood framework [Ball and
Torous, 1988].
Torous, 1988].
• Unclear how much extra power more painful ML approach
brings.
Gompers/Greenwood/Lerner 2009
Open issues
• Exogeneity of corporate announcements:
– Many decisions (e.g., acquisitions) are under the control of
management. t
– They can be expected to be “timed” to fall when they have
maximum positive impact on stock price.
p p p
– Simple analyses may thus be biased.
Gompers/Greenwood/Lerner 2009
Illustration
• Eckbo, Maksimovic, and Williams [1990] look at
announcements of corporate acquisitions.
– Estimate “naïve” OLS regressions, as well as maximum
likelihood regressions that account for timing of decision:
• With OLS, extent of synergies in acquisition do not explain
With OLS extent of synergies in acquisition do not explain
size of reaction.
• Once ML is used, explanatory power of synergy measures
increase dramatically.
Gompers/Greenwood/Lerner 2009
Open issues (2)
• Many announcements are at least partially anticipated by
the market.
• Danger in cross‐sectional analyses:
– May interpret characteristics as being related to event impact.
– Actually, may be related to extent of disclosure.
A ll b l d f di l
Gompers/Greenwood/Lerner 2009
Illustration
• Austin [AER, 1993] examines whether market rationally
responds to patent awards.
• Concludes that it does:
– Patent awards listed in Wall Street Journal have greater
importance and more positive reaction
importance and more positive reaction.
– But published patents may also represent more “news.”
Gompers/Greenwood/Lerner 2009
Conclusions
• Contrast to studies of long‐run returns:
– Well agreed‐upon methodology.
– Theoretical foundations.
• Huge published literature suggests that most “low‐
h i f it” h b
hanging fruit” has been plucked
l k d
• Continuing opportunities on finance‐economics
interface.
interface
Gompers/Greenwood/Lerner 2009
Next Class
• Long horizon returns
– Applications:
• Security offerings and repurchases
• M&A
• Settings where market efficiency is being called into
Settings where market efficiency is being called into
question
• CARs versus BHARs versus Wealth Ratios
pp g p
• Panel data approaches to clustering when events overlap
• Calendar time regression approaches
Gompers/Greenwood/Lerner 2009