Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Lecture 6 Event Studies
ET2013 Introduction to Econometrics
Giacomo Pasini
Ca’ Foscari University of Venice
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Lecture overview
1 Event studies
2 Event studies in the Stock market
3 Event studies with Regression
4 Time series forecasting
reference:NHK (2022) Ch. 17
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Outline
1 Event studies
2 Event studies in the Stock market
3 Event studies with Regression
4 Time series forecasting
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Time and events
The idea is this: at a certain time, an event occured, leading a treatment to be put
into place at that time.
Whatever changed from before the event to after is the effect of that treatment.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Back door
Treatment ← AfterEvent ← Time → Outcome
There is clearly a backdoor: anything that changes over time affects both what
happens after treatment and the outcome directly
because Time only affects Treatment through AfterEvent, there is the possibility
that we can close the back door by controlling for some elements of Time, while
still leaving variation in AfterEvent
In other words, we need Treatment to be the ONLY thing that’s changing
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Predicting the counterfactual
We should remove all the ways that moving from before to after the eveny Time
affects the outcome EXCEPT for the part due to Treatment
If we can assume that that whatever was going on before would have continued
doing its thing if not for the treatment,
...then we can use that before-event data to predict what we’d expect to see
without treatment (the counterfactual)
and then see how the actual outcome deviates from that prediction.
The extent of the deviation is the effect of treatment.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Predicting the counterfactual
We assume that whatever pattern we were seeing before the event would have
continued in the absence of the treatment
In (a) the time series is jumping up and down on a daily basis, but doesn?t seem to
be trending generally up or down or really changing at all.
In (b) we have a clear upward trend before the event occurs.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
HOW can we predict the counterfactual?
1 Compare before and after event data as they stand
2 Predict After-Event Data Using Before-Event Data
3 Use the relationship with other variables
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
1 Compare before and after event data
Just ignore the whole thing and simply compare after-event to before-event without
worrying about changes over time at all
Two main ways that this can be justified
1 The time series is extremely flat before the event occurs, with no trend up or down,
no change, no nothing.
2 If you’re looking only at an extremely tiny span of time. Super high-frequency data in
finance
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
2 Predict After-Event Data Using Before-Event Data
If there’s an obvious trend in the data, extrapolating that trend we end up with
some predictions that we can compare to the actual data.
In execution, there are many different ways to perform such a prediction. Later in
this lecture.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
3 Use the relationship with other variables
Instead of only looking for things like trends in the outcome, you also look for
relationships between the outcome and some other variables in the before-event
period.
Researchers doing an event study on stock X will look at how stock X?s price
relates to the rest of the market, using data from the before-event period (β in a
CAPM model)
they find that a 1% rise in the market index is related to a .5% rise in stock X?s price
before the event.
in the after period, they extrapolate on any trend they found in stock X?s price...
...But they also check the market index: if the market index happened to rise by 5%
in the after-event period, then we’d subtract that 2.5% rise before looking for the
effect of treatment.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Outline
1 Event studies
2 Event studies in the Stock market
3 Event studies with Regression
4 Time series forecasting
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Stock returns and information
In any efficient, highly-traded stock market, a stock’s price should reflect all the
public knowledge about a company.
The only reason the price should have any sort of sharp change is if there’s
surprising new information revealed about the company
Perfect setting for an Event study
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
How to perform an event study with stock market data
1 Pick an “estimation period” a fair bit before the event, and an “observation period”
from just before the event to just after the event.
2 Use the data from the estimation period to estimate a model that can make a
prediction R̂ of the stock’s return R in each period
3 In the observation period, subtract the prediction from the actual return to get the
“abnormal return.” AR = R − R̂
4 Look at AR in the observation period.
Nonzero AR values before the event imply the event was anticipated,
Nonzero AR values after the event give the effect of the event on the stock’s return.
If the market is efficient effects will generally spike and then fade out quickly.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
How to predict stock’s return
1 Means-adjusted returns model. Just take the average of the stock?s return in
the estimation period, R̂ = R̄
2 Market-adjusted returns model. Use the market return in each period, R̂ = Rm
3 Risk-adjusted returns model. Use the data from the estimation period to fit a
regression describing how related the return R is to the market portfolio:
R = α + βRm . Then, in the observation period, use that model and the actual Rm
to predict R̂ in each period.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Example
On August 10, 2015, Google announced that they would be rearranging their
corporate structure.
There would be a new parent company called “Alphabet” that would own Google
along with all other companies (Alphabet etc).
How did the stock market feel about that?
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Google returns
Daily return for Google and for the S&P 500 (price divided by the previous day’s
price, all minus 1).
Returns are very flat a lot of the time
Google’s returns and the S&P 500 returns tend
to go up and down at the same time
We don’t want to confuse a market move with a
stock move and think it’s just all about the
stock!
This is why two of our three prediction methods
make reference to a market index
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Event study
Estimation period: May-July; Observation period: August 6-August 24
Big spike just after August 10, no matter which
of the three methods we use
around August 20, the Mean line takes a dive,
while the Market and Risk lines stay near zero?
That?s a spot where the Google price dipped,
but so did the market.
The graph, and the movement around August
20, also nudge us towards using a narrow
window
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Outline
1 Event studies
2 Event studies in the Stock market
3 Event studies with Regression
4 Time series forecasting
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Event studies when the effect is long–lasting
The method used in finance is purpose-built for effects that spike and then, usually,
quickly disappear.
What if you’re interested in an event that changes the time series in a long-lasting
way?
1 Estimate one regression of the outcome on the time period before the event.
2 Then estimate another time series of the outcome on the time period after the event.
3 Then see how the two are different.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Segmented regression
This can be easily implemented with an interaction term:
Outcome = β 0 + β 1 t + β 2 After + β 3 t × After + ε
t is the time period and
After a binary variable equal to 1 in any time after the event.
This forces the time trend to be linear. You could easily add some polynomial terms
to allow the time series to take other shapes
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Segmented regression: pros and cons
Pros: you can get a more-precise estimate of the time trend than going day by day
(or second by second, or whatever level your data is at)
Cons: you need to be very careful with any sort of statistical significance testing.
If your data is autocorrelated then you’re extremely likely to find a statistically
significant effect of the event, even if it truly had no effect
This occurs because the data tends to be “sticky” over time, with similar values
clustering in neighboring time periods,
Take home message: Use heteroskedasticity- and autocorrelation-robust standard
errors
...or model directly autocorrelation: next topic
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Outline
1 Event studies
2 Event studies in the Stock market
3 Event studies with Regression
4 Time series forecasting
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Segmented regression: pros and cons
The concept behind an event study is that you have some actual data in the
after-event period and need, as a comparison, a prediction of what would have
happened in the after-event period if the event hadn’t occurred
We’re using the time series from before the event (and maybe some additional
predictors) to predict beyond the event. This is the kind of thing that time series
forecasting was designed for.
In a time series, observations are related across time. If a change happens in one
period, the effects of that change can affect not just that period, but the whole
pathway afterwards.
If we want to get serious about predicting that counterfactual,we better use a
forecasting model that takes the time series element seriously.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
AR models
AR stands for autoregression, meaning that the value of something today influences
its value tomorrow.
If I put a dollar in my savings account today, that increases the value of my account
by a dollar today...
...but also by a dollar tomorrow (plus interest),
and the next day (plus a little more interest),
and on into the future forever.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
AR models
My savings account would be a model with an AR(1) process - the 1 meaning that
the the value of Y in the most recent period is the only one that matters for
predicting Y in this period.
If I want to predict my bank balance today, then knowing what it was yesterday is
very important.
If I already know what yesterday was, then learning about the day before that won’t
tell me anything new
Yt = β 0 + β 1 Yt −1 + ε
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
MA models
MA stands for moving-average. Moving-average means that transitory effects on
something last a little while and then fade out.
I lend my friend ten dollars from my savings account and she pays me back over the
next few weeks
This is a ten-dollar hit to my account today...
...A little less next week as she pays some back
A little less the week after that as she pays more back
And then when she pays it all back, the amount in my account is back to normal
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
MA models
A MA(1) process - where only the most recent transitory effect matters:
Yt = β 0 + (ε t + θε t −1 )
Since we can’t actually observe ε t or ε t −1 , we can’t estimate θ by regular regression.
There are alternate methods for that.
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
ARMA models
Since the amount in my savings account clearly has both AR and MA features,
we’d say that the time series of my savings balance follows an ARMA process
Example: an ARMA(2,1) model, would include two autoregressive terms and one
moving-average term
Yt = β 0 + β 1 Yt −1 + β 2 Yt −2 + (ε t + θε t −1 )
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
AR, MA and ARMA at work
(a): completely random white noise
(b): same white noise + MA(1) in ε:
each period in time a bit closer to its
neighbors
(c): each value directly influences the
other: much smoother
(d): AR + MA combined, data even
more related
Event studies Event studies in the Stock market Event studies with Regression Time series forecasting
Other time series features
You can add an Integration term (I), which differences the data (i.e., instead of
evaluating the amount in my account, you evaluate the daily change in the
amount) and get an ARIMA
You can add other time-varying eXogenous (X) predictors, like how much I get paid
at my job, and get an ARIMAX.
You can account for Seasonality (S), like how my savings dips every Christmas, and
get a SARIMAX.
You can allow not only ε t to be related over time, but also σt (ε)
In this latter case we allow for conditional heteroskedasticity (CH): ARCH, GARCH,
EGARCH models and so on
Time series forecasting deserves (more than) a course on itself!
