An event study asks whether something predictable happens around a repeatable event. Take every occurrence of the event, line them all up so the event date becomes day zero, measure returns over a fixed window on either side, and compare the resulting distribution with the base rate.
The method is the same whether the event is a RELIANCE results announcement, a promoter pledge disclosure, an index inclusion, or a bulk deal crossing a size threshold.
The steps
- Define the event precisely, including the timestamp at which the information became public.
- Align every occurrence on day zero.
- Compute abnormal returns over the window, netting out the benchmark or model expectation.
- Aggregate across days into a cumulative abnormal return, then average across events.
- Compare with the base rate over the same period and universe.
- Compute the t-statistic on the difference.
Steps five and six are the ones most often skipped, and they are the ones that decide whether the study found anything.
Reading the output
| Column | What to check |
|---|---|
| Event count | Below about 50 the confidence interval swamps the estimate |
| Mean CAR | The headline effect over the window |
| Median CAR | If it differs sharply from the mean, a few events are carrying the result |
| Hit rate | Read against the base rate, never alone |
| Standard deviation | Drives the t-statistic and is usually larger than expected |
| Pre-event drift | Non-zero drift before day zero means leakage or a contaminated definition |
Pre-event drift deserves attention. If the cumulative abnormal return is already rising before day zero, either the market anticipated the event or your event date is wrong.
Caveats
Timestamps are where most event studies quietly break. An announcement filed at 18:40 IST cannot be traded until the next open, so day zero has to be the following session. Using the filing date as a tradeable date manufactures a return you could never have captured.
Event clustering inflates significance. If forty of your events are earnings announcements in the same fortnight, they share a market move, and treating them as independent observations makes the t-statistic look far better than the evidence supports.
Confounding events contaminate windows. A results announcement that lands alongside a dividend declaration and a block deal gives you a return you cannot attribute to any of them.
Selection of the event set matters as much as the analysis. If you build the event list from securities that are currently listed, you have dropped every company that was delisted after its event, which is survivorship bias entering through the sample rather than through the price data.
Thin liquidity distorts the result. An abnormal return on a stock with a wide spread and low delivery volume is a number on a screen rather than a return you could have realised.