A financial event study compares observed returns around news with a model-based benchmark to measure abnormal returns, subject to timing and inference limits.
An event study, in financial markets, compares a security’s observed returns around a defined event with the returns estimated by a normal-return model. The difference is an abnormal return. It helps researchers examine market reactions to news, but does not automatically prove that the event caused the entire difference.
This article covers short-horizon return event studies, such as an analysis around an earnings release or merger announcement. The broader term also appears in research using other outcomes and statistical designs.
The event date identifies the news being studied. The event window includes the return intervals being examined, while the estimation window supplies observations used to fit the normal-return model. Separating these windows helps avoid fitting that model to the very reaction being measured. MacKinlay’s Event Studies in Economics and Finance describes this framework.
Suppose a fictional company releases results after the exchange closes on Tuesday. With close-to-close daily returns, label Wednesday’s trading session day 0: its return interval includes the first opportunity for the announcement to affect the next closing price. Check holidays, time zones, and the actual release time rather than assigning dates mechanically.
One illustrative design is:
| Segment | Trading days relative to day 0 | Purpose |
|---|---|---|
| Estimation window | -130 through -11 | Fit the normal-return model using 120 observations |
| Gap | -10 through -2 | Leave 9 observations between model fitting and the event window |
| Event window | -1 through +1 | Examine 3 daily returns around the announcement |
These lengths are assumptions for this example, not a universal standard. A wider window may capture anticipation or delayed reaction, but also admits more unrelated news. A narrow window can miss a response that began before the selected timestamp.
A simple market model estimates the relationship between the stock’s returns and a chosen market benchmark:
Here, the estimated normal return uses an intercept, alpha, plus the stock’s estimated market sensitivity, beta, multiplied by the market return for the same interval. The coefficients are fitted in the estimation window; the market return is observed during the event window.
Abnormal return is the difference between the observed stock return and this fitted normal return:
Wharton’s WRDS event-study implementation illustrates this calculation and alternative benchmark models. The simple market model shown here uses returns directly; it is not automatically a CAPM regression of excess returns.
Assume the fictional company’s pre-event estimates are daily alpha of 0.05% and beta of 1.20. These are supplied example inputs, not estimates from a real dataset. Use arithmetic daily total returns with consistent currency, dividend, and split treatment for both stock and benchmark.
| Day | Market return | Stock return | Estimated normal return | Abnormal return |
|---|---|---|---|---|
| -1 | +0.50% | +0.40% | +0.65% | -0.25% |
| 0 | -1.00% | +2.00% | -1.15% | +3.15% |
| +1 | +0.25% | -0.10% | +0.35% | -0.45% |
On day 0, the model gives 0.05% plus 1.20 times -1.00%, or -1.15%. The stock actually returns +2.00%, so it exceeds the modeled normal return by 3.15 percentage points.
The positive abnormal return is therefore not just the stock’s 2% gain. It also reflects the fact that the benchmark-based model estimated a negative return that day.
Cumulative abnormal return, or CAR, sums the abnormal returns in the chosen window:
The example’s three-day CAR is +2.45%. It summarizes deviations from the model, not the actual return earned by holding the stock.
Using the stock-return column, its compounded three-day total return is:
The two numbers answer different questions. Compounding the daily abnormal returns would not turn them into an actual buy-and-hold stock return either.
Direction: The positive CAR means the stock outperformed this model across the specified three days.
Statistical uncertainty: The table alone cannot establish significance. An analysis needs an appropriate estimate of sampling variability, including uncertainty in the fitted benchmark. No estimation residuals or standard errors are supplied here, so there is no basis for reporting a t-statistic or p-value.
Economic importance: Whether the estimated difference matters depends on the question and its uncertainty. A detectable return difference is not necessarily large enough to cover trading costs.
Causation: Other information released in the same window can contribute to the measured return. A significant result does not identify the announcement as its sole cause.
The American Statistical Association’s statement on p-values explains why significance is not a measure of effect size or the probability that a hypothesis is true. Report the estimate and its uncertainty, not only whether a threshold was crossed.
For a study of many firms, explain how events are pooled and how dependence between observations is handled. Firms responding to the same policy announcement are not necessarily independent observations.
This explanation is educational, not personalized investment advice. Observing a reaction after publication does not mean a reader could have traded before it occurred or can profit from the next announcement.