Reading an event study without fooling yourself

An event study asks a simple question: did this news move this price? The answer depends on three choices that are easy to get wrong. I used the method in my paper on fake news and stock markets, and these are the places I learned to slow down.

The idea in four steps

1. Pick the event and its date. The moment the news became public.
2. Estimate normal behaviour. Using a window before the event, fit a model of what the return would usually be. The simplest is the market model:

$$R_{i,t} = \alpha_i + \beta_i R_{m,t} + \varepsilon_{i,t}$$

1. Compute the abnormal return. Subtract the model's prediction from the actual return:

$$AR_{i,t} = R_{i,t} - (\hat\alpha_i + \hat\beta_i R_{m,t})$$

1. Add them up over an event window. The cumulative abnormal return is $CAR_i = \sum_{t=t_1}^{t_2} AR_{i,t}$. Average across events and test whether it differs from zero.

Where it goes wrong

The event date is blurry. For earnings, the date is exact. For a rumour, it is not. Is the event when a false story first appeared, when it spread widely, or when it was corrected? Each choice measures something different, so say which one you used and test the others.

Other news arrives at the same time. An abnormal return is only abnormal relative to your model. If a company also announced results that day, you have measured both. Check for confounding events inside the window.

Events cluster. If many events fall on the same days, their abnormal returns are correlated, and standard errors that assume independence are too small. Cluster by date or use a test designed for this.

Prices move without the average moving. News can raise uncertainty without changing the expected return. Then the CAR is near zero while volatility jumps. A study of volatility needs a test on variance, not only on the mean.

Thin trading. Shares that trade rarely have stale prices, which distorts both the normal model and the event window.

A short checklist

• State the event definition and test a second one.
• Show results for more than one window, such as one day and three days.
• Look for other announcements in the window.
• Report standard errors that allow for clustering.
• Show the abnormal returns before the event. They should be flat. If they are not, the market may have known earlier, or the model is wrong.

None of this is exotic. The method is simple, and the value is in the discipline around it.