Event Study Methodology

A stock jumps 4% the day it announces an acquisition. Was that the deal, or was it just a strong day for the market? An event study is the standard answer to that question: estimate what the stock would have done absent the event, subtract that from what it actually did, and call the difference the abnormal return attributable to the event. It's the workhorse behind decades of academic finance — earnings announcements, stock splits, index additions, executive departures, merger announcements — and it's also a genuinely useful research tool for testing whether a signal you've found actually moves prices around a specific, dateable moment.

The logic: expected return, then subtract it out

Pick an event date, day zero — the announcement. Around it, define an event window, say five trading days before to five after, over which you'll measure the stock's reaction. Separately, define an estimation window well before the event (often 120 to 250 trading days, ending well clear of the event window) and use it to fit a normal-times model of the stock's return, most simply the market model:

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

In words: on a normal day, stock $i$'s return is some fixed baseline $\alpha_i$ plus its usual sensitivity $\beta_i$ to the market's return $r_{m,t}$, plus noise. Fit $\alpha_i$ and $\beta_i$ on the clean estimation window, then use them to predict what the stock should have returned on each day of the event window given what the market actually did. The abnormal return on each day is actual minus predicted:

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

Summing the abnormal returns across the event window gives the cumulative abnormal return (CAR) — the total move attributable to the event once ordinary market movement is stripped out.

Worked example. A stock has an estimated $\hat\alpha_i = 0$ and $\hat\beta_i = 1.2$ from its estimation window. On the announcement day the market returns $+0.5\%$ and the stock returns $+5.0\%$. The model's predicted return for that day is $0 + 1.2 \times 0.5\% = 0.6\%$, so the abnormal return is $5.0\% - 0.6\% = 4.4\%$. If the stock had already drifted up $1.1\%$ of abnormal return over the two days before the announcement (a hint of leakage) and gives back $0.3\%$ over the two days after, the three-part CAR is $1.1\% + 4.4\% - 0.3\% = 5.2\%$ — the total move the event window shows, net of what beta and the market already explain.

What the shape of the CAR chart tells you

The plot above is the actual output researchers look at, and its shape is diagnostic, not just decorative. Flat before day zero, then a clean jump, is the textbook signature of an efficient, unleaked announcement — the market didn't know until it knew, and it fully priced the news within the event window. A rising line before day zero means information was leaking ahead of the public announcement — insiders, analysts, or a savvy subset of traders acting early. A line that keeps drifting after day zero, rather than flattening, is evidence of underreaction — the market didn't fully digest the news immediately, which is precisely the pattern behind post-earnings announcement drift (see Post-Earnings-Announcement Drift).

What breaks an event study

The market model is a simplification, and the two most common failure modes are worth naming. First, contaminated events: if other news hits the same stock inside the event window (a sector-wide rally, an unrelated lawsuit), the CAR conflates two effects and the "abnormal return" isn't cleanly attributable to the event you're studying — this is why researchers exclude firms with other confounding announcements in the same window. Second, clustering: if many events happen on the same calendar day (all Fed decisions, all earnings on the same reporting date) their abnormal returns are correlated with each other through the shared market factor's estimation error, and a naive standard error that assumes independence across events will overstate statistical significance — the fix is to compute standard errors that account for cross-sectional correlation within event dates, not just across time.

In interviews

Walk through the market-model regression and the abnormal-return formula from memory — this is table stakes. Then show you understand the diagnostics: a CAR plot that slopes before the event date is leakage, one that keeps sloping after is underreaction, and either one is itself a finding worth reporting, not just noise to explain away. If asked how you'd apply this to a new signal, describe defining a clean event date, an estimation window that ends well clear of it, and checking the pre-event CAR for leakage before trusting the post-event number as tradeable.