Event Study: How to Read Abnormal Returns From News

A headline crosses the wire: a stock “jumped 4% after earnings.” It reads like a clean reaction, cause and effect in one line. Then you pull up the tape and see the whole market was up 3% that afternoon on a soft inflation print. So how much of that 4% actually belonged to the earnings, and how much was just the tide lifting every boat in the harbor? An event study is the tool built to answer that question. It’s a disciplined way to separate a company-specific move from the market it was floating on, and once you can do it, you’ll read both research papers and market headlines very differently.

What an event study actually measures

An event study asks something narrow and testable: around a specific event, did the price move more than a benchmark would have predicted? The event can be almost anything with a date attached. An earnings release, a macroeconomic data publication, an index addition or deletion, a dividend change, a spin-off. The method doesn’t care about the narrative. It cares about timing and size.

The engine is the abnormal return: the actual return over a chosen window minus the return the benchmark told you to expect. If a stock returns +3% on a day its benchmark predicted +1.2%, the abnormal part is +1.8%, and that’s the number the study is really about. The raw candle’s just context. It’s the same instinct behind reading price behavior around scheduled macro releases, where the reaction only means something relative to what the market had already priced in. Done at scale across many companies, this is also the classic test of market efficiency: how quickly and how completely prices absorb new public information.

The building blocks: windows, benchmark, and abnormal return

Five pieces do the work, and each one’s a design choice you can get wrong.

• The event date and timestamp. This is the anchor. Day 0 is the session the information reached the market, and the exact hour matters when news lands mid-session or after the close.
• The estimation window. A stretch of ordinary trading before the event, often 120 to 250 trading days ending well before day 0, used to describe how the stock normally behaves.
• The event window. A short band around the announcement, say day -1 to day +1, or a single session, over which you measure the reaction.
• The benchmark model. The rule for what return to expect. The simplest is the market model: expected return equals alpha plus beta times the index return, with alpha and beta fitted over the estimation window.
• Abnormal return and its running total. Abnormal return is actual minus expected on each day. Cumulative abnormal return, or CAR, sums those daily abnormal returns across the event window.

The benchmark is where most of the judgment lives. A stock with a beta of 1.3 is supposed to move 1.3 times the index, so on a strong tape it’ll rise with no news at all. Ignore that and you’ll credit the event for a move the market already owed you. That’s why the split between beta and idiosyncratic risk sits underneath the whole method: the abnormal return is your estimate of the idiosyncratic, company-specific piece.

Granularity matters here too. Daily bars are the norm, but if you aggregate up to weekly returns you can smear a sharp one-day reaction into the surrounding noise, and if you drop to intraday you inherit a different set of timing problems. The window you choose is a time-aggregation decision, and it quietly shapes every number that follows.

A worked example with an index benchmark

Take a hypothetical stock, call it Northwind, benchmarked against a broad index. Over the estimation window its beta comes out at 1.2 and its average daily alpha is close to zero, so I’ll treat expected return as simply 1.2 times the index return. This is a measurement example. It shows the arithmetic, and it isn’t evidence that any event caused anything or a suggestion to trade around announcements.

On the event day the index returns +1.0%. The model expects Northwind to return 1.2 times that, or +1.2%. Northwind actually returns +3.0%. The abnormal return is 3.0% minus 1.2%, which is +1.8%. Now widen the lens to three days. Say the abnormal returns land at +0.4% on day -1, +1.8% on day 0, and +0.3% on day +1. The cumulative abnormal return is +2.5% across that window.

I mark that +2.5%, not the raw three-day change, because the CAR is what strips the market back out. When someone asks whether a reaction was big, that’s the figure I reach for. A raw percentage with no benchmark beside it is a candle with no scale.

Why a green candle can hide a negative abnormal return

That’s the trap the raw move sets, and it catches careful people. Suppose a stock closes +3.0% on its earnings day and the recaps call it a strong reaction. If the index rose +3.5% that session and the stock’s beta is 1.0, the model expected +3.5%. The abnormal return is 3.0% minus 3.5%, or -0.5%. The stock underperformed what the day already justified. Green candle, negative abnormal return.

The reverse fools people just as often. A stock ticks up only +0.8% on a day the index fell -1.5%. With a beta near 1.0 the model expected roughly -1.5%, so the abnormal return works out to about +2.3%. A move that looked like a shrug was a large relative gain. I once logged a +3.0% session that felt like real strength until I checked the tape, saw the index had run +3.5%, and watched the abnormal return come out below zero. That single check is why the framework earns its keep.

An abnormal return marks a move as unusual relative to the benchmark. It does not prove the event caused anything. Whether the earnings, a leaked rumor, or an unrelated sector story did the driving is a separate question the abnormal return can’t settle alone. Even a clean gap at the open tells you where the reaction registered, without explaining why.

Where event studies break: leakage, overlap, and false positives

A clean-looking event study can still be wrong, and the failure modes deserve names.

Timestamps can be fuzzy. Tag day 0 as the release date when the wire actually crossed after the prior close, and your window is off by a session while the reaction bleeds into the wrong bar. Information can leak. When a stock drifts higher for a week ahead of a formal announcement, much of the reaction is already in the price by day 0, and a tight window misses it. Events can overlap. If earnings land the same week a company joins an index, the earnings effect and the index-reconstitution effect are tangled, and no single abnormal return can attribute the move cleanly. Here the honest answer is often that you can’t separate them, rather than a confident number.

The estimation window carries its own assumption: that the stock’s beta was stable while you measured it. A structural break in that window, a merger, a business-model shift, a jump in leverage, quietly corrupts the expected return you carry into day 0. And then there’s the benchmark choice itself. Swap a broad index for a sector index and every expected return shifts, which shifts every abnormal return downstream.

The quiet killer is repeated testing. Run the same test across hundreds of events and a handful will look significant by pure chance. The multiple-testing problem is exactly how noise gets dressed up as a finding, and it’s worth studying how one trader built an entire philosophy around being fooled by randomness, because an event study is a comfortable place for randomness to hide.

How to read an event-study claim in a headline or paper

When you meet an event-study claim, in a research abstract or a market recap, a short checklist tells you how much weight it can bear.

• What is the event window? A one-day window and a twenty-day window can tell opposite stories from the same data.
• What is the benchmark? A broad index, a sector index, or a matched peer group. The choice moves every number.
• How are dividends and splits handled? Returns should run off an adjusted price series, or a 3-for-1 split will masquerade as a 67% crash.
• How large is the sample? One event’s an anecdote. A few hundred, tested carefully, can be evidence.
• Was the result assessed statistically, or is it just a large-looking average? An abnormal return with no sense of its variability is half a sentence.
• Which confounding events were considered? Honest work names the other things happening in the window and argues why they aren’t the driver.

Run a live headline through that list and most of them thin out fast. “Stock soars 9% on upgrade” says nothing about the market that day, the benchmark used, or whether 9% is even unusual for a name that swings that hard on an ordinary Tuesday.

Turning the framework into a habit

The real payoff from an event study is a lens you carry into every reaction you read. Once you’re in the habit of asking what the benchmark expected before judging a move, the raw percentage stops running the show. A +3% day means one thing against a flat market and the reverse against a market that itself ran +3.5%.

You don’t need software to begin. A benchmark, a beta, an event window, and honest arithmetic will separate most of the signal from most of the noise. The discipline is in refusing to let the biggest candle win the argument on its own. Learn the pattern. Ride the trend. Keep the gains.

Educational content only. Not investment advice. Trading involves risk. You are responsible for your decisions.