What is an event study? How we test whether news moves markets

Every result on this site comes from the same method: an event study. It is a standard tool in economics and finance for answering a simple question — when a specific event happens, does a market react? This page explains, in plain language, how we run one, and why a few unglamorous details (a baseline, a significance test, a correction for testing many things at once) are what separate a real finding from a story.

The question an event study answers

Suppose someone tells you "the 10-year Treasury yield jumped after CPI." Maybe it did. But to know whether CPI actually causes a reaction, you need to answer two harder questions:

1. Did it move more than it would have on an ordinary day?
2. Did it move in a predictable direction, or was it just as likely to go the other way?

An event study is the disciplined way to answer both. You gather every instance of the event, measure the market's move around each one, and compare that to what "normal" looks like.

Step 1: define the event and the window

We take an event — say, every CPI release — and pull its exact release dates from the source agency. For each release we measure the market's change over a fixed window: the "release day" is the move from the prior day's close to the release day's close. We also look at a 5-day window to see whether any reaction lasts.

Crucially, the window starts at the prior close, before the number is public. That keeps the test look-ahead protected — we never accidentally use information that wasn't available when the event happened. Look-ahead leakage is one of the most common ways backtests fool themselves, and guarding against it is non-negotiable.

Step 2: build a baseline

This is the step most casual analysis skips, and it is the most important. A move of "8 basis points" means nothing on its own — you have to know what a normal move looks like. So for each market we compute the baseline: the distribution of moves over all trading days, not just event days. If the 10-year yield moves about 4.3 basis points on a typical day, then an 8-basis-point move on CPI day is roughly twice normal. Without the baseline, every event looks dramatic; with it, you can see which ones actually stand out.

Step 3: measure two different things

We deliberately separate the two questions, because they usually have different answers:

• Direction — the average signed move. If hot prints push a yield up and cool prints push it down, these cancel, and the average lands near zero. We test the average against zero with a t-test. A near-zero average with a high p-value means the direction is not predictable from the event.
• Size (volatility) — the average absolute move, regardless of direction, compared against the baseline. This asks whether the market moves more than usual, and we test it with a bootstrap. A ratio well above 1.0 with a low p-value means the event is a genuine volatility event.

Across our tests, the recurring pattern is: some events reliably increase the size of the move, but almost none give you a reliable direction. That distinction is the single most useful thing an honest event study reveals.

Step 4: a significance test on every claim

"It moved" is not enough — small samples throw up patterns by chance. So every result carries a p-value: roughly, the probability of seeing a result this strong if the event actually had no effect. A low p-value (we use 0.05 as the line) means the effect is unlikely to be a fluke. A result sitting right on the line is reported as not established, not rounded up to a win.

Step 5: correct for testing many things at once

Here is a subtle trap. If you test one event against one market, a p-value of 0.05 is meaningful. But if you test dozens of event-and-market combinations, some will clear 0.05 purely by chance — run twenty coin-flip experiments and one will look "significant." To guard against this, we apply a multiple-testing correction (the Benjamini-Hochberg false-discovery-rate method) across our whole grid of tests. A result is only called robust if it survives that correction. This is why some numbers that look elevated on an individual page are still labeled "not robust": they don't hold up once you account for how many tests were run. You can see the full corrected grid on our methodology and full-results page.

Why this matters

The point of all this machinery is to be honest about what the data does and does not show. Most market commentary asserts a reaction and a direction with total confidence. An event study, done properly, usually delivers a more modest and more trustworthy message: this event makes the market move more than usual, but you can't tell which way. That is not a limitation of the method — it is the method telling you the truth.

See it in action

• Does CPI move the 10-year Treasury yield? — a clear volatility effect with no predictable direction.
• Do weekly jobless claims move Treasury yields? — a confident null, thanks to a large sample.
• How we test — and the full results grid — every event and market in one corrected table.
• Glossary — definitions for the terms used above.

Historical statistics for informational purposes only, not financial advice.