What Is an Event Study? How Fed, CPI and Jobs Day Moves Are Measured

An event study measures how an asset's price behaves around a dated event, lining up the sessions before and after it against how it behaves on ordinary days.

Scheduled announcements are the easiest market events to study, because their dates are known in advance. This page covers what an event study contains, why the event date decides the answer, how a surprise differs from a direction, and what a sample of fifteen events can support. It is descriptive throughout.

What an event study measures

An event study has two halves. The first is a set of dates sharing one definition: every Federal Reserve decision day, every Consumer Price Index release, every Employment Situation release, the monthly report traders call the jobs report. The second is a rule for stacking those dates so returns from different years can be compared. That step is called event time: the axis reads minus five sessions, day 0, plus twenty one sessions, with one observation per event at each point.

The comparison is what turns a list of dates into a study. Returns around the event sit beside returns over the same number of sessions with no condition attached, the unconditional baseline. The question is never whether a market moved after CPI, because some move is the norm, but whether it moved differently from an ordinary stretch of the same length. The method goes back to Fama, Fisher, Jensen and Roll in 1969, who used it on stock splits.

The parts you have to specify

Every choice below changes the answer.

• Event definition and dating. A policy decision is dated on the day the statement lands. A data release is dated on the day it is published.
• Base date and day 0. The base is the last close strictly before the event, day 0 the first close on or after it. The one session figure therefore spans the announcement.
• Horizons. Counted in trading sessions: one, five, twenty one and sixty three stand in for a day, a week, a month and a quarter.
• Baseline. The same asset and horizon over its whole stored history, with no condition attached.
• Statistic. A difference of medians against that baseline, a hit rate (the share of events that closed positive), and a bootstrap interval around the difference.
• Path. The cumulative return from minus five sessions to plus thirty, showing whether a move arrived before the announcement, on it, or over the weeks after.

Why the event date decides the answer

The Federal Open Market Committee announces a decision, and the new target range takes effect the following day. The statement of 10 December 2025 lowered the target range to 3-1/2 to 3-3/4 percent, and its implementation note set that change effective 11 December 2025. Public rate series can carry either date, and the FRED target series changed convention partway through: every change up to 14 December 2016 is stamped on the statement day, and every change from 16 March 2017 on is stamped on the day the new range took effect, the session after the statement. A study keyed to that series therefore measures the session after a modern decision and drops the decision-day move.

That is the correction Kresmion made while rebuilding its macro event table: decision days now come from the Federal Reserve's statement calendar, no longer from the effective date in the rate series. The two conventions differ by exactly the session that carries the announcement. The dot plot is dated the same way, on the statement day that carries it.

Surprise and direction are two different things

A quarter point cut everyone saw coming and one nobody saw coming are the same action and different events. Markets price the anticipated part in advance, so the reaction tracks the part that was not, a decomposition associated with Kuttner's work on policy surprises. Fed funds futures are the usual measure of anticipation, and they are licensed data, so a free proxy helps: the change in the 2-year Treasury constant maturity yield across the decision, published daily on FRED as DGS2.

Bucketing on that proxy splits one action into different events. Sorted at the 33rd and 67th percentiles of the distribution over 219 FOMC decision days since 2000, the cut points fall at minus 3 and plus 2 basis points (measured 2026-09-08). Of the 33 quarter point hikes, 16 came with a 2-year yield that rose by at least 2 basis points, 6 sat in the middle, and 11 with a yield that fell by 3 basis points or more even though the target range went up. Of the 15 quarter point cuts, 7 read dovish, 1 in line and 7 hawkish. Of the 19 cuts of 50 basis points or more, 12 read dovish, 2 in line and 5 hawkish. Odds priced before a meeting are a separate lens, covered in how prediction markets price the Fed.

A worked example: 887 events and 44,082 outcome rows

The rebuilt study covers 887 events since 2000, read from the public payload generated 2026-09-14 at 03:10 UTC: 340 CPI release days (2000-01-14 to 2026-09-11), 328 Employment Situation release days (2000-01-07 to 2026-09-04) and 219 FOMC decision days. The decision days split into 145 holds, 33 hikes of 25 basis points, 7 hikes of 50 basis points or more, 15 cuts of 25 basis points and 19 cuts of 50 basis points or more. The most recent hold in the set falls on 2026-07-29. The most recent quarter point cut falls on 2025-12-10.

Crossed with 19 assets and four horizons, the payload holds 44,082 outcome rows, fewer than the full grid because an asset contributes nothing to an event that predates it. The assets are the ETFs SPY, QQQ, IWM, TLT, IEF, HYG, GLD and USO, the VIX index, whose outcomes are index point changes and not percent returns, the continuous futures ES, NQ, ZN, ZB, CL, NG, GC, SI and HG, and the dollar index DXY, with ETF closes ratio-spliced so each fund carries its own history back to its launch, which is why several assets start well after 2000. Kresmion publishes the result as a public page, the macro event impact tool: the conditional distribution beside the unconditional baseline, a bootstrap interval on the difference of medians, hit rates, the minus five to plus thirty session path, the instance table, and the next scheduled FOMC, CPI and Employment Situation dates.

Take one cell. Across the 340 CPI release days, the median SPY return over the single session spanning the release is 0.1339 percent, on that same 2026-09-14 03:10 UTC build. Set against the unconditional median for SPY at the same horizon, the difference is small and the bootstrap interval around it spans zero: by median return, the session containing a CPI print is not distinguishable from an ordinary session.

What a sample of fifteen can support

The count beside a figure carries as much information as the figure. A category with 15 or 19 members reads as a short list of cases, and one unusual instance can set its median. The tool flags any cell built on fewer than ten instances as anecdotal. The bootstrap interval reports the width of the plausible range around the difference, which a single point estimate hides.

Two further limits shape how the counts read. Windows overlap: FOMC meetings sit about six weeks apart, which is about thirty sessions, so consecutive 63 session windows share trading days and those observations are not independent. Regimes differ too, so a category drawn from one cycle describes that cycle, as in what happened to stocks, bonds and gold after Fed rate decisions. Holds count as events, and at 145 of the 219 decision days they are the largest category in the set.

Key takeaways

Frequently asked questions

Does an event study predict how markets will react to the next Fed decision?

No. It describes what already happened around past events of the same type, and it does not forecast the next one. Each category mixes different inflation, growth and positioning conditions, so the distribution is a historical record.

Why are horizons counted in trading sessions instead of calendar days?

Because calendar windows stretch and shrink with weekends and holidays, and macro events cluster around both. Twenty one sessions is the same amount of trading in every instance, while a thirty calendar day window holds however many sessions its weekends and holidays leave.

What is the difference between an event study and a backtest?

An event study measures the distribution of returns around a defined date against a baseline. A backtest simulates a rule that enters and exits positions, so it must also account for transaction costs, slippage and sizing. An event study makes no entry or exit decision, so its output stays a distribution of returns.

Why use the 2-year Treasury yield to measure a policy surprise?

The 2-year yield is the maturity that moves most with the expected policy path, and its change across a decision captures how much the announcement moved that path. Fed funds futures are the more direct measure, but they are licensed data. The proxy is coarse, and it is stated as a proxy.

How many events does a category need before its summary means anything?

There is no clean threshold, which is why the count is printed next to every figure. Below ten instances a cell reads as a short list of dates, and even at 15 or 19 a single unusual event can set the median. This page is information, not investment advice.

--- Source: Kresmion macro event study, public payload version 2 generated 2026-09-14 03:10 UTC, https://kresmion.com/api/tools/macro-event-impact ; FOMC implementation note of 10 December 2025, https://www.federalreserve.gov/newsevents/pressreleases/monetary20251210a1.htm ; FOMC calendars, https://www.federalreserve.gov/monetarypolicy/fomccalendars.htm ; FRED DFEDTARU, for the series dating convention, https://fred.stlouisfed.org/series/DFEDTARU ; FRED DGS2, https://fred.stlouisfed.org/series/DGS2 ; Fama, Fisher, Jensen and Roll (1969), https://www.jstor.org/stable/2525569 Kresmion Research.