Results of operations (earnings) · run dated 2026-08-14

How much does a stock typically move on the day a company reports earnings?

What the move actually looks like

These three are descriptive figures covering every filing in this category (11,235 of them), which is why their sample size differs from the validated figure below — that one is measured on the discovery slice and re-tested on two holdouts. No significance is claimed for the three above, and the pre-filing figure is shown against its own control window rather than as a finding in its own right.

The headline number

The typical move is +6.10% on the filing day itself, measured without regard to direction, across 11,241 filings (run dated 2026-08-14). The average direction over the same filings is -0.15%, which is close to nothing — large moves in both directions cancel out.

This is the dataset's positive control: earnings days are known in the finance literature to produce large moves, so if this measurement had come back small it would have meant the pipeline was broken rather than the literature being wrong. It also shows why direction and size are separate questions — the typical move is large while the average direction is close to nothing.

The figures

How this was measured

This category is the official SEC item code as the company filed it. No reading of the filing text is involved, so the category rests on filing metadata alone.

For each filing we compare the stock's actual return against what that specific stock would have been expected to return given how the market moved that day. The expectation comes from the stock's own prior trading history, fitted over a window that ends well before the filing, so the filing itself cannot influence its own benchmark. The difference between actual and expected is the abnormal return, and the figures above are averages of that difference, in percentage points.

Many categories are tested at the same time, so the results are corrected for multiple comparisons. Direction and size are corrected as separate families. The windows before the filing are corrected as a family of their own. Figures from different families are not comparable to each other.

Why the holdout columns matter

A pattern found by searching a dataset will often not survive being tested on data it was not found in. Every finding on this page was discovered on one slice, frozen as a written prediction, and then tested on two untouched slices: later filings the discovery step never saw, and a separate set of companies with no overlap. A finding appears here only if it held up on both. Findings that failed this test are kept in the record and reported as refuted rather than deleted.

Why the holdout number is often bigger — and why that is not good news

On most published findings the holdout figure exceeds the discovery figure. That is not evidence the effect strengthened, and we would rather say so than let it read as a selling point.

It is selection. Across all 154 hypotheses that were actually tested, the median holdout-to-discovery ratio is 0.461 — effects shrink out of sample, which is exactly what the winner’s curse predicts, because a finding gets picked for having looked large and some of that size was luck. Among the ones that replicated, the median is 1.638. Only the replicated ones are published. Conditioning on survival is what pushes the number above 1, not anything about the market. The holdout slices are also about a third the size of the discovery slices, so clearing significance there mechanically demands a large estimate; findings that shrank by the ordinary amount simply did not survive to appear on this page.

We checked the obvious alternative — that the most recent stretch of the sample was a more volatile market — and it does not carry the result: standardising absorbs roughly half of a pooled 1.13× gap, leaving most of the effect with selection and no step change in volatility at the split. So read the holdout columns as evidence that the direction and existence of an effect survived unseen data. They are not an estimate of its size. For size, the discovery figure is the conservative one, and the truth is plausibly below both.

Limitations you should read before citing this

• Averages, not predictions. These are historical averages across many filings. Individual filings vary enormously around them, and nothing here forecasts what any future filing will do.
• Survivorship. The company universe is built from current index membership, so it leans toward companies that still exist and are still in the index.
• Daily resolution. Reactions are measured on daily closing and opening prices, not intraday ticks.
• No expectations data. The dataset does not know analyst consensus, so it never claims a company beat or missed. Categories describe what a filing asserts, nothing more.
• Counts shift between runs. The universe is rebuilt each run and companies enter and leave, so event counts move even when nothing about the method changes. Every figure here is quoted with its run date for that reason, and effect sizes travel better than counts.
• US only, bounded window. One market, and a fixed multi-year window ending at the run date.
• Overlapping windows. Companies file often, so longer windows around different filings can overlap.
• The “quiet” window before a filing is not quiet. The two-weeks-before window is used throughout as the baseline to compare against, and 19.58% of events have another filing from the same company inside it. That pushes the baseline up, which makes it a cautious anchor for reaction claims and an unreliable one for any claim that nothing happens before filings.
• The later window measures volatility, not drift. The [+2,+10] window is nine days long, so it accumulates nine days of ordinary movement. Once each event is scaled by its own volatility, that window sits within a rounding error of the pre-filing baseline for every category with a usable sample, while the filing day sits far above it. Read a non-zero figure there as noise unless it clears its own baseline.
• Overnight and session pieces do not add up to the day. Each piece subtracts the same daily market-model constant that the daily figure already subtracts, so the two fall short of the total by a small fixed amount per day. Totals are unaffected and comparisons between the pieces are unaffected; their sum is not a breakdown of the whole.
• Filing-time slices exclude 384 events. Where a company filed more than once on the same day and the filings fall in different parts of the day, no single filing-time label fits, so those events are left out of the by-time slices rather than assigned to one arbitrarily. They are still counted everywhere else.

Cite this

Get the data

This page carries a single finding. The complete historical layer is a non-recurring, dated, versioned file. It holds every validated finding with the out-of-sample holdout columns, classification below the raw filing categories, the full aggregate grid, the data dictionary and the methodology.

A forward-looking version — the same classification applied to filings as they arrive — does not exist yet and has no date. If that is the one you would actually use, say so here. What people ask for is how the order of work gets decided.

Related questions

• What happens to a stock when a company lowers its guidance?
• What happens to a stock when a company raises its guidance?
• Does the SEC's earnings filing code tell you anything about how a stock will react?
• What happens to a stock when an earnings release describes results as weak?
• What happens to a stock when an earnings release describes results as strong?
• What happens to a stock when an earnings release reports mixed results?