How much does news actually move price? A method for measuring it

"That news moved the market" is one of the easiest claims to make and one of the hardest to support. Most attempts fail on the same handful of details: an ambiguous start time, a horizon chosen after seeing the chart, silently dropped observations, and no baseline to compare against. Each of those turns a measurement into an anecdote.

Fix the two endpoints before looking at anything

An event study needs two prices and a rule that produced them. A workable definition, written down before any data is pulled:

1. Base: the last candle close at or before the publication timestamp, and no more than one bar old.
2. Target: the first close at or after publication plus the horizon, and no more than one bar late.
3. Return: (target - base) / base, expressed in percent.
4. Timeframe: the finest stored resolution that can resolve the horizon, never coarser than the horizon itself.

The "no more than one bar" constraints are what stop the method from quietly lying. If the nearest close before publication is nine hours old, the base is not measuring the pre-event price and the row should be marked unusable rather than computed anyway. Keeping both endpoints and the timeframe alongside the result means the arithmetic is recomputable by anyone, which is the property that makes it a measurement rather than a claim.

Choose horizons in advance

Horizons must be fixed before the data is examined, because a horizon selected afterwards is a free parameter that guarantees an interesting result. A standard set — +15m, +1h, +4h, +24h — covers the reflex move, the settled move, the session move and the overnight repricing.

The right-hand column is not a footnote. At +24h the measured return contains every other thing that happened in a day, which is why long horizons produce large numbers and weak inference. The short horizons are cleaner attribution and noisier prices. Neither is the correct one; reporting both is.

Unmeasurable is a result, not a reason to delete

This is the step where most published event studies go wrong, and it is invisible to the reader. Free data feeds commonly store only daily candles for equities, FX and commodities. A +15m horizon on those assets can never be resolved. If those rows are silently dropped, every published aggregate tilts toward the assets that happen to have intraday history — usually crypto — and the result describes the data coverage rather than the market.

The fix is to publish the state. Every bucket should report the number scheduled, the number measured, the number unmeasurable and the number still pending, with the three summing to the total, plus the reasons anything could not be measured. A study that shows only its successful measurements is showing you a filter.

We publish ours on the calibration page, including the measurements that could not be taken and why. At the time of writing the majority are unmeasurable, almost entirely because the intraday candles do not exist for the non-crypto assets — which is exactly the bias described above, stated rather than hidden. The same figures are available over the public API.

Without a baseline, a number means nothing

Suppose the average absolute move after high-impact headlines is 1.4% at +1h. Is that large? The question is unanswerable until you know what the asset does in an average hour with no event attached. If the unconditional hourly move is 1.1%, the event added very little. If it is 0.3%, the effect is substantial.

Two honest ways to supply a baseline: sample random timestamps from the same period and compute the same statistic, or compare buckets against each other rather than against zero — do headlines scored 8 produce larger moves than those scored 5, measured identically? The second is weaker but requires no additional data, and it directly tests whether a scoring scale is calibrated, which is usually the question that matters.

Report the distribution, not the mean

Post-event returns are heavy-tailed. A mean is dominated by a handful of large observations, and a mean absolute move conceals whether the reaction had a consistent direction. Report the median, the interquartile range, and the share of observations in each direction. A bucket where 52% of moves are positive with a wide spread is telling you something very different from one where 78% are positive with a narrow one, even if the means match.

• Median return and median absolute return, per horizon.
• The 25th and 75th percentiles, so the spread is visible.
• The count in each bucket, always next to the statistic.
• The share of measurements resolved, so coverage is not hidden.

What a measured reaction is not

Even a careful study establishes association, not causation. Events cluster, markets anticipate, and the headline that appears at 14:31 is often the third-fastest report of something already trading. A measured post-event move is a description of what followed a timestamp on the assets the event was tagged with — nothing more, and it is not a strategy.

That constraint is exactly why the method is worth the trouble. A number produced under a published rule can be checked, argued with and shown to be wrong, which is the standard applied to any track record in how to check whether a signal record can be falsified. If you would rather run the analysis on your own data, the scored event stream is available through the API, and how an impact score is built explains what the fields mean before you aggregate them.