Do one-week tests hold their false-alarm rate at longer windows?
Study 10 · No reading (stopped by its own checks) · recorded 27 September 2026 · full paper with figures
Correction, 30 September 2026. Technical wording rewritten in shorter sentences under the readability rule; the meaning is unchanged.
Correction, 19 October 2026. This text was reviewed after the study finished, in a catch-up of the earliest studies; the result, outcome and dates are unchanged.
The question
On the same simulated draws, we compare each test's false-alarm rate at 10, 21 and 63 trading sessions after the event with its own rate at five sessions, against a target of 2.5% in each tail. We also ask whether the windows of different companies that overlap in calendar time move together and so inflate the tests' false alarms.
Rules fixed before the result
The rules for each possible answer were written down and fixed before any result was seen.
Data and method
We ran the tests chosen for one-week windows 40,000 times on data built to contain no effect, so that every rejection is a false alarm, at 5, 10, 21 and 63 sessions after the event. Each longer window's false-alarm rate was compared with the same tests' rate at five sessions on the same draws. We also measured how far the returns of different companies move together when their windows overlap on the calendar.
What it changes
The planning rules for one-week windows stay as recorded. No rule was extended to the longer windows of two weeks, one month or three months.
What remains uncertain
Whether the one-week tests keep their false-alarm rate at two weeks, one month and three months is still unknown. The tests do not treat dependence across companies whose windows overlap on the same dates, the uncertainty this study was built to measure. Only one pool of stock split events was used, the same one earlier methods studies used.