Do shares that gained most overnight in one month keep gaining most overnight the next month?
Study 57 · Changed a rule or plan · recorded 16 October 2026 · answer: Yes · full paper with figures
Correction, 19 October 2026. This reading was decided by the clipped fallback measure, whose false-alarm rate was not checked on this study's own design. The plain measure would have read the pattern as a closing-price artefact, so the reading rests on the clipped measure, which was untested here.
Correction, 8 November 2026. Corrected 7 November 2026. The answer said the most heavily traded third's overnight difference was above zero, but only by the clipped measure, whose range cleared zero by 0.03 points. From 31 October 2026 that measure no longer decides two-group comparisons in new designs, so the sentence gives the plain average, whose range holds zero. The overall Yes remains the clipped measure's reading; as noted on 18 October, the plain measure would read the pattern as a closing-price artefact.
The question
A share's daily return splits into an overnight part, from the close to the next opening price, and a daytime part, from the open to the close. Published work and our own scans of 2013-2017 found the overnight part persisting month to month while the daytime part gives it back, leaving no full-day premium; we tested that on 2018-2022.
Rules fixed before the result
Fixed in advance: at least 48 months with 750 shares each; 1.1 points of monthly return, a quarter of the 2013-2017 difference, as the size that matters; and a full-day difference whose range lies within 1.1 points of zero as no premium. The plain average decided unless its 95% range was wider than 1.54 points; it was 1.67 wide, so the clipped average, the measure we fall back on, decided, and no minimum or bar was changed.
Data and method
Each month we ranked shares priced at 5.00 or more in US currency by that month's summed overnight returns and compared the top fifth with the bottom fifth over the next month, on overnight, daytime and full-day returns, averaged over 58 months with 95% ranges. Two checks were fixed in advance: the two parts must multiply back to the full-day return on one price-adjustment basis, and dropping days of a cash payout or share split must move the difference little.
The answer
Yes. In 2018-2022 the top fifth beat the bottom fifth by 3.8 percentage points of summed daily overnight returns over the next month (95% range 3.1 to 4.5), wholly above the 1.1 points set in advance; the plain average gave 4.3 (3.5 to 5.1), against 4.4 in 2013-2017. The daytime difference was -4.1 (-4.8 to -3.4), giving the gain back, and the full-day difference -0.3 (-1.0 to 0.4) was inside its band. Among the most heavily traded third the overnight difference was 0.6 points by the plain average, with a range from -0.05 to 1.2 that holds zero, so its survival among the most traded shares is not settled. The two parts multiplied back to the full-day return on every share-day, and without days of a cash payout or share split the overnight difference was 3.8 (3.2 to 4.5).
Main estimate
0.038 fraction (95% range 0.031 to 0.045), from 58 observations
What it changes
By the rule set in advance, the overnight pattern enters our record as a finding within the development data, and the planned study of big one-day moves will measure from the opening price. The two other questions from the same exploration, on the day a share first trades without its payout and the day before a market holiday, keep their place with no added check. The plan expected this answer with medium confidence, at 2 to 3.5 points.
What remains uncertain
These years' prices were read before by studies eighteen and thirty-two for other questions, so this is a check within our development data, not a fresh test. The data carry one adjusted price basis from the vendor and no opening-auction prices or quoted spreads, so a stale or one-sided closing price could mimic part of the pattern; among the most heavily traded third the plain average's range reached below zero (-0.05 to 1.2). The months are read as independent.
How it was made
Run took 93 minutes; independent code review: approve; the recorded calculation reruns to the same result.
Fingerprints
plan a3794155b94b2407212420e50e82ffc8648221938979f42bbbfb81d80456be52 result 0559970844205aece3857df6f6e6ebfba5200ebf5b552ad9047d37c645754d29
Exploratory work on development data; not a performance claim.