Intraday losers beat intraday winners in the last half hour of trading. A published anomaly, sample ending 2019. We tested it on 2020–2026 under a pre-registration written before the run. New paper, free PDF below; replication archive on Zenodo, submitted to arXiv and SSRN.
Baltussen, Da and Soebhag (2025) documented an end-of-day reversal in U.S. stocks: the names that fell between the prior close and 3:00 p.m. outperform the morning's winners over the last half hour. Their sample ends in December 2019. Anomalies usually shrink once their sample ends, so the natural question is whether this one did.
We tested it on point-in-time S&P 500 members from 2016 to October 2026, with one deliberate cut: about 80 large, heavily traded stocks whose recent intraday data we had examined in earlier, unpublished work were excluded, so that nothing we had already seen could shape the result. The list is not released. The paper says why, and what the cut costs.
The registration came first
The estimators, the thresholds and the wording of each verdict were fixed before the run. A reproduction test came before anything else: the 2016–2019 data had to show the published sign at t ≥ 2 before any statistic linking the sort to the outcome was computed on later data. Only after that result was committed was the confirmatory script executed, once, on all periods, under a lock that blocks a second execution.
The design, the pre-registration and the decision rules were drafted by an AI model and adopted by us. An AI agent then ran an adversarial audit of the design and found two defects that would have let the paper overstate a change. Both were fixed before any primary-sample outcome was seen. Appendix B lists them.
What the registered test says
Measured to the last trade before the closing auction, the dollar-volume-weighted reversal is 3.13 bps a day in 2016–2019 and 3.05 bps a day in 2023–2026. The registered no-weakening verdict holds, and the one-sided 95% bound on the decline is 61% of the 2016–2019 effect. A proxy sample that anyone can rebuild from public data, added after the run and reported as descriptive, gives 2.81 and 3.86 bps.
What it depends on
Three things, and the paper puts each on the table.
The window. 44% of the later sessions fall in 2020–2022, where the reversal is indistinguishable from zero. Over all sessions from 2020 it is 1.69 bps, 46% below 2016–2019, and the bound on the decline becomes 91%. The 2023–2026 estimate rests on 2023, the strongest year; 2024–2026 alone cannot be told apart from 2020–2022.
The end point. Ending at 3:55 instead of the last trade before 4:00 raises the bound to 90%. The closing auction itself carries a reversal in every period, and it grew.
The sample. On the full index the reversal falls from 3.40 to 0.74 bps. The fall is short of the registered bar of t = −3, so under the registration it does not count as a weakening, but equal weights, square-root weights or dropping the five largest stocks each day restore it. The weakness sits in the very largest names, where dollar-volume weight became far more concentrated after 2019.
Mechanisms: none passes
Four pre-registered mechanism tests, one Holm family: a market component, an idiosyncratic component, the weekdays that gained same-day index options, and single-stock leveraged-ETF launches. None passes. The market component moved from slight continuation to reversal, in the direction a dampening story predicts, but not past the threshold. The leveraged-ETF test could not be judged: too few treated stocks sit in the sample, and the paper says so rather than reading a sign into twelve names.
One by-product may be useful elsewhere. We release a table of first single-stock leveraged-ETF launch dates for S&P 500 members, each sourced from issuer documents, because the aggregator we first used placed nine launches up to ten months too early.
The largest number in the paper
On the full index in 2023–2026, intraday winners beat intraday losers from 3:30 p.m. to the next open by 13.7 bps. It is not a split or dividend artefact, it rests on the weights, and it is not in the primary or proxy samples. We offer no mechanism and do not test why it occurs. It is reported because a reader who sees only the half-hour result would miss it.
Where to read it
The full paper, fifteen pages with the decision rules, the audit record and every post-run addition marked as such, is on its abstract page as a PDF, with BibTeX. The replication archive, with the scoring code, the point-in-time membership file, the launch-date table and public copies of the pre-registration documents, is on Zenodo as 10.5281/zenodo.24003093. The paper is submitted to arXiv and SSRN; links go on the abstract page when they exist.
Two sentences from the paper that we mean exactly as written. It makes no claim net of costs and does not propose a trading strategy. The authors do not trade a strategy based on the end-of-day reversal.
Keywords: end-of-day reversal, intraday returns, pre-registration, out-of-sample replication, leveraged ETFs, market anomaly.