Research note, sixth in the series. Data: one-minute bars of the SPY, QQQ, IWM and DIA ETFs, with the maximum history available per instrument. None of this is investment advice.
The previous note inaugurated the replication arc with an anomaly that died by decay. The present replication examines a more recent published case of finer mechanics: the market intraday momentum documented by Gao, Han, Li and Zhou in 2018, according to which the sign of the first half hour of the session predicts the return of the last half hour. The outcome, stated upfront in one line, is less kind than that of the previous case: replicated on the four main index ETFs, the effect net of costs is negative in every cell, and the only gross vestige of the phenomenon lives in the small instrument, before the paper’s publication, and below its friction cost. The note is divided into four sections: the paper and the replicated strategy; the net result; the gross decomposition, which is where the real story is; and two inference lessons the exercise left behind, worth more than the verdict itself.
The paper and the replicated strategy
The original finding holds that the return of the first half hour (9:30 to 10:00) positively predicts that of the last one (15:30 to 16:00), with a proposed mechanism of two sources: informed investors who resume their trading toward the close, and the mechanical rebalancing of leveraged ETFs, whose footprint is proportionally larger in small-company indexes. The implied strategy is transparent: observe the sign of the first half hour, take a position in that direction at 15:30 and close in the closing auction. The replication implemented exactly that specification, with no window variants nor adjustments, on SPY, QQQ, IWM and DIA with one-minute bars and the maximum history the source offers per instrument (since 2004 for IWM and DIA, 2011 for QQQ and 2014 for SPY), with a base round-trip cost of 2 basis points per traded day. It is pertinent to note that the original paper’s sample covers approximately 2004 to 2013, so that the replication contains that period in the two long-history instruments and adds more than a decade after the publication.
The net result
Net of costs, the strategy loses in all four instruments and in every sub-window examined. In the common panel from 2014 to 2026 the average is −2.30 basis points per day, with a day-clustered t of −5.17, which annualized is around −5.8%, and the proportion of winning days stays between 43% and 49%, below the coin flip in all four cases. The sign of the effect is, in fact, the opposite of the published one: SPY and DIA exhibit a slight intraday reversal (the first half hour predicts the last one with a negative and significantly estimated coefficient), not momentum. The replicated hypothesis is not merely unconfirmed; at today’s prices, executing it amounts to paying to systematize a small error.
The gross decomposition: the story in three layers
The net, however, tells less of the story than the gross, and the decomposition by instrument and by era is where the replication becomes informative. First layer: the effect barely exists, and where it exists is exactly where the paper’s mechanism predicts it. Only IWM, the small-cap ETF, shows gross intraday momentum, +1.40 basis points per day before 2018 (Sharpe 0.51), consistent with the footprint of leveraged rebalancing in small companies; SPY, QQQ and DIA do not show it in any era. Second layer: even at its peak, the effect is smaller than its cost. IWM’s +1.40 gross bp are left at −0.58 bp after the 2 bp round trip, and the sensitivity to an optimistic cost of 1 bp does not rescue it either. The effect, where it was real, was always a microstructure phenomenon below its own friction for an ordinary executor. Third layer: the decay after publication. After 2018, the year the paper appeared, IWM’s vestige falls from +1.40 to +0.08 bp per day, an almost total shutdown that replicates live the McLean and Pontiff pattern the previous note introduced. The complete sequence (a real effect in the predicted niche, below its cost, and arbitraged after being published) suggests that the market read the paper and ate the residue, without the retail reader ever having had anything to collect.
Two inference lessons
The first lesson was left by clustering. The panel of four correlated ETFs (cross correlations range from 0.85 to 0.95) produces a naive t of −8.21 that clustering by day reduces to −5.17: a 60% inflation in the statistic from treating as independent observations that are almost the same observation repeated four times. In this case both t values point to the same verdict and the error would have been harmless; the day they point to different verdicts, the difference between funding and archiving can live entirely in that 60%.
The second lesson was left by the correction for multiple comparisons, and it deserves to be put on record because it is a silent failure mode: the cells of this replication pass the two-sided significance filter, since a loss with a t of −5.17 is overwhelmingly significant, only in the direction contrary to the hypothesis. A pipeline that reads the significance mark without reading the sign would here celebrate a significantly losing strategy. The multiple correction protects against chance, though not against distracted reading, and no statistical threshold substitutes for the act of looking at where the effect points.
Limitations
The replication confines itself to the paper’s central specification (windows from 9:30 to 10:00 and from 15:30 to 16:00, position on the sign) and explores no window variants nor entry delays, by virtue of the base effect having turned out negative or absent, and multiplying specifications over that starting point would only open the door to false positives. The available one-minute history truncates SPY to 2014 onward, so that the paper’s sample period is only fully visible in IWM and DIA, which are, fortunately for the exercise, the instrument where the effect lived and one of its counterexamples. The 2 bp cost is a reasonable estimate for liquid ETFs traded twice per day, and the conclusion is robust to half that figure; executors with an effective cost close to zero face in any case a post-2018 gross of +0.08 bp, which does not pay for the infrastructure of measuring it.
Conclusions
The four sections are settled. The paper proposed an intraday momentum with a plausible and localized mechanism. The net result of the replication is negative in all four cells (−2.30 bp/day, t −5.17), with the predictive sign inverted in the large indexes. The gross decomposition confined the phenomenon to its theoretical niche (small caps, before 2018, +1.40 bp) and showed it below its cost even there, with almost total decay after publication. The two inference lessons, the 60% inflation of the t from cross correlation and the two-sided significance read without its sign, are recorded as the net gains of the exercise.
The replication arc thus accumulates two different outcomes with the same moral: the announcement premium was real and is extinct; intraday momentum was, for the common executor, illusory from the origin and decayed besides. In both cases, the academic publication marks the point after which the phenomenon ceases to be available, a reading that supports treating the publication date of a paper not as the beginning of an opportunity, but as the beginning of its end.
Series: pre-registered experiments on systematic equity signals. This program publishes methods and negative results; live signals are not published. Data: intraday prices (EODHD) are used under license and are not redistributed. None of this is investment advice.