Bitcoin Hourly Return Reversal: Frozen 2026 Holdout Test

The result in one sentence

The frozen BTCUSDT hourly holdout produced a slightly negative next-hour return coefficient, −0.0040, but its one-sided 95% HAC upper bound was +0.0286, above zero. The completed study therefore closed DID_NOT_MEET_CRITERIA.

That is not evidence that hourly reversal is absent in every market or period. It means this specific, predeclared BTCUSDT test did not establish the negative relationship it required.

What question did this test ask?

I tested whether the immediately preceding one-hour log return negatively predicted the next adjacent one-hour log return on Binance BTCUSDT spot, using a holdout frozen before acquisition.

I took the motivation from Xue and Zhou’s 2026 paper, Diagnostic Expectations and Cryptocurrency Returns. An owner-provided derived capture of its public publisher page reports negative short-horizon prediction from past one-hour returns for most coins, including BTC.

That capture is neither publisher HTML nor the paper’s full text. The present study did not obtain or reconstruct the paper’s complete method, market sample, coefficients, controls, or economic conclusions. A direct automated request for the publisher page returned HTTP 403, so the retained source record supports only that narrow motivating fact.

This is an original hypothesis validation, not a paper replication and not a trading-strategy backtest.

The test was fixed before its data were acquired

The study used 2025 BTCUSDT hourly archives only to build and check the parser, continuity checks, return construction, regression, HAC calculation, replay and evidence generation. Those development data were separate from the scientific holdout.

Before the 2026 data were acquired, the study froze:

Part of the contract	Fixed choice
Market and venue	Binance Spot BTCUSDT
Holdout	2026-01-01 00:00 UTC through 2026-09-01 00:00 UTC, end exclusive
Required data	5,832 contiguous one-hour bars from eight monthly official archives
Model	Intercept-inclusive regression of next-hour log return on prior-hour log return
Dependence adjustment	Bartlett HAC with lag 24
Success rule	Negative coefficient and a one-sided 95% HAC upper bound below zero

The method froze at 2026-10-19 17:32:14 UTC. The eight holdout archives and their publisher checksum sidecars were then acquired and hash-bound at 17:34:58 UTC. These timestamps record the before-capture boundary for this study. They do not establish that the January–August market observations occurred after the study was designed, nor that no one had seen them elsewhere.

How the result was calculated

For hourly close Cₜ, the study defined the completed-bar log return as:

rₜ = log(Cₜ / Cₜ₋₁)

It then estimated:

rₜ₊₁ = α + βrₜ + εₜ₊₁

Eligible pairs required each of the three closes used by the two returns to be inside the holdout. Missing, duplicate, malformed or nonpositive-price records would have stopped the study rather than being repaired or interpolated.

All eight monthly archives admitted successfully. They contained the required 5,832 contiguous bars, yielding 5,830 regression observations.

A negative estimate was not enough

The point estimate did lean in the hypothesized direction:

Quantity	Result
Intercept α	−1.94e−5
Reversal coefficient β	−0.0040
HAC standard error	0.0198
One-sided 95% upper bound	+0.0286
Left-tailed p-value	0.420
Frozen scientific outcome	DID_NOT_MEET_CRITERIA

The selected result artifact preserves the full recorded values.

The decision rule deliberately required more than a negative estimate. A claim of reversal required the entire one-sided 95% bound to lie below zero. Here, the upper bound remained positive. Under the frozen contract, the uncertainty was too large to establish the requested relationship.

The result was not rescued by changing the lag, selecting a shorter sample, adding filters, testing another horizon or converting the estimate into a trading rule.

What this result does and does not say

This study supports one bounded conclusion:

The frozen January–August 2026 Binance BTCUSDT holdout did not establish a negative next-hour relationship from the preceding hour’s log return under its specified OLS and HAC contract.

It does not establish:

that the source paper is wrong;
that no hourly reversal exists on BTC or elsewhere;
that a zero or positive relationship is true;
that a tradable strategy fails after costs, spreads, slippage or execution delay; or
that the study reproduced the paper’s data, estimator or wider claims.

The model uses adjacent completed hourly returns, with no positions, execution model, transaction costs, position sizing or profit-and-loss calculation. It is evidence about one conditional statistical relationship, not a recommendation to trade.

Evidence record

The sealed run was 20261019T210110Z-acdcc6bec2ac. Its result artifact is SHA-256 ac95653544d1414fe777d6a79af52233a8986b4607dd817f978cd3e10da6a93c; its study-owned semantic-replay artifact is SHA-256 1e7ae0bfc510656bf3bc8fd3d511555061832199f91807651d2db3304bf7c677.

The platform’s integrity check and the study’s semantic replay both passed before the handoff was prepared. The replay recomputes the result using the study’s recorded analysis implementation. Those checks establish that the selected sealed artifacts and recomputation agree with their recorded contract; they do not independently validate the source paper, market-data provider, inference assumptions or scientific interpretation.

The evidence links below are selected copies of those sealed artifacts. They do not republish the provider ZIP archives, the private retained study workspace or the complete source manifest.

A Frozen Bitcoin Hourly Reversal Test Found No Clear Signal

How to read these numbers

Hourly reversal coefficient

HAC standard error

One-sided 95% upper bound