Ethereum's 30-Minute Continuation Estimate Was Positive. The Test Still Failed.

The result in one sentence

After unusually large positive one-minute ETHUSDT moves, the mean price response over the next 30 minutes was +2.99 basis points in the 2026 historical validation sample. But the one-sided 95% lower confidence bound was −2.77 basis points, below the strictly positive bound required by the frozen rule. The completed experiment closed DID_NOT_MEET_CRITERIA.

That decision is not a claim that the underlying effect is absent. It means this particular test did not establish a positive mean response with the specified confidence requirement. I retained the positive estimate alongside the failed confidence criterion.

Where the question came from

Alan Ronald Chernoff’s 2026 study of Bitcoin and Ethereum price discovery examines several high-frequency market relationships using one-minute Coinbase data from 2022–2025. One finding in its Table 11 was narrower than the paper’s overall thesis: following the top 1% of extreme positive ETH moves, the mean 30-minute forward return was +4.76 bps across 10,611 reported observations (paper-reported p = 0.041).

The paper also considers BTC, negative shocks and other horizons. The present evaluation does not independently test that whole set of findings. I selected one question: does the ETH positive-extreme 30-minute response remain supported by the specified inference rule in a separate 2026 Binance sample?

The data venue, period and inference method differ from the paper. In particular, this article does not claim to reproduce the source paper’s original significance test.

Defining an extreme move and its response

For consecutive one-minute closes, the research calculates the log return log(close[t] / close[t−1]). Stage A fixed the nearest-rank 99th percentile of absolute one-minute log returns at 0.003380992401538842 (approximately 0.3381% on the log-return scale).

A Stage-B event occurs when a one-minute log return is positive and at least that fixed threshold. The 30-minute response is log(close[t+30] / close[t]): it starts at the trigger minute’s close, not at a hypothetical earlier trade price. An event whose forward response would extend beyond the sample’s endpoint is not evaluable. If two events have overlapping 30-minute windows, both are retained. Their returns are not independent observations by construction.

The study reports an equally event-weighted mean of those responses. There was no executable trading rule, position sizing, cost deduction or profit-and-loss evaluation.

Two stages, with different evidential roles

Stage A: same-era cross-venue calibration, 2022–2025. Using Binance ETHUSDT minute bars overlapping the original paper’s calendar period, Stage A measured 10,622 positive events and a +4.6054 bp mean 30-minute response. Its HAC standard error was 1.7895 bps, giving a +1.6620 bp one-sided 95% lower bound. It fixed the numerical extreme-event threshold carried into Stage B. Because this is calibration/same-era cross-venue reproduction, not an independent later-time test, its terminal scientific classification was INDETERMINATE; no Stage-A PASS/FAIL criterion had been frozen.

Stage B: separate historical-window validation, January–August 2026. The frozen Stage-A threshold was reused without recalibration on Binance ETHUSDT minute bars from 2026-01-01 00:00 UTC up to, but excluding, 2026-09-01 00:00 UTC. The source requirement was strict: exactly 349,920 consecutive calendar minutes, with no missing-minute repair, alternate-source substitution or shortened period. The retained source inventory reports eight official monthly Binance archives and eight matching published checksums. The admitted dataset met the specified complete-contiguous requirement.

These are different samples with different evidential roles. The table is descriptive: it is not an additional test of a change between periods.

The frozen decision and why it was not met

The primary question was whether the event-weighted mean 30-minute forward log return after positive extreme events was positive and had a strictly positive one-sided 95% lower confidence bound. The inference method was UTC-day cluster-score Bartlett HAC, lag 20, with at least 30 evaluable positive events required for inference.

Stage B had 1,485 evaluable positive events across a 243-day sample. Its reported primary quantities were:

The selected result artifact retains the full recorded precision. The lower bound and its corresponding upper bound form an approximate central 90% confidence interval, [−2.77, +8.75] bps under this specified HAC procedure. It is not a two-sided 95% interval. Zero lies inside it.

In ordinary language: the observed average was positive, but the uncertainty was too large to establish the strictly positive lower bound required for validation. This neither proves a positive effect nor establishes that the effect is exactly zero, negative or absent. The study did not respond to this outcome by changing the horizon, selecting a different threshold or adding another hypothesis test.

The calendar chronology matters

The original paper was published in September 2026, after the Stage-B market period had ended in August. Stage B was designated after its historical interval ended. Its scientific rules were frozen before the Stage-B minute-level data capture, but not before those market observations occurred. This is a historical out-of-sample evaluation with a before-capture freeze, not prospective preregistration or forward monitoring.

The Stage-B scientific decision addendum was frozen on 13 October 2026. The study also acknowledges that earlier research covered an overlapping 2026 ETHUSDT period; whether relevant earlier results had been inspected was not conclusively established. The absence of prior exposure to all overlapping-period information cannot be certified from the retained chronology.

The completed run reached REPORTED / COMPLETE. Strategy Lab’s separate offline terminal verifier was reported to pass; the public artifact copies and their hash checks below permit limited inspection, not an outside replication of the evaluator or its input dataset.

What the evidence cannot settle

The HAC inference allows for some within-day and across-day dependence, but its validity still depends on assumptions. Extreme events can cluster in time; their forward windows can overlap; and the fixed lag-20 bandwidth may fail to capture longer-range dependence. There were 243 daily clusters, and the estimated standard error in Stage B was large relative to the mean. The design may have limited power to detect a modest positive response. Therefore a failed confidence criterion is not a reliable measurement of the effect’s absence.

Neither the paper’s price-response statistic nor the present Stage-B mean is an abnormal return relative to a causal or risk-adjusted counterfactual. The experiment did not identify a market mechanism, establish causality, or model a realizable trading entry, exit, fees, spread, slippage or latency. It supplies no evidence of executable net trading profitability.

Why preserve this result?

The 2026 sample gave a positive estimate with a lower confidence bound below zero. I kept both in the record and closed the study DID_NOT_MEET_CRITERIA. Changing the horizon or threshold afterward would have created a different test; treating this result as proof of no effect would overstate it.

Implementation and evidence identities

The frozen implementation commit was d5ca5c31ae6a781e20ef7259915f50605d81f613. The Stage-B ResearchSpec SHA-256 was 6a81be7ca363aaf0a26871c9e6ac80dc39eef4021492c69c43330b19959ec8bd, and the executed ResearchPlan SHA-256 was 794f195b3e9061295bde457763b2f0477ee4cd1595d361a3bdf3a09422bd9ac3. The admitted canonical bars had SHA-256 4d0ee8b371c8cae6c9c6347778475d94a43fc1715d7fb935b119ebe94f22eb00.

The Stage-B result artifact is SHA-256 d776ceab5b6afd993c7367482464a779ecca136a14f356ab43965d2728411354. These identifiers bind retained contents, not an independently proven freeze date or scientific validity. The evidence downloads below are selected byte-identical artifacts rather than a complete reproducible dataset or a public copy of the private research software.

For the general approach to frozen research contracts, see Methodology. For a separate example of why a positive metric need not pass a frozen rule, see Why a +20.5% annualized return still failed the test.