Last Updated: 2026-10-26
The earlier report described a 1 August 2026 scan of 23,495 leaderboard entries across Hyperliquid, its vaults, OKX lead traders and Polymarket. It said 150 candidates received deeper analysis and 117 received the final statistical evaluation. That is not a test of all 23,495 traders' skill.
This page preserves the reported figures and the useful selection-bias question, while separating them from conclusions the record does not establish. The original row-level scan, exclusion ledger and exact execution revision are not linked in the retained article. The scores below have not been independently reproduced in this revision. They are historical reported outputs, not verified performance evidence or current rankings.
What the saved report says
Scroll sideways to see every column. Keyboard: focus the table region and use the arrow keys.
| Reported stage or output | Historical value | What remains to check |
|---|---|---|
| Leaderboard population | 23,495 | Distinct entries, scan coverage and duplicate identities |
| Candidates examined more deeply | 150 | Selection rule and per-candidate inclusion log |
| Candidates evaluated after gates | 117 | Exact exclusions, overlaps and missing records |
| PSR at or above 0.95 | 45 | Benchmark, return data, frequency and computation |
| DSR at or above chosen 0.60 threshold | 0 | Trial-count/variance assumptions and same candidate set |
| Largest reported DSR | About 0.069; companion article gave 0.0692 | Candidate identifier and unrounded result |
The prior headline said "zero passed" as though that established the absence of skill. At most, a verified version of this result would show that the evaluated records did not meet one chosen threshold under one specified model. It would not prove that nobody in the larger population could trade profitably.
The exclusion arithmetic needs a ledger
One version listed 32 concentration exclusions and 3 unusually smooth records, yet also reported 117 remaining from 150. Simple subtraction without overlap gives 115, not 117. Two overlapping exclusions could reconcile the counts, but overlap has not been established from a retained row ledger. Another version simply called the total 33. Do not manufacture an explanation to make the totals agree.
A concentration threshold such as one trade contributing more than 40% of gains is a screening choice, not proof that the trade was lucky. Likewise, no losing observed periods can warrant inspection for accrued fees, flows, sampling or stale values, but it is not proof of fraudulent or invalid accounting. Passing these gates would not certify that every remaining series is clean.
Preserve the sensitivity claim without turning it into a rule
The saved report supplied this table for a candidate, without an identifier and full return inputs alongside it:
Scroll sideways to see every column. Keyboard: focus the table region and use the arrow keys.
| Assumed trials | Reported DSR |
|---|---|
| 1 | 0.9735 |
| 5 | 0.0053 |
| 10 | 0.0000 |
| 150 | 0.0000 |
| 23,495 | 0.0000 |
These historical values are unreproduced, not a new calculation. A displayed zero may be rounding. The table does not justify saying that comparing five traders always invalidates a record. Nor does the highest PSR necessarily belong to the candidate with the highest DSR. Both identities and inputs must be matched before describing a before/after score for one person.
Why selection still matters
Choosing the best-looking result from many alternatives changes the question you are asking. The Bailey and López de Prado DSR paper addresses selection and non-normal returns using a benchmark that depends on the trials. It does not supply a universal posterior probability that a named trader is skilled.
The original method notes specified a minimum of 20 valid periods, full kurtosis where 3 denotes a normal distribution, trial count 23,495 and trial-Sharpe variance 1.0. Those are reported inputs, not independently justified estimates. Correlated traders and repeated versions of one strategy complicate the number of independent trials. The full population count cannot automatically stand in for that number.
The notes also said cumulative-PnL differences were "deposit-proof". A difference in dollar PnL is not by itself a normalized return. To evaluate that claim, supply the equity or capital denominator, cash-flow timing, valuation and fee treatment. Splitting long gaps also needs an explicit observation rule. Combining irregular per-trade returns with daily or annualized quantities can produce incompatible inputs.
What a reproducible replacement must include
Retain the original scan timestamp and raw source snapshots; identify each candidate and every exclusion. Provide the code revision, periodic return series, units, handling of deposits and withdrawals, missing data and fees. Record the benchmark, skewness/kurtosis estimators, trial-count/dependence model, trial variance and the preselected threshold. Publish all evaluated outputs, including failures and unavailable states.
Then recompute the population summary and candidate-specific sensitivity table. Changing the trial count requires a new calculation; a low score at 23,495 does not alone prove what happens at 150. Later observations can change estimates in either direction and may reflect a different strategy or market regime.
For a worked explanation with transparent hypothetical inputs, see how to read PSR and DSR without turning them into a skill verdict. That guide does not reproduce this scan. Until the missing study evidence is reconciled, these historical numbers should not be used to recommend copying or rejecting an identifiable trader.
Research correction, not investment advice. The historical account and its useful method questions are preserved; personal test claims and certainty interpretations are withdrawn. AITR may earn commissions from commercial links elsewhere. No affiliate recommendation or return promise is made here.