I Tested 23,495 Crypto Traders for Real Skill. Zero Passed.

Last updated: August 2026 · AI Trading Ranked

Last Updated: August 2026

Disclaimer: This article is for informational purposes only and is not financial advice. Crypto trading involves significant risk of loss. Never trade with money you cannot afford to lose. Always do your own research (DYOR).

Quick answer: I pulled the public track record of 23,495 traders across Hyperliquid, Hyperliquid vaults, OKX lead traders and Polymarket, and ran the top candidates through the Deflated Sharpe Ratio — the statistic hedge funds use to work out whether a track record is skill or the survivor of a big search. 45 of the 117 clean candidates looked statistically certain to have real skill. Zero survived. The best score was 0.069 out of 1.00, against a pass mark of 0.60. And the collapse does not need a big number: the best trader's evidence evaporates once you admit you compared as few as five candidates.

Every leaderboard you have ever looked at has the same defect, and it is not fraud. It is arithmetic.

The problem with every leaderboard

Pick the best-performing trader out of 23,495. Now ask: how surprised should you be that this person's record looks good?

Not very. If you flip 23,495 coins ten times each, some of them come up heads ten times running. Those coins are not skilled. They are the tail of a distribution you searched through. The leaderboard shows you the winners and silently discards the thousands who lost — so the number at the top is inflated by the act of searching itself.

This is called multiple-testing bias, and quant finance has a specific fix: the Deflated Sharpe Ratio (Bailey and López de Prado). It asks a sharper question than "is this record good?" It asks: given that I searched N candidates, how good would the luckiest one look by chance alone — and does this record beat that?

The correction is brutal, and it should be.

What I actually measured

Sources, all public and keyless:

Combined population: 23,495. From that I deep-judged the top 150 and applied forensic gates before any statistics ran:

GateWhat it killsKilled
One-hit wondera single trade is >40% of all gains32
Implausibly smoothno losing periods at all — not a P&L curve3
Insufficient historyfewer than 20 valid return periods

That left 117 clean candidates. Then two numbers for each:

The result

45 of 117 candidates scored PSR ≥ 0.95. On the naive reading, 45 traders are 95%+ likely to have genuine skill. That is the number a leaderboard shows you.

0 of 117 reached DSR 0.60. Best score in the entire population: 0.069.

The collapse is best seen in individual cases. My highest-PSR candidate:

MetricValue
PSR (naive)1.0000
DSR (deflated)0.0000
Sharpe per observation0.392
Observations44

A trader who looks statistically certain becomes statistically worthless — not because their trades changed, but because I admitted how many traders I looked at before picking them.

"You only got that because you deflated by 23,495"

This is the fair objection, so I tested it. Here is my single best clean candidate re-scored at every plausible search size:

Candidates searchedDSRVerdict
10.9735passes
50.0053fails
100.0000fails
1500.0000fails
23,4950.0000fails

The evidence dies at five. Not 23,495 — five. If you have ever compared a handful of traders before choosing one, you have already run enough of a search to invalidate the record that made you choose. You do not need my population size to reach my conclusion; you need only to be honest that you shopped around.

The one thing that ever passed was a bug

For months nothing cleared even the lowest tier. Then, on 1 August, something scored a perfect DSR 1.0000 and landed in the top tier.

It was an accounting artifact. The candidate was a vault reporting 45 observations, 100% positive periods and zero drawdown. No real strategy has never had a losing period. A curve with no down periods has almost no standard deviation, so its Sharpe ratio — and every statistic built on it, including the deflated one — saturates at 1.0 and sails straight to the top of the rankings.

Deflation cannot save you there, because the input was never a profit-and-loss series in the first place. Accrued-fee curves and smoothed vault accounting both produce exactly that shape.

I added a gate for it: reject anything with ≥99% winning periods or zero drawdown. It caught three. The first thing ever to pass my test was a bug in my test — which is the honest hazard of every backtest anyone has ever shown you, including mine.

What this does and does not prove

It does not prove nobody can trade. It proves that public track records of the length these platforms publish are too short to distinguish skill from luck once you account for having searched. Absence of evidence is not evidence of absence — but you should not risk money on the difference.

It does not mean copy trading is a scam. It means the leaderboard is not evidence. Those are different claims.

Consider what would have to be true for a leaderboard to work. The record would need to be long enough that luck averages out, and you would need to have found it without searching. Both are usually false.

How to test this yourself

Whenever anyone shows you a track record:

  1. Ask how many candidates they looked at before showing you this one. That number is the deflation penalty. If they will not say, assume it is large.
  2. Ask how long the record is. Under ~30 observations, almost nothing is distinguishable from noise.
  3. Ask what the worst drawdown was. If the answer is "none", it is not a P&L curve.
  4. Ask what share of the profit came from the single best trade. Over 40% and you are looking at one lucky event with commentary attached.

The maths is not hard and the code is public — the Deflated Sharpe implementation I used is a short, dependency-light Python function. If a platform's leaderboard cannot survive it, that tells you something the leaderboard was never going to.

Method notes

Returns were built from cumulative PnL differences rather than reported ROI, which makes them deposit-proof; periods with gaps over three days were split rather than zero-filled; kurtosis is full (3.0 = normal), not excess. Minimum 20 valid observations. `n_trials` equals the full scanned population, and variance of trials is set to 1.0. Every figure above comes from a single scan on 1 August 2026 and moves as the population does.

Free Cheat Sheet