Is that trading record skill, or could it be luck?
Also on Telegram, @SkillOrLuckBot (send an address, get the verdict), and as a Chrome extension that shows it on Hyperliquid, Binance and OKX copy-trading pages.
Got a performance claim instead of an address? Paste the claim →
Building or checking an AI trading agent? It ran N backtests: what is the best one worth →
Have 47,607 people call ten coin flips each, and about 46 will get all ten right. On a leaderboard, they would be at the top. Paste an address and see how much of its record luck alone could explain, once you count how many accounts it was picked from.
The records hardest to explain by luck, among 47,607 accounts
“Luck, picked blind” is how often a record this strong turns up with no skill behind it, if you had picked it without looking at any ranking. Picked as the best of 47,607, records this strong are expected by chance: that is why none clears 4.90 yet.
| # | Record | Length | t (bar 4.90) | t after BTC | Luck, picked blind | Notes |
|---|---|---|---|---|---|---|
| 1 | 0x58e1…132e#5 by 30-day ROI | 2.8 yr | 3.48 | 1 in 270,000 (0.00036%) | no losing week: check for market making or selling tail risk | |
| 2 | Hyperliquidity Provider (HLP)#10 by all-time PnL | 3.5 yr | 3.41 | 1 in 3,300 (0.03%) | Hyperliquid’s own market-making vault | |
| 3 | 0x0526…bb9a#46 by 30-day ROI | 0.9 yr | 3.31 | 1 in 2,400 (0.041%) | ||
| 4 | 0xd487…4982#34 by 30-day ROI | 2.3 yr | 3.06 | 1 in 790 (0.13%) | ||
| 5 | 0x3e7c…3909#15 by 30-day ROI | 0.2 yr | 2.72 | 1 in 560 (0.18%) | no losing week: check for market making or selling tail risk | |
| 6 | 0xff4c…463c#88 by all-time PnL | 2.7 yr | 2.66 | 1 in 360 (0.28%) |
From the all-time PnL and 30-day ROI top 100s. The verdict uses t; the BTC-adjusted t is shown alongside until a dated method change. Famous records, same ruler →
None of the top 100 has strong evidence yet. 9 are promising, not proven; 89 look like luck; 1 is losing, not by chance.
Why does no one clear it? Can it ever say yes?
Zero is what a board with no proven skill looks like. If none of the 47,607 accounts had any skill, luck would carry anyone at all past 4.90 only about 2% of the time: that is how the bar is set.
It can say yes; the records are short. The chance that a trader whose true Sharpe is S clears 4.90 after a given length of record:
| True Sharpe | 1 yr | 2 yr | 3 yr | 5 yr |
|---|---|---|---|---|
| Sharpe 2 | <1% | 2% | 8% | 33% |
| Sharpe 3 | 3% | 26% | 62% | 96% |
| Sharpe 4 | 18% | 78% | 98% | >99% |
| Sharpe 5 | 54% | 99% | >99% | >99% |
The median record in this top 100 is 1.7 years. If every one of these 99 had a true Sharpe of 3, about 22 would clear the bar by now, given their record lengths. 0 do.
How high the bar sits also depends on how independent the accounts are: 4.90 if each is a separate bet, about 2.6 by our rough estimate if many trade alike (3 clear that). Can it say yes · how the bar is set.
An agent can run ten thousand backtests in a weekend. Its best one has to clear the bar for all of them.
CICC’s factor loop tested 16,939 candidates; its best factor reached t = 3.37 against a bar of 4.69 for that many tries. An agent’s tries can be counted from its log, so the correction no longer depends on what anyone remembers.
Every record found by looking back was picked. So we froze some, and score only what happens next.
200 Hyperliquid wallets from today’s two top-100 boards, frozen 2026-10-14: first reading 2026-12-09. Papertrade, three groups of 100 including a random control, frozen 2026-10-16: readings 10-23 and 10-30. Each list is published with its hash when frozen, and every reading is published whatever it shows.
Published checks
Every number recomputed from public sources, dated, with the raw data kept.