Graveyard / Copy trading / Hyperliquid
Copying the Hyperliquid leaderboard
Does copy-trading top Hyperliquid traders work?
The idea
The Hyperliquid leaderboard is public and so are the fills. Copy the top 25 accounts, with a stop on every copy.
How we tested it
| Data | Public Hyperliquid fills of 25 top accounts. Live copies on two venues with $15 margin and a hard −$25 kill per bot. |
|---|---|
| Period | In-sample 15 July to 13 September 2026 (494 copies). Out-of-sample 10 to 20 October 2026 (291). Live 17 to 21 October 2026. |
| Costs | Live fills. Copy delay was 4-5 seconds and the entry was slightly in our favour. The long-only filter test used 19 bp per round trip. |
| Holdout | The out-of-sample cohort was rebuilt as of its start date, with no look-ahead. |
| Control | No random control. Returns were adjusted for BTC's own move instead. |
Result
| In-sample | +$1,162, profit factor 1.82 |
|---|---|
| Part of that from the market itself | +$654 |
| Out-of-sample | −$229, profit factor 0.52 |
| Live, bot 1 / bot 2 | −$34.97 / −$17.56 |
| Skill persistence per account | −0.07 |
| Long-only filter | 41 copies, −1.63% each |
Why it failed
The backtest edge was survivorship plus market beta: a long-leaning group of accounts in a 27% BTC rally. Past rank did not predict future rank. Copy delay was not the problem.
What would make it worth retesting
Only a test that does not select accounts by past results, run forward.
Test id: hlcopylb live copy (fixcopy diagnosis), hlcopylblong-perps-event-v1. Past results of a backtest say nothing certain about the future.