Evidence-adjusted leaderboard · No. 1 · October 2026

Public AI trading leaderboards rank by return. None of them can yet name a winner.

We take the equity history each board publishes and ask a different question of the same data: how much can the ranking tell you, and how long until it could show that the top entry is skilled rather than lucky?

0 / 41competition teams reach t = 2 (Liquidity Arena)

0 / 32top public agents with history reach t = 2 (AI-Trader)

~5 moearliest the arena could name a winner, if the best team’s true Sharpe were 5

5.7 yrfor AI-Trader at a true Sharpe of 2, against 23,376 entries

t = annualised Sharpe × √years. One result needs t ≥ 2; the best of N needs more. How this is computed.

Board	Entries	Window	t ≥ 2	Highest t	Best-of-N bar	Earliest winner*
Liquidity Arena 2026 Track A, Phase I	41	about 30 trading days per team (2026-07-20 to 2026-08-21)	0	1.63	3.26	5 months at Sharpe 5 · 2.7 years at Sharpe 2
HKUDS AI-Trader (ai4trade.ai), top 50 by return	23,376	16 return days (2026-07-08 to 2026-07-24) for agents with history	0	0.78	4.76	11 months at Sharpe 5 · 5.7 years at Sharpe 2

* Months of clean, forward history the top entry would need to clear the best-of-N bar, if its true Sharpe were the value shown. Counting every entry as independent overstates the charge when entries are correlated; the true Sharpe is unknown. The two cut in opposite directions.

41 entries · ranked by composite score

Liquidity Arena 2026 Track A, Phase I

about 30 trading days per team (2026-07-20 to 2026-08-21)

Live sandbox trading, so the model could not have seen the test period: closed-book by construction. The gap is length and selection. Audit 003.

0 / 41teams reach t = 2

1.63closest: Kinance, rank 11 by return. At its observed Sharpe, about 4.1 months of forward record would clear the bar.

11 · 7 · 17return ranks of the top three by evidence. The return leader is fifth by t.

Source: arena.liquiditytech.com/client-web-api/official/tracka/ranking/list?phase=PHASE_I, fetched 2026-10-12.

23,376 entries · ranked by cumulative return

HKUDS AI-Trader (ai4trade.ai), top 50 by return

16 return days (2026-07-08 to 2026-07-24) for agents with history

31 / 34agents with history moved under 1 point between 24 July and 12 October. A third-party report (#311, unconfirmed) finds the price feed stuck at about 24 July.

+271.6%the #1 agent, with no history published. So is #2 (+145.4%): no Sharpe or drawdown can be computed.

10 / 100of the top 100 are one parameter grid (OC-rot-lb126-rb30-…), ranked as separate winners.

History is July only, then one live snapshot, so annualisation is rough. 28 of 34 already had ≥ 90% of their return when the history begins. 46 of the top 100 show zero trades alongside non-zero returns; reported as published.

Source: ai4trade.ai/api/profit/history?days=365&limit=50&include_history=true, fetched 2026-10-12.

For operators

What would bring the date forward

LONGER WINDOWS

Carry records across phases

A month of live trading is too short for any realistic Sharpe. Carrying Phase I into Phase II doubles each team’s evidence at no cost.

PUBLISH THE HISTORY

Rank only full equity curves

A single snapshot supports no Sharpe, no drawdown and no test. An entry with no history should not sit at the top.

GROUP THE VARIANTS

One owner, one family

Parameter sweeps are one attempt with many tries. Show them as a family against one bar, not as separate winners.

Data

Use the numbers

Per-entry return, days, Sharpe and t, built by tools/build_leaderboard.py from each board’s public endpoint. CC BY 4.0.

Download 2026-10.json Suggest a board to rescore Text version

Operate one of these boards and think a number is wrong? Open an issue; corrections are published in the next edition. Next: TradeRank seasons 1–6 and Liquidity Arena Phase II.