FinzData Research · October 2026
AI Trading Tournament
Eleven AI models, seven served by NVIDIA NIM and four from Claude, each designed one US equity strategy with one rule: use FinzData data. Ranked on out-of-sample Sharpe from January 2019 to September 2026, then graded by three judges drawn from the field.
Results
Click a column to sort, click a row to highlight that model in the charts. Final score = 0.7 × performance percentile + 0.3 × mean judge score ÷ 10. A judge never scored its own entry (—).
| # | Model | Maker | Book | Universe | Rebalance | Names | OOS Sharpe | In-sample Sharpe | CAGR | Max drawdown | Turnover / yr | Deflated Sharpe prob. | Judge: Haiku | Judge: Fable | Judge: Nemotron Super | Judge mean | Final score | Signal | Strategy |
|---|
Benchmarks, out of sample: equal-weight top 500 Sharpe 0.63, equal-weight top 1000 Sharpe 0.57. 10 bps costs on every trade. In-sample 2011–2018. Haiku and Kimi K3 hold identical portfolios (un-winsorised z-scores let momentum dominate both).
Out-of-sample Sharpe
January 2019 to September 2026, after costs
Did the judges just grade the results?
Mean judge score vs out-of-sample Sharpe. Almost a straight line: outcome bias.
Growth of $1
Each strategy re-run on the Hub from January 2019 to October 2026, 10 bps costs. Hover for values; click a table row or a button to highlight.
Curves run to 12 October 2026, so a few Sharpe ratios differ in the second decimal from the tournament window. Historical simulations on FinzData data. Not investment advice.
Fork a strategy
Every entry is a public, forward-tracked page on the FinzData Hub with its exact spec. Change a line and re-run it on the same point-in-time data.