FinzData Research · October 2026
AI Futures Championship
Twelve AI models, nine served by NVIDIA NIM and three from Claude, each designed three systematic strategies for 48 futures markets: equity indexes, rates, currencies, energy, metals and agriculture. Designed on 2009 to 2018, scored once on January 2019 to September 2026, then graded by three blind judges.
Results
Click a column to sort, click a row to highlight that model in the charts. Each model's median out-of-sample draw of three counts. Final score = 0.7 × performance percentile + 0.3 × mean blind judge score ÷ 10.
| # | Model | Maker | Rebalance | OOS Sharpe | In-sample Sharpe | CAGR | Max drawdown | Turnover / yr | IR vs trend | Deflated Sharpe prob. | Judge: Claude Opus | Judge: Mistral Large 4 | Judge: StepFun Step-5 | Judge mean | Final score | Avg corr. | Signal |
|---|
Benchmarks, out of sample: 12-month trend sign(ret(252)) Sharpe 0.65; long every market Sharpe 0.39. Same engine for all: 48 back-adjusted futures, 10% volatility target, next-session execution, 3 bps per unit of turnover. In-sample 2009–2018. Avg corr. = average return correlation with the other eleven; lowest is most original (GPT-OSS-20B, 0.04).
Out-of-sample Sharpe
January 2019 to September 2026, after costs
What did the judges reward?
Mean blind judge score vs out-of-sample Sharpe. Judges ranked with the in-sample record at 0.75, with the test at 0.40.
Growth of $1
Month-end values, January 2019 to September 2026, after costs. Dashed: 12-month trend benchmark. Hover for values; click a button to highlight.
Historical simulations on FinzData futures data. Not investment advice.
Strategy
Click a row or a button to read that model's hypothesis and falsification test.
Test your own futures strategy
139 futures markets with every contract since 2007, roll calendars and back-adjusted series, in Python, REST or through the MCP server for AI agents.