# AI Futures Championship 2026 Results Source: https://finzdata.com/research/ai-futures-championship-2026 Twelve AI models designed three systematic strategies for 48 futures markets. One model beat the 12-month trend benchmark on unseen almost eight-year data. [FinzData Research](https://finzdata.com/research/) · Published 31 October 2026 # AI Futures Championship 2026: can AI models beat trend following? After our US stocks championship, we moved to the harder macro problem. Twelve AI models from 9 labs each designed three long/short strategies for 48 futures markets across six sectors, built on 7,041 futures contracts since 2007. We scored them once on almost eight years none of them had seen, against the toughest simple benchmark in futures: 12-month trend following. One beat it. [Download the deck (PDF, 14 slides)](https://finzdata.com/research/ai-futures-championship-2026.pdf) ### Get FinzData research free, every morning The daily brief: what moved, the macro releases that matter, positioning and liquidity, with charts. One email a day, free, unsubscribe anytime. **0.75**winning out-of-sample Sharpe (Claude Haiku) **0.65**12-month trend benchmark, same period **36**strategies from 12 AI models, 9 labs **7,041**futures contracts across 48 markets, six sectors ## Contents 1. [Why futures, and why this test matters](https://finzdata.com/#context) 1. [Breadth and depth of the data](https://finzdata.com/#breadth) 1. [Breadth of the field and depth of the test](https://finzdata.com/#field) 1. [How the championship worked](https://finzdata.com/#design) 1. [Final standings](https://finzdata.com/#standings) 1. [Growth of $1 out of sample](https://finzdata.com/#curves) 1. [Sharpe ratios with confidence intervals](https://finzdata.com/#sharpe) 1. [In sample versus out of sample](https://finzdata.com/#isoos) 1. [Turnover and costs](https://finzdata.com/#turnover) 1. [Herding into carry and trend](https://finzdata.com/#herding) 1. [What the blind judges rewarded](https://finzdata.com/#judges) 1. [The winner: Claude Haiku](https://finzdata.com/#winner) 1. [Lessons and Round 2 changes](https://finzdata.com/#lessons) 1. [Frequently asked questions](https://finzdata.com/#faq) ## 1. AI Futures Championship 2026 We gave twelve AI models the same brief: design one systematic strategy for 48 liquid futures markets, write down why it should work and how it could fail, and express it in a fixed formula language. Every model drew three independent strategies, 36 in all. We ran them all through one engine on 2009 to 2018 data, then scored them once on January 2019 to September 2026, a period no model had seen. **Claude Haiku won** with an out-of-sample Sharpe ratio of 0.75. It was the only entry to beat the simplest serious benchmark in futures, a 12-month trend-following rule, which made 0.65 over the same period. ## 2. Why futures, and why this test matters In October we ran our first AI championship on US stocks: eleven models each designed an equity strategy, and Nemotron-3 Ultra won with an out-of-sample Sharpe of 0.81 against 0.63 for an equal-weight benchmark. Stocks are a stock-picking problem. Futures are a macro problem, and a harder one. A futures strategy can be short as easily as long, and it has to decide at the same time about Treasury yields, the Bund, crude oil, gold, the yen and soybeans. Futures also come with the most demanding simple benchmark in systematic investing. A 12-month trend rule, long what has risen and short what has fallen, is one line of code with a long published record across a century of markets, and trend following is the core of the managed-futures industry. An AI model that cannot beat it has not added anything a junior quant could not write in a minute. The test window, January 2019 to September 2026, was unseen by every model and covers very different markets: the 2020 crash and recovery, the 2022 inflation shock that drove the worst bond sell-off in decades, the higher-for-longer years that followed, and the 2026 oil shock. A strategy that held up across all of them earned its place. ## 3. Breadth and depth of the data The championship runs on 48 of the 139 futures markets FinzData covers, chosen for depth and liquidity across six sectors and three continents: nine equity indexes from the S&P 500 to the Nikkei 225, US Treasuries and German government bonds from two to thirty years, eight currencies, crude, products and natural gas, precious and industrial metals, and twelve grains, softs and livestock markets. The depth matters as much as the breadth. Behind the 48 continuous series sit 7,041 individual futures contracts, every one listed since 2007, and 3.7 million daily contract records. They are part of the 139 futures markets and 17,610 individual futures contracts FinzData serves. That depth is what makes carry measurable: on each of 4,873 sessions we read the front contract and the next one out, as they were that day, and annualise the gap. Returns come from back-adjusted series on our roll calendar, so no strategy is paid for the jump at a roll. Sector by sector: equity indexes: S&P 500, Nasdaq-100, Russell 2000, Dow Jones, Euro Stoxx 50, DAX, FTSE 100, CAC 40, Nikkei 225 (1,054 futures contracts); government bonds: US 2-year, US 5-year, US 10-year, US 30-year, Schatz, Bobl, Bund, Buxl (624 futures contracts); currencies: euro, yen, Australian dollar, pound, Canadian dollar, Mexican peso, New Zealand dollar, Swiss franc (1,300 futures contracts); energy: WTI crude, Brent crude, gasoil, natural gas, gasoline, heating oil (1,373 futures contracts); metals: gold, silver, copper, platinum, palladium (1,208 futures contracts); agriculture: corn, soybeans, wheat, soybean oil, soybean meal, KC wheat, sugar, coffee, cocoa, cotton, live cattle, lean hogs (1,482 futures contracts). ## 4. Breadth of the field and depth of the test The field was built to be broad: twelve models from 9 labs (Anthropic, DeepSeek, Google, Meta, MiniMax, NVIDIA, OpenAI, Poolside, Z.ai), from frontier systems to small open models that run on a single GPU. Nine were served through NVIDIA NIM and three were Claude models. Each received the same brief and the same formula language, and each drew three strategies independently, so we could see whether a model had a view or just a lucky guess. The test was built to be deep. Every draw was run on the same engine over 227,913 market-days of returns and 233,253 carry readings, scored on Sharpe, return, drawdown, turnover, information ratio against trend and correlation with every other entry, given a 95% confidence interval from 2,000 block-bootstrap resamples, and a deflated Sharpe that accounts for all 36 tries. Three blind judges from three labs scored the reasoning separately. ## 5. How the championship worked The universe is 48 futures markets: nine equity indexes, eight government bond futures, eight currencies, six energy contracts, five metals and twelve agricultural markets. Returns come from back-adjusted continuous contracts, and carry comes from the price gap between the front and next contract, annualised. Each entry is a signal formula built from a short list of functions: returns, volatility, moving averages, breakouts, carry, time-series and cross-sectional ranks and z-scores. The engine turns the signal into positions, scales the whole book to 10% expected volatility, trades at the next session and charges 3 basis points per unit of turnover. Because the engine is the same for every model, the only thing that differs is the idea. Each model's three draws are scored and the median out-of-sample draw counts, so a lucky draw cannot carry a model. The final score is 70% performance rank and 30% the mean of three blind judges, who saw the hypothesis, test and signal but no results. ## 6. Final standings | # | Model | Access | OOS Sharpe | IS Sharpe | Judges | Final | Signal | | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | Claude Haiku | Anthropic | 0.75 | 0.36 | 7.3 | 0.92 | `zscore(ret(252)/vol(252)) + 0.5*zscore(carry())` | | 2 | MiniMax M3.1 Flash | MiniMax via NIM | 0.64 | 0.68 | 8.0 | 0.88 | `rank(breakout(252)) + 0.5*rank(ts_mean(carry(),126))` | | 3 | Claude Sonnet | Anthropic | 0.48 | 0.69 | 8.0 | 0.75 | `0.5*sign(ret(252)) + 0.25*sign(ret(63)) + 0.25*clip(zscore(carry()),-1,1)` | | 4 | Nemotron-3 Ultra | NVIDIA NIM | 0.48 | 0.05 | 5.3 | 0.73 | `ts_z(ret(252), 252) + 0.5 * ts_z(carry(), 252)` | | 5 | Muse Glimmer 30B | Meta via NIM | 0.36 | 0.47 | 7.3 | 0.67 | `rank(ret(126)) + 0.4*rank(carry())` | | 6 | Claude Fable | Anthropic | 0.35 | 0.93 | 7.7 | 0.61 | `sign(ret(252)) + sign(carry())` | | 7 | Gemma 4 31B | Google via NIM | 0.30 | 0.68 | 6.3 | 0.51 | `rank(ret(126)) + rank(carry())` | | 8 | GLM-5.3 Flash | Z.ai via NIM | 0.06 | 0.57 | 8.3 | 0.44 | `sign(ret(63)) + sign(ret(126)) + sign(ret(252)) + 0.5 * rank(carry())` | | 9 | Laguna XS 2.1 | Poolside via NIM | 0.06 | 0.07 | 5.3 | 0.41 | `0.6*rank(ts_z(ret(60),60)) + 0.4*rank(carry())` | | 10 | GPT-OSS-20B | OpenAI via NIM | 0.05 | -0.42 | 4.3 | 0.26 | `rank(ret(252)) + 0.5*rank(carry()) + 0.5*ts_z(breakout(30),20)` | | 11 | Nemotron-3.5 Lightning | NVIDIA NIM | -0.04 | -0.17 | 3.3 | 0.16 | `carry() - ts_mean(carry(), 63)` | | 12 | DeepSeek V4.1 Flash | DeepSeek via NIM | -0.31 | 0.44 | 5.3 | 0.16 | `rank(carry()) - 0.5*rank(ret(5))` | Benchmarks over the same test window: 12-month trend 0.65, long every market 0.39. Only one model beat trend; four of twelve beat buying everything. ## 7. Growth of $1 out of sample From January 2019 a dollar in the winning strategy grew to 1.88, against 1.60 for the trend benchmark and 1.35 for holding every market long. 2022, the inflation shock, was the best year for every leader, up 22% to 27% as trends in energy, rates and grains ran long. The winner's lead over trend came later: it was flat in 2023 while trend lost 7%, and made 25% in 2024 against 10%. ## 8. Sharpe ratios with confidence intervals Seven to eight years of out-of-sample data is not a lot for strategies with Sharpe ratios below one. The 95% bootstrap interval for the winner runs from 0.08 to 1.48; the runner-up's from 0.03 to 1.27. Every interval overlaps the trend benchmark. The ranking is real as a record, but the gaps between neighbours are within noise. ## 9. In sample versus out of sample The rank correlation between in-sample and out-of-sample Sharpe across all 36 draws was 0.26. Claude Fable's counted draw had the best design-window record of the twelve, 0.93, and fell to 0.35; Nemotron-3 Ultra did the reverse, 0.05 to 0.48. A good backtest on ten years of futures data told us little about the next eight. ## 10. Turnover and costs The 11 weekly-rebalancing draws averaged an out-of-sample Sharpe of -0.16 with median turnover of 72 times a year. The 25 monthly draws averaged 0.45 at 22 times. Across all draws the rank correlation between turnover and out-of-sample Sharpe was -0.72, the strongest single pattern in the championship. Costs were modest at 3 basis points, so most of the drag was faster signals catching noise, not fees. ## 11. Herding into carry and trend All 36 draws used carry, and most paired it with a 6 to 12-month trend term. That is a sensible prior, since both are among the best-documented premia in futures, but it made the field crowded: the top three counted strategies had return correlations of 0.83 to 0.86 with each other. GPT-OSS-20B was the outlier, with an average correlation of 0.04 to the rest, from a breakout and carry mix rebalanced weekly. ## 12. What the blind judges rewarded Three judges, Claude Opus, Mistral Large 4 and StepFun Step-5, scored each strategy from 0 to 10 without seeing any results. They agreed with each other closely (pairwise correlations 0.89 to 0.93). Their scores ranked with the out-of-sample record at 0.40 but with the in-sample record at 0.75: a well-argued idea looked like a good backtest more than a good forecast. Claude entries scored 1.7 points higher on average than others with every judge. Part of that follows performance, and part may be clearer falsification tests, but some of it may be a family preference, which is why Round 2 changes the panel. Monthly designs scored 6.7 against 5.4 for weekly ones, so the judges did reward the lower-turnover choice. ## 13. The winner: Claude Haiku The winning signal is `zscore(ret(252)/vol(252)) + 0.5*zscore(carry())`: twelve-month return divided by twelve-month volatility, standardised across markets, plus half-weight carry, rebalanced monthly. It is close to the trend benchmark (correlation 0.81) but added an information ratio of +0.23 on top of it. Its stated test was that the blend must clearly beat the trend benchmark out of sample; it did, narrowly. After 36 trials, its deflated Sharpe probability is 0.41, well short of the 0.95 that would mark it as skill rather than selection. We report that plainly: this is a championship result, not a proven strategy. ## 14. Lessons and Round 2 changes Four things shaped the results more than model size. The example signal in the brief pulled seven draws toward it. Weekly trading cost more than any idea gained. Judges rewarded clear writing that matched in-sample results. 3 signals were submitted twice by the same or different models. Kimi K2.6 could not be reached and was left out, and StepFun Step-5 replaced Gemini on the panel after two timeouts. Round 2 will run with no example signal, turnover reported in the brief, a judge panel with no model from the same family as any entrant, and redraws for duplicates. ## Frequently asked questions ### Which AI model is best at trading futures? In this championship Claude Haiku won with an out-of-sample Sharpe of 0.75 from 2019 to September 2026, the only one of 12 models above the 12-month trend benchmark (0.65). MiniMax M3.1 Flash was second at 0.64. ### Did any AI model beat trend following? One did, narrowly: Claude Haiku added an information ratio of +0.23 over the trend benchmark. The difference is within statistical noise over this sample. ### What signals did the AI models use? All 36 strategies used futures carry (roll yield), and most combined it with 6 to 12-month trend. Monthly rebalancing did far better than weekly. ### How was the test kept fair? Every strategy ran on the same engine, markets, 10% volatility target, next-session execution and costs. Models designed on 2009 to 2018 and were scored once on 2019 to September 2026. Judges saw no results. ### What data did the championship use? 48 futures markets in six sectors (equity indexes, government bonds, currencies, energy, metals and agriculture), built from 7,041 individual futures contracts listed since 2007 and 3.7 million daily contract records, with back-adjusted continuous series and daily carry from the front and next contracts. ### Can I test my own futures strategy on this data? Yes. FinzData covers 139 futures markets with every contract since 2007, roll calendars and back-adjusted series, by Python, REST or the MCP server for AI agents. [Download the deck (PDF, 14 slides)](https://finzdata.com/research/ai-futures-championship-2026.pdf) [More research](https://finzdata.com/research/) [Download the full paper (PDF, 14 slides)](https://finzdata.com/research/ai-futures-championship-2026.pdf) [Open the live results dashboard](https://finzdata.com/ai-futures-championship) ## Related - [AI Trading Tournament: US stocks](https://finzdata.com/ai-trading-tournament) - [AI Trading Championship 2026 paper (stocks)](https://finzdata.com/research/ai-trading-championship-2026) - [Continuous futures data](https://finzdata.com/continuous-futures-data) - [Why most backtests lie](https://finzdata.com/research/why-backtests-lie) **Run it yourself:** `pip install finzdata` · 139 futures markets with every contract since 2007, roll calendars and back-adjusted series; 16,000+ US stocks since 2000; 18,000+ macro series with every vintage. [See plans](https://finzdata.com/shop). Built for AI, quants, data scientists and financial analysts who backtest. Clean data. Real insights. Real alpha. finzdata.com. This material is for information only and is not investment advice. Data sources: FinzData futures (139 markets, every contract since 2007); FinzData AI Futures Championship engine. ### Get FinzData research free, every morning The daily brief: what moved, the macro releases that matter, positioning and liquidity, with charts. One email a day, free, unsubscribe anytime.