Quant-Arena
An end-to-end quantitative research benchmark built on real A-share market data. It requires an LLM to construct features autonomously, write scoring code, and freeze a strategy, which is then evaluated out-of-sample on sealed-year data. A public leaderboard accompanies the benchmark.
Timeline (last 90 days)
Quant-Arena is an end-to-end quantitative research benchmark built on real A-share data, and it also comes with a public leaderboard.
Quant-Arena requires models to construct features themselves and write stock scoring code, giving them feedback only on 2022–2023 data, with strategies frozen and evaluated on the sealed year 2024.
Quant-Arena's evaluator uses a cash account model, imposing trading constraints and costs, including market impact estimated from a simulated order book.
Quant-Arena's score is dominated by realized excess returns, supplemented by signal quality and the diversity of features used by the strategy as secondary dimensions.
The Quant-Arena paper evaluates nine frontier models, with GPT-6 Astra scoring the highest.
None of the models beat the equal-weight market benchmark in the sealed 2024 test.
Quant-Arena uses a public leaderboard and open submissions, and any model or agent harness can participate.
Quant-Arena currently focuses only on the A-share market, and whether the conclusions transfer to other markets still needs verification.
Quant-Arena uses only a single sealed year (2024), and one year's market environment involves chance, so strategies of different styles may perform very differently depending on market conditions.
Quant-Arena's score is dominated by excess returns, which may lead participants to over-pursue returns while neglecting risk-adjusted performance.
2 more timeline events
All Facts (12)
Quant-Arena is an end-to-end quantitative research benchmark built on real A-share data, and it also comes with a public leaderboard.
50%UnverifiedQuant-Arena requires models to construct features themselves and write stock scoring code, giving them feedback only on 2022–2023 data, with strategies frozen and evaluated on the sealed year 2024.
50%UnverifiedQuant-Arena's evaluator uses a cash account model, imposing trading constraints and costs, including market impact estimated from a simulated order book.
50%UnverifiedQuant-Arena's score is dominated by realized excess returns, supplemented by signal quality and the diversity of features used by the strategy as secondary dimensions.
50%UnverifiedThe Quant-Arena paper evaluates nine frontier models, with GPT-6 Astra scoring the highest.
50%UnverifiedNone of the models beat the equal-weight market benchmark in the sealed 2024 test.
50%UnverifiedQuant-Arena uses a public leaderboard and open submissions, and any model or agent harness can participate.
50%UnverifiedQuant-Arena currently focuses only on the A-share market, and whether the conclusions transfer to other markets still needs verification.
50%UnverifiedQuant-Arena uses only a single sealed year (2024), and one year's market environment involves chance, so strategies of different styles may perform very differently depending on market conditions.
50%UnverifiedQuant-Arena's score is dominated by excess returns, which may lead participants to over-pursue returns while neglecting risk-adjusted performance.
50%UnverifiedModels may have learned the form of "looking like quantitative research" without acquiring genuine market insight that can survive cycles.
50%UnverifiedAt least for now, frontier models have not managed to independently complete the closed loop from hypothesis to deployable strategy.
50%