Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research
None of the nine models tested beat the equal-weight market in the sealed 2024 evaluation.
GPT-6 Astra ranked highest on Quant-Arena’s overall score. The benchmark asks models to build features and stock-scoring code using real A-share data, with feedback restricted to 2022–2023 before each freezes a strategy for 2024. Evaluation includes trading constraints and costs, with market impact estimated from a simulated order book. Realized excess return dominates the score, supplemented by signal quality and feature diversity.
HF Daily Papers' note
GPT-6 Astra ranked highest on Quant-Arena’s overall score. The benchmark asks models to build features and stock-scoring code using real A-share data, with feedback restricted to 2022–2023 before each freezes a strategy for 2024. Evaluation includes trading constraints and costs, with market impact estimated from a simulated order book. Realized excess return dominates the score, supplemented by signal quality and feature diversity.
HF Daily Papers' note
score 5