金融研究Hugging Face Daily Papers21 Oct, 04:00 am SGTSIGNAL F1272D

Quant-Arena: An End-to-End Benchmark for LLM Quantitative Research on Real A-Share Data

ORIGINAL / Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research

Existing benchmarks score formulas, code, or single trading decisions, leaving open whether LLMs can complete the full quantitative research loop from hypothesis to executable signal. Quant-Arena provides a benchmark and public leaderboard built on real A-share data: models construct features and stock-scoring code, receive feedback only on 2022–2023, freeze one strategy, and are assessed on the sealed year 2024 with trading constraints and costs, including market impact esti

01 ABSTRACT

Quant-Arena is an end-to-end quantitative research benchmark that asks LLMs to complete feature construction, stock-scoring code, historical backtesting, and strategy freezing on real A-share data. Models receive feedback only from 2022–2023, and the final strategy is evaluated on the sealed year 2024 using a cash-account evaluator with trading constraints and costs. Among nine frontier models, GPT-6 Astra scores highest, but no model beats the equal-weight market in 2024.

02 KEY FINDINGS

The benchmark covers the full quantitative research loop from hypothesis to executable signal, rather than scoring formulas or single decisions alone.
Evaluation uses real A-share data with 2024 as a sealed test year, and incorporates trading constraints, costs, and market impact estimated from a simulated order book.
Among nine frontier models, GPT-6 Astra scores highest, but none beats the equal-weight market in 2024.

03 RELEVANCE

The benchmark advances LLM quantitative research evaluation from fragment-level tasks to an end-to-end pipeline, and tests out-of-sample strategy performance under a sealed year and trading cost constraints. Its results show that model-generated strategies do not yet reliably outperform a simple equal-weight baseline, providing a more realistic reference for future research and evaluation.

04 EVIDENCE BOUNDARY

Evidence comes only from a single benchmark evaluated on specific A-share data; the single sealed test year 2024 and the set of nine models limit generalization to other markets, asset classes, or longer horizons. Evaluation relies on a simulated order book for market impact, which may deviate from real trading environments. The leaderboard is open to submissions, and comparability across agent harnesses is affected by implementation details.

Return to the primary source

AI GENERATED SUMMARY / DISCOVERED BY HUGGING FACE DAILY PAPERS · ARXIV CS.CL

Read original ↗