Research map / Multi-agent systems and games / LLM agents
R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes R&D-Agent(Q), a multi-agent LLM framework that automates quant factor and model research and development with a bandit scheduler, reporting gains over factor libraries and deep time-series models on CSI300, CSI500, and NASDAQ100.
Problem
Quantitative research pipelines for asset return prediction require heavy manual work in factor mining and model design, and factor and model optimization are fragmented and poorly interpretable.
Approach
A data-centric multi-agent loop (Specification, Synthesis, Implementation, Validation, Analysis units) in which LLMs generate factor/model hypotheses, the Co-STEER code agent implements them with DAG-scheduled tasks and knowledge reuse, Qlib backtests validate them, and a contextual Thompson sampling bandit chooses between factor and model optimization.
Key results
- R&D-Agent(Q) with o3-mini reaches IC 0.0532, ARR 14.21% and IR 1.74 on CSI 300, versus the strongest baselines such as Alpha 158 (ARR 5.70%) and TRA (ARR 6.49%).
- R&D-Factor with GPT-4o achieves ARR 14.61% on CSI 300 while using roughly 22% of the factors of the baseline libraries.
- On out-of-sample 2024–mid-2025 tests, R&D-Agent(Q) with o4-mini gets the best IR on CSI 500 (2.17 vs 1.40 for the best baseline LSTM... ) and NASDAQ 100 (IR 1.77), with the highest IC on CSI 500 (0.0288).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| jointly optimize factors and model | R&D-Agent(Q) o3-mini · CSI300 test | IC: 0.0420 -> 0.0532 | Helped |
| jointly optimize factors and model | R&D-Agent(Q) o3-mini · CSI300 test | ARR: 0.0570 -> 0.1421 | Helped |
| use Bandit rather than random action selection (ablation) | R&D-Agent(Q) o3-mini · CSI300 | IC: 0.0445 -> 0.0532 | Helped |
| use Bandit rather than random action selection (ablation) | R&D-Agent(Q) o3-mini · CSI300 | ARR: 0.0897 -> 0.1421 | Helped |
| use Bandit rather than LLM-based action selection (ablation) | R&D-Agent(Q) o3-mini · CSI300 | IC: 0.0476 -> 0.0532 | Helped |
| use Bandit rather than LLM-based action selection (ablation) | R&D-Agent(Q) o3-mini · CSI300 | ARR: 0.1009 -> 0.1421 | Helped |
Surprising findings
- Time-series forecasting models such as PatchTST and Mamba underperform simple ML and general deep learning models on stock return prediction.
- Bandit scheduling outperformed the LLM-based scheduler, which incurred higher per-step overhead and completed fewer iterations.
Limitations
- Relies solely on the LLM's internal financial knowledge, without external data diversity or domain priors.
- Results come from a single backtest period and specific LLM backends, with reproducibility depending on proprietary APIs.
Experimental setup
Datasets: CSI 300 Alpha 20 CSI 500 NASDAQ 100
Benchmarks: CSI 300 CSI 500 NASDAQ 100
Metrics: IC ICIR Rank IC ARR IR MDD Calmar ratio
Baselines: Alpha 101 Alpha 158 Alpha 360 AutoAlpha LightGBM XGBoost CatBoost DoubleEnsemble GRU LSTM ALSTM Transformer
Builds on: Qlib LightGBM Thompson sampling chain-of-thought
Closest papers by meaning
- AlphaQCM: Alpha Discovery in Finance with Distributional Reinforcement Learning (ICML 2025 · similarity 0.87)
- FinRpt: Dataset, Evaluation System and LLM-based Multi-agent Framework for Equity Research Report Generation (AAAI 2026 · similarity 0.86)
- HiveMind: Contribution-Guided Online Prompt Optimization of LLM Multi-Agent Systems (AAAI 2026 · similarity 0.86)
- AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors (AAAI 2025 · similarity 0.86)
- MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model (ICLR 2025 · similarity 0.86)
- Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading (ICLR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.