Skip to content

R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization

🕒 Published (v1): 2025-05-21 06:20 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

R&D-Agent-Quant (R&D-Agent(Q)) is a multi-agent framework that automates the full quantitative finance R&D pipeline through coordinated factor mining and model optimization. It closes a hypothesis–implementation–validation–feedback loop driven by LLMs, achieving ~2× higher annualized returns than static factor libraries while using 70% fewer factors, at a compute cost under $10.

Problem

Quantitative research pipelines suffer from three compounding limitations: (i) extensive manual intervention in hypothesis generation, coding, and tuning; (ii) LLM-based trading agents that produce signals without grounded factor construction, risking hallucinations and poor explainability; (iii) siloed optimization where factor mining and model training are decoupled, preventing cross-stage feedback and joint performance gains.

Method

R&D-Agent(Q) decomposes the quant R&D pipeline into five LLM-powered units operating in a continuous closed loop:

  1. Specification Unit — formalizes task context as a tuple \(S = (B, D, \mathcal{F}, \mathcal{M})\) (background, data interface, output format, execution environment), enforcing \(\forall x \in D,\ f_\theta(x) \in \mathcal{F}\) within the backtesting environment.
  2. Synthesis Unit — maintains a knowledge forest over experiment trajectories \(\{h_t, f_t\}\) and generates the next hypothesis via a stochastic mapping \(h^{(t+1)} = G(H_t^{(a)}, F_t^{(a)})\), conditioned on action type \(a_t \in \{\text{factor}, \text{model}\}\).
  3. Implementation Unit (Co-STEER) — a code-generation agent that builds a DAG \(G=(V,E)\) of task dependencies, derives a topological execution order \(\pi_S\), and maintains a growing knowledge base \(K^{(t+1)} = K^{(t)} \cup \{(t_j, c_j, f_j)\}\). Similar past solutions are retrieved via \(c_{\text{new}} = \arg\max_{c_k \in K} \text{similarity}(t_{\text{new}}, t_k) \cdot c_k\).
  4. Validation Unit — deduplicates candidate factors by computing per-factor \(\text{IC}^{(n)}_{\max} = \max_m \mathbb{E}_t[\text{IC}^{(t)}_{m,n}]\) against the SOTA pool, discarding factors with \(\text{IC}^{(n)}_{\max} \geq 0.99\), then backtests survivors via Qlib.
  5. Analysis Unit — evaluates outcomes and selects the next optimization arm (factor vs. model) via contextual Thompson sampling over an 8-dimensional performance state vector \(x_t \in \mathbb{R}^8\) with Bayesian linear posteriors.

Key Contributions

  • First end-to-end, data-centric multi-agent framework for automated quantitative R&D with verifiable, interpretable outputs.
  • Co-STEER: a knowledge-evolving code-generation agent with DAG-guided scheduling and transferable failure-recovery via a persistent task-code-feedback knowledge base.
  • A contextual two-armed bandit (linear Thompson sampling) for adaptive factor/model optimization direction selection.
  • Joint factor–model co-optimization that outperforms either component optimized independently.

Results

  • R&D-Factor: IC up to 0.0497, ARR up to 14.61% (vs. Alpha 158 ARR 5.70%), using over 70% fewer factors than static libraries.
  • R&D-Model: best Rank IC of 0.0546 and lowest MDD of −6.94% across all baselines.
  • R&D-Agent(Q) (joint): IC 0.0532, ARR 14.21%, IR 1.74 with o3-mini; Calmar ratio 1.915, outperforming the best standalone model (MASTER CR 1.053) and best factor library (Alpha 360 IC 0.0420, ARR 4.38%).
  • Total compute cost reported as under $10 USD.
  • Evaluated on CSI 300 A-share dataset, test period Jan 2017–Aug 2020.

Limitations

  • Evaluated on a single market (CSI 300, Chinese A-shares); generalization to other markets, asset classes, or regimes is untested.
  • Cost figure (\(<\)10) likely reflects a single run with o3-mini; production use with repeated iterations or GPT-4o would be substantially higher.
  • Hypothesis similarity analysis covers only factor generation (R&D-Factor); model-side exploration dynamics are not analyzed.
  • The knowledge base and SOTA pool grow unboundedly; no strategy is described for pruning stale or market-regime-irrelevant entries.
  • Backtesting via Qlib may not fully capture real-world execution frictions (e.g., market impact, liquidity constraints).

Relevance to Harnesses / Meta-Harnesses

R&D-Agent(Q) is a domain-specific meta-harness: it orchestrates five specialized agents (Specification, Synthesis, Implementation, Validation, Analysis) within a closed feedback loop, directly instantiating the harness pattern of automated hypothesis generation → execution → result routing → next-iteration scheduling. The Co-STEER implementation agent's persistent knowledge base (recording task-code-feedback triples) and DAG-guided task sequencing are particularly relevant architectural primitives for harness designers dealing with interdependent, stateful subtasks. The contextual multi-armed bandit for direction selection demonstrates a principled, learnable meta-scheduler—an alternative to static pipeline ordering that harness researchers can adapt for resource allocation across concurrent research threads.