EVOQUANT: Self-Evolving Verifier-Guided Strategy Optimization for Robust Quantitative Trading
EVOQUANT introduces an LLM-driven, verifier-closed-loop framework for quantitative strategy optimization, achieving a mean test Sharpe ratio improvement from -0.298 to 0.538 across seven representative strategies (four A-share and three crypto), with the best strategy showing a 199% relative gain—effectively mitigating hallucinated edits, strategy drift, and backtest overfitting inherent in direct LLM-based strategy rewriting.
Core Contribution: Addressing LLM Trustworthiness Bottlenecks in Strategy Optimization
EVOQUANT does not rely on LLMs to directly rewrite trading code. Instead, it establishes a four-stage closed loop: diagnosis (LLMs deeply analyze performance bottlenecks, e.g., drawdown sources under specific market regimes), generation (semantically constrained candidate edits, e.g., modifying take-profit logic without altering position-sizing models), verification (multi-stage pipeline including simulation consistency checks, risk metric robustness evaluation, and out-of-sample generalization testing), and distillation (encoding successful optimization patterns into structured, reusable knowledge rules for continual self-improvement).
Empirical Results: Robust Gains Across Asset Classes
- Evaluation covers seven representative strategies: four from the A-share market (CSI 500 timing, CSI 300 momentum, sector rotation, high-frequency price-volume) and three from the crypto market (BTC/USDT mean-reversion, ETH/USDT volatility arbitrage, stablecoin yield spread arbitrage);
- Key metric: Mean test Sharpe ratio improves from -0.298 to 0.538 (Δ = +0.836); best-performing strategy achieves 199% relative improvement;
- Ablation studies confirm module necessity: removing the verifier reduces Sharpe by 42.7%; disabling knowledge distillation degrades subsequent iteration efficiency by 68%.
Methodological Positioning: A New Paradigm of RL–LLM Synergy
Published on arXiv:2607.12455v1, EVOQUANT treats LLMs as intelligent explorers over strategy space and verifiers as critics and reward shapers—realizing LLM-Augmented RL for Algorithmic Trading. It diverges from end-to-end fine-tuning or prompt engineering by emphasizing explainable, auditable, and cumulative strategy evolution paths.