EvoQuant: How Large Language Models Can Self‑Evolve as Quantitative Researchers

Abstract

Optimizing quantitative strategies has traditionally required extensive manual effort from domain experts, who must identify weak signals, tune risk‑control rules, and iteratively back‑test and refine code. While large language models (LLMs) promise acceleration, naïvely applying them often leads to hallucinated modifications and over‑fitted back‑test results. The paper proposes EvoQuant, a self‑evolution + validator‑guided framework that first diagnoses performance bottlenecks, then generates controlled candidate edits, and finally filters them through a multi‑stage verification pipeline. Experiments on seven representative strategies—four from the Chinese A‑share market and three from the cryptocurrency market—show average test‑set Sharpe improving from ‑0.298 to 0.538, with the best strategy achieving a 199 % relative gain.

Background

Quantitative strategy development is an iterative scientific process: formulate a testable market hypothesis, implement it as trading rules, rigorously back‑test, diagnose failures, revise the hypothesis, and repeat. Traditional optimization suffers from expert scarcity, high repetitive workload, low iteration efficiency, and difficulty breaking performance bottlenecks. LLM agents could theoretically automate code generation, hypothesis testing, and iterative refinement, but direct use risks hallucinated rules, unverifiable edits, and opaque causal explanations.

Problem Description

The formal goal is: given a user‑provided strategy s, find an improved strategy g* that significantly outperforms s on out‑of‑sample risk‑adjusted returns (Sharpe ratio) while satisfying three constraints—auditability (each edit is traceable), verifiability (edits must pass out‑of‑sample validation), and non‑drift (edits must not deviate excessively from the original logic). This frames strategy optimization as an evidence‑constrained program evolution problem.

Method

EvoQuant consists of four modules (Module 1–4) that embody a "execute‑rather‑than‑pure‑generate" philosophy.

Module 1: Strategy Ingestion & Representation

The framework parses the user’s strategy code into an abstract syntax tree, tags trading semantics (signals, entry/exit conditions, risk controls, position sizing), decomposes the strategy into hierarchical editable components, and recompiles a "strategy genome" g = (g_sig, g_risk, g_pos, g_entry, g_exit, ξ) where ξ stores provenance, family identifiers, and mutation history.

Module 2: Baseline Evaluation & Evidence Construction

Using real market data, the evaluator runs the current genome on train, validation, and test splits, recording scalar metrics M_X (return, Sharpe, max‑drawdown, trade count, win rate), behavioral portraits P_X^trade, and regime sensitivities P_X^regime. Additional evidence includes out‑of‑sample decay δ_oos, risk anomalies H_risk, and other diagnostics, forming an evidence package E(g) = {M_X, P_X^trade, P_X^regime, …}.

Module 3: Strategy Optimization Pipeline

From the evidence, a bottleneck diagnosis z_t = B(E_t, M_t) produces a search plan p_t = (o_t, ℓ_t, O_t, F_t, K_t), specifying the target objective, edit hierarchy (repair → bridge → redesign → family migration), allowed operators, forbidden shortcuts, and candidate budget. Candidate generation prompts the LLM with the current genome, compressed evidence, diagnosed bottleneck, and retrieved memory cases. Each candidate c_{t,i} must satisfy type‑checked edit contracts and semantic drift constraints d_G(c_{t,i}, g_t) ≤ ε_{ℓ_t}. The verifier then applies hard gates (minimum validation performance, out‑of‑sample decay thresholds, max‑drawdown limits, drift limits, executability) and scores surviving candidates with a composite metric

Q(c) = w_v·ΔS_val + w_o·ΔS_oos + w_r·ΔR_oos - λ_d·P_dd - λ_g·P_gap - λ_s·P_stress

. Candidates are classified as adopt, incubate, or reject based on this score.

Module 4: Iterative Refinement & Output

The loop tracks the best validated genome g*_t = argmax_{g∈H_t} Q_{audit}(g), where H_t contains all adopted and incubated candidates. If lower‑level repairs repeatedly fail, the controller escalates to bridge, redesign, and finally family migration. The process stops when the iteration budget is exhausted, the audit score converges, or no viable search level remains. The final output includes the optimized strategy code, before‑and‑after back‑test reports, and the full optimization trajectory.

Experiments

All data come from the AKShare database covering A‑share equities and Bitcoin (2020‑01‑01 to 2025‑12‑31) with daily frequency. The LLM component uses DeepSeek‑R1, and each optimization runs for 20 iterations.

Experimental Setup

A‑share tasks : EMA‑RSI trend‑continuation, Bollinger‑RSI mean‑reversion, volume‑breakout, low‑volatility high‑quality stock strategies. For each strategy, 30 random stocks are selected, optimized independently, and results are aggregated via equal‑weight portfolio.

Cryptocurrency tasks : MACD‑RSI‑Bollinger, ATR trend‑breakout, oversell‑reversal strategies, testing framework adaptability to high‑volatility assets.

Main Results

Across seven strategies, EvoQuant consistently improves test Sharpe (average from ‑0.298 to 0.538). Notable A‑share improvements: Bollinger‑RSI Sharpe 0.351→0.966, EMA‑RSI Sharpe ‑0.465→0.367, low‑volatility Sharpe ‑0.500→0.386, volume‑breakout Sharpe ‑0.493→0.489. Corresponding annual returns also rise dramatically (e.g., EMA‑RSI from 1.64 % to 16.83 %). Cryptocurrency results show similar trends, with MACD‑RSI‑Bollinger Sharpe ‑1.118→0.265 and oversell‑reversal Sharpe ‑0.838→0.679.

Ablation Study

On the A‑share volume‑breakout task, the full pipeline yields a Sharpe gain of 0.982. Removing or simplifying components shows their contributions: Simple threshold (0.053), Parameter‑only (0.274), Random mutation only (0.645); removing promotion engine (0.128), rule diagnosis (0.693), or memory (0.782). These results confirm that diagnosis, candidate generation, promotion engine, and memory each provide indispensable gains.

Robustness & Walk‑Forward Validation

Under doubled transaction cost and slippage, all A‑share strategies retain positive Sharpe improvements (volume‑breakout Sharpe drops to 0.133 but stays positive). Shifted data‑split tests also preserve gains. Walk‑forward tests on two rolling windows (2020‑2023, 2021‑latest) show EMA‑RSI Sharpe 0.440 and Bollinger‑RSI Sharpe 0.602, indicating that EvoQuant’s improvements are not tied to a specific data split.

Conclusions

The study highlights two core insights: (1) improvements stem from evidence‑driven, targeted edits rather than random perturbations, as demonstrated by ablation; (2) verification is the cornerstone of meaningful gains, framing LLM‑driven strategy optimization as a "generation + validation" problem where only edits passing explicit evidence gates are adopted. EvoQuant thus transforms costly manual trial‑and‑error into an automated, auditable, and verifiable iterative process.