AlphaCrafter: A Full‑Stack Multi‑Agent Framework for Adaptive Cross‑Sectional Quant Trading

Background

Financial markets are high‑dimensional, non‑linear dynamic systems with heavy tails, volatility clustering, and complex cross‑sectional dependencies. Traditional quantitative strategies rely on static models and periodic manual recalibration, which incurs high maintenance costs when market structures shift.

Large language models (LLMs) enable automation of the quantitative pipeline, but existing LLM‑based systems either stop after factor generation or separate factor discovery from execution, resulting in fragmented workflows that assume a static environment.

Problem Definition

Three deficiencies of current quantitative methods are identified:

Static models and manual calibration cannot adapt to changing market structures.

LLM‑based approaches either terminate after factor generation or decouple factor discovery from execution, preventing a fully adaptive pipeline.

Agent‑centric designs that mimic human committees introduce inference latency and behavioral bias, while pure code‑driven agents lack dynamic market‑state awareness.

A unified, full‑stack framework that continuously discovers factors, selects market‑state‑sensitive factors, and enforces risk constraints is required.

Method

Environment Formalization

The trading environment is defined as a tuple (M, Z, \Pi, T, J, H) where: M – market‑state space (overall conditions, asset features, macro indicators, sentiment signals, etc.). Z – dynamic factor library; each factor maps historical observations to an N -dimensional cross‑sectional signal vector. \Pi – admissible strategy space parameterized by portfolio construction rules and risk‑constraint parameters. T – set of discrete trading days. J – evaluation function measuring risk‑adjusted performance. H – shared memory channel that stores observations, factor validation results, and execution feedback for all agents.

Miner Agent

The Miner continuously generates candidate factors guided by an LLM, validates them using information coefficient (IC), IC stability, turnover, and decay curves, and integrates qualified factors into Z. Exploration stops when internal quality and diversity thresholds are satisfied. Algorithm 1 (illustrated in the paper) details the generation‑validation‑maintenance loop.

Screener Agent

The Screener evaluates the current market state M, ranks factors in Z based on relevance to observed trends, volatility, and correlation structure, and selects a diversified subset E_f with directional weights. Algorithm 2 describes the ranking, correlation‑aware selection, and memory‑update process.

Trader Agent

The Trader combines the selected factor set E_f with market‑state assessment R_f, applies self‑determined risk constraints and proprietary logic, and optimizes hyper‑parameters to generate a trading strategy \pi_t. Algorithm 3 outlines the reference‑strategy‑based search, risk‑exposure control via parameters \beta and \gamma, and feedback‑driven memory update.

System Objective

On each decision day t, the three agents produce a strategy \pi_t that maximizes the evaluation function J, which penalizes drawdown and volatility while encouraging stable capital growth.

Experiments

Experimental Setup

Dataset – Comprehensive data covering China A‑share CSI 300 constituents and US S&P 500 constituents, including daily price‑volume, fundamental ratios, quarterly financial statements, and alternative data (news, announcements). The time split follows Table 1 of the paper.

Metrics – Annualized return (AR), Sharpe ratio (SR), maximum drawdown (MDD) for portfolio performance; IC and ICIR for factor‑level evaluation.

Baselines – Five families: (1) Quantitative methods (MACD, grid trading); (2) Machine‑learning methods (LightGBM, XGBoost); (3) Deep‑learning methods (LSTM, Transformer, TRA); (4) Traditional trading‑agent methods (TradingAgents, TradingGroup); (5) Quant‑trading‑agent methods (RD‑Agent, AlphaAgent). All baselines use original hyper‑parameters and share the same reference strategy \pi_{ref}. LLM‑based baselines employ three backbone models (GPT‑5.3 Codex, Claude Opus 4.6, Gemini 3.1 Pro); results report the best‑performing backbone.

Settings – Daily‑frequency cross‑sectional trading framework; portfolio weights updated at market close. Transaction costs are ignored. Each LLM‑based configuration is repeated ten times; metrics are averaged over the inter‑quartile range to mitigate outliers. Real‑time trading windows are strictly out‑of‑sample relative to LLM training data.

Main Results

AlphaCrafter is the only method achieving positive risk‑adjusted returns in all phases and markets. On CSI 300, back‑test Sharpe = 1.5322, real‑time annualized return = 5.70 % with Sharpe = 0.7002 and max drawdown = ‑5.31 %. On S&P 500, back‑test max drawdown = ‑7.86 %, real‑time max drawdown = ‑3.95 %, real‑time annualized return = 9.26 % and Sharpe = 0.7212, indicating reduced market‑state sensitivity and over‑fitting.

Stability Study

Overall Performance Stability – Ten independent runs show AlphaCrafter’s return distribution is tightly clustered with a narrow inter‑quartile range and stable median in both markets, whereas several baselines exhibit high variance or negative outliers.

Model Robustness – Across the three backbone LLMs, AlphaCrafter’s back‑test performance remains consistent, confirming low sensitivity to the specific LLM choice. Claude Opus 4.6 yields slightly better factor mining and strategy execution.

Alpha Decay Analysis

Alpha decay is evaluated over four consecutive half‑year windows (Jan 2024 – Jan 2026). Baseline Alpha158 methods show divergent behavior: a periodically updated top‑20 factor set maintains high IC, while a static top‑20 set’s IC fluctuates and becomes negative, highlighting sensitivity to market state. Agent‑based methods (RD‑Agent, AlphaAgent, AlphaCrafter) keep IC stable between 0.015–0.025, with AlphaCrafter exhibiting the smallest decay in later windows.

Ablation Study

Three ablations replace each agent with a non‑adaptive alternative: w/o Miner: static Alpha158 factor set replaces the Miner. w/o Screener: uniform random sampling and equal weighting replace the Screener. w/o Trader: fixed reference strategy replaces the Trader.

All ablations degrade every metric, confirming each component’s contribution. Removing the Miner lowers annualized returns, highlighting the value of LLM‑driven factor generation; removing the Screener increases max drawdown the most, underscoring its risk‑mitigation role; removing the Trader yields the lowest Sharpe ratio, showing the importance of adaptive execution.

Conclusion

AlphaCrafter demonstrates that a closed‑loop, LLM‑augmented multi‑agent architecture can continuously adapt to evolving market conditions, outperforming a wide range of static and learning‑based baselines in both back‑test and real‑time settings while maintaining stability across random seeds, LLM backbones, and market regimes.