AlphaCrafter: A Full-Stack Multi-Agent Framework for Cross-Sectional Quantitative Trading

In one sentence

Summary

Financial quantitative trading faces significant challenges due to market non-stationarity and the rapid decay of alpha signals. Traditional approaches often decouple factor discovery from execution or rely on static factor sets that fail to adapt to shifting macroeconomic regimes. AlphaCrafter addresses this fragmentation by introducing a closed-loop, fully automated pipeline that continuously adapts to evolving market dynamics without manual intervention. The framework is designed to maintain predictive power and risk-adjusted returns in non-stationary environments by integrating three specialized agents that operate in a coordinated daily rotation, ensuring that signal generation, selection, and execution are holistically aligned with current market conditions. The system comprises three distinct agents: a Miner, a Screener, and a Trader. The Miner utilizes LLM-guided search to autonomously generate, validate, and maintain a dynamic factor library, pruning ineffective signals to prevent alpha decay. The Screener assesses prevailing market regimes—such as trend direction, volatility, and liquidity—and constructs a regime-conditioned factor ensemble by selecting and weighting factors based on their suitability for the current environment. The Trader then translates this ensemble into executable quantitative strategies, optimizing portfolio construction and risk constraints through hyperparameter tuning and backtesting. Experiments were conducted on the CSI 300 and S&P 500 indices using daily-frequency data from 2016 to 2026, including price-volume, fundamental, and alternative data sources. The framework was evaluated against traditional technical strategies, machine learning models, deep learning architectures, and other LLM-based trading agents, using metrics such as Annualized Return, Sharpe Ratio, and Maximum Drawdown. AlphaCrafter consistently outperformed state-of-the-art baselines in both backtesting and live trading phases, achieving the highest risk-adjusted returns and the lowest cross-trial variance. Notably, it was the only method to deliver positive live trading returns on both markets, whereas many baselines suffered from severe overfitting or regime sensitivity. The framework exhibited robust stability across different backbone LLMs and demonstrated superior factor longevity compared to static factor sets. However, the study acknowledges limitations related to the daily trading frequency, which may not capture high-frequency microstructure effects, and the potential for LLM inference latency in real-time deployment. The results confirm that an integrated, adaptive factor-to-execution design yields more reliable and robust trading performance than decoupled or static approaches.

Research problem

Financial markets are non-stationary, causing static quantitative strategies and decoupled factor mining frameworks to suffer from alpha decay and regime sensitivity. Existing LLM-based trading systems often rely on role-playing agents that introduce behavioral noise or fail to unify factor discovery with execution. The core problem is the absence of a fully automated, rationality-driven framework that continuously adapts the entire quantitative pipeline—from signal generation to risk-constrained execution—to evolving market conditions without manual recalibration.

Core contributions

• Proposes AlphaCrafter, the first full-stack multi-agent framework unifying LLM-driven factor discovery, regime-sensitive selection, and risk-constrained execution in a single adaptive pipeline.
• Introduces a regime-conditioned factor ensemble construction mechanism that dynamically reweights signals based on market states without retraining or human intervention.
• Demonstrates empirical robustness across CSI 300 and S&P 500 markets, showing consistent outperformance in risk-adjusted returns and lowest cross-trial variance compared to strong baselines.
• Validates the closed-loop hypothesis-validation-execution design, proving that integrated adaptive systems mitigate overfitting and alpha decay better than static or decoupled methods.

Method

• Miner Agent: Uses LLM-guided search to generate candidate factors, validates them using IC, ICIR, turnover, and decay metrics, and maintains a dynamic factor library with periodic re-validation.
• Screener Agent: Assesses market regimes (trend, volatility, liquidity) and selects a diversified factor ensemble by assigning weights and directions based on regime suitability and correlation constraints.
• Trader Agent: Optimizes portfolio construction by tuning hyperparameters (exposure, weighting, rebalancing) via backtesting, then executes strategies under explicit risk constraints.
• Closed-Loop Architecture: Agents share a persistent memory H to propagate performance diagnostics and conditional guidance, enabling holistic adaptation.
• Evaluation: Daily-frequency trading with Z-score normalization, slippage/latency negligible, and live trading on out-of-sample data to prevent look-ahead bias.

Results & findings

• AlphaCrafter achieved a live AR of 5.70% (SR 0.7002) on CSI 300 and 9.26% (SR 0.7212) on S&P 500, outperforming all baselines in live trading.
• Baselines like LSTM and XGBoost showed strong backtesting AR but collapsed in live trading (e.g., LSTM live AR -7.74% on CSI 300), highlighting severe overfitting.
• AlphaCrafter exhibited the lowest cross-trial variance, indicating reliable adaptation rather than overfitted artifacts.
• Factor decay analysis showed AlphaCrafter maintained stable IC (0.015-0.025) over time, whereas static Alpha158 factors suffered severe volatility and negative IC.
• Ablation studies confirmed each agent contributes meaningfully: removing Miner reduced AR, removing Screener increased MDD, and removing Trader lowered Sharpe ratios.
• Performance was stable across different backbone LLMs (GPT, Claude, Gemini), confirming the framework's model-agnostic design.

Data & experiments

• Datasets: CSI 300 and S&P 500 constituents with daily OHLCV, fundamental indicators (PE, PB, etc.), financial statements, and alternative data (news, announcements) from 2016 to 2026.
• Baselines: Traditional strategies (MACD, Grid), ML (LightGBM, XGBoost), Deep Learning (LSTM, Transformer, TRA), and LLM agents (TradingAgents, TradingGroup, RD-Agent, AlphaAgent).
• Metrics: Annualized Return (AR), Sharpe Ratio (SR), Maximum Drawdown (MDD), Information Coefficient (IC), and IC Information Ratio (ICIR).
• Experimental Design: Daily-frequency cross-sectional trading with backtesting and live trading phases; 10 independent trials for stability analysis; ablation studies for each agent component.
• Live Trading: Evaluated on data strictly outside the training cutoff of backbone LLMs to eliminate look-ahead bias and memory effects.

Limitations & caveats

• Daily-frequency trading may not capture high-frequency microstructure effects or arbitrage opportunities.
• LLM inference latency could be a barrier for real-time deployment, though not explicitly quantified in the paper.
• Performance relies on the quality of the underlying LLM backbone, though stability across models is claimed.
• Live trading period is short (4 months), which may not fully capture long-term regime shifts or tail events.
• Factor generation is LLM-driven, which may introduce biases or non-interpretability compared to traditional factor models.
• Slippage and execution costs are assumed negligible, which may not hold for larger capital allocations or different market conditions.

Curator notes

Discovery status: institutional_trading_round3_promoted. Source links: AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay