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

Researchers introduce AlphaCrafter, a multi-agent AI framework that automates quantitative trading by continuously discovering trading factors, adapting to market regimes, and executing trades with risk constraints. Tested on CSI 300 and S&P 500 indices, the system outperforms existing baselines in risk-adjusted returns, addressing a critical gap in fully automated, adaptive trading pipeline design.

AlphaCrafter represents a meaningful advance in automating quantitative trading by abandoning static factor discovery in favor of continuous market adaptation. Traditional quantitative systems treat factor identification as a one-time problem, then apply those factors across all market conditions—an assumption that breaks down when macroeconomic regimes shift or market microstructure changes. The framework's three-agent architecture (Miner, Screener, Trader) creates a closed-loop system where factors are continuously regenerated, evaluated against current market conditions, and deployed only when statistically appropriate.

The research addresses a persistent tension in AI-driven trading: most systems either over-engineer factor discovery without considering execution constraints, or they anthropomorphize trading processes through role-playing agents that introduce behavioral bias rather than systematic logic. AlphaCrafter sidesteps this by maintaining rationality throughout the pipeline—each agent performs a discrete, quantifiable function without simulating human traders.

For the quantitative finance industry, this work validates that multi-agent systems can effectively handle non-stationary financial environments when agents are task-specialized rather than role-playing. The emphasis on regime-adaptive selection is particularly relevant given recent market volatility driven by geopolitical uncertainty and policy shifts. Cross-trial variance reduction suggests the system produces consistent results despite market noise, a critical property for institutional deployment.

Looking ahead, the key question is whether LLM-guided factor discovery genuinely uncovers novel alpha or simply optimizes existing factor families. Replication studies on out-of-sample data will determine whether this represents genuine methodological innovation or incremental improvement in existing quantitative approaches.

• →AlphaCrafter's three-agent architecture continuously discovers, selects, and executes trading factors adapted to current market regimes rather than relying on static factor assumptions.
• →The system outperforms baselines on CSI 300 and S&P 500 indices with both superior risk-adjusted returns and lowest cross-trial variance, indicating robust generalization.
• →LLM-guided factor mining enables automated alpha discovery without manual intervention, reducing dependency on human trader intuition and behavioral bias.
• →Regime-conditioned factor ensembles allow the system to adjust strategy composition when market conditions shift, addressing non-stationarity in financial markets.
• →The framework's rationality-driven design avoids anthropomorphic role-playing common in prior multi-agent trading systems, focusing instead on discrete quantifiable functions.