New Framework Uses AI to Discover Better Market Signals at Scale

A team of researchers has developed a novel approach to uncovering predictive financial signals—known as "alphas"—by merging large language model reasoning with code-based evolutionary search. The Cognitive Alpha Mining Framework (CogAlpha) addresses a longstanding challenge in quantitative finance: finding reliable trading patterns within massive datasets where actual signals are rare and easily drowned out by noise.

The Problem#

Finding effective alphas has historically demanded choosing between competing tradeoffs. Deep learning models can process vast amounts of data but typically produce predictions that are opaque and prone to overfitting. Symbolic or formula-based methods create interpretable results but often generate redundant expressions that don't hold up well when applied to new market data. Neither approach can conduct the kind of broad, structured exploration that mimics human reasoning—balancing systematic logic with creative insights.

How CogAlpha Works#

Rather than limiting the search to a narrow slice of possibilities, the framework treats large language models as reasoning agents that can iteratively refine alpha candidates. The system combines three key elements:

• Code-level representation of alphas, making them executable and economically grounded
• Multi-stage LLM prompts that guide refinement and mutation of candidate formulas
• Evolutionary search paired with financial feedback loops to recombine and improve promising approaches

This synergy enables the models to explore a much larger portion of the search space while maintaining interpretability and logical consistency.

Validation Across Markets#

The researchers tested CogAlpha on stock data from five datasets spanning three different markets. The results showed the framework consistently produced alphas with stronger predictive accuracy, more reliable performance across different time periods, and better ability to transfer to untested data—compared to existing deep learning and genetic programming approaches.

The work underscores the potential of combining evolutionary optimization with LLM-based reasoning to automate the discovery of trading signals that are both powerful and explainable.