Discovering effective predictive signals, or alphas, in high-dimensional financial data with an extremely low signal-to-noise ratio remains a difficult open problem. Deep learning, genetic programming, and more recent LLM-based methods have improved factor generation, but they still explore only a narrow region of the vast alpha search space.
Neural models often produce opaque and fragile patterns. Symbolic and formula-based methods can generate redundant expressions that lack economic grounding and generalize poorly. Although these approaches differ in form, they share a central limitation: none conducts broad, structured, human-like exploration while balancing logical consistency with creative leaps.
Code as a richer alpha representation
CogAlpha expands the effective search space by representing alphas as code rather than restricting them to fixed formulas or predefined operator sets. Code can express more complex structures, transformations, and conditional logic while remaining executable, inspectable, and open to systematic validation.
This representation gives the system room to explore beyond familiar factor templates without abandoning interpretability. Candidate signals remain concrete research objects that can be read, tested, challenged, and refined.
LLM-driven reasoning meets evolutionary search
The framework treats large language models as adaptive cognitive agents. Through multi-stage prompts and continuous financial feedback, these agents iteratively refine, mutate, and recombine alpha candidates rather than generating one-off expressions.
Evolutionary optimization provides breadth and selection pressure, while LLM reasoning contributes structural variation, logical coherence, and economically meaningful changes. Their combination enables deeper exploration, richer diversity, and more interpretable alpha discovery.
Evidence across markets
Experiments across five stock datasets from three stock markets show that CogAlpha consistently discovers alphas with stronger predictive accuracy, robustness, and generalization than existing methods.
The findings highlight the promise of aligning evolutionary optimization with LLM-based reasoning for automated and explainable alpha discovery. The complete methodology, experiments, and author information are available in the paper on arXiv.