A diagram showing the working process of the RL-LLM hybrid alpha generator HARLA. The RL-based optimization loop and the LLM-based one share the same alpha pool to continuously and alternately improve upon it. Notice that the procedure of the two feedback loops closely resembles each other
In quantitative trading, transforming historical financial data into predictive signals, known as alpha factors, is fundamental. Formulaic alphas, expressed through explicit mathematical formulas, are highly valued for their transparency and interpretability compared to opaque deep learning models. However, traditional discovery methods rely heavily on Genetic Programming (GP), which often searches blindly through vast symbolic spaces. Without financial intuition, these methods suffer from low search efficiency and struggle to identify factors with deep underlying logic, limiting the development of effective trading strategies.
To address this, the team proposed HARLA, which incorporates LLMs into the optimization loop. The framework alternates between LLM-based sampling and evolutionary optimization. The LLM acts as a “knowledge guide,” suggesting promising formula skeletons based on financial priors, while evolutionary operators refine these candidates through fine-tuning and crossover. This hybrid approach synergizes the high-level reasoning of LLMs with the robust search capabilities of evolutionary algorithms, effectively balancing exploration of new ideas with the exploitation of existing successful patterns.
Empirical tests on the A-share market demonstrate that HARLA significantly outperforms traditional GP-based baselines in key metrics like Information Ratio (IR) and Information Coefficient (IC). Beyond numerical performance, the framework consistently identifies factors with sound financial logic, reducing the risk of statistical overfitting. This research provides a new paradigm for intelligent quantitative research, showcasing the potential of LLMs in specialized symbolic regression tasks and accelerating the development of transparent financial signals.
