Cognitive Alpha Mining via LLM-Driven Code-Based Evolution

Jul 16, 2026

Upload by Patrick

Discovering effective predictive signals (“alphas”) from high-dimensional financial data with extremely low signal-to-noise ratio remains challenging. Existing deep learning, genetic programming, and LLM-based factor generation often explore only a narrow portion of the alpha search space: neural methods are opaque and fragile, while symbolic or formula approaches can be redundant and economically ungrounded. CogAlpha addresses this by combining code-level alpha representation with LLM-driven reasoning and evolutionary search. Experiments on five stock datasets show consistently stronger predictive accuracy, robustness, and generalization.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages35

Reading Time1 h 28 min

Introduction

Alpha mining challenges and evolution

Manual, machine learning, and symbolic approaches

LLM-based alpha mining and the remaining gap

Cognitive Alpha Mining via LLM-Driven Code-Based Evolution Fengyuan Liu2 ,3 * Yi Huang 1 Sichun Luo2 ,3 Yuqi Wang2 ,3 Yazheng Yang2 ,3 Xinye Li2 ,3 Zefa Hu 1 Junlan Feng 1 Qi Liu2 ,3 * 1Jiutian Research, China Mobile 2 School of Computing and Data Science, The University of Hong Kong 3 Grace Investment Machine Correspondence: [oxfengyuan@gmail.com](oxfengyuan@gmail.com), liuqi@cs.hku.hk arXiv :2511 . 18850v4 [ cs .CL] 11 Jul 2026 Abstract Discovering effective predictive signals, or “alphas,” from financial data with high dimensionality and extremely low signal-to-noise ratio remains a difficult open problem. Despite progress in deep learning, genetic programming, and, more recently, large language model (LLM)–based factor generation, existing approaches still explore only a narrow region of the vast alpha search space. Neural models tend to produce opaque and fragile patterns, while symbolic or formula-based methods often yield redundant or economically ungrounded expressions that generalize poorly. Although different in form, these paradigms share a key limitation: none can conduct broad, structured, and human-like exploration that balances logical consistency with creative leaps. To address this gap, we introduce the Cognitive Alpha Mining Framework (CogAlpha), which combines codelevel alpha representation with LLM-driven reasoning and evolutionary search. Treating LLMsas adaptive cognitive agents, our framework iteratively refines, mutates, and recombines alpha candidates through multi-stage promptsand financial feedback. This synergistic design enables deeper thinking, richer structural diversity, and economically interpretable alpha discovery, while greatly expanding the effective search space. Experiments on 5 stock datasets from 3 stock markets demonstrate that CogAlpha consistently discovers alphas with superior predictive accuracy, robustness, and generalization over existing methods. Our results highlight the promise of aligning evolutionary optimization with LLM-based reasoning for automated and explainable alpha discovery. 1 Introduction Alpha mining is the process of discovering predictive financial signals, or “alphas,” from financial markets such as the stock market to forecast future asset returns. However, since financial markets are * Corresponding author characterized by high dimensionality, time-varying volatility (Engle, 1982), and a low signal-to-noise ratio, it remains challenging to identify explainable, reliable, and diverse alphas that support sustainable profitability and effective risk management. Over the decades, alpha mining has undergone several major transformations: from manual construction, to machine learning–driven automation, and more recently, to generative and reasoning-based exploration using LLMs (Guo et al., 2024a) . In the earliest stage, alpha factors were manually designed by financial experts, grounded in economic intuition and empirical observation. Classic examples include the Fama–French factors (Fama and French, 1992) and various documented financial anomalies (Harvey et al., 2016 ; Hou et al., 2017) . These human-crafted alphas are interpretable and theoretically sound. However, the design process is inherently labor-intensive and inefficient. As financial markets became increasingly complex and data-rich, manual approaches struggled to scale, resulting in diminishing returns and crowding among similar strategies. To enhance efficiency, researchers began leveraging machine learning models for alpha discovery. Some studies directly employed neural networks (Duan et al., 2022 ; Xu et al., 2021a,b) to implicitly extract complex and nonlinear alpha structures from market data through deep learning. These neural approaches demonstrate strong predictive power and the ability to capture highdimensional and nonlinear dependencies. However, they also suffer from inherent weaknesses: such models often behave as black boxes, making it difficult to trace the

FAQs

What problem does CogAlpha address in alpha mining?

CogAlpha targets the difficulty of discovering explainable, reliable, and diverse predictive signals from high-dimensional, noisy financial data, where existing methods explore only a limited portion of the alpha search space.

How does CogAlpha combine LLMs with code-based evolutionary search?

CogAlpha treats LLMs as adaptive cognitive agents and iteratively refines, mutates, and recombines alpha candidates using multi-stage prompts together with financial feedback to guide evolution.

What do experiments on stock datasets show about CogAlpha?

Across five stock datasets from three stock markets, CogAlpha consistently finds alphas with superior predictive accuracy, robustness, and generalization compared with existing methods.

Jul 16, 2026

Upload by Patrick