A survey on large language model-based alpha mining

Abstract

Alpha mining, which refers to the systematic discovery of data-driven signals predictive of future cross-sectional returns, is a central task in quantitative research. Recent progress in large language models (LLMs) has sparked interest in LLM-based alpha mining frameworks, which offer a promising middle ground between human-guided and fully automated alpha mining approaches and deliver both speed and semantic depth. This study presents a structured review of emerging LLM-based alpha mining systems from an agentic perspective, and analyzes the functional roles of LLMs, ranging from miners and evaluators to interactive assistants. Despite early progress, key challenges remain, including simplified performance evaluation, limited numerical understanding, lack of diversity and originality, weak exploration dynamics, temporal data leakage, and black-box risks and compliance challenges. Accordingly, we outline future directions, including improving reasoning alignment, expanding to new data modalities, rethinking evaluation protocols, and integrating LLMs into more general-purpose quantitative systems. Our analysis suggests that LLM is a scalable interface for amplifying both domain expertise and algorithmic rigor, as it amplifies domain expertise by transforming qualitative hypotheses into testable factors and enhances algorithmic rigor for rapid backtesting and semantic reasoning. The result is a complementary paradigm, where intuition, automation, and language-based reasoning converge to redefine the future of quantitative research.

1. Introduction

The construction of quantitative investment strategies is an inherently iterative and data-driven process, involving stages such as data preprocessing, alpha factor design, model training, portfolio optimization, trade execution, and performance attribution. At the core of this process lies alpha mining, the systematic discovery of signals that predict future cross-sectional returns. From a computational perspective, an alpha refers to any variable, either constructed or observed, which can be used to rank financial assets by their expected relative performance. Rather than forecasting exact price levels, an alpha aims to distinguish between likely outperformers and underperformers over a given horizon.

A key feature of alpha mining is its practical orientation. Therefore, in this paper, we adopt a broad and practical definition: an alpha is any data-derived signal that exhibits empirical predictive power for excess returns. It may originate from financial ratios, technical indicators, sentiment scores, or other structured or unstructured information sources. This view extends beyond the classical asset pricing notion, where the alpha denotes returns unexplained by known risk factors. In addition, the alpha mining research field concentrates mostly on single-factor alphas, i.e., structured, interpretable signals that encode specific return hypotheses. These alphas are typically defined in symbolic or rule-based form, and may depend on one or multiple features.

Compared to black-box models that extract latent structure from high-dimensional inputs, single-factor alphas provide transparency, control, and attribution. Each single-factor alpha can be tested independently, evaluated statistically, and combined modularly to form larger alpha libraries. These properties make them suitable for systematic validation and scalable deployment. Alpha mining methodologies have evolved through multiple stages. Traditionally, the process has relied on human-driven alpha design, where candidate signals are manually constructed and peer-reviewed based on economic theory or domain expertise. Despite growing automation, such handcrafted alphas remain a cornerstone of quantitative research, valued for their interpretability, domain alignment, and theoretical grounding. With the surge of artificial intelligence (AI), algorithm-based alpha mining emerges as a scalable alternative. Early systems leverage heuristic search, automated feature construction, deep learning, and reinforcement learning to uncover patterns across large feature spaces. These approaches expand the search frontier and reveal signals that often elude manual exploration.