Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Factor Mining

TL;DR

The paper tackles the challenge of discovering interpretable, high-performing alpha factors in quantitative finance, where traditional formulaic approaches suffer from poor interpretability and search inefficiency. It proposes an LLM-guided Monte Carlo Tree Search framework that iteratively generates and refines symbolic alphas, with backtesting feedback steering exploration and a Frequent Subtree Avoidance mechanism to diversify structures. Empirical results on Chinese A-shares and U.S. markets show superior predictive power (IC and RankIC) and trading performance (AER, IR) relative to baselines, along with favorable interpretability and search efficiency. The approach is flexible with respect to LLM backbones and demonstrates a practical, scalable path toward more transparent and effective formulaic alpha mining in finance.

Abstract

Alpha factor mining is pivotal in quantitative investment for identifying predictive signals from complex financial data. While traditional formulaic alpha mining relies on human expertise, contemporary automated methods, such as those based on genetic programming or reinforcement learning, often struggle with search inefficiency or yield alpha factors that are difficult to interpret. This paper introduces a novel framework that integrates Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS) to overcome these limitations. Our framework leverages the LLM's instruction-following and reasoning capability to iteratively generate and refine symbolic alpha formulas within an MCTS-driven exploration. A key innovation is the guidance of MCTS exploration by rich, quantitative feedback from financial backtesting of each candidate factor, enabling efficient navigation of the vast search space. Furthermore, a frequent subtree avoidance mechanism is introduced to enhance search diversity and prevent formulaic homogenization, further improving performance. Experimental results on real-world stock market data demonstrate that our LLM-based framework outperforms existing methods by mining alphas with superior predictive accuracy and trading performance. The resulting formulas are also more amenable to human interpretation, establishing a more effective and efficient paradigm for formulaic alpha mining.

Figures (20)

• Figure 1: A high-level schematic of our proposed alpha mining pipeline. The pipeline features an iterative Alpha Search loop where an LLM, guided by MCTS, generates and refines formulas. Effective alphas are collected in an Alpha Zoo before being used in the final Strategy Building stage.
• Figure 2: Overview of our LLM-powered MCTS framework. The process begins with node selection via UCT. A refinement dimension is then chosen based on the node's multi-dimensional evaluation scores. The LLM first proposes a conceptual refinement suggestion for that dimension and then translates it into a concrete formula. The new formula is backtested, and its performance results are used to expand the tree with a new node.
• Figure 3: Illustration of the Frequent Subtree Avoidance (FSA). The set of effective alphas from the Alpha Repository is mined for frequent subtrees. The most frequent ones are identified, and the LLM is subsequently instructed to avoid generating new formulas containing these common structural motifs.
• Figure 4: The average predictive performance of LightGBM and MLP models trained on alphas mined by different methods.
• Figure 5: Analysis of search dynamics. (a, b) Alpha mining efficiency, measured as the count of effective alphas found versus total generated. (c, d) Average in-sample and out-of-sample RankIC of the top 50 alphas over generations.