AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay DOI: XXXXXXX.XXXXXXX ISBN: 978-1-4503-XXXX-X/18/06 CCS: Computing methodologies Multi-agent planning CCS: Computing methodologies Natural language generation CCS: Applied computing Forecasting
Ziyi Tang, Zechuan Chen, Jiarui Yang, Jiayao Mai, Yongsen Zheng, Keze Wang, Jinrui Chen, Liang Lin · 2025
Read AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay DOI: XXXXXXX.XXXXXXX ISBN: 978-1-4503-XXXX-X/18/06 CCS: Computing methodologies Multi-agent planning CCS: Computing methodologies Natural language generation CCS: Applied computing Forecasting by Ziyi Tang, Zechuan Chen, Jiarui Yang, Jiayao Mai, Yongsen Zheng, Keze Wang, Jinrui Chen, Liang Lin (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-16789.txt.
Introduction
Tang and co-authors present a 2025 paper on mining trading signals with language model agents. In quant investing, an alpha is a formula that predicts asset returns, but good alphas fade as crowds copy them. Older search tools overfit and grow too complex, while language models recycle known ideas. Their system, AlphaAgent, guards against all three failure modes at once.
One guard scores each new formula's shape against known alphas to demand novelty. A second checks that the formula matches the market idea it claims to test. A third caps structural complexity to block overfitting. Together the guards steer the search toward fresh, durable signals, the authors report.
Catalog record
| CREATOR | Ziyi Tang, Zechuan Chen, Jiarui Yang, Jiayao Mai, Yongsen Zheng, Keze Wang, Jinrui Chen, Liang Lin |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-16789
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay DOI: XXXXXXX.XXXXXXX ISBN: 978-1-4503-XXXX-X/18/06 CCS: Computing methodologies Multi-agent planning CCS: Computing methodologies Natural language generation CCS: Applied computing Forecasting Ziyi Tang, Zechuan Chen, Jiarui Yang, Jiayao Mai, Yongsen Zheng, Keze Wang, Jinrui Chen, Liang Lin Abstract . Alpha mining, a critical component in quantitative investment, focuses on discovering predictive signals for future asset returns in increasingly complex financial markets. However, the pervasive issue of alpha decay—where factors lose their predictive power over time—poses a significant challenge for alpha mining. Traditional methods like genetic programming face rapid alpha decay from overfitting and complexity, while approaches driven by Large Language Models (LLMs), despite their promise, often rely too heavily on existing knowledge, creating homogeneous factors that worsen crowding and accelerate decay. To address this challenge, we propose AlphaAgent, an autonomous framework that effectively integrates LLM agents with ad hoc regularizations for mining decay-resistant alpha factors. AlphaAgent employs three key mechanisms: (i) originality enforcement through a similarity measure based on abstract syntax trees (ASTs) against existing alphas, (ii) hypothesis—factor alignment via LLM-evaluated semantic consistency between market hypotheses and generated factors, and (iii) complexity control via AST-based structural constraints, preventing over-engineered constructions that are prone to overfitting. These mechanisms collectively guide the alpha generation process to balance originality, financial rationale, and adaptability to evolving market conditions, mitigating the risk of alpha decay. Extensive evaluations show that AlphaAgent outperforms traditional and LLM-based methods in mitigating alpha decay across bull and bear markets, consistently delivering significant alpha in Chinese CSI 500 and U.S. S&P 500 markets over the past four years. Notably, AlphaAgent showcases remarkable resistance to alpha decay, elevating the potential for yielding powerful factors. 1. Introduction Factor investing is a key strategy in modern quantitative finance, focusing on the systematic identification and exploitation of alpha factors (a.k.a., alphas) — quantifiable characteristics that can predict asset returns. However, alpha decay, or the decline of factor returns over time, arises from two main challenges. First, overfitting through excessive data mining ("p-hacking") leads to the emergence of spurious factors that appear significant in backtests but decay rapidly in real-world applications ( Fama and French 1992 ) . Second, factor crowding occurs when too many investors adopt similar strategies, which can accelerate alpha decay and trigger sudden reversals during market stress ( Asness et al. 2013 ; Falck et al. 2022 ; Chen and Cheon 2024 ) . This was evident in early 2024 with the size factor’s underperformance in China’s A-share market ( Wool 2024 ; CICC Research 2024 ) , highlighting the risks of concentrated positioning in popular factors. Thus, it is rather crucial to counteract alpha decay during alpha mining. Traditional methods for alpha mining mainly build on genetic programming (GP) ( Lin et al. 2019 ; Schmidt and Lipson 2010 ; Zhaofan et al. 2022 ; Patil 2023 ; Yu et al. 2023 ; Cui et al. 2021 ) and reinforcement learning (RL). However, they struggle to effectively address the alpha decay challenge. Traditional GP and RL approaches tend to over-emphasize the optimization of historical performance metrics while neglecting the underlying financial and economic rationale. Without sufficient consideration of financial soundness and economic intuition, these methods often produce alpha factors that show strong historical performance but experience rapid alpha decay when deployed in live markets. Large Language Models (LLMs) offer promising potential in factor mining due to their extensive understanding of financial knowledge, which could help generate and evaluate factors beyond pure statistical significance. Despite their versatility, LLMs’ direct application to alpha factor mining remains underutilized, revealing significant challenges in addressing alpha decay. The fundamental limitation lies in the lack of effective constraints in current LLM-based factor mining frameworks. Without proper theoretical and empirical constraints to guide factor mining, LLMs tend to overly rely on well-documented financial knowledge and established factors during construction, such as traditional technical indicators (e.g., RSI ( Ţăran-Moroşan 2011 ) ) and widely studied market anomalies (e.g., momentum, value, and size effects). This constrained-free approach exacerbates factor crowding - LLMs predominantly generate factors that capture the same market inefficiencies as existing ones, which are already heavily exploited by market participants. In rapidly evolving markets like the U.S. stock market, LLM-generated factors thus fail to identify novel sources of alpha, leading to suboptimal investment performance. This highlights the need for innovative approaches that constrain LLMs’ factor generation process by encouraging the exploration of unique factor constructions that balance theoretical soundness with market adaptability. To address these issues, we propose a novel paradigm that effectively constrains LLM-based factor generation to mitigate alpha decay, addressing key limitations of conventional approaches. At its core, AlphaAgent introduces three critical regularization mechanisms to guide LLM-based factor generation: (1) complexity control through symbolic expression trees and parameter counting, (2) hypothesis alignment via LLM-evaluated semantic consistency between market hypotheses and generated factors, and (3) novelty enforcement through a similarity measure based on abstract syntax trees (ASTs) against existing alpha libraries (e.g., Alpha101). AlphaAgent formalizes factor construction through an operator library and abstract syntax trees (ASTs), implementing a pairwise subtree isomorphism detection mechanism to quantify factor originality while using LLMs to verify financial intuition alignment through consistency scoring between hypotheses, descriptions, and expressions. These constraints guide LLMs to explore novel market inefficiencies while maintaining theoretical soundness, alleviating the alpha decay challenge. Based on these constraints, AlphaAgent builds an autonomous workflow encompassing hypothesis proposal, factor construction, factor development, backtesting, and feedback mechanisms. The framework begins with the idea agent for the hypothesis proposal, with the first market insight provided by domain experts. Grounded in these hypotheses, the factor agent constructs parsimonious and original factors to explore unexploited market inefficiencies, applying regularization mechanisms to balance complexity, novelty, and hypothesis alignment. The eval agent then rigorously validates factors’ executability and numerical stability while backtesting assesses their predictive effectiveness on historical data. The feedback mechanism evaluates performance metrics and theoretical soundness, guiding iterative refinement. This closed-loop process progressively derives a family of factors that capture emerging rather than overcrowded market inefficiencies, promoting alpha mining by balancing theoretical soundness with factor originality. Extensive experiments demonstrate AlphaAgent’s effectiveness in generating factors resistant to alpha decay. Through comprehensive evaluations in the Chinese CSI 500 and U.S. S&P 500 markets from January 2021 to December 2024, our framework demonstrates stable factor performance across different market regimes. Through comprehensive evaluations in the Chinese CSI 500 and U.S. S&P 500 markets from January 2021 to December 2024, our framework achieves an average annual excess return of 11.0% (IR=1.5) and 8.74% (IR=1.05) respectively after accounting for transaction costs, while maintaining robust performance across different market regimes. While traditional factors exhibit substantial decay in predictive power due to market crowding and overfitting, AlphaAgent’s alphas maintain stable predictive effectiveness throughout the past 5 years, demonstrating stronger resistance to alpha decay. The framework also displays efficiency in mining alphas, achieving an 81% higher effective factor ratio (hit ratio) while consuming 30% fewer tokens. These results substantially outperform traditional alpha mining approaches and existing LLM-driven ones in terms of both performance and performance persistence. Our main contributions can be summarized as follows: • We propose a systematic regularization mechanism to counteract alpha decay, combining originality enforcement, hypothesis alignment, and complexity control, facilitating the