Key AI-Driven Quantitative Finance Papers from KDD2025

AlphaAgent: LLM‑Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay

Link: https://arxiv.org/pdf/2502.16789

Authors: Ziyi Tang, Zechuan Chen, Jiarui Yang, Jiayao Mai, Yongsen Zheng, Keze Wang, Jinrui Chen, Liang Lin

Abstract: Alpha mining in quantitative investing aims to discover signals that predict asset returns, but it suffers from alpha decay. Traditional methods such as genetic programming and reinforcement learning struggle to address this problem, while existing LLM‑based approaches lack constraints, leading to factor convergence and accelerated decay. AlphaAgent integrates an LLM agent with three regularization mechanisms: (1) an AST‑based originality constraint, (2) hypothesis‑factor alignment evaluated by the LLM, and (3) AST‑based structural complexity control. These mechanisms balance originality, financial principles, and market adaptability, thereby reducing decay risk. Experiments on US and Chinese markets show AlphaAgent outperforms traditional and LLM baselines, maintaining significant alpha and strong resistance to decay.

Learning Universal Multi‑level Market Irrationality Factors to Improve Stock Return Forecasting

Link: https://arxiv.org/pdf/2502.04737

Authors: Chen Yang, Jingyuan Wang, Xiaohan Jiang, Junjie Wu

Abstract: Existing deep‑learning models for stock investment ignore specific irrational factors. The UMI model learns irrational factors at both stock and market levels. At the stock level, it estimates a rational price and treats the deviation from the actual price as an irrational factor. At the market level, abnormal synchronous volatility defines irrational behavior, which is incorporated into the market representation vector via two self‑supervised tasks. Experiments on US and Chinese markets demonstrate the model’s effectiveness and the universality of the learned factors.

Progressive Dependency Representation Learning for Stock Ranking in Uncertain Risk Contrasting (PDU)

Link: https://dl.acm.org/doi/pdf/10.1145/3690624.3709189

Authors: Li Huang, Yanzhe Xie, Qiang Gao, Kunpeng Zhang, Guisong Liu, Xueqin Chen

Abstract: To address insufficient coupling of time‑correlation dependencies and inadequate modeling of uncertain risk in stock ranking, PDU introduces a progressive dependency (PD) module that captures time dependency, relational dependency, and multi‑term dynamics, together with an uncertain risk contrast mechanism that enhances robustness to noise and missing data. Experiments on four major stock datasets (NASDAQ, NYSE, S&P 500, SZSE) and the Nikkei 225 show PDU significantly outperforms baselines such as LSTM, GCN, and RT‑GCN on Sharpe ratio, mean reciprocal rank, NDCG@5, and downside deviation.

Pre‑training Time Series Models with Stock Data Customization (SSPT)

Link: https://dl.acm.org/doi/pdf/10.1145/3711896.3737005

Authors: Mengyu Wang, Tiejun Ma, Shay B. Cohen

Abstract: Stock selection predicts prices and identifies high‑profit potential stocks. Existing work focuses on model architecture and graph construction, while pre‑training strategies are under‑explored. SSPT proposes three stock‑specific pre‑training tasks—stock code classification, sector classification, and moving‑average prediction—and builds a two‑layer Transformer‑based Stock‑Specific Pre‑training Transformer. Extensive experiments on five stock datasets covering four markets and two time periods show SSPT consistently achieves higher cumulative return and Sharpe ratio than market baselines and prior methods. Simulated data experiments further analyze the underlying mechanisms.

Enhancer: A Distribution‑Aware Framework with Temporal‑Relational Meta‑Learning for Stock Prediction

Link: https://dl.acm.org/doi/pdf/10.1145/3711896.3736934

Authors: Weijun Chen, Shun Li, Heyuan Wang, Tengjiao Wang

Abstract: Accurate stock prediction requires models that adapt to market changes. Existing temporal‑relational models struggle with distribution shifts in time and relation. Enhancer is a model‑agnostic framework that introduces a Time Meta‑Learner (TML) with Reactive Point Process Attention (RPPsAtt) to capture fine‑grained temporal points, and a Relation Meta‑Learner (RML) with an Approximate Intervention (Ant) mechanism to mitigate relational distribution shift. Experiments on four long‑term stock datasets for trend prediction and investment recommendation show Enhancer improves profit by an average of 29.3% and Sharpe ratio by 18.54% over baselines.