ACT: Anti-Crosstalk Learning for Cross-Sectional Stock Ranking via Temporal Disentanglement and Structural Purification

arXiv cs.LG / 4/23/2026

📰 NewsModels & Research

共有:

Key Points

The paper studies cross-sectional stock ranking using deep learning that models both time-series dynamics and inter-stock dependencies, but identifies a major drawback in prior graph-based methods: crosstalk between predictive factors.
It distinguishes two types of crosstalk—temporal-scale crosstalk (entangled trend/ fluctuation/ shock patterns that spread non-transferable local information) and structural crosstalk (heterogeneous relations fused together that hide relation-specific signals).
The proposed Anti-CrossTalk (ACT) framework tackles these issues by decomposing each stock’s sequence into trend, fluctuation, and shock components, then using dedicated branches to decouple non-transferable patterns.
ACT also introduces a Progressive Structural Purification Encoder to sequentially reduce structural crosstalk on the trend component after addressing temporal-scale crosstalk, followed by an adaptive fusion module for ranking.
Experiments on CSI300 and CSI500 report state-of-the-art performance, including improvements up to 74.25% on CSI300, along with better portfolio outcomes.

Abstract

Cross-sectional stock ranking is a fundamental task in quantitative investment, relying on both temporal modeling of individual stocks and the capture of inter-stock dependencies. While existing deep learning models leverage graph-based approaches to enhance ranking accuracy by propagating information over relational graphs, they suffer from a key challenge: crosstalk, namely unintended information interference across predictive factors. We identify two forms of crosstalk: temporal-scale crosstalk, where trends, fluctuations, and shocks are entangled in a shared representation and non-transferable local patterns contaminate cross-stock learning; and structural crosstalk, where heterogeneous relations are indiscriminately fused and relation-specific predictive signals are obscured. To address both issues, we propose the Anti-CrossTalk (ACT) framework for cross-sectional stock ranking via temporal disentanglement and structural purification. Specifically, ACT first decomposes each stock sequence into trend, fluctuation, and shock components, then extracts component-specific information through dedicated branches, which effectively decouples non-transferable local patterns. ACT further introduces a Progressive Structural Purification Encoder to sequentially purify structural crosstalk on the trend component after mitigating temporal-scale crosstalk. An adaptive fusion module finally integrates all branch representations for ranking. Experiments on CSI300 and CSI500 demonstrate that ACT achieves state-of-the-art ranking accuracy and superior portfolio performance, with improvements of up to 74.25% on the CSI300 dataset.