A1937
Title: Temporal information extraction for cross-sectional stock return prediction
Authors: Hsin-Yu Chiu - National Central University (Taiwan)
Kai-Xiang Chang - National Central University (Taiwan) [presenting]
Abstract: The aim is to examine whether temporal information from the joint evolution of firm characteristics and market states improves cross-sectional stock return prediction. While recent empirical asset-pricing studies increasingly rely on large predictor sets, such high-dimensional settings may cause static models to miss nonlinear interactions and dynamic patterns in rolling histories. To address this issue, a two-stage framework is proposed that separates temporal representation learning from final cross-sectional prediction. In the first stage, a temporal convolutional network processes twelve-month rolling sequences of firm characteristics and market-state variables to extract a sequence-based signal for future stock ranking. In the second stage, this signal is combined with contemporaneous predictors and incorporated into a tree-based model to predict next-month returns. Using U.S. equity data from 2000 to 2023, the results show that the framework generates economically meaningful out-of-sample long-short portfolio performance and significant Fama-French five-factor alpha. The learned temporal signal provides information beyond conventional predictors and remains relevant in interpretability analyses. Overall, the evidence suggests that temporal representations can enhance cross-sectional stock return prediction, although strong single-stage benchmarks remain highly competitive.
