DAE-SRP: A Deep Auto-Encoder Framework for Denoised Stock Return Prediction

Abstract

Predicting stock returns remains challenging because financial time series are inherently noisy and driven by unobserved dynamic factors. Although numerous deep learning models exist, they typically process raw price data end-to-end. This direct processing makes them prone to overfitting and unable to capture underlying dynamic states. To address this issue, we propose the Deep Auto-Encoder for Stock Return Prediction (DAE-SRP), a framework that learns denoised representations of stock dynamics. DAE-SRP employs an auto-encoder architecture to effectively distinguish meaningful movement patterns from chaotic market signals. Specifically, the encoder combines a stock-wise convolutional layer, a GRU layer with an attentive distinctive generative filter, and a stock-mixing layer to capture unique local dynamics, long-term temporal dependencies, and cross-stock correlations, respectively. These latent representations simultaneously feed into an MLP for return prediction and a decoder module for input reconstruction. Extensive experiments on four benchmark datasets show that DAE-SRP outperforms existing baselines in portfolio trading, improving Sharpe and Sortino ratios by up to 11.9% and 17.4%, respectively.