Target-Oriented Causal Representation Learning for Robust Cross-Market Return Prediction

Abstract

This paper addresses key challenges in cross-market return prediction, including feature redundancy, limited information transfer, and distributional instability. A target-oriented causal representation learning method is proposed to tackle these issues. Centered on the target variable, the method incorporates causal structure modeling to guide the extraction of latent representations that are highly causally relevant to the prediction task from multi-source market data. This enhances both the effectiveness and robustness of the modeling process. Specifically, the model constructs a latent causal space and jointly optimizes three loss functions: minimizing prediction error, maximizing target causal mutual information, and aligning cross-domain representations. This framework balances task relevance, causal interpretability, and cross-market transferability. Extensive experiments are conducted on real-world financial datasets to evaluate the effectiveness of the proposed method. These include comparisons with different models, causal regularization ablation, robustness tests under noise perturbations, stability analysis across time windows, and transfer experiments across different market combinations. The results show that the proposed method consistently outperforms traditional deep models and existing transfer approaches across all performance metrics. It demonstrates clear advantages in handling distribution shifts, non-stationary data, and multi-market heterogeneity, validating the modeling potential and practical applicability of causal-oriented representation learning in cross-market financial prediction tasks.