Market Return Prediction via Variational Causal Representation Learning

Abstract

This paper proposes a multi-factor market return prediction model based on variational causal representation learning. The goal is to improve prediction accuracy, robustness, and interpretability. The method adopts a variational inference framework to learn latent causal structures from high-dimensional factor data. It incorporates a causal regularization term and a counterfactual consistency loss to enhance the model's resistance to spurious correlations and data perturbations. The model consists of three components: a variational encoder, a generator, and a predictor. It is trained end-to-end to jointly learn causal representations and perform return prediction. In multiple experiments, the proposed method outperforms representative existing models in terms of mean squared error, causal consistency, and robustness. It also shows strong adaptability in transfer learning tasks across different economic cycles. Ablation studies confirm the contribution of each module to overall performance. These results further demonstrate the value of causal modeling in improving the stability of financial prediction models.