Multimodal News–Price Fusion for Event-Driven Return Prediction with Causal Debiasing

Abstract

The prediction of financial asset returns has long been a central challenge in computational finance, characterized by non-stationarity, high volatility, and a low signal-to-noise ratio. While recent advancements in deep learning have enabled the fusion of quantitative market data with qualitative textual streams—such as news articles and earnings reports—most existing multimodal architectures rely on associative correlations rather than causal mechanisms. This reliance renders models susceptible to spurious correlations driven by confounders, such as global market sentiment or macroeconomic shocks, which simultaneously influence both news content and asset prices. This paper introduces the Multimodal News–Price Fusion with Causal Debiasing (MNPF-CD) framework, a novel architecture that integrates structural causal models into deep neural networks to mitigate confounding bias. We propose a backdoor adjustment mechanism implemented via a variational intervention layer, allowing the model to learn invariant representations of asset dynamics. By mathematically separating the causal effect of idiosyncratic news events from systemic market noise, MNPF-CD achieves superior generalization capabilities. Our extensive experimental evaluation on S&P 500 constituents demonstrates that the proposed method significantly outperforms state-of-the-art baselines in terms of Information Coefficient (IC) and Sharpe Ratio, particularly during periods of market regime shifts.