Lightweight Multimodal Fusion Architectures for Intraday Abnormal Return Reversal Prediction of S&P 500 Constituent Stocks: A Literature Review

Abstract

Integrating lightweight deep learning models with multimodal fusion techniques provides a promising approach to complex predictive tasks in resource-constrained environments. Drawing on recent literature, this paper systematically reviews research in three major areas: lightweight deep learning, multimodal fusion, and intraday reversal prediction and quantitative trading strategy optimization for S&P 500 constituent stocks. Empirical studies in non-financial domains show that lightweight neural architectures can balance predictive accuracy and computational efficiency. However, their adoption in financial forecasting remains limited. Most multimodal fusion methods integrate information at the feature level. The intraday reversal effect in S&P 500 constituent stocks has been empirically confirmed. However, existing prediction models typically rely on single-modal inputs or complex architectures, without combining lightweight design and multimodal fusion, making them unsuitable for real-time intraday trading. Accordingly, this paper identifies several key research gaps and proposes hypothesis and key insights to support the practical deployment of quantitative trading.