Stock price crash prediction based on multimodal data machine learning models

Abstract

This study introduces multimodal data machine learning framework to predict stock crashes. It encapsulates market data, graph data cultivated from industry affiliations through node2vec, and text data derived from sentiment analysis. The LightGBM is utilized, marking an improvement by 7.13% over preceding studies, achieving 75.85% balanced accuracy. An innovative long-short portfolio construction approach is articulated, demonstrating the practical significance of the predictions, with a 4.75% portfolio return in 2022 — a 27.26% advancement over the CSI 300. This endeavour in leveraging multimodal data machine learning for stock crash prediction offers a promising performance, serving as a valuable reference for investors.