Multimodal Fusion-Based Explainable Model for Capital Market Intelligent Audit Risk Assessment

Abstract

Amid global digital transformation and heterogeneous financial data explosion, traditional audits fail to effectively utilize unstructured information and identify dynamic fraud. This study proposed an intelligent audit risk assessment model integrating multimodal data fusion and explainable stacking ensemble learning. Using China Securities Index 300 companies' data (2018–2022), the study constructed dynamic financial indicators and text features and adopted random forest, eXtreme Gradient Boosting, bidirectional long short-term memory with SHapley Additive exPlanations for interpretability. Experimental results showed the model's accuracy reached 95.3%, outperforming eXtreme Gradient Boosting (92.1%) and logistic regression (88.4%). It effectively detected hidden financial fraud, providing technical support for capital market supervision.