Design and Implementation of FARPM-Net Model for Financial Risk Prediction and Automated Auditing in Enterprises

Abstract

This paper addresses the complex demands of enterprise financial risk prediction and automated auditing by proposing an innovative deep learning model—FARPM-Net—based on multimodal fusion and multi-level temporal modeling. The model integrates a multi-level Mamba module, temporal convolutional network (TCN), and an attention-based cross-modal fusion module to achieve integration of structured financial data, unstructured textual information, and external market factors, while capturing dynamics across multiple time scales. Experimental validation on two public datasets, SEC EDGAR 10-K and Yahoo Finance, demonstrates that FARPM-Net attains accuracies of 91.5% and 90.2%, respectively, representing a 4.7% improvement over mainstream models; F1 scores increase by up to 6.1%, and mean absolute error decreases by more than 16%, showcasing excellent capabilities in risk identification. Ablation studies confirm the contributions of each key module to the performance, verifying the synergistic advantages of multimodal fusion and multi-level temporal modeling. This work enhances the accuracy and stability of financial risk prediction and provides technical support for intelligent analysis of multimodal financial data. Future research will focus on model optimization, cross-modal fusion strategies, and interpretability to promote practical applications in auditing and risk management.