Predicting financial distress using current reports: A novel deep learning method based on user-response-guided attention

Abstract

Effective financial distress prediction (FDP) can discover a company's potential financial risks and support relevant decisions in a timely manner. Previous studies on FDP have mostly focused on using financial indicators and periodic reports. Compared with periodic reports, current reports disclose major events in a timelier manner. But leveraging the information in current reports involves the critical challenges of capturing the complex semantics and measuring the importance of heterogeneous events. To this end, we propose a novel deep learning method, a user-response-guided deep attention network (URGDAN), to predict financial distress using current reports. In the proposed method, we construct a deep learning architecture to integrate financial indicators, current report texts, and user responses. URGDAN leverages the user responses to current reports to guide the semantic feature representation of the reports, it also identifies event information that has a significant correlation with company financial distress. Empirical evaluation shows that URGDAN significantly improves predictive performance and can accurately determine the importance of different current reports. Our work provides practical implications for creditors and investors.