Diagnosis with incomplete multi-view data: A variational deep financial distress prediction method

The proliferating of big data from both financial and non-financial aspects, has been flourishing multi-view-data-based financial distress prediction. However, when various data views, e.g., report texts, forum posts, and legal judgments, are jointly utilized, modeling challenges, such as heterogeneities in distribution and completeness among data views, may be inevitably raised. To this end, we propose a variational deep financial distress prediction method (VDFDP). The proposed method consists of three modules: a view-specific encoder module to learn a latent representation for each view, a view fusion module to learn a joint representation by transferring knowledge from all views considering different degrees of completeness, and a financial distress decoder module to map joint representation to financial distress status. Empirical evaluation using Chinese listed company data shows that VDFDP significantly outperformed all benchmarked financial distress prediction methods. It can more effectively leverage incomplete multi-view data and more accurately predict financial distress. Our study also provides valuable insights and practical implications for stakeholders, such as investors and companies themselves, to effectively identify risk signals and make risk management decisions.