Leveraging Multiview Data Through Discrete and Regularized Deep Learning for Dynamic Financial Risk Prediction

The dramatic surge of demand for predictive insights into the dynamics of financial risk and the rich, yet entangled, information brought by proliferating multiview data have spawned a new research problem, multiview data-based dynamic financial risk prediction (DFRP). Although previous studies have provided diverse methods for financial risk prediction, how to effectively extract information from entangled multiview data and accurately predict the financial risk of a company over time remains challenging. To better leverage multiview data for DFRP, we take a general risk event-time modeling approach and propose a discrete and regularized deep learning (DRDL) method. We design a disentangled multiview learning module, inspired by the chunking theory, to learn a discrete and disentangled representation using a tailored discrete multiview recoder and extract complementary information using a tailored mask-based view fusion decoder. We also design a monotonicity-aware multiperiod risk prediction module to guarantee time-wise monotonicity using a tailored risk accumulation function, accommodate instance-wise monotonicity using a tailored focal survival loss, and avoid conflicts between these two objectives using a tailored adaptive gradient balancing method. We have evaluated DRDL on two types of companies across three markets in China and the United States, using yearly, quarterly, and monthly prediction windows, with three specific types of financial risk (i.e., fundamental distress, regulatory listing risk, and market tail risk). Evaluation at the model level, in terms of time-to-risk prediction performance and out-of-time prediction performance, and impact analysis at the application level, in terms of identification performance and profitability performance, demonstrate advantages of DRDL over benchmarked classic and state-of-the-art methods on all fronts. Mechanism-level analyses further reveal the core drivers underlying the utility of DRDL.

History: Jeffrey Parsons, Senior editor; Gene Moo Lee, Associate Editor.

Funding: Financial support from the National Natural Science Foundation of China [Grants 72471076 and 72101073], the Natural Science Foundation of Anhui Province [Grant 2508085J045], the University Synergy Innovation Program of Anhui Province [Grant GXXT-2023-063], and the Fundamental Research Funds for the Central Universities [Grant JZ2025HGPA0246] is gratefully acknowledged.

Supplemental Material: The online appendix is available at https://doi.org/10.1287/isre.2024.1417.