Non-financial information, especially information carried in disclosure reports, plays an important role in conveying financial distress signals. Considering the rise of generative AI (GenAI) and its potential in capturing both surface and latent meanings of disclosure reports, we initiate a new research avenue, GenAI-enhanced financial distress prediction. We position GenAI as an information intermediary and propose a functional analogy framework to conceptualize the process of leveraging disclosure reports with four functions: perception, extraction, reasoning, and evaluation. We then provide a guideline with three GenAI use strategies (i.e., prompt engineering, knowledge injection, and fine-tuning) and design a deep learning method featuring a function-based bidirectional representation module, which explicitly and separately extracts representations for the emphasis information produced by the extraction function and insight information produced by the reasoning function, guided by tailored convergent and divergent mutual information criteria, respectively. Empirical evaluation at the model level and impact analysis at the application level demonstrate advantages of the proposed method over benchmarked state-of-the-art methods on all fronts. Mechanism-level analyses further reveal the core drivers underlying the utility of the proposed method.

With the rapid development of fintech, the need for dynamic credit risk evaluation is becoming increasingly important. While previous studies on credit scoring have mostly focused on single-period loan default prediction, we call for a new avenue—multiperiod default prediction (MPDP)—to depict risk profiles over time. To address the challenges raised by MPDP, such as monotonic default probability prediction and complex relationship accommodation, we propose a novel approach, hybrid and collective scoring (HACS). We design a hybrid modeling strategy to predict whether and when a borrower will default separately through a default discrimination model and a default time estimation model, respectively, and synthesize them through a probabilistic framework. To accommodate various possible patterns of default time and measure the distribution of default probability over successive time intervals, we propose a joint default modeling method to train the default time estimation model. Empirical evaluations at the model (time-to-default prediction performance and discrimination performance) and mechanism (identifiability and discriminability) levels, as well as impact analyses at the application (granting performance and profitability performance) level, show that HACS outperforms the benchmarked survival analysis and multilabel learning methods on all fronts. It can more accurately predict time-to-default and provide financial institutions and investors better decision-support in granting loans and selecting loan portfolios.