【报告主题】Predicting Bank Credit Losses Using Machine (Deep) Learning
【主讲嘉宾】苗斌，香港中文大学(深圳)经管学院会计学副教授、会计学博士项目主任
【报告时间】2025年3月18日 10:00
【报告地点】会计学院108室
【内容提要】
The transition in banks’ loan loss accounting from the incurred loss model to the expected loss model (CECL) requires a forward-looking approach to estimating credit losses over the lifetime of loans. Developing reliable long-term credit loss predictions presents a significant challenge due to the inherent uncertainty in future economic conditions. This study applies deep learning methods to long-term bank credit loss prediction and evaluates their performance against existing linear models. Our findings show that Long Short-Term Memory (LSTM) networks and Feature Tokenizer Transformer (FT-Transformer) significantly improve predictive accuracy, particularly in capturing long-term dependencies. In contrast, decision-tree-based models offer no advantage over linear methods. The predictive gains of deep learning models are strongest in longer-term forecasts, financial crises, and banks with heterogeneous loan portfolios. Using explainable AI techniques, we find that the relative importance of real GDP increases with the forecasting horizon, while the significance of current net charge-offs declines. Finally, our results suggest that CECL provisions may help reduce the procyclicality of bank lending, offering valuable insights for financial institutions, regulators, and investors.