Research Seminar by Tsai Ming-Feng | Multistage Corporate Default Prediction – A Domain Knowledge-tailored Neural Network

About The Talk

In this work, we propose a domain-knowledge-tailored neural network for multi-period default prediction. By incorporating economic domain knowledge, we customize conventional neural networks to enhance the model's performance and mitigate overfitting. We validate the effectiveness of our approach through experiments performed on a substantial US corporate default dataset spanning from 1994 to 2021. The results show that our proposed model surpasses the state-of-the-art econometric model and proves more robust than traditional neural networks. Moreover, the model exhibits consistent performance across various dataset subgroups. Although its predictive power declines in high credit risk years for long-term horizons, the outcomes remain justifiable. Our method is applicable to most neural networks and offers valuable insights for ongoing machine learning research in financial applications.

About the Speaker

Ming-Feng Tsai is currently an Associate Professor in the Department of Computer Science at National Chengchi University. In 2006, he was at Microsoft Research Asia as an intern with the Web Search & Mining Group, and was awarded by the research institution the Best Intern of the Year, and invited to visit the headquarters of Microsoft and Bill Gates’ house in Redmond. He received his Ph.D. degree from National Taiwan University in 2009. After receiving his Ph.D. degree, he worked at National University of Singapore as a Research Fellow, participating in a research project related to machine translation. In 2010, sponsored by Taiwan National Science Council, he joined University of Illinois at Urbana-Champaign as a visiting scientist, working on the projects associated with advanced Web search and mining. His research interests span the areas of information retrieval, recommender systems, natural language processing, machine learning, artificial intelligence.