Use of Neural networks and Deep Learning for survival analysis in Credit Risk

This thesis explores the application of survival analysis and machine learning techniques to credit risk modeling, focusing specifically on predicting the probability of default (PD) over a discrete-time horizon. A model based on the DeepHit architecture is developed for PD estimation and benchmarked against three other models: CoxTime, Random Survival Forest (RSF), and the standard Cox Proportional Hazards (CPH) model. The models are trained and evaluated on a real-world dataset of credit card clients from Taiwan, using survival analysis metrics such as the C-index and Integrated Brier Score (IBS). The results show that while DeepHit is capable of producing meaningful PD outputs, the RSF and CoxTime models outperform it in both discrimination and calibration. The models' mean survival predictions are further compared against empirical Kaplan–Meier estimates to assess the alignment. A series of hypotheses are formulated and statistically tested using bootstrapped confidence intervals. The study concludes that modern machine learning methods improve upon the baseline CPH model, but their practical deployment depends on interpretability and stability across real-world settings.