Abstract
Survival models have received great attention in credit risk with their ad vantages of predicting not just "if" but also "when" a borrower will default. This thesis focuses on integrating discrete-time survival models (DTSM) to machine learning and deep learning frameworks, exploring the explainabil ity for the black-box of the neural network, and improving the predictive capability compared with the state-of-the-art models. To our knowledge, this is one of the first novel works in this area. Through extensive investiga tion, the following three main important research gaps have been identified and resolved in this thesis:1. How to combine DTSM with machine learning models to enhance pre dictive performance while ensuring that complex black-box neural networks remain interpretable and analyzable. We introduce the neural network with Age-Period-Cohort(APC) model to address the challenges. We developed a series of embedded discrete sub-networks with a DTSM to predict the temporal evolution of default events. Our proposed model uses the neu ral networks to capture nonlinear feature to enhance the model predictive performance and incorporate the APC model to decompose the predicted default risk and provide the black box network with a transparency and explainability.
2. Given that DTSM is inherently a time-series model, we pair it with the long short-term memory (LSTM) recurrent network to better capture the non-linear temporal patterns between consecutive time points. How ever, after the investigation, we found that achieving a proper integration of DTSM and LSTM to better capture temporal patterns from the time series loan transaction while comprehensively addressing the challenges of instable initialization problem and long-term dependency issue introduced by LSTM remains an unresolved but unavoidable focal point. To address this gap, we introduced an adapted LSTM-based attention to replace the linear matrix multiplication in tradional attention mechanism with LSTM based transformation to help the model focus on important transactional time point without disrupting the nonlinear temporal interactions of the data and use a washout phase to resolve the initial noise which further enhances the performance of the attention mechanism.
3. Since survival analysis was commonly used as a powerful model in for eign countries, it will be practical and interesting to apply the DTSM to analyze the local credit risk in Chinese datasets. In this thesis, we first cooperated with Ningbo Big Data Bureau and apply DTSM to water con tract dataset which is the preliminary works of the research study in the thesis. We utilized our proposed survival analysis to explore the nonlinear hidden features between the variables and predict the default risk behaviors for the companies, which effectively reflect the local special phenomenon in Ningbo.
In conclusion, beyond merely improving model performance, we address new issues and challenges posed by the incorporation of machine learning and deep learning methods, such as black-box model interpretability, long term dependency in recurrent neural networks, and instability in model initialization. These findings and solutions provide scholars, practitioners, and financial regulators with more effective practical tools and a broader perspective, advancing credit risk modeling towards a clearer and more efficient future.
| Date of Award | 15 Mar 2026 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Anthony Graham Bellotti (Supervisor) & Ruibin Bai (Supervisor) |