Not logged in.
Quick Search - Contribution
Contribution Details
| Type | Master's Thesis |
| Scope | Discipline-based scholarship |
| Title | Machine Learning Applications for Reverse Stress Testing |
| Organization Unit | |
| Authors |
|
| Supervisors |
|
| Language |
|
| Institution | University of Zurich |
| Faculty | Faculty of Business, Economics and Informatics |
| Number of Pages | 61 |
| Date | 2020 |
| Abstract Text | Reverse stress testing is a novel idea that intends to identify scenarios that can lead a financial institution to unviability without the cognitive biases that traditional stress testing imposes. In this thesis, we propose a reverse stress testing framework that is based on historical bootstrapping simulation and machine learning techniques. To achieve this, we create a mini-bank balance sheet and map its components to risk factors that are generated by a historical bootstrapping methodology. The mini-bank balance sheet enables us to create a framework that is applicable to real-world problems. The choice of modeling risk factors with a non-parametric bootstrapping method provides us with an additional degree of realism since we capture many stylized facts observed in financial time series. Furthermore, we are able to generate a vast amount of unexplored scenarios to train and test our machine learning methods. To verify the validity of our approach and to present the structure of our problem we start with a simplified version of our problem. To demonstrate the capability of our framework, we introduce non-linear risk factors by including CoCo bonds and a more complex credit risk modelling approach that includes jumps. For the complex problem, we use Support Vector Classifiers, Tree-based models, Ensemble-based models, Linear models, and Neural Networks (Multilayer Perceptron and Convolutional Neural Networks). The best performing model was the Multilayer Perceptron. The Support Vector Classifier, Random Forest, and Decision Tree were the next best performing models with small differences compared to the champion model. After testing our models under different conditions, we conclude that machine learning techniques can be used for reverse stress testing purposes. As a final task, we investigate the performance of the best performing models in a data set that was generated by a balance sheet for which its allocations are changed up to 15%. We find, that even though the model was trained and tested in datasets that were generated by different balance sheets, the performance remained surprisingly high. This indicated two things. The first is that some scenarios are so severe that even with a 15% less participation they can still cause unviability. The second is that our model is robust since it can make good generalizations. |
| Export |
BibTeX
|