Predicting loss given default

The topic of credit risk modeling has arguably become more important than ever before given the recent financial turmoil. Conform the international Basel accords on banking supervision, financial institutions need to prove that they hold sufficient capital to protect themselves and the financial system against unforeseen losses caused by defaulters. In order to determine the required minimal capital, empirical models can be used to predict the loss given default (LGD). The main objectives of this doctoral thesis are to obtain new insights in how to develop and validate predictive LGD models through regression techniques. The first part reveals how good real-life LGD can be predicted and which techniques are best. Its value is in particular in the use of default data from six major international financial institutions and the evaluation of twenty-four different regression techniques, making this the largest LGD benchmarking study so far. Nonetheless, it is found that the resulting models have limited predictive performance no matter what technique is employed, although non-linear techniques yield higher performances than traditional linear techniques. The results of this study strongly advocate the need for financial institutions to invest in the collection of more relevant data. The second part introduces a novel validation framework to backtest the predictive performance of LGD models. The proposed key idea is to assess the test performance relative to the performance during model development with statistical hypothesis tests based on commonly used LGD predictive performance metrics. The value of this framework comprises a solution to the lack of reference values to determine acceptable performance and to possible performance bias caused by too little data. This study offers financial institutions a practical tool to prove the validity of their LGD models and corresponding predictions as required by national regulators. The third part uncovers whether the optimal regression technique can be selected based on typical characteristics of the data. Its value is especially in the use of the recently introduced concept of datasetoids which allows the generation of thousands of datasets representing real-life relations, thereby circumventing the scarcity problem of publicly available real-life datasets, making this the largest meta learning regression study so far. It is found that typical data based characteristics do not play any role in the performance of a technique. Nonetheless, it is proven that algorithm based characteristics are good drivers to select the optimal technique. This thesis may be valuable for any financial institution implementing credit risk models to determine their minimal capital requirements compliant with the Basel accords. The new insights provided in this thesis may support financial institutions to develop and validate their own LGD models. The results of the benchmarking and meta learning study can help financial institutions to select the appropriate regression technique to model their LGD portfolio's. In addition, the proposed backtesting framework, together with the benchmarking results can be employed to support the validation of the internally developed LGD models