Fast approximation methods for credit portfolio risk calculations

Abstract

Credit risk is one of the main risks financial institutions are exposed to. Within the last two decades, simulation-based credit portfolio models became extremely popular and replaced closed-form analytical ones as computers became more powerful. However, especially for non-homogenous and non-granular portfolios, a full simulation of a credit portfolio model is still time consuming, which can be disadvantageous within some use cases like credit pricing or within stress testing situations where results must be available very quickly. For this purpose, we investigate if methods based on artificial intelligence (AI) can be helpful to approximate a credit portfolio model. We compare the performance of AI-based methods within three different use cases with suitable non AI-based regression methods. As a result, we see that AI-based methods can generally capture portfolio characteristics and speed-up calculations but - depending on the specific use case and the availability of training data - they are not necessarily always the best choice. Particularly, considering the time and costs for collecting data and training of the complex algorithms, non-AI-based methods can be as good as or even better than AI-based ones, while requiring less computational effort.