Scenario Reduction in Debt Simulations Using Recurrent Autoencoders : Finding Meaningful Patterns in Stochastic Processes

Abstract: This thesis studies how improvements can be made to a simulation model used to analyse debt portfolios. Today, the simulation model needs to evaluate a portfolio over a large sample of scenarios to get accurate results. This can be time-consuming if the portfolios consist of many different and complex securities. The proposed improvement method uses recurrent neural network autoencoders together with clustering for scenario reduction. Different architectures of the networks and a different number of hidden features are tested, to study what impact this has on the performance of the method. The autoencoders are also trained with auxiliary tasks in the form of clustering in the latent space and classifying shuffled samples. The results show that the scenario reduction method outperforms random sampling on all experiments conducted, in some cases more than halving the number of scenarios needed. Furthermore, the scenario reduction method performed well for a majority of the architectures tested. Indicating that the method is robust and requires little work in terms of hyperparameter optimisation to improve the simulation model. The views and opinions expressed in this thesis are those of the author and do not necessarily reflect the official policy or position of the Swedish National Debt Office.