An Integrated Risk Capital Aggregation Framework Using Nonlinear Classification by Random Forests and Its Proximity Measures

Abstract

This paper proposes an integrated risk capital aggregation methodology based on the non-linear classification algorithm Random Forests. Random Forests offers the advantages of handling a large amount of data, allows for over-specification and requires neither assumptions about the distributions nor knowledge of the models behind the included risk indicators and capital figures of different risk types. All interactions in the aggregation model are estimated by an adaption of the Random Forests proximities conditional on the risk indicators used to build the forest. The proposed proximity based aggregation model is shown to be accurate yet requires less risk capital in comparison to established aggregation methodologies and reservation models. Additional results are its strong stability over time and that the adapted proximities represent an alternative to the usage of correlation.

Keywords: Random Forests, risk aggregation, nonlinear modeling, benchmark, integrated risk analysis, diversification, big data

JEL Classification: C14, C38, G20