A new neuro-fuzzy inference system for insurance forecasting

Abstract

Insurance forecasting is a matter of vital importance to insurance companies for analyzing of annual income, premium and loss reserving, loss payment, etc. Recent years have also seen increasing discussion within the actuarial community of the need for insurance forecasting techniques that are more solidly grounded in rigorous machine learning methodologies. Taking advantages of knowledge-reuse and learning capability for dealing with uncertainties, hybridization of neural networks and fuzzy logic could enhance the accuracy of forecasting for insurance applications. In this paper, we propose a novel neuro-fuzzy inference system for insurance forecasting. It uses multiple parameter sets where each set is responsible for a small subset of records. The aim of each parameter set is to minimize Mean Square Error within records of the subset. The learning strategy and a rule reduction method are also proposed. Empirically validation on the benchmark and real insurance datasets show the advantages of the new system.