The Evaluation of Dutch Non-Life Insurance Companies A Comparison of an Ordered Logit and a Neural Network Model

This paper describes the development of two models which determine the financial solidity of Dutch non-life insurers: an ordered logit model and a neural network model. Since bankruptcies of Dutch insurance companies are very rare, the annual assessment texts made by the supervisor are used to classify the companies into one of the three possible groups: strong, moderate, or weak. Both models use the same six variables, which were selected by means of a stepwise logistic selection procedure. These variables cover three aspects: solvency, profitability, and investments. Both models are estimated on a 1992 data set which contains 195 companies, and they are tested on a 1993 data set containing 193 companies. The ordered logit model correctly classified 85% of the test set, 95% of the strong companies and 85% of the weak companies are classified correctly. The neural network model correctly classified 86% of the test set, 96% of both the strong and the weak companies are classified correctly. A combination of the outcomes of both models leads to an overall score of 87%, with 97% of the strong and 96% of the weak companies classified correctly. Neither the ordered logit model nor the neural network model are able to adequately recognize moderate companies. Some possible reasons for the problems with moderate companies are given.