Insolvency modeling with generalized entropy cost function in neural networks

Abstract

In this study, we analyzed empirically the customer insolvency problem in the telecommunication industry. The goal was to provide objective, data-driven, and exhaustive recommendations that can support the decision-making process of the telecom operator. In this context, we comprehensively examined a q-generalized function based on the Tsallis statistics as an alternative error measure in neural networks. Many error functions have been proposed in the literature to achieve a better predictive power of neural networks. However, there is no direct implementation of the Tsallis statistics as the error function, although it was successfully applied to other fields. Our results indicate that the proposed entropy as a cost function can be applied successfully to neural networks yielding satisfactory results. The proposed neural network models, which were derived numerically, performed well and depending on the q-parameter, could detect significantly, in top deciles, a large number of insolvent customers along with the amount due. We believe that the applicability of the proposed approach can be extended to any business, where customers can purchase goods or services on credit (without paying cash) and paying for them later.