Data mining for assessing the credit risk of local government units in Croatia

Abstract

Over the past few decades, data mining techniques, especially artificial neural networks, have been used for modelling many real-world problems. This paper aims to test the performance of three methods: (1) an artificial neural network (ANN), (2) a hybrid artificial neural network and genetic algorithm approach (ANN-GA), and (2) the Tobit regression approach in determining the credit risk of local government units in Croatia.

The evaluation of credit risk and prediction of debtor bankruptcy have long been regarded as an important topic in accounting and finance literature. In this research, credit risk is modelled under a regression approach unlike typical credit risk analysis, which is generally viewed as a classification problem. Namely, a standard evaluation of credit risk is not possible due to a lack of bankruptcy data. Thus, the credit risk of a local unit is approxi-mated using the ratio of outstanding liabilities maturing in a given year to total expendi-ture of the local unit in the same period. The results indicate that the ANN-GA hybrid approach performs significantly better than the Tobit model by providing a significantly smaller average mean squared error. This work is beneficial to researchers and the  govern-ment in evaluating a local government unit’s credit score.