Bankruptcy forecasting: A hybrid approach using Fuzzy c-means clustering and Multivariate Adaptive Regression Splines (MARS)

During the last years, hybrid models have proven to be a promising approach for the design of classification systems for the forecasting of bankruptcy. In the present research we propose a hybrid system which combines fuzzy clustering and MARS. Both models are especially suitable for the bankruptcy prediction problem, due to their theoretical advantages when the information used for the forecasting is drawn from company financial statements. We test the accuracy of our approach in a real setting consisting of a database made up of 59,336 non-bankrupt Spanish companies and 138 distressed firms which went bankrupt during 2007. As benchmarking techniques we used discriminant analysis, MARS and a feed-forward neural network. Our results show that the hybrid model outperforms the other systems, both in terms of the percentage of correct classifications and in terms of the profit generated by the lending decisions.

Research highlights
► A major fallacy of some of the prior research on bankruptcy prediction is the manner in which the sample is drawn and the model accuracy is defined. In these studies, each one of the bankrupt companies is matched with a known non-bankrupt company from the same time period. Then, a model that predicts better than 50% (the assumed rate of chance) is thought to outperform random guessing.
► We considered a real setting. That is, we used a database made up of the annual accounts of 59,474 Spanish firms, 138 of them bankrupt. Therefore only a 0.232% of the companies went bankrupt.
► The present research presents a hybrid approach using fuzzy c-means clustering and Multivariate Adaptive Regression Splines (MARS). In a first stage, clusters are created using fuzzy c-means. The clusters are classified into two groups: those that contain bankrupted companies and those that not. Then, a MARS model is created using such clusters as a part of the input information.
► The performance of the proposed model is better than those obtained with the following benchmark techniques: MARS, discriminant analysis and neural networks.