Model Selection and Feature Ranking for Financial Distress Classification

In this paper we apply several learning machine techniques to the problem of financial distress classification of medium-sized private companies. Financial data was obtained from Diana, a large database containing financial statements of French companies. Classification accuracy is evaluated with Artificial Neural Networks, Classification and Regression Tress (CART), TreeNet, Random Forests and Liner Genetic Programs (LGPs). We analyze both type I (bankrupted companies misclassified as healthy) and type II (healthy companies misclassified as bankrupted) errors on two datasets containing balanced and unbalanced class distribution. LGPs have the best performance accuracy in both balanced data and unbalanced dataset. Our results demonstrate the potential of using learning machines, with respect to discriminant analysis, in solving important economics problems such as bankruptcy detection. We also address the related issue of ranking the importance of input features, which is itself a problem of great interest. Elimination of the insignificant inputs leads to a simplified problem and possibly faster and more accurate classification of financial distress. Experiments on Diana dataset have been carried out to assess the effectiveness of this criterion. Results show that using significant features gives the most remarkable performance and performs consistently well over financial datasets we used.