The purpose of this paper is to do a performance comparison of three different data mining techniques.
Logit model, decision tree and random forest are applied in this study on British, French, German, Italian, Portuguese and Spanish balance sheet data from 2006 to 2012, which covers 446,464 firms. Because of the strong imbalance with regard to the solvency status, classification trees and random forests are modified to adapt to this imbalance. All three model specifications are optimized extensively using resampling techniques, relying on the training sample only. Model performance is assessed, strictly, based on out-of-sample predictions.
Random forest is found to strongly outperform the classification tree and the logit model in almost all considered years and countries, according to the quality measure in this study.
Obtaining reliable estimates of default propensity scores is of immense importance for potential credit grantors, portfolio managers and regulatory authorities. As the overwhelming majority of firms are not listed on stock exchanges, annual balance sheets still provide the most important source of information. The obtained ranking of the three models according to their predictive performance is relatively robust, due to the consideration of several countries and a relatively long time period.
CitationDownload as .RIS
Emerald Publishing Limited
Copyright © 2017, Emerald Publishing Limited