The Influence of Sampling Method and Variable Selection Approach on the Prediction Ability of Financial Distress Prediction Models

Recent dynamic development of the economic environment causes steadily growing number of enterprises, which declare bankruptcy, in like manner enterprises in the Slovak Republic. Therefore the examination and early detection of financial difficulties, as well as corporate failure prediction take an important place in the economic research area. Several. issues determining the quality of final prediction models arise during its creation. The goal of the paper is to examine the influence of some different factors determining the overall accuracy of prediction models. The creation of prediction models will be realized using neural networks CART and logistic regression prediction methods for the selected data set of enterprises operating in the Slovak dynamic economic environment. Our experimental results shows that there is no the best generally accepted combination of the training/testing proportion, feature selection method and the classification technique guaranteed the highest classification ability over the 36 data sets. Therefore, the relevance of performing the feature selection depends on the chosen classifier. However the highest financial distress prediction ability in one year horizon was achieved on datasets combining 75:25 training to testing sample proportion with CART feature selection method.