The effect of feature selection on financial distress prediction

• Filter and wrapper based feature selection are employed for financial distress prediction.
• In addition, two bankruptcy prediction datasets and two credit scoring datasets are used.
• On average, GA and LR perform better over the credit and bankruptcy datasets respectively.
• However, there is no exact winner over the four datasets.
• In addition, performing feature selection does not always improve the models’ performance.

Financial distress prediction is always important for financial institutions in order for them to assess the financial health of enterprises and individuals. Bankruptcy prediction and credit scoring are two important issues in financial distress prediction where various statistical and machine learning techniques have been employed to develop financial prediction models. Since there are no generally agreed upon financial ratios as input features for model development, many studies consider feature selection as a pre-processing step in data mining before constructing the models. However, most works only focused on applying specific feature selection methods over either bankruptcy prediction or credit scoring problem domains. In this work, a comprehensive study is conducted to examine the effect of performing filter and wrapper based feature selection methods on financial distress prediction. In addition, the effect of feature selection on the prediction models obtained using various classification techniques is also investigated. In the experiments, two bankruptcy and two credit datasets are used. In addition, three filter and two wrapper based feature selection methods combined with six different prediction models are studied. Our experimental results show that there is no the best combination of the feature selection method and the classification technique over the four datasets. Moreover, depending on the chosen techniques, performing feature selection does not always improve the prediction performance. However, on average performing the genetic algorithm and logistic regression for feature selection can provide prediction improvements over the credit and bankruptcy datasets respectively.