Firm failure prediction using genetic programming generated features

Many studies on predicting firm failure have focused on finding new features that improve the accuracy of the models. In this paper, genetic programming (GP) is used for this purpose. The main problem in GP is to specify a function that evaluates the fitness of the feature. Direct optimization of a machine learning (ML) model that uses a generated feature in most cases leads to high computational costs since evolving a population of N programs over G generations while evaluating each model using K-fold cross-validation requires N*G*K model learning cycles. Thus, many researchers use scores that measure the relationship of the generated features to the class label. However, our empirical analysis shows that most such scores correlate poorly with ML model performance. The novelty of our work is that we introduce several ways of combining different scores into a single measure of expected model performance. Experimental results on data from Hungarian firms (7167 observations, class imbalance 9.37) using five ML models (Logistic Regression, Random Forest, Gradient Boosting, Histogram Boosting, and AdaBoost) prove that the proposed way of setting the fitness function increases the ROC AUC of the listed models by 6.6%, 5.2%, 6.8%, 5.5% and 5.2% respectively. Moreover, by applying the found formula to the data from Czech firms (3872 observations, class imbalance of 74.92), which were not used for the feature search, we obtained increases in ROC AUC by 13.1%, 11.8%, 14.9%, 9.9%, and 8.2%, respectively. This indicates that the proposed method allows to find universal features, which opens the way to build effective models in case of insufficient data (small number of observations, extreme imbalance, etc.).