An Explainable Hybrid Model for Bankruptcy Prediction Based on the Decision Tree and Deep Neural Network

Abstract

For investors seeking solutions to optimize their portfolio of assets to generate profits and minimize losses, choosing a sophisticated model to evaluate the risk of corporate financial distress will be crucial to support their asset management and investment decisions. Both the machine learning and the newly developed deep learning techniques have been employed to construct bankruptcy prediction models for decades. However, applying the deep learning models might increase the predictive accuracy in exchange for losing model interpretability, because the structure and parameters of the model are not easy to provide accountability for investors. In this study, a hybrid approach integrating the decision tree with the deep neural network was proposed to provide a compromise solution for investors. The decision tree was adopted as the primary model to provide explainable ability, while the deep neural network was chosen to improve predictive accuracy. The decision fusion of two models was designed with the compensatory and non-compensatory approaches. The hybrid model was implemented by concatenating the deep neural network to the selected branches of decision tree that perform poor predictive accuracy during model training. The empirical results showed that the predictive accuracy of the deep neural network and the decision tree were 80% and 87% respectively, and the overall accuracy was improved to 91% by the hybrid model.