Consumers’ Financial Distress: Prediction and Prescription Using Interpretable Machine Learning

Abstract

This paper shows how transactional bank account data can be used to predict and to prevent financial distress in consumers. Machine learning methods were used to identify the most significant transactional behaviours that cause financial distress. We show that Random Forest outperforms the other machine learning models when predicting the financial distress of a consumer. We obtain that Fees and Interest paid stand out as primary contributors of financial distress, emphasizing the significance of financial charges and interest payments in gauging individuals’ financial vulnerability. Using Local Interpretable Model-agnostic Explanations, we study the marginal effect of transactional behaviours on the probability of being in financial distress and assess how different variables selected across all the data point selection sets influence each case. We also propose prescriptions that can be communicated to the client to help the individual improve their financial wellbeing. This research used data from a major South African bank.