Predicting Firm-Level Bankruptcy in the Spanish Economy Using Extreme Gradient Boosting

Abstract

We apply a machine learning (ML) algorithm in order to predict bankruptcy rates among companies within the Spanish economy from 1992 to 2016. The model identifies some relevant variables when predicting bankruptcy: such as the ratio total liabilities to total assets or current liability to financial expenses along with size factors such as the log of sales. Additionally, the model allows us to analyse firms individually: the marginal contribution of a given variable to the firm’s prediction depends on all its other observed characteristics. This can be particularly useful in analysing case by case lending decisions within financial institutions. An exercise on the cost of extending the forecasting horizon up to 4 years ahead is also provided, as financial institutions are naturally interested in the early detection of bankruptcy. We also compare XGBoost to a number of ML models, such as a Logistic Model, Support Vector Machine, Neural Network, Random Forest and LightGBM.