Identifying Important Variables in Bankruptcy Prediction Models with Interpretable Machine-Learning – The Difference between Crisis and Non-Crisis Periods

The share of companies that file for bankruptcy is a countercyclical variable, as it increases during a recession. Although some bankruptcy prediction models also include macroeconomic factors as input variables, the literature about the differences between the companies that file during a crisis period and those which file during an economic boom is still limited. We analyze which factors determine if companies will fail during an economic crisis, and if they are different than those during the non-crisis period. Therefore, we divide the data, which is obtained from Compustat North America and the UCLA LoPucki Bankruptcy Research Database, into two datasets. One set contains the observations from the crisis periods, and the other contains observations from the non-crisis periods. We apply different machine learning models for bankruptcy prediction on both datasets separately and analyze the variable importance of the individual models. Our results show that the leverage ratio, as well as growth and profitability ratios are more important during and non-crisis period, while the liabilities to assets ratio and the inventory intensity ratios are more important during an economic crisis