Default Prediction for Russian Food Service Firms: Contribution of Non-Financial Factors and Machine Learning

Abstract

The food service industry's instability due to COVID-19 and sanctions has heightened the need for developing an efficient tool to assess default risks in this industry. Default prediction modelling relies heavily on how well a model fits the specific environment. Due to that, some adjustments have to take place in order to adapt the classical default prediction models to the Russian food service industry. We build hypotheses that adding non-financial factors and employing modern prediction methods can increase the accuracy of the models significantly. The aim of this study is to determine the effect of non-financial factors' inclusion and modern modelling methods on the accuracy of default prediction for the food service industry in Russia. Tests for a sample of 1241 firms for the period from 2017 to 2021 have shown that creating a prediction model with modern methods, such as Random Forest and XGBoost increases the accuracy of the prediction from 70% to about 80%, compared to standard Logit model. The addition of non-financial factors to the models also increases the accuracy slightly, which however, does not provide a significant effect. The most important metrics in predicting default turned out to be Current Liquidity Ratio and the ratio of Working Capital to Total Assets. The most important non-financial factors are Total Assets and Age. Our results correspond with existing research in this field and form a new knowledge layer due to being applied to a specific industry. The results can be used by banks or other counterparties that interact with food service industry firms in order to assess their credit risk.