Probability of default model using transaction data of Russian companies

Shevelev A., Buzanov G.

The purpose of this study is to test the usefulness of transaction data from the Bank of Russia Payment System (BRPS), to predict the default probabilities of Russian companies. To fulfil this purpose, we build probability of default models for each industry group using machine learning methods based on annual accounting data. Thereafter, we add features generated from transaction data to the models and improve their forecast quality according to the ROC AUC metric.

Additionally, we train our probability of default models for each industrial group using a Random Forest based only on BRPS data. The forecast quality of this is a little worse on average according to the ROC AUC metric, but these estimates can be obtained at least three months earlier than estimates using annual accounting statements.

Our results confirm that BRPS transaction data are useful for improving the quality of forecasting the default probabilities of Russian companies. In addition, the Random Forest feature importance shows that the main sources of this additional information are payroll taxes and social payments.

Probability of default model using transaction data of Russian companies