Forecasting Model for Non-Performing Loans of Banks Using Random Forest Algorithm

Liquidity management is one of the biggest challenges of the banking system. One of the problematic factors in liquidity management is that some loans, which are defined as non-performing loans, do not return to the liquidity cycle. The prediction of non-performing loans helps in the optimal management of banks' cash. Forecasting is possible by extracting the determining factors and using a suitable forecasting model. Therefore, this study was conducted with the aim of predicting the management of cash flow in non-performing loans of banks in three stages. In the first stage, all related factors were extracted and categorized using the meta-synthesis method and extensive background review (review of 194 articles). The categories included 10 classes of determining factors (95 factors), 11 mitigation strategies and 12 consequences. Due to the multiplicity of factors, it was not possible to include all of them in the prediction model. Therefore, in the second stage, the most important determining factors were selected by conducting three Delphi rounds and using experts' opinions. The output of this stage was the selection of 4 macroeconomic variables and 5 bank-specific variables. In the third step, the data of model variables for the years 1389 to 1400 were extracted and three machine learning algorithms, random forest, K nearest neighbor and Tobit logistic regression were used for prediction. The results showed that the prediction accuracy of random forest algorithm is higher than other used models.