Predictive Models of Non-Performing Loans: The Case of Greece

This paper investigates the effectiveness of various predictive models in identifying non-performing loan (NPL) ratios within the Greek banking system. The study utilizes quarterly aggregate data encompassing performing and non-performing loans across four loan categories: mortgages, consumer loans, corporate loans, and a total loan portfolio view. The timeframe analyzed spans from 2002 to 2022. Eight distinct predictive models are constructed to assess the NPL ratio for each loan category and the total portfolio. The models leverage two primary modeling techniques: (a) ordinary least squares (OLS) regression and (b) neural network models. To enhance predictive power, a meticulous selection process identifies a limited set of explanatory variables with maximum forecasting potential. These variables include macroeconomic indicators such as unemployment rate, GDP growth, house price index growth, and consumer price index. Additionally, lagged versions of the NPL index itself are incorporated into the models (autoregression). The models undergo rigorous validation procedures to assess their effectiveness within the sample period and out of sample for forecasting purposes. The paper presents all eight models in detail, along with comprehensive analyses of their validity and explanatory power. Accompanying tables and graphs visually demonstrate the model construction, validation process, and key findings. The paper concludes by summarizing the effectiveness of each model and highlighting their potential applications within the Greek banking sector for NPL risk management.