Impact of Low Events and Class Imbalances in PD Models

This research article investigates the challenges faced by financial institutions in estimating risk parameters for low-default portfolios (LDPs) characterized by a low or zero-default nature. It addresses the common issue of class imbalance in data, where default events are minimal and underrepresented. The study employs a simulation design that reflects real-world LDP scenarios to assess the effects of limited event occurrences and class imbalance on model performance metrics such as prediction accuracy, discrimination ability, and classification cut-offs. Key findings indicate that the degree of association between response variables and predictors significantly influences the F1/P4 score and Gini coefficient. The research concludes with practical guidelines for achieving optimal model performance metrics based on the characteristics of the development data, emphasizing the importance of understanding the interplay between event rates, sample sizes, and classification thresholds.