The impact of class imbalance in logistic regression models for low-default portfolios in credit risk

Battle Count: 74

The paper is primarily focused on credit risk modeling and loan default prediction, which is a different domain from quantitative trading. However, the findings about class imbalance effects on logistic regression, the distinction between classification accuracy and ranking (Gini) performance, and the practical guidelines for sample sizes could be tangentially relevant to trading signal classification problems with rare events (e.g., detecting rare market regimes or credit events in trading portfolios). The methodology is not directly applicable to trading strategy development.

The core methodology (logistic regression with WoE) is straightforward and well-established in credit risk. The simulation framework is clearly described with 7 steps. However, implementing the full study requires: (1) generating class-conditional distributions for multiple configurations, (2) fitting logistic regression with WoE, (3) grid search for optimal cut-offs, (4) computing multiple metrics (F1, P4, Gini), and (5) running 500 MC iterations per configuration. The 16 additional configurations add complexity. No code is provided, so implementation from scratch would require moderate effort.

The simulation procedure is well-described with clear steps (7-step Monte Carlo iteration), specific sample sizes, event rates, and configurations with exact class-conditional distributions provided in tables. However, no code or software implementation is provided. The use of standard logistic regression with WoE (SAS default theta=0.5) aids reproducibility. The 16 additional configurations are described but their exact class-conditional distributions are not fully tabulated.