Abstract

Credit risk assessment is central to modern lending because it supports default prevention and improves financial decision-making. Traditional credit scorecards often rely on binary logistic regression, which is simple and interpretable but limited in that it does not model the timing of default. This paper develops a monthly behavioral scorecard framework that combines discrete-time logistic regression with survival analysis. Using a large longitudinal dataset of 30-year fixed-rate mortgages from Freddie Mac covering 2018-2024, the study constructs an exploded panel representation to capture time-varying risk profiles. To address the computational burden of very large panels, a backward progressive weighting scheme reduces the effective sample size while preserving statistical power. The model integrates static borrower attributes, dynamic loan lifecycle variables transformed with multivariate adaptive regression splines (MARS) to capture nonlinear effects, and time-varying macroeconomic indicators to reflect systemic volatility. A calibration method is then presented to convert discrete monthly hazard rates into cumulative default probabilities, followed by an offset adjustment to produce a points-to-double-the-odds (PDO) behavioral scorecard. Backtesting across in-sample, holdout, and out-of-time samples demonstrates strong discrimination, with in-sample AUC of 0.82 and out-of-time AUC of 0.70. The resulting framework supports practical, lifecycle-based credit interventions and regulatory capital applications.

This work is licensed under a Creative Commons Attribution 4.0 License.