What is the end-to-end scorecard development process for credit risk?

Question

AcadiFi · Accepted Answer

Building a scorecard follows a disciplined eight-step process. First, sample design: select a development window (typically 18 to 24 months of originations), a performance window (12 to 24 months after booking), and define good vs bad (e.g., 90+ days past due within performance window). Reject inference is applied to declined applications to avoid survivorship bias.

Second, data preparation: pull application data, bureau data, and performance outcomes into an analytical base table, handle missings, and winsorize extreme values. Third, variable binning: coarse-classify continuous variables using techniques like CHAID or manual monotonic bins, targeting smooth weight-of-evidence curves. Fourth, attribute selection using information value (IV > 0.1 is useful, > 0.3 is strong) and correlation filtering.

Fifth, modeling: fit a logistic regression on WoE-transformed predictors, often with stepwise or L1 selection. Sixth, scaling: convert log-odds to a points-based score using points-to-double-odds (PDO, typically 20) and an anchor odds/score pair. Seventh, validation: check Gini/AUC (aim for > 0.60 on out-of-time), Kolmogorov-Smirnov, calibration, and population stability index across segments. Eighth, deployment and monitoring: implement in the origination platform, set up PSI tracking, characteristic analysis, and override rate monitoring.

For an Arrowglen Credit Union retail portfolio, expected development time is 4 to 6 months including SR 11-7 model risk governance sign-off.

What is the end-to-end scorecard development process for credit risk?

Master Part I with our FRM Course

Related Questions

Practice Questions