An outlier is an unusually extreme value, such as very high income or a large transaction. In credit data, extreme values can distort binning and Weight of Evidence (WoE) calculations or cause rare observations to dominate a model. Treatment should preserve genuine risk signals without overfitting to sparse cases. section 4.4.4
Common treatments include:
- Cap or winsorize values in the top—and possibly bottom—1%–2%. section 4.4.4
- Apply a log or other transformation to reduce skew while preserving rank order. section 4.4.4
- Create a dedicated outlier bin when extreme observations are too sparse for a reliable distribution. section 4.4.4
Why extreme values need scrutiny
A handful of extremely wealthy borrowers could lead a model to treat wealth as a sign of safety, even though wealth alone does not guarantee willingness to repay. section 4.4.4
