Imputation replaces missing values with estimated or designated values. In credit risk, incomplete applications and withheld fields can bias a model or distort its assessment of risk. section 4.4.4
Choose a treatment that reflects the data pattern and preserves meaningful signals:
- For numeric fields, use a mean, median, or model-based estimate, while recognizing that an average can hide risk when data is missing not at random (MNAR). section 4.4.4
- Create a dedicated missing bin when the absence itself may carry information; for example, omitted employment status can signal higher default risk. section 4.4.4
- Consider listwise deletion when only a small fraction of the dataset is missing; this is an alternative to imputation. section 4.4.4
Missing-value treatment can affect binning and Weight of Evidence (WoE) and Information Value (IV) calculations. Document why assumptions are reasonable: auditors and regulators may scrutinize treatment choices, particularly when missingness relates to sensitive attributes. Transparent rules support confidence in model fairness. section 4.4.4
A BFSI modeling setup may use scikit-learn’s SimpleImputer, but the import alone does not specify an imputation procedure or strategy. section 5.3.1
