This article, part three of a series on robust credit scoring models, tackles critical preprocessing steps: managing outliers and missing values in borrower data. The author uses a Kaggle dataset of 32,581 loan records with 12 variables to demonstrate techniques. A key innovation involves creating an artificial 'time' variable using credit history length to enable proper train/test/out-of-time splits, crucial for evaluating model stability over time.

The core focus remains on data quality, emphasizing that preprocessing methods must ensure the model generalizes to new borrowers, not just fits historical data. The dataset's artificial time variable allows for meaningful assessment of risk driver stability across different periods, addressing a fundamental methodological question in credit modeling.