A sentinel value is a placeholder code for a special condition, such as missing or inapplicable information, not an ordinary measurement. In credit data, codes like -9999 or 8888 may represent no bureau record or an unknown employment type. Treating them as valid numeric values can skew distributions and model performance. section 4.4.4
Handle sentinel-coded fields explicitly:
- Identify candidate codes during data exploration and confirm their meaning with the data provider. section 4.4.4
- Keep sentinel values separate from valid numeric ranges, for example by assigning them to a distinct model bin. section 4.4.4
- Check default rates for each sentinel group; the group may carry risk information rather than being neutral. section 4.4.4
- Document the treatment and its effects so the approach can be assessed for consistency with fair-lending principles. section 4.4.4
Sentinel codes are one form of special-value challenge, alongside missing values and outliers. Do not assume every sentinel code has the same meaning or risk profile; validate the code and examine its group before deciding how to model it. section 4.4.4 section 4.4
