Hidden Redlining: How Zip Codes Become Proxies for Caste in Credit Models

In India, historical patterns of residential segregation mean that geography encodes protected attributes with surprising precision. **Example: Mumbai zip codes and community composition** A study of credit applications in Mumbai found that 23 postal codes in the city have community composition concentrations above 65% for a single religious group. A credit model trained on historical data from these areas will learn that the postal code predicts creditworthiness — but the creditworthiness signal is confounded with the protected attribute signal. **How proxy features enter models:** 1. **Data inheritance** — Training data reflects historical lending patterns that were themselves discriminatory. The model learns the historical pattern. 2. **Correlation mining** — Modern ML models (gradient boosting, neural networks) are extremely good at finding subtle correlations. A feature that humans would not consider a proxy can be identified by the model as predictive. 3. **Interaction effects** — Two individually innocuous features (city + employment type) can interact to create a powerful proxy for a protected attribute. **The Feature Correlation Matrix:** Claris's RAI Module computes pairwise correlations between all features and all protected attributes (including self-reported and proxy-inferred attributes) as part of the model validation process. Features with Cramér's V correlation above 0.15 with any protected attribute are flagged for human review.