Credit risk models need to beware of proxy variables: zip codes may implicitly contain racial information, and even if the model does not directly use protected features, it may produce indirect discrimination through proxy variables.