Separate explainability from compliance
Stratyfy markets tools for credit decisioning and lending compliance, including feature analysis and approaches to improve transparency. The company’s public descriptions and customer stories are vendor claims. They do not alone demonstrate that a model is unbiased, that a lender has satisfied ECOA, or that reasons are correct for a particular deployment. [1][2]
Regulation B requires creditors to provide specific principal reasons for adverse action, including when complex models are used. A post-hoc explanation should faithfully reflect the actual decision process; selecting plausible reason codes after the fact is not enough. The creditor remains accountable for data, policy, notices and monitoring. [3]
Test the full decision path
A bank should evaluate input features, transformations, model score, policy rules, cutoffs, overrides and notice generation as one chain. Test reason-code fidelity on edge cases and monitor approval, pricing and error rates across legally permissible groups. A feature that appears neutral may proxy for protected status or reflect historical exclusion; removing one variable does not remove all proxy effects. [3]
For an illustrative test, select a stratified sample of denials and independently reconstruct the leading reasons from the production model and rule stack. Compare those reasons with the notices actually sent, then quantify unsupported or unstable reasons by product and channel. This is a hypothetical control design, not a reported Stratyfy customer result.
Trade-offs and proof
Transparent features can help analysts identify fragile variables and improve governance, but reducing model complexity may reduce predictive power. The goal is not a simple model at any cost; it is a well-controlled decision system whose performance, reasons and consumer effects can be tested. Require documentation, reproducibility, versioning and independent validation. [1][2]
Public vendor pages do not provide enough detail to establish independent accuracy, fair-lending impact or pricing. Ask for reproducible evidence using the bank’s population and compare outcomes against an approved baseline. Attribute performance claims and avoid treating a product label such as “unbiased” as a legal conclusion.
Sources
- Stratyfy — credit decisioningSourceBack to text: ↑1↑2
- Stratyfy — lending complianceSourceBack to text: ↑1↑2
- CFPB — Regulation B, adverse-action requirementsOfficial textBack to text: ↑1↑2