Model risk tiers and independent validation
Models are tiered by materiality. The tier decides how deeply Model Risk Management validates, how often it revalidates, and where a person must decide.
For: Builder / Publisher, Model Risk & Governance, MLOps Engineer, Consumer · 1 min read
The three tiers
| Tier | When | Validation | Revalidation |
|---|---|---|---|
| Tier 3 | No effect on a customer outcome; low materiality | Reproduced: the platform re-runs the builder's tests; a reviewer checks coverage | Every 3 years or on material change |
| Tier 2 | Informs a decision about a customer that a person still makes | Independent: scored on an out-of-time sample Model Risk holds back | Every 2 years or on material change |
| Tier 1 | Decides or materially shapes credit, pricing or a customer outcome | Challenged: challenger model, fairness tests, explainability review, stress scenarios | Annually, with continuous monitoring |
Why validation is independent
Model Risk Management sits in Group Risk, the second line. The people who build a model never validate it, and builders do not see the sample it is validated on. A named validator signs the opinion before a Tier 1 or Tier 2 model is used. Internal Audit, the third line, reviews that the process was followed.
Credit decisions
- Fairness testing on protected groups is required for any model that feeds a credit decision.
- Every credit score comes with reason codes, so an adverse outcome can be explained to the customer.
- Champion and challenger models run side by side before a new version replaces the old one.
Read next
Was this helpful?