Learn / Regulation
How to tier your credit models by risk
Tiering means giving each model a risk rating, such as high, medium or low. The rating decides how often the model is validated, who has to approve it and how closely it is watched. RBI's draft makes tiering the lever for everything else, and leaves the scoring method to you. Here is a simple one.
What are the two questions?
The draft names two main factors. You can score both with a plain question.
- Impact (the draft calls it materiality): if this model is wrong, how much money, how many customers or how much regulatory trouble is at stake?
- Complexity: how hard is it to see why the model gave an answer? A short logistic scorecard is easy to read. A model on text, or one that retrains itself, is not.
For AI models the draft adds two more: how heavily people rely on the output, and how much of the decision the model makes without a person.
What rule do people get wrong?
The draft says that combining the factors must not let one offset or dilute the other. A simple model that drives a large decision cannot be tiered low just because it is simple. One easy way to follow that: take the higher of the two ratings, not the average. That method is ours, not RBI's.
A worked example
Illustrative only. Each lender sets its own tiers in its own framework.
| Model | Impact | Complexity | Tier (higher of the two) |
|---|---|---|---|
| Personal loan application scorecard | High | Low | High |
| Provisioning (ECL) model | High | Medium | High |
| Collections propensity model | Medium | Medium | Medium |
| Marketing response model | Low | Medium | Medium |
| Customer chatbot using generative AI | Medium | High | High |
The scorecard is simple, but it decides who gets a loan, so it is high. The marketing model can be wrong without much harm, but it is hard to explain, so it stays medium.
What should the tier change?
| Tier | Validation | Approval | Monitoring |
|---|---|---|---|
| High | Full independent validation before go-live and on a short cycle | Board risk committee | Monthly, with set limits for drift |
| Medium | Independent review before go-live, lighter periodic checks | Senior management | Quarterly |
| Low | Light review | Model owner and risk | Yearly |
The specific cycles above are examples. The draft leaves the frequencies to your own framework.
When should you re-tier a model?
- At least once a year.
- When the model is used for a new product or customer group.
- When it is retrained or its inputs change.
- After a serious incident or a validation finding.
Try it on your own tables
The free tools give you the prompt, the checks and how to present the result. You enter table names only.
See the free toolsFrequently asked questions
What is model tiering?
Ranking each model by how much harm a wrong result could do, so that the most important models get the most checking.
Should I average impact and complexity?
Not if one factor would dilute the other. The draft says they must not offset each other, and taking the higher rating is a simple way to meet that.
Are vendor models tiered too?
Yes. They need a tier like any other model, and the draft says the Board's risk committee gives vendor models extra oversight whatever their tier.
How often should tiers be reviewed?
At least once a year, and sooner when something material changes.
Keep reading
RBI's draft model risk rules: what they mean for your credit models
RBI has drafted rules on how banks, NBFCs and credit bureaus manage the risk of their models. What is in the draft, what it means for scorecards and AI, and what to do now.
RegulationWhat is a model inventory, and what should it hold?
A model inventory is the single list of every model your institution uses. RBI's draft says no model may be used unless it is on it. Fields to include, with an example.
ModelsHow to validate a credit scorecard: Gini, KS and PSI explained
Gini, KS and PSI explained with one worked table and tested SQL. What each measure tells you about a scorecard, and the checks that catch what they miss.