Model Risk Governance
Model-risk governance defines intended use, inventories material models, assigns accountable owners and subjects models to effective challenge. Validation considers conceptual soundness, implementation and outcomes; monitoring detects when a model no longer fits its use.
The method at a glance
| Concept | How it works | Watch for |
|---|---|---|
| Intended use | Specify the decision, population, horizon and limitations. | Reusing a model for a materially different decision without assessment. |
| Validation | Challenge assumptions, data, implementation and out-of-sample performance. | Equating developer approval with independent challenge. |
| Monitoring | Define performance thresholds, escalation and remediation. | Leaving a deteriorating model in service with no documented response. |
01Intended use
How it works: Specify the decision, population, horizon and limitations.
Watch for: Reusing a model for a materially different decision without assessment.
02Validation
How it works: Challenge assumptions, data, implementation and out-of-sample performance.
Watch for: Equating developer approval with independent challenge.
03Monitoring
How it works: Define performance thresholds, escalation and remediation.
Watch for: Leaving a deteriorating model in service with no documented response.
Original worked example
A good development fit fails on new data
A loss model has development RMSE of 1.2 and holdout RMSE of 3.6 in the same units. The holdout covers a different market regime.
- STEP 1
The holdout error is three times the development error.
- STEP 2
Investigate overfitting, data differences and regime change rather than choosing a cause from the ratio alone.
- STEP 3
Compare a simpler benchmark and determine whether restricted use, adjustment or redevelopment is justified.
The performance gap triggers investigation; it does not by itself prove overfitting or prescribe one remedy.
Illustrative assumptions and calculations for learning. Not observed market data or an official exam item.
Apply the method and check its limits
Before you start: Model assumptions, validation data and the decision the output supports.
Challenge concept, implementation and use
A model can fail because its theory is unsuitable, its code implements the theory incorrectly, or its users apply it outside its intended scope. Validation should examine all three. Reproducing a calculation does not demonstrate that its assumptions are appropriate. Document population, horizon, inputs, limitations and who can approve material changes.
Choose a response tied to a diagnosed weakness
A data-quality failure may require correction and rerun; unstable relationships may require recalibration or redesign; use outside scope may require a restriction. Specify monitoring triggers and an accountable remediation owner. Compare with a simple benchmark so that a complex model must demonstrate usefulness beyond fitting its development sample.
Separate current supervision from exam readings
The Federal Reserve’s April 2026 SR 26-2 supersedes SR 11-7 for current US supervisory guidance. That update is external context, not a statement that GARP changed a prescribed reading midyear. For the exam, follow the learning objectives and assigned edition; for a real institution, check the applicable regulator, scope and implementation requirements.
Work through the topic
Evaluate whether a model is fit for a stated decision, organize independent challenge and monitoring, and make limitations consequential for how the output is used.
Define the model’s use and materiality
Model risk arises when decisions are harmed by incorrect or misused model outputs. Governance starts with the intended use, the decisions influenced and the potential consequence of error. A model suitable for a rough planning estimate may be unsuitable for approving a large exposure. Treating its output as authoritative outside its validated purpose can create risk even when the underlying calculation works as designed.
Create an inventory entry with ownership, inputs, outputs, implementation, intended users, limitations and change history. Materiality should reflect the importance of the decision and exposure to error, not just mathematical complexity. A simple spreadsheet used for a major funding decision can matter more than a complex experimental model with no production use.
The linked Federal Reserve SR 26-2 guidance was issued in April 2026 and supersedes SR 11-7 as current US supervisory background. It does not rewrite GARP’s prescribed exam-year readings, and its scope should not be presented as a universal rule for every institution. Distinguish current practice references from the source and assumptions specified in an exam question.
Challenge design, implementation and outcomes separately
A design review asks whether the model’s concepts and assumptions suit its purpose. An implementation review asks whether data and code reproduce that design accurately. Outcome analysis asks how predictions compare with relevant experience. Strong evidence in one area does not automatically resolve weaknesses in the others. A correctly coded model can omit a material economic mechanism.
Consider an original credit model trained on a portfolio of established borrowers and then proposed for newly formed businesses. Implementation tests reproduce the formula exactly, but the new population has different data availability and financial characteristics. The main concern is applicability and performance in the new use, not whether the arithmetic matches the specification. Relevant validation needs to examine the new population and the uncertainty caused by limited outcomes.
Independent challenge needs sufficient expertise and authority to question use, require remediation or recommend restrictions. Record findings with severity, evidence, owner and a due date. A limitation such as “weak performance in the new segment” should lead to an observable control, such as restricted exposure, additional review or a decision not to use the model there. A disclaimer that changes nothing is a weak response.
| Review dimension | Evidence to inspect | Decision it supports |
|---|---|---|
| Conceptual suitability | Relationship between training population and new use | Whether the model applies to the segment |
| Implementation | Input reconciliation and reproducible calculations | Whether the deployed process matches the design |
| Outcomes | Calibration and errors in relevant populations | Whether observed performance is acceptable |
| Use controls | Limits, overrides and accountable review | How unresolved uncertainty affects decisions |
Original new-borrower model review
- Conceptual suitability
- Evidence to inspect: Relationship between training population and new use
- Decision it supports: Whether the model applies to the segment
- Implementation
- Evidence to inspect: Input reconciliation and reproducible calculations
- Decision it supports: Whether the deployed process matches the design
- Outcomes
- Evidence to inspect: Calibration and errors in relevant populations
- Decision it supports: Whether observed performance is acceptable
- Use controls
- Evidence to inspect: Limits, overrides and accountable review
- Decision it supports: How unresolved uncertainty affects decisions
Make monitoring thresholds lead to a response
Monitoring should detect changes relevant to the model’s purpose: input distributions, missing data, population composition, calibration, overrides and downstream decisions. A stable overall average can hide deterioration in a material segment. Choose thresholds with a rationale and specify the action they trigger. The threshold itself should not become a target that encourages relabeling observations.
For an original governance exercise, an institution sets a review trigger when a material segment’s observed default rate exceeds the model estimate by more than two percentage points over a defined, sufficiently informative review window. If the estimate is 3% and the observed rate is 6%, the three-point gap triggers review under that policy. It does not by itself establish statistical significance or identify the cause. Sample size, uncertainty and changes in the segment must be assessed.
Changes to data sources, parameters, implementation or use need appropriate review and a reproducible record. Preserve the earlier model and relevant results so that improvements can be evaluated without rewriting history. If a model is restricted or retired, identify how users will make the affected decisions and whether a replacement introduces new dependencies. Governance covers the lifecycle, including the consequences of stopping use.
Does validation mean a model is correct?
Validation supplies evidence about suitability, implementation and performance for a defined use. It does not eliminate uncertainty or prove correctness under every future condition. Record limitations and connect them to monitoring and use restrictions.
Can a human override remove model risk?
An override can correct a particular weakness or introduce another source of error. Record its rationale, approval and outcome, then review patterns. Frequent overrides may indicate model limitations, misuse or a process problem that needs investigation.
Try a changed assumption
Independent validators reproduce the developer’s exact output. Does that alone establish fitness for use?
Show answer and reasoning
No. It is evidence about implementation consistency. Conceptual soundness, data quality, outcomes and appropriateness for the intended decision still need assessment.
Original learning check. No email or account needed.
Does validation guarantee a model is correct?
No. Validation assesses fitness for a defined use and exposes limitations. Ongoing monitoring and accountable decisions remain necessary as data and conditions change.
Sources and scope
Links checked 2026-10-10. GARP establishes exam scope; the technical references support the methods identified below. These are selected study notes and original examples, not complete curriculum coverage or official exam questions.
- GARP FRM study materials and 2026 learning objectives
Exam-domain scope and the official route to the annual learning objectives. This page does not certify coverage of every reading.
- Federal Reserve: revised model-risk guidance, SR 26-2
Current US supervisory background. Issued April 2026; supersedes SR 11-7. This does not change the prescribed FRM exam-year readings.
- MIT OpenCourseWare: quantitative lecture index
Further reading across quantitative methods. Specific technical references are linked beside the explanations below.
- MIT: Regression Analysis, lecture 6
Least squares, coefficient interpretation, model assumptions and inference.
Prepared by MiloPrep editorial team. MiloPrep is independent of GARP. References do not imply endorsement or independent review of these notes.