MMM governance: making control proportionate to the decision risk.
Key distinctions
Three conditions before calculating.
Authorised use
The same model can be fit to explain trends, compare scenarios or design a test without being fit to impose an allocation. Each use must be approved separately.
Independent challenge
Validation examines conceptual coherence, data, out-of-sample results, sensitivity, calibration and the stability of the decisions. The validator must be able to restrict or refuse the use.
Monitoring and withdrawal
Performance, data drift, market change, gap to the experiments and actual decisions are tracked. Thresholds determine review, overlay, restriction or withdrawal.
Method
Putting the model under control throughout its life cycle
- 01Inventory the model
Document version, owner, developer, data, objectives, uses, users, dependencies, assumptions, validation date and decisions influenced.
- 02Class the risk
Score financial exposure, uncertainty, frequency and reversibility with common criteria; the level obtained determines independence, depth of testing and approval authority.
- 03Pass the gates
Require data review, technical validation, comparison with benchmarks, sensitivity testing and business approval before use; keep the reservations and limits visible in the outputs.
- 04Monitor the use
Compare recommendations, decisions taken and results; trigger a review after a major change, drift, experimental gap or exposure overrun.
The MMM risk register and approval gates
The model receives a level from exposure, uncertainty and reversibility; each level triggers precise controls.
€800k reallocated
iROAS 0.8–2.6
Monthly review
A reallocation of €800,000 is limited to a pilot of €160,000
An MMM proposes moving €800,000. Exposure receives 3 out of 4, uncertainty 4 because the iROAS ranges from 0.8 to 2.6, and reversibility 2 because the budget can be revised monthly. The weights are 40%, 40% and 20%.
Risk score: 3 × 0.4 + 4 × 0.4 + 2 × 0.2 = 3.2 out of 4, that is a high level. The rule requires independent validation, sensitivity, experimental comparison and a pilot limited to 20%: 800,000 × 20% = €160,000.
The full move is not authorised. The pilot measures the direction of the effect and the operational capacity; the extension requires the results to stay within the approved range.
Acceptance conditions
What must be true to act.
- 01
Every version has an owner, authorised uses and limits.
- 02
The level of control increases with exposure, uncertainty and irreversibility.
- 03
High-risk uses require an independent challenge.
- 04
Drift, experimental gap and major change trigger a review.
Limits
What this analysis does not prove.
- The grid proposed is not a regulatory standard for marketing.
- A good governance process does not repair a conceptually weak model.
- The risk score must not replace the documented judgement of the committee.
References
Works cited.
- Board of Governors, OCC & FDIC (2026), SR 26-2, Revised Guidance on Model Risk Management.
- Chan, D. & Perry, M. (2017), “Challenges and Opportunities in Media Mix Modeling”, Google Research.
- Zhang, Y. et al. (2024), “Media Mix Model Calibration With Bayesian Priors”, Google Research.