The bar for the complexity
MMM earns its keep when experiments cannot cover the allocation question alone: channels that resist geo-testing (brand, sponsorship, above-the-line), interaction effects between channels, and diminishing-returns curves that decide the next unit of budget. If your spend is concentrated in two testable performance channels, run the tests and skip the model — a decision that saves a quarter of data assembly and a permanent maintenance obligation.
The inputs that decide the output
| Input | Standard | iGaming-specific note |
|---|---|---|
| Outcome series | From the warehouse: registrations, FTDs, cohort-anchored value | Never platform-claimed conversions — the model would learn the platforms' self-grading |
| Spend & activity | Weekly per channel per market, including free variation (pauses, pulses) | Affiliate cost enters as spend too — CPA is a media cost with a different invoice |
| Confounders | Seasonality, pricing, distribution | The fixture calendar, the promo calendar, regulatory events, product launches — the four that dominate this industry's demand curve |
| Calibration anchors | Holdout results where they exist | The experiment registry is the anchor store; a model without anchors is opinion with priors |
Reading the model like an operator
- Response curves over point estimates. The useful output is the shape — where each channel's next unit of spend stops paying — not a league table of average ROIs.
- Uncertainty travels with the recommendation. A reallocation whose confidence interval includes zero is a hypothesis for the next experiment, not an order for the next budget.
- Stability across refits is the trust metric. Coefficients that swing wildly between quarterly runs are telling you the data lacks variation or the confounders are incomplete — believe the swing, not either endpoint.
- The model proposes, the experiment disposes. Big reallocations the model suggests get validated by a geo test before they become permanent — closing the loop that keeps both instruments honest.
Governance, briefly
The model is a metric with extra steps, and it gets the dictionary treatment: a named owner, versioned specifications, a documented refit cadence, and every published number labelled with its model version and calibration date. The quarterly review reads the allocation recommendation next to the experiment registry and the unit-economics view — three instruments, one decision, no single number pretending to be the truth alone.
Continue reading: CAC, LTV and payback — the unit economics the allocation serves. The measurement stack — where MMM sits among the three clocks.