Most MMM stories start the same way.
A vendor shows a model that fits the data “almost perfectly.” The predicted line hugs the actuals. Peaks and troughs line up. The decomposition looks tidy. The fit metrics are clean enough to end the conversation.
And then the model is used for what it was purchased for—budget allocation, marginal ROI, forecasting—and something feels off. The ROIs don’t pass a basic smell test. The response curves imply behavior the business doesn’t recognize. The recommendations look confident, but not trustworthy.
That gap—between statistical fit and decision safety—is not a vendor problem. It is a property of the setup.
Because in MMM, fit is easy to manufacture.
Fit is not verification
MMM is built on a seductive idea: if a model explains historical sales, it can guide future spend.
The danger is that many different stories can explain the past—especially when you allow the model enough freedom to shape the story.
So “great fit” often becomes a certificate of plausibility, not a proof of truth.
1) The model can fit by using variables that are too close to the outcome
The fastest path to an impressive fit is to include inputs that correlate with the outcome for the wrong reason—because they are downstream of it, mechanically related to it, or effectively a repackaged version of it.
A non-MMM example makes this hard to ignore.
In a churn / lifetime model for a service business, a vendor used “number of past churns” as a predictor of future tenure. The model looked strong in-sample. But the coefficient came out positive—more churns implied longer lifetime.
That wasn’t insight. It was a bookkeeping effect.
Customers who stay longer have more chances to churn and return. “Past churns” didn’t cause longevity; it was a side effect of longevity. Fit improved, but the interpretation broke.
MMM allows the same mistake, just with better charts: the model is rewarded for grabbing late-stage demand proxies that sit close to conversion and therefore correlate tightly with sales. The proxy then absorbs variance that should have been attributed to upstream drivers—marketing, pricing, competition, macro demand—and the model becomes confident in the wrong place.
A manufacturing example: “days of supply” as disguised sales
A subtler version shows up in manufacturing and durable goods.
I’ve seen “days of supply” or “weeks of supply” included as an input into MMM or forecasting models. The fit looked excellent.
If you’ve worked in this environment, the issue is immediate: near-term sales are often predictable from recent sales dynamics. So any metric that is tightly coupled to recent sales will look powerful, whether or not it contains causal information.
Days-of-supply measures often move with recent sales unless inventory policy is changing meaningfully. When inventory levels are relatively stable, “days of supply” becomes a convenient way of importing recent sales structure into the model.
So the model “learns” that days of supply explains sales and delivers a strong in-sample fit.
But what has really happened is simpler: the model has been allowed to predict sales using a proxy that is partially constructed from sales.
That isn’t measurement. It’s a form of leakage.
And because the proxy sits so close to the outcome, it doesn’t just improve fit—it distorts attribution. It absorbs variance that should have been assigned to true drivers. The model looks precise while becoming less informative.
2) The response functions are flexible enough to impersonate “truth”
Even when the inputs are clean, MMM can still earn fit for the wrong reason: the response functions themselves are powerful enough to synthesize behavior the data never truly supported.
Adstock and saturation are not just transformations. They are behavioral assumptions. They define what the model is allowed to believe about carryover, decay, and diminishing returns.
In principle, that flexibility is a feature. Consumer response is delayed and nonlinear.
In practice, it creates a failure mode: the model can produce plausible-looking dynamics because it is rewarded for any dynamics that reduce error.
I’ve seen saturation curves that—once combined with coefficients—effectively become a step function. Not because demand suddenly becomes binary, but because the functional forms can create sharp thresholds, and the optimizer is happy to use them if the data doesn’t strongly resist.
This gets worse because MMM is often parameter-rich relative to the data.
Weekly time series sounds large until you subtract reality: you don’t have that many independent “marketing experiments.” Channels move together. Seasonality is strong. Promotions are lumpy. Competitive behavior is imperfectly observed. Macro conditions drift.
So the system ends up with a small number of effective observations supporting a large set of interacting assumptions.
At that point, “good fit” can simply mean the model found a clever way to draw a line through the past—not that it identified a stable response relationship.
And you won’t catch that from fit metrics. You only catch it by inspecting what the model is implying: the carryover lengths, the saturation shapes, and the marginal curves that drive decisions.
3) Seasonality is a hidden place to “absorb variance”
Seasonality is one of the easiest places to improve fit without looking like you’re doing anything aggressive.
If the seasonal component is flexible enough, it can absorb whatever the model struggles to explain. Residuals shrink. And the model looks “stable.”
But that stability can be purchased at a price: the seasonal term quietly takes credit that should have gone to marketing, pricing, competition, or macro shifts.
This gets especially dangerous when history is short. With limited time, highly flexible seasonal terms can track idiosyncrasies in your specific period and present them as timeless structure.
The model doesn’t have to be wrong to be unsafe. It just has to be too free to explain the past.
And the more seasonality explains, the less the model is actually measuring.
Why this creates false confidence
None of this requires fraud. Most of it happens inside sincere work.
Fit is clean. Fit is legible. Fit gives stakeholders permission to stop asking questions.
But MMM is not a curve-fitting contest. It’s a decision system built on weak identification, correlated inputs, and large modeling discretion.
So “perfect fit” often means only this:
The model found a story that tracks the past.
The question is whether that story remains true when money moves.
The uncomfortable ending
Most organizations say they want MMM to tell them what is true.
What they often need, operationally, is something easier: a stable narrative that lets budgeting proceed without conflict.
Fit is an efficient way to produce that stability.
Which is why “perfect fit” isn’t reassuring. It’s a warning sign that the model may have become persuasive before it became trustworthy.
And the spending moves first—long before anyone can tell whether the story was real.