Marketing Mix Models in their second year:
what to do next.
AMarketing Mix Model gets more valuable the more it is used. The saturation curves are a live claim about how your category works: which channels have room to grow, which are approaching their ceiling, where a reallocation produces the most return. That claim is worth testing, regularly, against actual outcomes. Most brands do not test it. This note is about why, and what to do.
In year one, the model is new and the team treats it seriously. By year two, the channel mix has shifted, the saturation points no longer match current spend levels, and the model is producing recommendations that do not quite land. The model is not wrong. It is stale. And staleness is a different problem with a different fix.
The fix is to close the feedback loop. Connect the model to a system that measures whether its predictions are coming true, flags when they stop, and feeds that signal back into the next training run. When that loop is closed, the Marketing Mix Model compounds. Each run is calibrated by the outcomes of the last one. The model gets sharper each year, not weaker.
The staleness problem.
A Marketing Mix Model trained in October 2024 was calibrated against a media market with specific CPMs, specific competition, specific creative saturation across channels. By October 2025, all three have changed. The model’s saturation curves are drawn on the wrong axis.
Most brands do not notice, because they are not measuring model performance. They present the model’s output, they approve a plan that looks coherent, and they compare actuals to plan at quarter end. If the plan was off, the explanation is usually ‘market conditions’. The model is rarely interrogated.
The way to catch staleness is to measure the model’s predictions against outcomes in a holdout period. If the model predicted a 12% lift from a reallocation and the lift was 4%, the model is miscalibrated. That signal should go back into the next training run.
The feedback loop most brands skip.
A Marketing Mix Model is not a static artefact. It is a claim about how marketing works in your category at a specific point in time. That claim should be tested, regularly, against the actual outcomes of decisions the model recommended.
Most Marketing Mix Model vendors do not close this loop. They deliver the model, they deliver the report, and they return the following year with a new model. The brand has no mechanism for tracking whether last year’s model was right, and therefore no mechanism for improving the model’s quality over time.
Acera Labs closes the loop by measuring every approved decision against the model’s prediction. Variance is logged and surfaced. At the next model refresh, the brand has a structured file of predictions and outcomes to feed into the training run. The model learns from what it got wrong.
The model is not wrong. It is stale. Those are different problems with different fixes.
Three signs your model is past its use-by date.
One: the channel mix has changed by more than 15% since the model was trained. New channels added, old ones dropped, or spend proportions shifted materially. The model was not calibrated on the current mix.
Two: the model’s predictions are consistently off in the same direction on the same channel. Systematic bias means the channel’s saturation curve is drawn wrong. That is not noise; it is miscalibration.
Three: the model is being used to plan for a spending level outside its training range. A model trained at $2m per quarter cannot reliably recommend allocation at $5m. The curve shapes extrapolate poorly beyond the range they were fit on.
What to do.
All three signs point at the same fix: measure the model’s predictions against actual outcomes, regularly, and feed that variance back into training. Most Marketing Mix Model vendors do not offer this. Acera Labs closes the loop by logging variance on every approved decision. The model compounds because the loop is closed. Year three is sharper than year two, which was sharper than year one.