Last week I closed with a promise. I said the error in measurement does not stay in measurement.
Today I want to show you exactly how it travels:
• What it does when it arrives at the forecast.
• What number appears on the other side once that propagation gets measured against known ground truth.
The measurement bias travels by two routes.
1. Parameter inheritance. The forecast uses the same elasticity and the same baseline shape the model learned during measurement. If that decomposition was biased toward paid, the projection carries that bias inside as a premise.
2. Structural ceiling. The response curves most standard models use have a saturation cap built in by design, not by evidence. That second layer decides how fast your model believes a channel is allowed to grow.
The two routes compound. The first one inflates the estimate of how much paid is currently earning. The second one decides that the same channel cannot climb much further.
What arrives at planning is a version of your business where paid has already done almost everything it can, and there is no headroom to bet harder.
The rocket with the skewed telemetry
Picture a rocket whose telemetry has a 3% bias on the first reading. Small in magnitude. Small in immediate consequence.
That reading gets used to calculate the next maneuver. The maneuver executes on a slightly wrong position, and the next reading carries the bias plus the effect of the maneuver.
As the rocket climbs, the error compounds. By the time it reaches apogee, the trajectory is nowhere near the planned orbit.
The forecast is the trajectory. Measurement is the telemetry. If the telemetry drifts at peak, the trajectory built on top of it fails exactly when the business can least afford it.
Where the ceiling comes from
Most standard MMMs use parametric response curves: Hill, Weibull, S-curves. They model a real phenomenon, saturation, where the marginal revenue per dollar decays past a certain spend level.
What is debatable is how the ceiling gets chosen.
The saturation cap is not learned freely from the data. It is bounded by priors the modeling team set, and the model picks the value within that range that best fits the observed period.
The result: "can this channel keep growing" gets answered by a ceiling the model decided on before seeing the data.
If the prior was tight, the forecast projects saturation even when the channel still has runway.
The proof
Here is the number I promised.
Across the ten ground-truth datasets in the OMEN paper, the average forecast error at peak for standard models ran 2x to 4x the error of the mechanistic model.
The direction is consistent: too-early saturation on channels with runway, sustained growth on channels already near their cap.
When that biased projection feeds optimization, you get the 81% BFCM overspend recommendation I cited last week. Measurement bias goes in, structural ceiling gets applied, optimization recommends on top of both.
The lived example is MaryRuth Organics.
A model with default saturation priors would have flagged Pinterest as approaching its cap. A forecast built on those parameters would have recommended against scaling.
MaryRuth scaled Pinterest 8x. Pinterest's CAC stayed nearly 50% more efficient than the portfolio average, even at the new scale. Sixty-seven percent of the impact landed as halo on other surfaces of the brand.
The ceiling the model had projected was not in the real channel. It was in the structure of the model that generated the forecast.
- Where the math fails, across the OMEN paper
Layer Standard MMM Prescient Measurement 3x average error baseline Forecast at peak 2x to 4x error baseline Optimization 32-45% error, up to 81% BFCM overspend 5.6%, within ~1%
- Three failures, one structural cause.
Two questions to ask your team this week
If you want to check whether this cascade is operating in your own data, two questions are worth asking.
First. When was the last time the team reconciled the four-week forecast against what actually happened? If the average delta is 15 to 20 percent or more during peak weeks, the forecast structure is carrying a bias.
Second. What is the gap between the model's budget recommendation and the decision the team ultimately made? If the team systematically deviates from the model at peak, the team is manually correcting something the structure does not capture.
That correction is information about what the model would need to learn on its own.
The Halo
Exclusive insights, every week.
Subscribe to The Halo for sharper marketing thinking.