A smooth curve can hide an unsupported recommendation
A fitted response curve extends beyond the spend values that generated the data. Software can calculate a prediction there, but the data may provide little evidence for its shape. Budget optimisation is especially exposed because an optimiser can search precisely where a curve looks most favourable.
- Mark observed spend
- Compare plausible curves
- Bound the scenario
- Collect new evidence
Use the saturation guide to understand the function. Then separate interpolation within observed operating conditions from extrapolation into a new range.
The response you infer depends on transformation choices and what the data can separate. The how marketing mix models work learning path connects this idea to the assumptions, practical checks and next steps.
To work through the connected decisions, read MMM budget constraints: turn an optimum into a feasible plan and Robyn hyperparameters: investigate estimates at the bounds.
Compare two curves over the same history
Imagine a channel usually spends between £9,000 and £11,000 per week. Two hypothetical curves both predict about 1,000 incremental units at £10,000. At £30,000, one predicts 1,300 units and the other 2,000. The historical data near £10,000 may not distinguish those high-spend predictions.
Plot the observed spend distribution underneath the curves. Mark both the historical range and the proposed allocation. A visually attractive optimum is not evidence that the response at that point has been measured.
Check whether the operating conditions also change
Moving outside the historical range can alter audience composition, frequency, auction price and creative delivery. Even a mathematically well-estimated historical curve may not transfer unchanged. The relevant uncertainty is not only the confidence interval around fitted parameters.
Ask what the next increment of spend buys. If the plan requires a new audience or much higher frequency, record that mechanism as part of the scenario. Keep the cost and exposure assumptions separate so a stakeholder can see why a spend recommendation changed.
Use constraints as a decision safeguard
Set lower and upper bounds based on observed support, operational commitments and the size of a change the business can evaluate. These are not arbitrary obstacles to the optimiser; they express the decision's feasible domain. The allocation guide shows how constraints affect the solution.
Compare the recommendation across plausible curves and cost scenarios. If a modest increase performs reasonably across them, it may support a staged test. If the result depends on an optimistic extrapolated tail, collect evidence before committing the full move.
A useful approval note states current spend, proposed spend, historical support, uncertainty, constraints and a stop or review rule. After the test, use the new variation to update the model rather than claiming the original optimum was validated simply because sales increased.
Sources and further practice
This article’s worked example is illustrative, not a measured client result. Use the linked methodology to check the assumptions before applying it.