When a Forecasted-Control MMM Can (and Can't) Answer a Budget Question
On this page
A forecasted-control MMM combines a fitted sales model with predictions of future external conditions. In an illustrative $30 million planning scenario, the team needs future employment and temperature inputs. Luca Fiaschi's March 17, 2026 PyMC Labs demonstration tests this approach with simulated sales and real historical controls, rather than a reported client deployment.
The planning question is whether the fitted relationships remain useful under future conditions. Forecasting controls supplies inputs; it does not establish the model's causal assumptions or validate a budget intervention.
Pairing a causal model with a forecasting model
One way to close that gap is to let two different models do two different jobs. A Bayesian MMM (built in a framework like PyMC-Marketing) keeps the causal structure under its identifying assumptions, such as no unobserved confounding and media spend that isn't itself driven by expected demand: it encodes how advertising drives sales through mechanisms like adstock and saturation, and it returns a coefficient for each control variable. A separate time-series foundation model, for example Chronos-2, a transformer-based forecaster Amazon released for general-purpose time series prediction, projects the values of upcoming control variables like temperature or employment from historical patterns (Introducing Chronos-2: From univariate to universal forecasting).
The pipeline runs in three steps: fit the MMM on historical media spend and controls to learn the causal coefficients, forecast the upcoming controls with the foundation model, then feed the planned media spend plus the forecasted controls back through the fitted MMM to get a sales prediction for the future quarter. The MMM's own causal structure, separating channel effects from control effects such as competitive activity, is itself an active area of methods work; a recent framework addresses channel-influence bias and cross-channel effects directly inside the MMM specification (A new framework for Marketing Mix Modeling).
Why doesn't a forecasting error always wreck the prediction?
What matters isn't whether Chronos-2 forecasts controls well, but how much a forecasting error in those controls moves the final sales prediction. Each MMM prediction is the sum of a media-channel contribution and a control-variable contribution, where the control contribution is coefficients multiplied by control values. Swap in a forecasted control instead of the true one, and the resulting error in the sales prediction scales with two things: how large the forecasting error is, and how much that control actually matters to the model.
For an additive model y_hat = m + sum(beta_j * x_j), holding the media contribution and coefficients fixed gives delta_y_hat = sum(beta_j * (x_forecast_j - x_j)). Each term has outcome units because the coefficient converts control units into outcome units. Absolute error is bounded by the sum of absolute coefficient-weighted control errors. Relative errors require a stated, nonzero outcome denominator. A contribution-share multiplication is only an approximation in this specification; negative contributions, near-zero outcomes and changing coefficients can make it misleading.
The rule assumes the control variables behave in a relatively stable, well-patterned way. A structural break, such as a pandemic, a policy shift, or a new competitor entering the market, can throw off even a strong forecasting model, and a control with a clean seasonal pattern will forecast far better than one without.
What can and can't this pipeline decide?
The hybrid approach fits a specific set of planning problems: allocating a budget three to six months out under expected market conditions, and running scenario analysis across optimistic and pessimistic economic assumptions to see how conditions shift the forecast, while a proposed media split may require extrapolation beyond the historical spend support.
Forecasting controls alone does not identify the effect of a new intervention. A model can predict an unobserved spend value under its response and transport assumptions, but unsupported extrapolation needs additional evidence. A market-entry, price or creative effect absent from the fitted specification is not supplied by a more accurate temperature or employment forecast.
Where the decision has to move from forecasting to testing
Control forecasting and intervention identification answer different questions. A budget forecast needs valid relationships, supported spend ranges and propagated control uncertainty. A lift experiment can test a relevant channel effect; better input forecasting alone cannot establish an untested effect.
A simulated experiment can estimate a modeled response to a new action, but that is not yet real-market budget evidence. Check audience representation, treatment support and a matched held-out human or market effect before transporting it. Preserve the stimuli and intended estimand when comparing sources of evidence. See validation methodology or discuss the planning question.