When a Forecasted-Control MMM Can (and Can't) Answer a Budget Question
A marketing analytics lead just allocated $30 million for next quarter's campaigns. The allocation plan depends on employment rates, temperature, and other external conditions the underlying Marketing Mix Model needs as inputs, for a quarter that hasn't happened yet.
That's the scenario-planning paradox at the center of most Bayesian MMM work. The model is good at explaining what already happened: how much of past sales each channel caused, once seasonality and other controls are accounted for. It is not, by itself, able to supply the future values of those controls.
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: 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 a forecasting error doesn't 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.
That gives a practical rule of thumb: multiply a control's forecast error rate by its share of the sales variation it explains, and the result approximates the added error in the sales prediction. A control that swings the forecast but barely explains sales doesn't matter, and neither does one that explains sales but forecasts well. Only a control that is both influential and poorly forecast should worry a planner.
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 this pipeline is built to decide, and what it isn't
The hybrid approach fits a specific set of planning problems: allocating a budget three to six months out under expected market conditions, projecting into a new geography using historical control data even without local sales history, and running scenario analysis across optimistic and pessimistic economic assumptions to see how the optimal media split shifts.
What it cannot do is tell a team whether an action the historical model has never observed will work. The MMM's coefficients are learned from what already happened. Forecasting the controls more accurately doesn't create a coefficient for entering a market the company has never entered, testing a price point it has never charged, or launching creative unlike anything already in the training data.
Where the decision has to move from forecasting to testing
That's the boundary that separates this pipeline from a controlled experiment. When the question is "how will this quarter's known dynamics play out given expected conditions," a well-calibrated MMM-plus-forecast pipeline is the right tool, and running a full causal experiment instead would be redundant with what the historical model already answers. When the question is "will this untested action work," there is no historical coefficient to extrapolate, and no amount of better control forecasting fixes that.
A controlled causal experiment on a simulated population is built for that second case. Subconscious runs the tested action, such as a new market, a new price, or a new campaign, against a defined population and reads out the effect, rather than projecting a coefficient that doesn't exist yet. Subconscious can test or validate those studies with real human participants as well, moving from a simulated read to a human-validated one without changing the underlying causal question. Teams weighing which side of that boundary a given planning question falls on can see how a study gets designed and run.