A somewhat unpopular opinion we hold at Recast: yes, GDP, consumer sentiment, inflation, and interest rates do impact business performance — but no, you probably shouldn’t try to control for all of them in your media mix model.
That instinct to “control for everything” is understandable. It feels like the most responsible, data-driven approach. You want your model to be robust. You want to explain every possible shift in performance.
But this mindset creates two major problems.
First, it’s statistically impossible.
There are a ton of things that impact every business, and no statistical model — no matter how sophisticated — can control for all of them. If your ability to generate causal inferences requires controlling for every single macroeconomic and exogenous factor, you’ve already failed.
Even worse, this approach breaks down when you use your model for forecasting. Because if your model depends on exogenous inputs like inflation or interest rates, you then need to predict those variables to run your forecast. But how are you going to do that?
You can’t. No one knows what inflation will be six months from now. If you could forecast those variables with precision, you wouldn’t be building an MMM — you’d be running a hedge fund.
Second, it introduces bias.
When the analyst gets to choose which variables to control for — and how — it creates far too many degrees of freedom. The more knobs to turn, the easier it becomes to “optimize” toward the answer a stakeholder wants to see.
That leads to model cherry-picking, where ten versions of the MMM get built, but only the most flattering one makes it to the CMO.
Your media mix model shouldn’t be a cover-your-ass tool. It should be a measurement system grounded in statistical integrity and rigor, and the more you rely on complex external controls, the more likely you are to overfit or make decisions based on biased assumptions.
The recast approach: flexibility over control
At Recast, we don’t believe in building a model that explains everything. We believe in building a model that helps marketers make better decisions.
That’s why our approach doesn’t rely on hard-coded exogenous controls or pre-baked assumptions about how macro variables affect performance.
Instead, we use a flexible time-series model that evolves over time and adapts to changing baselines and marketing effectiveness.
We’ve stress-tested this approach through real-world shocks, and one of the clearest examples is Covid.
For some Recast customers, COVID meant a dramatic drop in sales. For others — especially those in home fitness or e-commerce — it triggered a surge in demand. That divergence wasn’t something we modeled explicitly. We didn’t need to include a “Covid” variable in our MMM.
Why? Because we saw those effects show up organically in two places:
- Baseline performance (i.e., organic demand)
- Marketing ROI (i.e., channel effectiveness)
For a consumer brand that benefited from Covid, our model picked up an increase in baseline sales and a simultaneous improvement in ROI. And that insight translated into a clear action: spend more on marketing while the environment favors you. The time-series structure of our model allowed us to observe the change and act on it — without overcomplicating the system.
That same flexibility applies to any macro shift: inflation, interest rates, even changing brand awareness. These dynamics show up in the model as time-varying changes in effectiveness or demand. We don’t pretend to know why the baseline moved — we just measure that it did, and we let the model adjust.
When and how to include external variables (carefully)
We’re not dogmatic. There are cases where including external variables in an MMM makes sense — but only when done with a clear causal rationale and the right modeling structure.
Here are a few situations where this can be valid:
1. platform-level shocks
Events like the iOS14 privacy update or sudden changes in ad platform policy clearly affect how paid media performs. These are structural shifts in how marketing works.
In these cases, a thoughtfully constructed variable (e.g. a post-iOS14 flag or a time-varying slope for channel performance) can help the model capture reality more accurately.
But even then, you need to be careful. These variables should be introduced to explain marketing performance, not to smooth residuals or juice R-squared.
2. observable supply-side disruptions
Let’s say your business had a major inventory shortage that suppressed sales for a few weeks. If that disruption was real, measurable, and not due to marketing, it may be appropriate to include a covariate that “mutes” sales during that window.
Again, the point isn’t to make your model look prettier. The point is to isolate the part of sales that marketing actually influenced.
3. known drivers of long-term effects
Sometimes, slow-moving macro factors like price or brand awareness have consistent, measurable effects on marketing ROI. Recast supports modeling these as context variables — but only when they’re observed over time and their relationship with performance is interpretable.
For example, we’ve seen that as brand awareness rises, marketing ROI improves across the board. That’s not noise — that’s a causal relationship that can be modeled and acted on.
But we only include those variables if:
- They’re measured consistently (e.g. monthly brand tracker)
- The relationship is conceptually and statistically clear
- They interact with marketing, not just baseline demand
In every case, the burden of proof lies with the variable. If it doesn’t improve decision-making — by sharpening attribution, reducing bias, or improving forecast utility — it doesn’t belong in the model.
Controls should never be cosmetic. If you can’t defend why a variable is in your model, you shouldn’t include it. Because every extra input adds complexity, introduces fragility, and increases the chance of misinterpretation.
The cleanest models are often the most honest ones. And in MMM, honesty is what drives a better strategy.
