Most marketing mix models break because they assume that marketing effectiveness is constant over time – but that’s not how marketing works in the real world.
Marketing effectiveness changes constantly. Sometimes dramatically.
It changes when you update creative. It changes when your audience saturates. It changes when new competitors enter the market or when external events (like iOS 14.5 or a global pandemic) meaningfully alter consumer behavior. It even changes when tracking methods degrade or when a platform tweaks its ad delivery algorithm.
Then, when marketing effectiveness fluctuates, and the model doesn’t allow for that, it compensates by “twisting” other parts of the model to fit the observed data. That leads to very unstable forecasts that “jump around” depending on the time window used. And when this happens, budgets get shifted based on whatever version of the model happened to run last, and teams lose confidence when results don’t hold up.
If your model assumes marketing ROI never changes, then your model is wrong – full stop. To solve for this, you need time-varying covariates.
What Time-Varying Covariates Actually Are
Time-varying covariates are exactly what they sound like: variables in a model that are allowed to change over time. In the context of media mix modeling, they’re what allow us to reflect the idea that marketing effectiveness itself is not static.
A model that treats ROI as a fixed coefficient will tell you that Meta drives, for example, $3.20 in incremental revenue for every $1 spent.
Time-varying covariates allow the model to say: “In Q1, Facebook’s ROI was 3.2x. But in Q2, after we refreshed creative and TikTok started eating into our audience, that ROI dropped to 2.1x.”
A common misconception we’ve heard is that this is about modeling seasonality. It’s not. Seasonality explains when demand changes due to external patterns (holidays, weather, etc.). What we’re talking about here is fundamentally different: the elasticity of marketing channels changing in response to controllable and uncontrollable factors.
Without time-varying covariates, the model can’t see the slope. It fits a straight line to data that’s clearly curving, and the result is forecasts that make no sense.
What you’re looking for is stability – when performance changes, does the model attribute that shift to plausible causes? Does the model tell a story that makes sense? Without time-varying covariates, that story often falls apart.
Why Everyone Isn’t Building This
If time-varying covariates are so important to getting accurate, stable marketing mix models, why don’t more vendors support them?
The short answer is: it’s really, really hard.
From a modeling perspective, introducing time-varying ROI estimates adds multiple layers of complexity. It’s not five additional points of complexity, it’s to the fifth power more complex.
There are two major hurdles here. The first is statistical: designing a model that can reliably estimate marketing performance as it evolves over time. That means grappling with identifiability, regularization, and the math behind time series decomposition.
These aren’t solved problems in standard marketing analytics workflows. We had to hire many PhD statisticians and iterate on model structure from the ground up.
The second hurdle is computational. Once you allow ROIs to shift over time, your model becomes exponentially slower to run. Very quickly, we hit the limits of what could be done on a laptop. We had to build dedicated cloud infrastructure, optimize underlying C code, and come up with new workflows just to get runtimes into a usable range. Even today, these models are significantly more resource-intensive than simpler approaches.
But this isn’t a “nice-to-have” feature that we could afford not to build. It makes or breaks anything we do. So our suggestion is that you ask all MMM vendors if they support time-varying effects and, if so, how they are baked into the core of their modeling system.
TL;DR
- Most MMMs assume static ROI, but marketing effectiveness changes constantly. Your model is inherently wrong if it doesn’t show that.
- Time-varying covariates allow models to reflect real shifts in performance driven by creative, competition, or platform changes.
- Models without this capability often produce unstable, misleading results that will bias your decision-making and waste budget.
- Implementing time-varying covariates is hard, complex, and expensive – but absolutely necessary.
