Any MMM that doesn’t account for time-varying ROIs is fundamentally flawed. It’s that simple.
Ask any marketer, and they’ll tell you the same thing: marketing performance changes constantly. The truth is, everything about marketing is dynamic:
- New creative launches influence how consumers respond to ads.
- Consumer attitudes shift with trends, economic conditions, and societal changes.
- Competitive pressures can alter platform dynamics and bidding strategies.
- And that’s just the beginning – platform delivery patterns, seasonality, and countless other factors play a role.
If your MMM assumes that marketing ROI is static (as most open-source tools do), it’s not just a little off –it’s dramatically wrong. And static models can lead to millions of dollars in misallocated spend. Period. End of story.
This is a huge issue that your MMM needs to get right. This article will cover how Recast approaches time-varying ROIs, but let’s give some context first:
A Historical Perspective
In the 1980s, marketing performance was far more stable:
- Creative was often static—one set of ads for an entire campaign.
- Media buys were planned months in advance and executed without major changes.
- Consumer behavior was more predictable.
Back then, time-varying ROIs weren’t a problem, because performance didn’t change as frequently.
But in 2024, marketing is a completely different game: new creatives are launched daily, competitors dramatically shift their bidding strategies, platform delivery patterns change, consumer attitudes shift…
Again, you should be thinking about this.
What Happens When Models Get It Wrong
If your MMM doesn’t account for time-varying ROIs, it will produce bad results – which means you won’t be able to trust your model. At best, you’ll ignore it. At worst, you’ll follow its recommendations and waste your money.
For example, imagine a scenario where a creative’s performance starts to decline due to fatigue. A static ROI model might blame the channel instead of recognizing the need for fresh creative. The result? You pull the budget from a high-performing channel and misallocate it somewhere else.
This happens all the time. And here’s why:
The Complexity of Building Time-Varying ROI Models
Building models that account for time-varying ROIs isn’t easy—it’s one of the hardest problems we’ve tackled at Recast. Here’s what it took to make it work:
The first hurdle was figuring out how to model time-varying ROIs in a way that’s both accurate and statistically rigorous. This meant designing entirely new model specifications and working through questions like:
- How do we represent ROI changes over time?
- How can we ensure the model remains identifiable (i.e., the results are valid and interpretable)?
Time-varying ROI models are exponentially more complex than static models. Early versions of our models took 48+ hours to train because of the sheer amount of computation required.
This wasn’t something we could run on a laptop –we had to build an entirely new cloud infrastructure to handle it.
Even then, we had to continuously optimize the model at the underlying C code level to improve run times without sacrificing rigor. It was an enormous effort for all of our team.
This doesn’t add “five” additional points of complexity. It’s actually a “to the fifth power” more complex.
Why Recast Invested in Time-Varying ROI Modeling
So, why did we do it? Because we couldn’t not do it.
We believe every model that doesn’t get this right is wrong. And not just a little wrong –catastrophically wrong. For businesses spending millions on marketing, the cost of getting it wrong is enormous.
We know – you know – that marketing performance changes over time. Every marketer knows this, and marketing science backs it up. So, if we know this to be true, why are other models ignoring it?
The answer is that it’s hard to solve.
But we decided to tackle it anyway because the stakes are too high to ignore. Time-varying ROIs are not a “nice to have” feature – they’re a foundational “must have” feature if you want a platform that reflects the real-world dynamics of marketing performance.
Closing Thoughts
- Time-varying ROIs are not optional anymore. Models that ignore this are wrong and will waste millions of dollars on misallocated spend.
- Incorporating them into your modeling is not easy –it requires years of research, collaboration, and innovation for us at Recast. But it was worth it.
