Why traditional MMMs fall apart during peak sales

Try to pick out one singer's voice when two people are hitting the exact same note, at the exact same volume, in perfect unison. It doesn't matter how good your ear is. You can't separate what you can't tell apart.

That's essentially what happens to a traditional marketing mix model (MMM) during Black Friday, a big holiday push, or a major product launch. The model is trying to hear two different signals, ad spend and natural customer demand, and during peak periods, those two signals hit the exact same note at the exact same time.

And, yes, it hits hardest when the pressure is already on for marketing teams. Peak season is exactly when budget decisions carry the highest stakes, since that's when the most money is moving and the cost of getting the split wrong is at its highest.

Key takeaways

• Standard MMMs try to split revenue into two pieces, baseline demand and media effects, but during peak periods, ad spend and organic demand spike at the same time, making it mathematically impossible to tell them apart.
• This creates baseline leakage, where the model can overstate organic holiday lift by 20% to 30% while misreading how well ads are actually working.
• Adding more historical data doesn't solve this, because it's a flaw in how the model is built, not a sample-size problem. More data just makes the model converge more confidently on the wrong answer.
• Statistical fixes like Ridge regression or Bayesian priors don't restore accuracy either. They just force the model to pick one arbitrary answer out of many equally plausible ones, and different prior settings can produce very different results from identical data.
• In benchmark testing, following a standard Bayesian MMM's peak-season budget recommendation cost businesses 45.1% of achievable revenue, and a Ridge regression model cost 32.2%.
• These models led to extreme misallocation: one overspent on saturated channels by around 81%, while another underspent by around 11.5%, leaving profitable growth on the table.
• Prescient’s model gave up just 5.6% of achievable revenue in the same test, landing within about 1% of the truly optimal spend plan.

When spend and demand spike together

Standard MMMs try to decompose revenue into two separate pieces: baseline (trend, seasonality, and holidays) and media effects (the conversions ad spend is driving). That works reasonably well when spend and demand move somewhat independently of each other.

The trouble starts at peak season, and it starts on purpose. Marketing leaders intentionally scale ad spend right as customer demand is naturally spiking too. When those two things move together this closely, the model hits a wall called structural non-identifiability: because spend and baseline demand are so strongly correlated, there are infinitely many ways to split the revenue between them that all fit the data equally well. The model genuinely can't tell whether sales rose because the ads worked or because customers were going to buy anyway for the holiday.

The result is baseline leakage. The model mistakenly credits natural holiday demand to ad spend, or the other way around, and in practice, that means overstating organic holiday lift by 20% to 30% while getting true ad effectiveness wrong in the process.

Why more data and statistical tweaks don't fix it

When an MMM starts spitting out unstable or implausible peak-season numbers, the standard advice is to throw more historical data at it or apply statistical penalties like Ridge regression or Bayesian priors. Neither one actually solves the problem:

• More data doesn't help. Non-identifiability is a flaw in the model‘s structure, not a sample-size issue. Adding more historical data reduces the model's estimation variance, but it just converges more confidently around the wrong answer that its underlying assumptions favor.
• Regularization doesn't restore identifiability. Statistical penalties and Bayesian priors don't fix the underlying ambiguity. They just force the model to land on one arbitrary solution out of an infinite pile of indistinguishable options, which is why different prior settings can produce wildly different attributions from the exact same historical data.

This costs brands up to 45% of revenue at peak times

None of this stays theoretical once those numbers feed into a budget allocation tool. Benchmark testing measured model-recommended peak-season budgets against the true optimal spend plan, and the gap was significant.

• Real revenue lost. Following a standard Bayesian MMM's budget recommendation (Baseline B) caused businesses to give up 45.1% of achievable revenue during peak demand. A Ridge regression budget (Baseline A) still lost 32.2%.
• Severe misallocation. One standard model overspent on already-saturated channels by around 81%, wasting capital that wasn't generating additional return. Another underspent by around 11.5%, leaving profitable growth on the table.
• Our model’s difference. In the same test, our model gave up just 5.6% of achievable revenue and landed within roughly 1% of the optimal spend allocation, holding up even as budgets scaled rapidly during peak demand.

Where Prescient comes in

Peak season is the worst possible time to be working off a model that's guessing at the split between your ads and your organic demand. Our model's system-based architecture is built to stay stable exactly when spend and demand are moving together the fastest, so the budget recommendations you're acting on during your biggest revenue weeks are actually grounded in what's happening, not in which arbitrary answer the math happened to settle on.

If you want to see how the platform helps you strategize an optimal plan, book a demo and we'll walk through it.

FAQs

What is baseline leakage in a marketing mix model? 

Baseline leakage happens when a model mistakenly credits natural, non-marketing-driven demand (like holiday shopping) to ad spend, or the reverse. It's especially common during peak periods, when ad spend and organic demand rise at the same time, making the two nearly impossible to separate with a standard model.

Why can't adding more historical data fix peak-season attribution? 

Because the issue isn't a lack of data. It's a flaw baked into the model itself called non-identifiability, where multiple different attributions fit the historical data equally well. Adding more data makes the model more confident, but it doesn't help the model choose the right answer over a wrong one that fits just as neatly.

Do Bayesian priors or Ridge regression solve this problem? 

No. Both are common fixes vendors suggest, but neither restores true identifiability. They just force the model to pick one answer out of many equally valid ones, which is why changing the prior settings can produce very different attribution results from the exact same underlying data.

How much revenue is actually at stake from bad peak-season budget decisions? 

In benchmark testing, following a standard Bayesian MMM's peak-season budget recommendation cost businesses 45.1% of achievable revenue, and a Ridge regression model cost 32.2%. Both led to real misallocation, including significant overspending on already-saturated channels.

How does Prescient’s model avoid these peak-season pitfalls? 

Our model’s state-space architecture is designed to stay stable even when ad spend and organic demand move together, which is exactly when standard models break down. In the same benchmark test, our model gave up just 5.6% of achievable revenue and landed within about 1% of the optimal spend plan.