Why 90% of Price Models, MMMs and Driver Analyses Are Wrong

Machine Learning is built to predict. Causal AI is built to explain. Confuse the two, and your elasticities, media ROIs and brand drivers are biased — and nobody in the room can smell it.

If you do price modeling, marketing mix modeling or driver analysis, chances are 90% that you do it wrong.

Ninety percent. That's not a throwaway line — it's the estimate that emerged when we ran a deep AI research sweep across how these models are actually built in practice.

How can that be? How can reputed vendors, expensive consultancies and smart data science teams get it wrong at that scale?

Because of a simple, uncomfortable mechanism: you find the results plausible enough to believe them, and nobody runs a Randomized Controlled Trial to prove them wrong.

Plausibility is the perfect camouflage. A biased model doesn't announce itself. It produces numbers that fit the story you already carry in your head. The deck gets approved. The budget gets shifted. And the error compounds, invisibly, quarter after quarter.

The Brown Paint Metaphor

Here is the whole problem in one picture.

Imagine a painting. It was painted using brown ink — maybe a bit of black too. But here's the detail that matters: the brown jar was previously mixed from a green jar and a red jar.

Now hand the painting to a Machine Learning algorithm and ask it to reproduce the picture using the colors available on the table: brown, red, green, black.

Machine Learning is designed to create a similar painting to the original. Predictions — that's its job. So it will happily use brown, but also red and green, mixing them right at the screen. Why not? Red and green blend into something brown-ish. The reproduction looks great. The prediction error is tiny.

But ask the algorithm how the painting was made, and it will tell you a false story: "The artist used brown, red and green."

Causal AI is designed differently. It is built to explain. It tells you: the painting was painted using the brown jar. Full stop. Red and green only exist inside the brown — they never touched the canvas.

And here is the beautiful side effect: once you have the true explanation, your future predictions get more accurate too. Because now the model uses the true mechanics of the world — not a coincidental color mix that happens to work on this one painting.

Prediction is a side product of explanation. It never works the other way around.

Now let's find the brown paint jar in your models.

Price Modeling: The Jar Is the Price Itself

In price modeling, the brown paint jar is the price variable.

Prices are not set independently. No company rolls dice to set them. Prices follow a calculation — costs, competitor moves, demand expectations, promotion calendars. The price you observe in your data is already a mixture: a blend of market conditions and management reactions to those conditions.

Feed that mixed variable into a standard model and the model can't tell what causes what. The result: derived price elasticities are biased — often too small.

Think about what that means. An elasticity that looks smaller than it is whispers a seductive message: "Go ahead, raise the price, demand barely reacts." Or the mirror image: it understates how much a smart price move could actually earn you. Either way, you're steering with a bent compass.

Marketing Mix Modeling: The Jar Is Media Spend

In marketing mix modeling, the brown paint jar is media spend.

Spends are not set independently either. They follow a regime. You know how it goes: "Brand awareness is down? We need TV ads." Sales soft in Q3? Push more performance budget. Big season coming? Everyone spends more.

So spend and outcomes dance together — but not because one purely causes the other. The regime that sets the spend is reacting to the very outcomes the model is trying to explain.

A conventional MMM reads this dance as effectiveness. The result: derived media dollar impacts are biased — often too high. You conclude TV "works" partly because you always turn TV on exactly when something else is about to move.

Then you reallocate millions based on that number.

Driver Analysis: The Jar Is Perception

In driver analysis, the brown paint jar is brand perceptions — or satisfaction with certain aspects.

These are the least independent variables of all. Perceptions influence each other. "Trust" colors "quality." "Quality" colors "value for money." All of them are pushed around by context, category habits, brand power and more.

Run a naive driver analysis on this tangle and the derived impact strengths come out biased — often implausibly so. You've seen it: the model tells you that satisfaction with service has a negative impact on loyalty. Everyone in the room laughs nervously, the analyst mumbles something about multicollinearity, and the slide gets quietly removed.

The slide was the honest one. It was showing you that the method is broken.

This is exactly the problem Universal Structure Modeling was built to solve: a Causal AI approach that models how drivers influence each other — and only then reads off what truly moves the outcome.

Why Nobody Notices

Here's the part that should worry you most.

Most data "scientists" have no clue about any of this. Not because they're lazy — because they were trained on prediction. Their toolkits, their benchmarks, their Kaggle instincts all optimize one thing: reproduce the painting. The question "which jar did the artist actually use?" never appears in the curriculum.

And if the specialists can't see it… how can we expect CEOs and CMOs to smell that something is off?

It's hard. Genuinely hard. You need to trust your team.

But — and this is the point — it's your responsibility to gather the right team.

Nobody ever got fired for buying the big-name vendor. That's precisely why the big-name vendor doesn't have to be right.

The Questions to Ask Your Modeling Team

You don't need to become a statistician. You need three questions.

One: "How does your model account for the fact that prices, spends and perceptions are not set independently?" If the answer is a blank stare or the word "controls," dig deeper.

Two: "How was this model validated — against a holdout, an experiment, anything beyond plausibility?" A model that has only ever been validated by nodding heads has never been validated.

Three: "If this number is wrong by half, what decision changes?" This one isn't for the analyst. It's for you. It tells you how much truth the decision actually requires.

The vendors won't volunteer this conversation. A Randomized Controlled Trial that could falsify their model is not in their commercial interest. It should be in yours.

Machine Learning reproduces the painting. Causal AI finds the jar. Your price elasticities, your media ROIs, your brand drivers — they all depend on which of the two your team is actually doing.

Demand explanation, not just prediction. This is how you 10x your insights.

Marketing models and Causal AI: frequently asked questions

What is the difference between Machine Learning and Causal AI?

Machine Learning is designed to predict: it reproduces outcomes from whatever correlated inputs it finds, even if those inputs are proxies rather than causes. Causal AI is designed to explain: it uncovers the true mechanism that generated the data. SUPRA uses Causal AI because a model that captures the real mechanics of the market delivers unbiased price elasticities, media impacts and brand drivers — and, as a side product, more accurate predictions too.

Why are most price elasticity estimates biased?

Because prices are not set independently. Companies set prices through calculations that respond to demand, costs and competition, so the price variable in your data is entangled with the very outcomes it is supposed to explain. Standard models mistake this entanglement for causation, and the derived price elasticities come out biased — often too small. Causal AI models the price-setting mechanism explicitly, which is how SUPRA recovers elasticities you can actually act on.

Why do marketing mix models overstate media impact?

Media spend follows a regime, not a random schedule. When brand awareness dips, budgets shift to TV; when sales look strong, spend often rises with them. A conventional marketing mix model reads this feedback loop as media effectiveness, so the derived dollar impacts are biased — typically too high. A causal marketing mix model separates the spend-setting rule from the true media effect, which is the correction SUPRA builds into its MMM work.

How can a CMO detect a flawed driver analysis?

Look for three warning signs: driver impacts that flip negative for attributes that obviously matter, results that merely confirm what everyone already believed, and vendors who validate on plausibility instead of holdout prediction or experiments. Brand perceptions influence each other and share common causes, so naive driver analysis is structurally biased. SUPRA addresses this with Universal Structure Modeling, a Causal AI approach built to untangle interdependent drivers.

Dr. Frank Buckler is the founder of SUPRA and a pioneer in Causal AI for marketing. He has applied implicit research methods across FMCG, pharma, financial services, and insurance for over 25 years.