A pricing model can predict your sales almost perfectly and still tell you to set the wrong price.
That sounds like a contradiction until you see that forecasting sales is one job and knowing what a price change will do is another. Most pricing models are only built for the first.
Why does my pricing model forecast sales well but give bad price recommendations?
Because the model learned when you raised prices. It never learned what the price did.
Companies raise prices when demand is strong. That is sensible management. But look at what it leaves behind in your historical data: high prices sit next to high sales. Again and again, the two show up together.
A predictive model learns exactly that pattern. And the pattern is real, which is why the forecast is good. As long as you keep pricing the way you always have, high prices will keep appearing next to high sales, and the model will keep being right.
The trouble starts when you use the same model for a decision. Ask it what happens if you raise the price, and it answers from the only pattern it knows. High price, high sales. It gets the price response wrong, and if you use it to decide on a price increase, it will mislead you.
So you own a model that passes every accuracy check and fails the one question you bought it for.
What is the difference between a predictive and a causal pricing model?
A predictive pricing model forecasts sales from patterns in past data. A causal pricing model estimates what happens to sales when you change the price. Three terms make the difference clear.
What is a predictive model?
A predictive model is built to forecast an outcome from patterns in historical data. It is judged by how close the forecast lands. For that, it is enough that a pattern repeats, whatever the reason behind it.
What is a causal model?
A causal model is built to estimate what changes in the outcome when you change one thing. In pricing that means: what happens to sales if you move the price, with the other influences kept apart from it. That is the question behind every price decision.
What is confounding in pricing data?
Confounding means that one factor drives both the thing you decide and the result you measure. In pricing data that factor is demand. Strong demand makes you raise the price. Strong demand also lifts sales. Price and sales rise together, and that correlation alone doesn't prove that one caused the other.
So the same data can be read in two ways:
- "We raised prices because demand was high."
- "Customers buy more because the price is high."
A forecast works the same under either reading. A price increase depends on which one is true.
Causal methods can separate the two readings, provided the data and the assumptions hold. I mean that condition seriously. A causal method works when your data carries the information and when the assumptions behind the model are true for your market. Anyone who sells it to you without that sentence is selling you something else.
If the mechanism sounds familiar: it is the same trap as in the sunscreen paradox. The correlation is real, and the conclusion still points the wrong way.
How to tell if correlation means causation in your pricing data
Ask why the price moved. If prices went up because demand was up, then "high price next to high sales" tells you how your company sets prices. How customers react to them is still an open question.
A correlation between price and sales is a fact about the past. A price decision needs a statement about cause. To get from one to the other, something has to keep the reason for the price change apart from the effect of the price change. A predictive model has nothing in it that does this. That separation is the job causal methods were developed for.
From Yule in 1897 to Causal AI: why business is decades late
The problem is old. So is the work on it.
In 1897 George Udny Yule was working out multiple regression. He did it because he wanted to separate the effect of one variable from the effects of others. That was an early step toward causal analysis.
Econometrics built on it, six decades later. Structural equation modelling took off in the 1960s and 70s. Pearl and others formalised causal inference in the 1990s and 2000s, and that work is what I and others now call Causal AI.
I see the same pattern again and again. Early science moves. Mainstream science follows later. And business needs many decades before it starts using what is already there.
Pricing shows why that delay costs money. The mechanism above is the bill: a company decides on a price increase with a model that reads the data the wrong way round, while the methods to read it correctly have been sitting in the research.
By the time a method shows up in textbooks, training courses and consulting decks, research usually has something better for your specific problem. Most of your competitors wait for the textbook version. If you work with what research offers today, you can be a decade ahead of them on the decisions that move margin.
And the tools have improved a lot in the last few years. You can now estimate what a decision will do before you put budget behind it.
If you buy what's "accepted", and what AI tells you when you simply ask it, then you are decades behind. I wrote about that reflex in Consensus Is the Riskiest Strategy.
Checklist: does your pricing model predict, or does it separate causes?
You don't need to be a statistician for this. Take these seven questions to whoever built or sold you the model.
- What was the model built for: to forecast sales, or to estimate what happens when we change the price?
- How was it tested? If the only proof is forecast accuracy, you know that it forecasts. The price response is still untested.
- Why did our prices change in the past? If we raised them when demand was strong, that habit is in the data.
- Does the model know those reasons? What in it keeps "demand was high" apart from "the price was high"?
- Does it ever suggest that customers buy more because the price is higher? Then ask which of the two readings it has learned.
- Which assumptions does the price response rest on? Can the team name them in plain words? If nobody can, nobody has checked whether they hold.
- How old is the method? Is it the textbook version, or what research offers today?
If the honest answers to questions 1 and 2 are "forecast" and "accuracy", keep the model for forecasting, and for forecasting only.
I have written before about why pricing needs causal understanding and what decision intelligence for pricing requires. For your own price range, that work happens in our Pricing Lab.
One question to take into your next pricing meeting: is the number on the slide a forecast, or an answer to "what happens if we raise the price"?
Predictive vs. causal pricing models: frequently asked questions
Why does my pricing model forecast sales well but give bad price recommendations?
Because forecasting and deciding are different jobs. Companies raise prices when demand is strong, so in the historical data high prices sit next to high sales. A predictive model learns exactly that pattern. It forecasts well, gets the price response wrong and will mislead you if you use it to decide on a price increase.
How to tell if correlation means causation in market data?
Ask why the variable moved. In pricing, prices often went up because demand was strong. Then the correlation between high prices and high sales describes how the company sets prices and leaves open how customers react. Causal methods can separate "we raised prices because demand was high" from "customers buy more because the price is high", provided the data and the assumptions hold.
What is the difference between a predictive and a causal pricing model?
A predictive pricing model forecasts sales from patterns in past data and is judged by forecast accuracy. A causal pricing model estimates what happens to sales when you change the price, with the other influences kept apart. The first can be accurate and still wrong about the price response, because of confounding: demand drives both the price and the sales.
Can I estimate the effect of a price increase before I roll it out?
Yes, with causal methods, provided the data and the assumptions hold. Dr. Frank Buckler points out that the tools for this have improved a lot in the last few years, so you can now estimate what a decision will do before you put budget behind it. A purely predictive model is the wrong tool for that question, however well it forecasts.
Dr. Frank Buckler advises leadership teams on growth decisions that depend on why customers choose. He founded SUPRA, holds a PhD in marketing and was among the first to apply causal AI to marketing questions. His latest book is THE TOP 5%.