What is uncertainty in marketing models?
A model that says a channel returns a ROAS of 2.6 is giving its best single guess. Behind that guess is a range of values the data can't rule out, maybe 1.5 to 3.8. That range is the model's uncertainty, and it's usually the most useful thing a model can say.
Marketing models carry a lot of it. They have limited history, channels that move together, and sales that swing for reasons the model can't see. Showing the range keeps a precise-looking number from being treated as a fact.
What are the types of uncertainty, and how are ranges built?
| Type | Where it comes from | Can it shrink? |
|---|---|---|
| Epistemic | Not enough data or knowledge about the effect | Yes, with more data, tests, or better priors |
| Aleatoric | Real randomness in what customers do | No, it's part of the world |
| Model | The choice of model form and inputs | Partly, by comparing several plausible models |
| Range | Comes from | How to read it |
|---|---|---|
| Confidence interval | Classical statistics | A method that captures the true value in a set share of repeated samples |
| Credible interval | Bayesian models | The range that holds the true value with a stated probability, given the model and data |
| Prediction interval | Forecasts | Where a future outcome, such as next month's sales, is likely to land |
| Conformal prediction | Any model, checked on held-out data | A range that holds a stated coverage rate, under mild assumptions |
Calibration is the test of a range. If a model's 80% intervals contain the actual result about 80% of the time, they're calibrated. If they contain it half the time, the model is overconfident.
What is an example of model uncertainty?
Two channels with the same estimate
A model gives search and online video the same ROAS estimate of 2.8. Breakeven is 2.5.
Search's range is 2.6 to 3.0, entirely above breakeven, so the team funds it. Video's range is 1.4 to 4.2, which could mean losing money or doing very well. Instead of scaling it, the team runs a geo-lift test, and the next model run narrows video's range with real evidence.
Point estimate vs. range
| Point estimate | Range | |
|---|---|---|
| Says | The single most likely value | Every value the data can't rule out |
| Feels | Precise | Honest |
| Good for | Ranking options when ranges are narrow | Deciding when a lot of money is at stake |
| Risk | False confidence | Being ignored if it isn't explained |
Why does uncertainty matter?
Budgets get moved on single numbers that the model itself isn't sure of. Showing uncertainty changes the question from "which channel has the highest ROAS?" to "which channels are we confident pay back, and which do we need to test?"
- Better calls. Fund what clears breakeven across its whole range, and cut what falls short across its whole range.
- Smarter tests. Spend test budget where the range is widest and the money is biggest.
- Steadier trust. When a refresh moves a number inside its old range, nobody needs to panic.
How does Alli do this?
Alli by PMG's Scenario Planner is built for planning under uncertainty. It gives modeled estimates with the assumptions in view, so teams can compare options and defend the choice before money moves.
Every forecast blends a media response model built on historical sales with a brand equity model built on current signals, and two-stage estimation is designed to reduce the instability common in traditional models, so forecasts stay steadier as data changes.
Explore Alli Scenario PlannerUncertainty in marketing models FAQ
What is the difference between a confidence interval and a credible interval?
A confidence interval comes from classical statistics and describes how often the method would capture the true value. A credible interval comes from Bayesian models and gives the probability that the true value lies in the range, given the model and data.
What is model calibration?
A check on whether a model's stated uncertainty is honest. A calibrated model's 80% intervals contain the actual outcome about 80% of the time.
What is the difference between epistemic and aleatoric uncertainty?
Epistemic uncertainty comes from limited data or knowledge and can shrink with more evidence. Aleatoric uncertainty is real randomness in outcomes and can't be removed.
How do you reduce uncertainty in a marketing model?
Add more varied data, calibrate with experiments such as holdout or geo-lift tests, use informed priors, and avoid asking the model about spend levels far outside its history.
Last updated Back to the glossary