What is a confidence range?

A Bayesian model does not produce a number for a channel's effect; it produces a distribution. The confidence range (statisticians say credible interval) is the slice of that distribution the model considers plausible, usually the central 90%. "Most likely 3.2, probably between 2.4 and 4.1" is a complete answer; "3.2" is the same answer with the uncertainty removed.

A narrow range means the history had a lot to say about that channel: plenty of variation in spend, clear movement in sales. A wide range means the data cannot settle the question, typically because the channel has always run at the same level, has little spend, or runs at the same time as a larger channel. Neither is a flaw in the model. The wide range is information: it says where the model would learn most from a test.

Ranges propagate. A forecast for a budget scenario carries a range because every channel effect behind it does, and the ranges around marginal returns are what make a reallocation recommendation a recommendation rather than a certainty.

The range is what makes a target plannable at a chosen level of certainty. If a quarterly revenue target has to be hit, plan for the budget that puts the target at the cautious end of the forecast range, not in the middle. That budget is higher, and the gap between the two is the price of certainty. A question Odins customers ask the model directly: what does it take to be 80% sure of reaching the target?

Odins reports a range on every recommendation and every scenario, and where the range is too wide to act on, the recommendation is a structured test rather than a budget move.

• Posterior: The model's answer: the prior after the data has had its say, with a most likely value and a range.
• Bayesian marketing mix modeling: A model that starts from what you already know and lets the data correct it, with a range on every answer.
• Scenario planning: Forecasting what a budget would return before committing it, and comparing several futures side by side.
• Pulse test: Vary one channel's spend over time, on and off or up and down, so the model can see how it responds.
• Forecast verification: Forecast first, compare with actuals afterwards. The only honest way to know whether a model is right.

Go one level deeper: How much marketing to hit a revenue target with 80% certainty. Or go back to all terms in the glossary.