Adstock in Marketing Mix Models: Carryover and Decay

Adstock is a modeling transformation that represents the idea that advertising can influence outcomes after the original exposure period. A campaign may create awareness or consideration that persists, so a model may distribute some of a period’s media input into later periods rather than treating each week as isolated.

A simple adstock equation

One common geometric form is Aₜ = Xₜ + λAₜ₋₁, where Xₜ is media input in the current period and λ is a decay parameter between 0 and 1. If λ is 0, the transformation carries no prior input forward. A larger λ implies a slower modeled decay. This is a modeling choice, not a directly observed law of consumer behavior.

Why the transformation matters

Without a carryover term, a model may attribute delayed outcomes to the wrong period or underestimate a channel whose effects persist. With an unsuitable decay assumption, it can instead spread effects too far and create misleading estimates. The appropriate structure depends on the channel, purchase cycle, data frequency, campaign pattern, and model specification.

• Use a time interval that fits the data and buying cycle.
• Estimate or compare decay assumptions with transparent validation, not intuition alone.
• Check whether media inputs and outcomes are aligned to the same calendar and geography.
• Model saturation separately; carryover and diminishing returns describe different behaviors.

Interpret adstock estimates cautiously

Adstock values are not a direct measure of how long an individual remembers an ad, and a fitted parameter should not be generalized to every campaign or market. Media channels often have different exposure patterns, while overlapping campaigns and seasonal demand can make carryover difficult to identify.

Reading a decay parameter

With an illustrative λ of 0.5, half of the transformed prior-period input carries into the next period under this simple equation; subsequent periods carry progressively less. That arithmetic helps explain the transformation, but the value is not a recommended default. A model should test plausible alternatives and show whether channel contribution estimates change materially when the decay assumption changes.

Report the transformation, decay assumptions, data interval, and sensitivity checks alongside the model result. Compare predictions with held-out periods or other validation evidence where possible. Adstock can help a marketing mix model represent persistence, but it does not by itself establish causality or prove that a specific channel deserves more budget.