AMC playbook: Media performance optimization
In short: Media performance optimization builds diminishing-returns curves from historical advertising data, projects expected performance at other budgets and calculates the budget distribution that maximises a chosen KPI. According to Amazon it is a performance projection, not a forecast of future returns.
What it is about
The curves are built with a non-linear least squares model on at least six months of history, emphasising recent data. The optimiser distributes the budget across the inputs with constrained non-linear optimisation.
The playbook uses historical data from the past year and weights recent data 10 percent more heavily through a time decay. Amazon stresses: the model is a projection based on past campaigns, not a time series forecast.
Which questions the playbook answers
- How do you analyse the cost to KPI relationship and how does it change at higher spend?
- How do you build diminishing-returns curves from it?
- How do you project custom spend levels to expected return?
- How do you determine the optimal budget distribution for a KPI?
Flow
- Data preparation: choose the dimensions and inputs of the model.
- Run the diminishing returns model.
- Run the optimisation model.
- Budget planning and visualisation.
Prerequisites according to Amazon
Our take
The curves are only as good as the history: campaigns that never had high budgets say nothing about high budgets. The model works as a basis for discussing budget shifts, not as a guarantee.
Amazon documents the playbooks in technical English with SQL and Python. This page summarises what the playbook does; the queries and scripts are in the original.