Budget Optimization Forecast

[sonriks6]

Hey, was testing Meridian a lot, now that I got the basis with the modeling phase I need some support or recommendations about budget optimization.
Business asks about a budget proposal to plan next quarter for example, how could we forecast incremental KPI or revenue based on a given budget? Makes sense to select a previous period as reference?
My suggestion would be to input a planned budget and start_date / end_date in the near future then the BudgetOptimizer module to provide a recommended budget split across different channels.

[alexwakim]

Budget optimization for media spend in future quarters is not currently supported. You can approximate this by doing the following,

• Select some previous time period represented in your data that's the same duration as the future quarter (selected_times argument in BudgetOptimizer.optimize see here for more info).
• Keep fixed_budget=True (the default) in BudgetOptimizer.optimize
• Set budget arg in BudgetOptimizer.optimize to the planned future budget

This is an approximation of a forecast. The optimizer tells you how to optimize spend across media channels in a given time period. When the optimizer suggests more or less spend, it assumes that media units in each geo and time period are scaled proportionally to the total spend and allocated according to the historical flighting pattern across geos and time periods. The flighting pattern can be relevant if media spend is allocated to, say, really effective geos. The flighting pattern can also be relevant if lots of media spend happened during a short time period, leading to an over-saturated media channel. To summarize, the optimizer does not tell you how to spend across geos and time. It tells you how to spend across media channels.

Meridian also assumes lagged effects of media. If a media channel has lots of spend immediately before a given time period, and that media channel has a strong lagged effect, that media channel may already be over-saturated during the given time period. Technically, this paragraph is only true if hill_before_adstock=False (the default) in ModelSpec.

When selecting a previous time period, we recommend picking a time period that

• Has a similar flighting pattern to the flighting pattern you anticipate in the future quarter
• Had normal spend volume on all channels preceding the time period.

[sonriks6]

Yes, this solution worked for us, we could create several future budget scenarios selecting a quarter period in the past as reference; it's important to select a similar campaign we run the past and a fixed budget also in same range.