Problem
An advertiser building a new campaign (an "adset": ads, audience, target cost, budget and goal) wants to know before launch the displays, clicks, sales and cost it will get. The adset has no history, so its own past performance cannot be used. The metrics depend on each other (there are more visits than sales), so predicting them independently breaks the funnel. The older forecast used machine learning only for displays and per-country statistics for the rest, which lost accuracy on sales, cost of sales and return on ad spend (ROAS).
Example from the article: users of the Commerce Growth Media-Planner try as many setting combinations as they like and immediately see estimated delivery and cost, with no impact on running adsets.
Solution in brief
Split the forecast into a cascade of gradient-boosted models: displays first, then click-through rate (CTR), conversion rate (CR) and landing rate (LR), each applied to the previous metric. Train them on past delivery of all adsets. To make forecasts comparable across budgets and targets, fit a smooth increasing function to the model output and run the cascade on functions instead of single values.
Questions this case answers
- How to forecast the outcome of a new entity with no history of its own by generalizing from all running entities
- How to forecast several dependent funnel metrics so that they stay mutually consistent and monotonic in spend
- How to forecast delivery and cost of a new ad campaign that has no performance history?
- How to predict several linked funnel metrics (displays, clicks, sales) without training one oversized model?
Tools and methods
- XGBoost — gradient-boosted tree models for displays, CTR, CR and LR
- Cascade of metric models — each funnel metric is a rate applied to the previous metric
- Fitted elasticity function — smooth approximation of displays versus cost
Google or a one-time email code. No password.