End-to-End Data Analytics
Forecasting ROAS and success for the next 15 days — by campaign, product and creative — so budget moves before results slip.
15 days
Horizon
3
Granularities
Regression · k-NN · cohorts
Models
Overview
With Amazon Ads and sales data, I built predictive models that measure performance at three granularities — campaign, product (ASIN) and creative — and forecast how each will perform over the next 15 days of the month. The output is a dashboard that compares past ROAS with the prediction and flags what to scale, hold or reduce. Visuals use illustrative data; the structure mirrors the production model.
Scope of work
Feature engineering
Linear regression
k-Nearest Neighbors
Cohort analysis
Ensemble & backtesting
Forecast dashboard
01 · The challenge
Budget decisions were made after the fact — by the time a campaign, product or creative underperformed, the spend was gone.
02 · The solution
A feature store in BigQuery (rolling ROAS, pacing, days since launch, CTR/CVR trends, seasonality) feeding three complementary models: linear regression for trend and spend elasticity, k-nearest neighbors to predict new items from similar past ones, and cohort curves for expected ROAS by age — blended into a weighted ensemble validated with rolling backtests.
03 · Results
Screens
Next project
