Predicting Modeling in E-Commerce Marketing Based on User Journeys

Traditional marketing attribution models evaluate past campaign performance but offer limited input for future actions. This thesis bridges that gap by developing predictive models that use actual user path data to forecast revenue and guide future campaign decisions. We address three key questions: (1) predicting revenue from completed marketing paths, (2) identifying the most effective campaign to convert a mid-journey user, and (3) recommending the next-best campaign step for revenue maximization. Using e-commerce user path data from the Google Merchandise Store, we apply Random Forest, XGBoost, and ensemble techniques across various user journey segments. Our analysis demonstrates high predictive accuracy in revenue forecasting and mid-journey campaign identification. While the third model provides directional insights, its use of proxy labels presents methodological limitations that are thoroughly discussed. This research concludes with recommendations for future work in real-time personalization, counterfactual modeling, and the operational deployment of predictive marketing systems.