Stacking Ensemble Revenue Predictions in Digital Marketing: A SHAP-Based Analysis of Price and Quantity as Key Predictors

Abstract

This study focuses on developing a digital marketing conversion prediction model using an ensemble stacking approach combined with explainable artificial intelligence (XAI) methods to improve model transparency. The primary objective of this study is to investigate the impact of price and product quantity on revenue predictions, as well as to gain a clearer understanding of the factors that influence customer purchasing behaviour in the context of digital sales. The methodology used includes data collection from a Kaggle dataset containing 3,000 records and 15 features related to customer demographics, product information, and marketing channels. The preprocessing stage ensures data quality, followed by feature engineering and model development using an ensemble stacking model consisting of Logistic Regression, Gaussian Naïve Bayes, and Support Vector Classification. Model evaluation was conducted using precision, recall, F1-score, and ROC-AUC metrics, with performance improvements achieved through cross-validation and probabilistic calibration. The study results showed that model accuracy reached 0.97, with significant contributions from price and product quantity features, as seen in the SHAP analysis. The ensemble stacking model provided stable and reliable predictions. These findings underscore the importance of effective pricing strategies and product volume optimisation in driving revenue growth. The use of SHAP enhances interpretability, enabling businesses to make more informed decisions. This research contributes to the development of transparent and practical machine learning applications in digital marketing, providing valuable implications for business strategy optimization.