Enhancing Predictive Models in E-Commerce: A Comparative Analysis using LightGBM, CatBoost and Prophet for Sales and Demand Forecasting

Abstract

The Hasty growth of e-commerce platforms has increased the need for accurate predictive models to forecast sales, understand customer demand, and optimize inventory management. This study presents a comparative analysis of three advanced predictive approaches LightGBM, CatBoost, and Prophet using the Amazon Sales Dataset to enhance sales and revenue forecasting in an e-commerce environment. The dataset consists of product-level attributes such as category, pricing, ratings, and textual reviews, which provide high-dimensional structured information for predictive modeling and trend analysis. LightGBM and CatBoost are employed to capture complex relationships among product features and demand indicators, while Prophet is utilized to analyze time-series patterns, including seasonality, promotional spikes, and festival-based sales variations. The experimental setup is implemented using the Python ecosystem on a Windows-based system, with evaluation conducted through classification and forecasting metrics such as Accuracy, Precision, Recall, F1-score, AUC-ROC, RMSE, and MAPE. The findings demonstrate that accurate sales prediction models significantly improve inventory planning, demand forecasting, and peak-season preparedness. The models effectively capture seasonal fluctuations, discount-driven demand surges, and customer engagement factors, enabling e-commerce platforms to increase stock availability during high-demand periods and improve operational efficiency. This research highlights the importance of machine learning-driven forecasting in enhancing revenue prediction, optimizing resource allocation, and supporting strategic decision-making. This work analysis further shows that LightGBM and CatBoost provide strong performance in structured data modeling, while Prophet excels in identifying time-based trends and seasonal patterns.