Comparative Predictive Analytics Framework for Demand Forecasting in Online Retail

Abstract

This research aimed to develop and evaluate a comparative predictive analytics framework for demand forecasting in online retail. Utilising an online retail sales dataset, the study followed the CRISP-DM framework to implement and compare 7 forecasting models, including Linear Regression, Random Forest, Gradient Boosting, SARIMAX, and hybrid approaches. The methodology involved data preprocessing and extensive feature engineering of lag variables and rolling averages. Experimental results revealed that Linear Regression achieved the highest predictive accuracy and strongest explanatory power, followed by Gradient Boosting. Statistical and baseline methods demonstrated significantly lower performance. The findings suggested that simpler, interpretable feature-driven machine learning models can outperform more computationally complex architectures for structured retail data. This study concluded that high-quality feature engineering is more critical for accuracy than model complexity. The proposed framework provided practical value for businesses seeking data-driven solutions for inventory optimisation and demand planning in dynamic e-commerce environments.