Retail Sales Forecasting Using Tree-Based Machine Learning Models: An Empirical Study of Preprocessing Configurations and Feature Ablation

Sales forecasting plays a critical role in business decision-making, including financial planning, resource allocation, and inventory management. Despite the growing adoption of machine learning in forecasting applications, controlled evaluations of how preprocessing decisions and contextual feature availability affect model performance remain limited. This study presents a controlled benchmarking evaluation that systematically examines preprocessing strategies and contextual feature availability across seven representative tree-based machine learning models, using two publicly available retail datasets: BigMart and Rossmann. Multiple preprocessing configurations are evaluated on both datasets, and ablation studies are conducted on the Rossmann dataset to quantify the independent contribution of seasonal and promotional features and the Customers variable to forecasting accuracy. Results show that GradientBoosting achieved the best performance on BigMart (R2 = 0.611), while RandomForest and ExtraTrees led on Rossmann R2=0.884 under realistic conditions excluding the Customers variable, and R2=0.965 in an ablation scenario where Customers, typically unavailable at forecast time, is included. Ablation analysis confirmed that seasonal and promotional features are critical drivers of forecasting accuracy, and that the Customers variable is the single most influential predictor in the Rossmann dataset. These findings highlight the importance of preprocessing design and contextual feature availability in retail sales forecasting and provide practical insights for building effective forecasting pipelines.