A systematic benchmarking pipeline for demand forecasting and inventory-aware supply chain decision support

Demand forecasting has been shown to be one of the key facilitators of successful supply chain and inventory management. In this work, we introduce a benchmarking pipeline for demand forecasting in retail that includes data preprocessing, comparative benchmarking of different machine learning/deep learning (ML/DL) approaches, and decision-focused evaluation. Unlike other works that compare predictive performance of ML/DL models and end there, our pipeline links reproducible ML/DL benchmarking and decision-focused evaluation through cost of inventory management in order to derive costs from forecasting error/bias. We also present a scalability study on another retail dataset. Several standard forecasting models, such as Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), CNN-LSTM, and XGBoost, are compared using a unified experimental pipeline on benchmark retail datasets. The experimental results reveal that while the deep learning models efficiently identify temporal dependencies, tree-based ensemble algorithms, especially XGBoost, consistently deliver better results in terms of forecasting accuracy on structured retail demand datasets. A scalability case study on the Rossmann Store Sales benchmark evaluates the top-performing model on a larger, richer dataset. Moreover, the study shows the connection between forecast accuracy and the subsequent inventory performance by performing a cost-based evaluation of inventory outcomes of forecasts made using various models.