Enhancing Supply Chain Forecasting with Machine Learning

Abstract

Accurately forecasting unconstrained demand is essential for effective supply chain planning and inventory management. This thesis examines the performance of both traditional statistical methods and modern machine learning models in predicting unconstrained demand, with a particular focus on comparing global and local modeling strategies across multiple products. Using real-world sales data, models such as ARIMA, XGBoost, and LightGBM were evaluated based on their ability to produce accurate forecasts.

The comparative analysis of forecasting models across multiple products reveals that model performance varies significantly depending on product characteristics and the chosen modeling approach. Overall, modern machine learning approaches, particularly LightGBM and XGBoost, consistently outperformed the traditional ARIMA benchmark in terms of predictive accuracy. However, their effectiveness varied by modeling approach and product type.

The results indicate that local models, especially those based on LightGBM, tend to perform better for products with lower demand, as they are more capable of capturing product-specific patterns and fluctuations. In contrast, global models show stronger performance for high-demand products, where shared patterns across products can be more effectively leveraged. Although global models may not always match the accuracy of well-tuned local models, they offer significant practical advantages, including simplified model management, reduced training time, and better scalability. These findings highlight the trade-offs between accuracy and efficiency, and provide valuable insights into when global forecasting models may serve as viable alternatives to individualized local models.