Applying Machine Learning to Demand Forecasting: A Case Study in Industrial Supply Chain Manangement

The application of time series forecasting spans a wide range of domains, including weather, finance, energy, and healthcare. Regardless of the field, forecasting plays a vital role in optimizing resources and preparing for future challenges. One domain where forecasting is particularly critical is supply chain management, where it is frequently used for demand forecasting. Demand forecasting serves as the foundation for all supply chain planning and helps balance inventory levels, reduce costs, and meet customer expectations. This thesis explores the application of machine learning to demand forecasting at a case company that has recently experienced greater fluctuations in demand and increasing variations in customer behavior. The limitations of the current approach have prompted the company to explore alternative methods, one of which being machine learning. The thesis involves two experiments forecasting demand for two product families, A and B, using historical demand data. The experimental setup was designed to reflect the company’s current forecasting conditions. The process included data collection, preprocessing, analysis, feature engineering, model development, and model evaluation. Three machine learning models were developed: Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM). Model performance was evaluated using Mean Square Error (MAE), Root Mean Square Error (RMSE), and Symmetric Mean Absolute Percentage Error (SMAPE), and compared against a 12-month moving average baseline. The results show that LightGBM and XGBoost outperformed the baseline for Product Family A, while none of the models outperformed the baseline for Product Family B. The thesis concludes that further investigation and refinements are needed before the models can be deployed into production.