Umeå universitets logga

Forecasting future demand is a central challenge in supply chain planning and business decision-making, where inaccurate forecasts may lead to inefficient inventory management, increased costs and difficulties in production planning. This thesis investigates whether machine learning-based forecasting models can generate more accurate and systematic forecasts compared with the manual sales budgeting process currently used by the dairy company.

The study was based on historical sales data for two different product subgroups. A com prehensive feature engineering process was conducted, where lagged variables, rolling statistics, cyclical encodings and holiday-related variables were used to capture temporal dependencies and seasonal effects. Several machine learning models were evaluated, including Multiple Linear Regression, Elastic Net, Random Forest, XGBoost and CatBoost. Both a standard forecasting approach and a recursive forecasting approach were investigated.

The results showed that machine learning models can generate more accurate forecasts than the company’s current manual budgeting process. All evaluated machine learning models consistently outperformed the manual budgeting process for both product subgroups. The results further showed that relatively simple and interpretable modelsperformed competitively despite their less complex structure.