Data-Driven Machine Learning Approaches to Demand Forecasting in Supply Chain Management

Abstract

Demand forecasting is a critical function in pharmaceutical supply chains, where inaccurate estimates may result in shortages, excess inventory, and financial losses. This thesis evaluates forecasting performance using statistical methods (Moving Average, Single Exponential Smoothing, ARIMA) and machine learning models (Decision Tree, Random Forest, XGBoost) applied to a pharmaceutical case study with two years of monthly data. The dataset was divided into 21 months for training and 3 months for testing, and model performance was assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The results show that statistical models consistently outperformed machine learning models. Moving Average, SES, and ARIMA achieved testing errors of around 18– 19% MAPE, supported by stable MAE and RMSE values. Machine learning models performed poorly without lag variables (MAPE > 320%), but showed improvements when lags were introduced. Among them, Random Forest and XGBoost performed best, yet their errors remained substantially higher (MAPE ≈ 234–237%) compared to statistical methods. While statistical approaches were more reliable under limited data conditions, their accuracy still lacks strong economic value due to the short historical period analyzed. Future research should therefore expand the dataset and incorporate richer explanatory variables to improve predictive performance and support decisionmaking in pharmaceutical supply chains.