Hybrid machine learning forecasting for resilient and sustainable pharmaceutical supply chains under regulatory and seasonal disruption.

Introduction: Demand forecasting in pharmaceutical supply chains is not a simple task. In regulated markets it becomes more difficult, because seasonality, epidemic waves, and also policy changes can make demand behavior unstable. This study proposes a hybrid residual learning approach for forecasting pharmaceutical demand in Türkiye. Methods: The model uses Support Vector Regression (SVR) together with Deep Neural Networks (DNN). In this structure, SVR estimates the main or baseline part of demand, and DNN tries to learn the remaining nonlinear residual variation. The framework was tested with real ERP-based demand data, including 10 critical pharmaceutical products in a 24-month period. For evaluation, temporal holdout testing was used. The model was also compared with ARIMA, Prophet, Random Forest, XGBoost, and LSTM. Statistical validation was done by paired t-tests and Diebold-Mariano tests. Results: The proposed SVR-DNN framework performed better than the standalone SVR and DNN models. It also stayed as the strongest model when it was compared with the wider benchmark models. The hybrid model gave the lowest forecasting errors, and it showed the highest explanatory performance among the tested models. Discussion: The results show that when baseline demand estimation is separated from residual correction, forecasting performance can be improved in pharmaceutical demand conditions that are sensitive to disruptions. SHAP analysis also helped the interpretation of model, because it showed the main factors that affect demand variation. The proposed framework gives an interpretable and context-aware forecasting model for pharmaceutical supply chains working under regulatory and seasonal disruption.