Frugal machine learning for pharmaceutical SME supply chain resilience: a real empirical case study

Purpose This study addresses a critical gap in pharmaceutical supply chain management for small and medium-sized enterprises (SMEs) in developing economies: the absence of validated, context-appropriate and practically deployable demand forecasting tools. Investigate which machine learning approach provides the most accurate, interpretable and economically impactful forecasts and how it can be transformed into an operational decision support tool. Design/methodology/approach A two-phase case study was conducted with a Tunisian pharmaceutical SME using 484,245 transactional records (2017–2022), aggregated to 72 monthly observations per product. Six forecasting models were evaluated under strict temporal validation: autoregressive integrated moving average (ARIMA) (classical baseline), linear regression, recurrent neural network, long short-term memory (LSTM), Random Forest and XGBoost. All models accessed the same information set; neural networks used a six-month look-back window selected on the 2021 validation set. The best model was deployed in an open-source Streamlit web application integrated with Power BI dashboards. Three novel analytical layers were added: segmented performance analysis, feature importance extraction and economic impact quantification. Findings XGBoost achieved the best performance (R2 = 0.64, root mean squared error (RMSE) = 36,147), substantially outperforming LSTM (R2 = 0.12) and ARIMA (R2 = 0.38). The segmented analysis reveals R2 = 0.71 for wholesalers (83.84% of revenue) versus R2 = 0.44 for export clients. The Kruskal–Wallis test confirms statistically significant differences between models (H = 296.50, p < 0.001). Historical lag features dominate importance (48.9%). XGBoost delivers an estimated 483,710 MDT annual savings (−41.9%) versus manual forecasting, providing 10× the economic value of LSTM. Research limitations/implications The study covers a single SME in one therapeutic domain; multi-case replication is required. Economic estimates are indicative projections. Future work should address multistep forecasting, external demand drivers, ethical artificial intelligence (AI) implementation, SHAP-based explainability and formal usability evaluation. Practical implications SME managers should prioritise gradient-boosting ensembles over deep learning when historical data are limited, apply forecasts autonomously for high-reliability segments (R2 > 0.65) and invest in data quality over model complexity. The system is deployable at near-zero marginal cost using open-source tools. Social implications Improving demand forecasting in pharmaceutical SMEs reduces drug shortages and treatment delays in developing economies, contributing to the sustainability of the health system and equitable access to essential medicines. Originality/value To our knowledge, no previous published work has combined five-model benchmarking, segmented performance analysis, feature importance interpretability, economic impact quantification and full operational deployment for pharmaceutical SMEs in an African developing economy. The frugal digital innovation blueprint is replicable across MENA and sub-Saharan pharmaceutical contexts.