Data-driven forecasting of pharmaceutical sales: distinguishing promotional vs. daily scenarios

This study presents an enhanced temporal fusion framework for pharmaceutical demand forecasting, designed to handle both baseline consumption and promotion-driven sales fluctuations. The approach integrates heterogeneous data sources - such as therapeutic seasonality and catchment health indicators - into a multivariate feature space, and incorporates a knowledge-guided attention mechanism within a temporal fusion transformer to decouple regular and promotional demand. Systematic factor analysis further quantifies the influence of product, store, and temporal variables. Using a dataset of 1.2 million retail transactions, the proposed model reduces forecasting error by 23.6% over traditional methods. Ablation studies confirm the importance of future promotion signals (accuracy loss of 15.4%, p < 0.001), while analysis reveals that seasonal effects vary significantly by drug type. Store-level factors such as deprivation index and competition proximity also significantly affect promotional effectiveness. The framework offers three key contributions: adaptive feature engineering for retail contexts, integration of domain knowledge into temporal modelling, and empirical identification of demand drivers. Practical deployment yielded 31% fewer stock-outs and 27% lower inventory costs, demonstrating its value for resilient and data-driven pharmaceutical supply chain optimisation.