Research pick: Reformulating pharma supply chains with AI - "Data-driven forecasting of pharmaceutical sales: distinguishing promotional vs. daily scenarios"

Research in the International Journal of Data Mining and Bioinformatics discusses a new approach to demand forecasting for the pharmaceutical retail sector based on an artificial intelligence model. The findings hold promise for improving accuracy in one of the industry’s most persistent logistical challenges: managing sales that swing sharply during promotional periods. The new system works better than traditional models by distinguishing between routine demand and the short-term surges driven by marketing campaigns.

The team has built their forecasting system using a machine-learning framework known as the Temporal Fusion Transformer. This deep-learning model is designed specifically to analyse time-series data, such as daily sales figures or seasonal illness rates. Where conventional systems might smooth over the spikes and troughs in this kind of data, the new model can interpret such fluctuations and offer a more nuanced analysis for more reliable forecasting.

One of the underlying factors leading to this improved approach is the model’s use of multivariate feature construction. This method can be used to integrate diverse types of data into a single analytical framework. Rather than relying solely on historical sales figures, the model can use public health trends, seasonal disease prevalence, and promotional calendars. By working with such an enriched dataset, the model can detect complex patterns and anticipate demand with much greater precision.

In addition, the system uses a knowledge-guided attention mechanism. This process allows the system to prioritize the most relevant data depending on the sales scenario. For example, during an outbreak of influenza, the model may focus more heavily on regional health reports, whereas during a promotion, it shifts emphasis toward marketing schedules and in-store behaviour. This flexibility allows it to treat routine and promotional demand as fundamentally distinct processes, and so make better predictions about demand.

The researchers have tested their system on data from over 1.2 million retail transactions. The model reduced forecasting errors by almost a quarter compared to traditional methods. When tested in a commercial setting, it led to an almost one-third improvement in medication stock availability and just over a quarter reduction in excess inventory. Such improvements are not merely operational gains. Both outcomes are central to ensuring access to essential medicines while minimising waste in pharmaceutical supply chains.

Zeng, Z., Guo, Y., Ji, Y., Shi, Y. and Feng, T. (2025) ‘Data-driven forecasting of pharmaceutical sales: distinguishing promotional vs. daily scenarios’, Int. J. Data Mining and Bioinformatics, Vol. 29, No. 5, pp.1–26.