Mikhail Polubaryev, a student at the Russian Technological University (RTU MIREA), has built a pilot AI model for forecasting demand in the pharmaceutical market. The hybrid approach combines gradient boosting for short‑term forecasts (one month) and neural networks with direct multi‑step forecasting for medium‑term planning (a quarter). This combination enables adaptation to Russia’s volatile pharmaceutical market, where traditional methods often fail, RTU MIREA said in a press release.
Russia’s drug market is highly volatile, with frequent launches of new generics that lack historical sales data. Under these conditions, traditional methods accumulate critical errors by the third month, making them unsuitable for inventory management. This is why a hybrid approach combining both methods proves more reliable for supply planning.
Gradient boosting ensembles performed well for one‑month forecasts, with a WMAPE error of 16‑17%. However, when extending the forecast to a full quarter, the error grew to 23‑26% by the third month.
Polubaryev then deployed neural networks with direct multi‑step forecasting, which produced more stable results without a sharp rise in errors. This led to an adaptive strategy: boosting ensembles with custom feature tuning for short‑term horizons, and drift‑resistant neural networks for medium‑term planning.
The project’s academic supervisor, Professor Andrey Gorshenin, said the architecture is ready for integration into drugmakers’ ERP systems. It includes logarithmic variance stabilisation, cyclical encoding of timestamps and blocked cross‑validation – a set of tools essential for industrial deployment.
One Russian pharmaceutical company has already expressed interest in the model.
Earlier, US and Chinese bioinformaticians developed an AI model that can predict whether a patient will benefit from immunotherapy based on gene activity data and tumour histological characteristics.
