Abstract: E-pharmacy demand forecasting is highly affected by drug attributes and promotional activitiesproposed by the e-pharmacy platform. Atimeseries-machine learning hybrid model that integrates price discount and coupons is proposed to better analyze sales improvement brought by promotional activities, based on which more accurate forecasting results can be obtained. Traditional demand forecasting research decomposes demand under promotional activities into a linear combination of baseline sales and promotional lifting sales, while the drug’s treatment cycleis considered in this model, and SARIMA model is used to predict the baseline sales.Finally,predicted baseline sales data and promotional features are put into XGBoost model for integrated learning to further analyze the promotional effects. Sales data from a Chinese leading e-pharmacy is used to test the model’s effectiveness, results indicate that this proposed hybrid model performs better compared to the other three widely used forecasting models. At the same time, the hybrid model’s efficiency under different price discount, as well as promotional information and data pooling strategy is verified.Results show that the hybrid model performs better when price discount varies,promotional information can sufficiently reduce the forecasting error by at least 40% when is added into the proposed hybrid model, while data pooling strategy can help the hybrid model reduce forecasting error by around 10%. The proposed hybrid model is confirmed to be applicable and useful, which sheds light on e-pharmacy’s demand forecasting with promotional activities.

Key words: e-pharmacy; demand forecasting; promotional activities; timeseries-machine learning hybrid model