A Drug Sales Forecasting Method Based on Multiple Models and Parameter Optimization

Drug sales forecasting models can help chain pharmacies find potential users of targeted drugs. Affected by factors such as region, disease characteristics, user characteristics and drug types, it is difficult to make sales forecasts for specific drugs in different stores and in different granularity across the country. This paper presents a two-layer structure method for drug sales forecasting. It is optimized from two dimensions of model and parameter. Building a deep learning multi-model structure based on bagging allows different models to be used in different regions. An improved genetic algorithm was used to design the selection strategy and variation mode for drug marketing. SS-KSMOTE is designed based on k-mean, SMOTE and stratified sampling to solve the problem of sample sparsity. The experiment proves that the drug sales forecasting method under the two-layer optimization structure can effectively meet the drug sales forecasting needs under the complex scenario of chain pharmacies.