Forecast in the pharmaceutical area – Statistic models vs deep learning

Abstract

The main goal of this work was to evaluate the application of statistical and connectionist models for the problem of pharmacy sales forecasting. Since R is one of the most used software environment for statistical computation, we used the functions presented in its forecast package. These functions allowed for the construction of models that were then compared with the models developed using Deep Learning algorithms. The Deep Learning architecture was constructed using Long Short-Term Memory layers. It is very common to use statistical models in time series forecasting, namely the ARIMA model, however, with the arising of Deep Learning models our challenge was to compare the performance of these two approaches applied to pharmacy sales. The experiments studied, showed that for the used dataset, even a quickly developed LSTM model, outperformed the long used R forecasting package ARIMA model. This model will allow the optimization of stock levels, consequently the reduction of stock costs, possibly increase the sales and the optimization of human resources in a pharmacy.