Forecasting pharmacy purchases orders

Abstract

Inventory represents the largest asset in pharmacy products distribution. Forecasting pharmacy purchases is essential to keep an effective stock balancing supply and demand besides minimizing costs. In this work, we investigate how to forecast product purchases for a pharmaceutical distributor. The data contains inventory sale histories for more than 10 thousand active products during the past 15 years. We discuss challenges on data preprocessing of the pharmacy data including cleaning, feature constructions and selections, as well as processing data during the COVID period. We experimented on different machine learning and deep learning neural network models to predict future purchases for each product, including classical Seasonal Autoregressive Integrated Moving Average (SARIMA), Prophet from Facebook, linear regression, Random Forest, XGBoost and Long Short-Term Memory (LSTM). We demonstrate that a carefully designed SARIMA model outperformed the others on the task, and weekly prediction models perform better than daily predictions.