Time-series sales forecasting for an Enterprise Resource Planning system

Resale businesses and product suppliers rely on enterprise resource planning (ERP) systems to manage their inventory. One of these systems is Microsoft Business Central (BC). BC can create purchase orders for sold products, in order to keep stock at a constant level. The aim of this thesis is to implement a model that can forecast the sales before they happen and thus create the purchase orders before the product runs out of stock or is sold. The dataset used in the thesis is from a product supplier consisting of product sales data for about a thousand different products on a time period of three years. The thesis presents three different forecasting algorithms: autoregressive integrated moving average (ARIMA), long short-term memory (LSTM) and neural networks. The different models were compared using root mean squared error (RMSE) values based on the datasets values against the predicted values. The results show that the best suited model for most of the product sub-datasets is ARIMA. Many of the products however had too little data to be reliably modeled. Further study and development is proposed to be done in testing the presented architecture with another dataset as well as optimizing the LSTM and neural network models