Demand forecasting in Python: Deep learning model based on LSTM architecture versus statistical models

Abstract

Demand forecasting for business practice is one of the biggest challenges of current business research. However, the discussion on the use of forecasting methods in business is still at the beginning. Forecasting methods are becoming more accurate. Accuracy is often the only criterion for forecasting. In the reality of business practice or management is also influenced by other factors such as runtime, computing demand, but also the knowledge of the manager. The goal of this article is to verify the possibilities demand forecasting using deep learning and statistical methods. Suitable methods are determined on based multi-criteria evaluation. Accuracy according to MSE and MAE, runtime and computing demand and knowledge requirements of the manager were chosen as the criteria. This study used univariate data from an e-commerce entity. It was realized 90-days and 365-days demand forecasting. Statistical methods Seasonal naive, TBATS, Facebook Prophet and SARIMA was used. These models will be compared with a deep learning model based on recurrent neural network with Long short-term memory (LSTM) layer architecture. The Python code used in all experiments and data is available on GitHub (https://github.com/mrnavrc/demand_forecasting). The results show that all selected methods surpassed the benchmark in their accuracy. However, the differences in the other criteria were large. Models based on deep learning have proven to be the worst on runtime and computing demand. Therefore, they cannot be recommended for business practice. As a best practice model has proven Prophet model developed at Facebook.