Comparison of Statistical Models with Deep Learning in Business Practice

Abstract

This diploma thesis focuses on the use of time series prediction in business practice. The main aim of this work is to compare selected prediction models on demand data from a selected company and to evaluate, which of the models is best suited for use in business practice. The partial goals are to define the basic concepts and concepts that are associated with the time series forecasting. Furthermore, to describe several prediction models that are currently used (e.g TBATS, Facebook Prophet, SARIMA, Deep Learning models). The work will be elaborated using a descriptive method, analysis method and synthesis method. Another important method used in the work will be the comparison method, which will be applied in the comparison of individual prediction models. Results of this diploma thesis confirm the results of a study already carried out by Makridakis. One of his latest studies pointed out that machine learning/deep learning methods cannot itself beat the statistical models in the accuracy of the prediction. The best model for demand predicting 90 days into the future is Facebook Prophet. This model had the lowest MAE and MSE and its performance requirements were very low. However, Prophet showed significantly worse results in the prediction of 365 days into the future, where TBATS became the best model. In predictive accuracy, it significantly outperformed other models and its only disadvantage was the complexity of the necessary computing power. Overall, TBATS is the best model in all assessment criteria and regardless of the length of the forecast period. Demand forecasting with TBATS model can help the company with increasing customer satisfaction. This advantage of forecasting in business will help to predict product demand so that enough product is available to meet customer orders on a timely basis. Furthermore, the implementation of the model can help to reduce the time span of inventory in the warehouse and therefore it can reduce the cost of the company. Furthermore, it can also help with effective production scheduling, lowering safety stock requirement and also with reducing product obsolescence costs, because obsolesced inventory will be reduced.