A Thorough Evaluation of Demand Prediction Models: Machine Learning, Deep Learning, and Statistical Techniques for Import Businesses

Abstract

Nowadays, managing demand in companies is crucial to avoid storage overcosts, stockouts and to improve the service level of companies. To address this scenario, demand predictions through models and algorithms emerge. Therefore, this research aims to evaluate the performance of seven prediction techniques applying machine learning, deep learning, and statistical methods. To validate our experiments, we used Dickey–Fuller, Shapiro–Wilk, Friedman, and Wilcoxon post-hoc statistical tests on the predictions of the models using demand records from a Peruvian import company. The results indicated that deep learning and statistical models have significantly better predictions than machine learning models. In particular, the LSTM, CNN, ARIMA, and Holt-Winters models significantly improve accuracy compared to the Ridge Regression, Random Forest Regressor, and Decision Tree Regressor models. Compared to machine learning models, statistical and deep learning models improve accuracy in a range from 66.01 to 86.10%. These results highlight the statistical advantage of deep learning and statistical models in demand prediction, with the LSTM model showing the lowest error.