NARX and NARMAX Models for Time Series Forecasting Using Shallow and Deep Neural Networks

The present work focuses on the study of NARX and NARMAX models generated by mechanical techniques of automatic learning, such as “Support Vector Machine” (SVM), “Multilayer Perceptron” (MLP) and “Extreme Learning Ma-chine” (ELM), as well as deep neural networks such as “Long Short- Term Memory” (LSTM), “Gated Recurrent Unit” (GRU), Transformers and “Convolutional Neural Network” (CNN). The aim is to carry out an analysis of the predictive capacity of each model. The purpose is to find the best technique and provide recommendations for the use of each one, taking into account the characteristics of the series, including its complexity, which is calculated using the MF-DFA method. The hypothesis is that models generated with deep learning techniques outperform shallow techniques. The results show that the hypothesis is not fulfilled for problems of low complexity, however it is true for problems of medium complexity.