SARIMA-ANN hybrid models for forecasts of SARS-CoV-2 contagion in Perú

• Autor: Ordoñez Mercado, Alipio Francisco
• Formato: artículo
• Fecha de Publicación: 2021
• Institución: Universidad Nacional de Ingeniería
• Repositorio: Revistas - Universidad Nacional de Ingeniería
• Lenguaje: español; inglés
• OAI Identifier: oai:oai:revistas.uni.edu.pe:article/1332
• Enlace del recurso: https://revistas.uni.edu.pe/index.php/iecos/article/view/1332
• Nivel de acceso: acceso abierto
• Materia: Modelos ARIMA; Redes Neuronales Autoregresivas; Perceptron Multicapas; Modelos híbridos NNAR-ARIMA; Modelos híbridos MLP-ARIMA; Models; Autoregressive Neural Networks; Multilayer Perceptron; Hybrid models NNAR-ARIMA; Hybrid models MLP-ARIMA

Hybrid ANN-ARIMA models have been built by remodeling, to make the forecasts of the new cases of infections by Covid-19 in Peru, for this the confirmed cases of Covid-19 were extracted and used between the period 06/03/20 until 02/28/21, from the open data platform of the Ministry of Health. The results found indicate that the 02 best models correspond to the multiplicative hybrid model NNAR (27, 1, 6) * ARIMA (3, 0, 2) (1, 0, 1), and to the additive hybrid model NNAR (27, 1, 6) + ARIMA (1, 0, 1), whose values of the mean absolute percentage error (MAPE) differ by only 0.575%, thus providing almost the same forecasts. Considering the average of the MAPE values for the 03 best models of each modeling category, it has been determined that the NNAR-ARIMA hybrid models are better than the MLP-ARIMA hybrid models, that the NNAR + ARIMA additive hybrid models have a superiority of 1.20 % on the multiplicative hybrid models NNAR * ARIMA; while the superiority of the MLP + ARIMA additive hybrid model over the MLP * ARIMA multiplicative hybrid model reaches 2.31%.