Utilización de Redes Neuronales Artificiales (RNA) como alternativa para el pronóstico del Índice de Precios al Consumidor (IPC) y contraste de resultados con modelos estadísticos tradicionales (ARIMA y VAR) para el Ecuador periodo 2000-2017

Abstract

In the present investigation the quality of forecast offered by three different models is evaluated. Two of them of traditional use in economic sciences (model ARIMA and VAR) and a model of artificial neural networks (ANN) for the series of the consumer price index (CPI) of Ecuador for the period 2000-2017. The models that offer the best results within each family are the SARIMA model (1,1,1), VAR model with a single lag and the ANN model with 1 hidden layer, 8 hidden neurons and linear activation function in each layer. As the main findings, it is highlighted the great prognosis efficiency that the SARIMA model (1,1,1) has, both for the short and medium term. It is also important to highlight the contribution that the ANN model makes to the quality of multivariate prognosis when used together with the VAR model, improving the root mean quadratic error. To analyze the prognosis quality of the different models, in addition to the traditional statistical indicators (mean quadratic error and its root, mean percentage error, average absolute percentage error and its deviation and Theil U1) a recurrent prognostication is also made. This prognostication shows that all the models which the prognosis has been made with are stable