Redes neuronales artificiales y modelos arima aplicadas a la modelación y predicción del tipo de cambio euro-dólar

Abstract

The present degree work aims to model and predict the EURO-DOLLAR exchange rate, for which the data was obtained on the virtual platform Investing.com whose access is free; an exploratory analysis was performed to identify the behavior of the data. Three techniques were applied: Box-Jenkins, Recurrent Elman-type recurrent neural networks and Long Short-Term Memory to model and predict the series, the scale and percentage errors, Theil's U coefficient and the Diebold-Mariano test were obtained to identify which technique has more accurate predictions. The data was obtained in the period from January 2, 1998 to December 31, 2019, the series was divided into training and validation sets. Data from January 2, 1998 to November 29, 2019 and from December 2, 2019 to December 31 of the same year for the validation set were used for the training package for The Training Set for January 29, 2019 and December 31, 2019. Because the time series has random walk characteristics an ARIMA(0, 1, 0) model is determined, the prediction in time t is the time data t-1, predictions of 22 days to the future were made and compared to predictions made with the two types of recurrent neural networks. The U coefficients of Theil for predictions with the model ARIMA, Elman network and Long Short-Term Memory network gave 0.04743, 0.002625 and 0.001808 respectively, identifying that the third technique presents more accurate forecasts for this type of data, being corroborated by the Diebold-Mariano test.