Autoregressive Moving Average Recurrent Neural Networks Applied to the Modelling of Colombian Exchange Rate

Abstract

Modeling and prediction of time series has required in recent times a lot of attention, due to the necessity to have to make with accurate tools a right decision and to surpass theoretical, conceptual and practical limitations of the traditional approaches. In this sense, the neural networks have demonstrated to be a valuable tool, because they allow to represent nonlinear relationships, which are not well captured by other models. The investigations on neural networks have led to the development of different topologies, which adapt better to diverse problems. It is how, it seems to be that by the prediction problem, neural networks with some types of recurrence exhibit better approaches than other models, because they conserve a long memory of the series behavior. This paper proposes the use of autoregressive moving average recurrent neural networks (ARMA-NN) in the modeling and prediction of the Colombian exchange rate, evaluating its performance by the contrast with an autoregressive integrated moving average (ARIMA) model and a traditional feed-forward neural network (NN). ARIMA models and traditional feed-forward NN models are often compared with mixed conclusions in terms of the superiority in forecasting performance. In this paper, the results are in favor of the use of ARMA-NN models, every time that the prediction displays a better approach to the values of the series, which stimulates the use of such models in similar series and the research of other topologies of recurrence that allow better results.