Exchange rate forecasting using echo state networks for trading strategies

Because of the diversity of portfolios based on assets throughout international markets, exchange rate prediction plays an important role in risk management, asset allocation, and trading strategies. This paper aims to investigate the use of a recent paradigm of recurrent neural networks, echo state networks (ESNs), applied to forecasting and trading currency exchange rates. It does so by benchmarking the statistical and trading performance of ESNs against a naïve strategy, an Autoregressive Moving Average (ARMA) model, and a multilayer perceptron neural network. One can interpret ESNs as a recurrent structure that provides both the simplicity of the resulting mathematical model and the ability to express a wide range of nonlinear and time-varying dynamics. As an application, this paper carries out computational experiments that include the Brazilian Real, the European Union Euro, the Japanese Yen, and British Pounds with American Dollar exchange rates from January 4, 2000, through December 31, 2012. The results reveal that the ESN and the ARMA model provide similar results, statistically outperforming the other models in terms of accuracy. However, when trading indicators are considered, the performance of the ESN is superior to that of the alternative approaches.