Forecasting daily and monthly exchange rates with machine learning techniques

The difficulty and the intense economic interest to forecast the behavior of exchange rate times series has led to the creation of many approaches that thrive to beat the random walk model. Some employ autoregressive or structural models, advanced econometric techniques such as VARs or GARCH or new and emerging in the field of economics techniques from the area of signal processing and machine learning that are proving to be very successful [see for example the use of Support Vector Regression (SVR) in forecasting exchange rates, e.g. Ince and Trafalis, 2005; Brandl et al, 2009]. In this paper, we combine SVR with a signal processing decomposition method (Ensemble Empirical Mode Decomposition - EEMD). The basic concept of our approach is decomposing the exchange rate time-series in a trend and a remaining fluctuation and forecast each sub-series separately. The initial input variable dataset for both sub-series consists of commodity prices, moving averages and stock market indices, basic macroeconomic indicators and eight exchange rate series, spanning the period from 04/01/1999 to 31/10/2011. The variable set is refined using the Multivariate Adaptive Regression Splines method. The SVR models are trained and then used for out-of-sample forecasting. The described scheme outperformed state-of-the-art methods in terms of statistical measures (RMSE, RMSPE, MAPE) and annualized returns, such as Differential EMD-SVM proposed by Premanode et al (2012) and adaptive Neural Networks proposed Sermpinis et al  (2012, 2013). Extending our research framework we applied our method on FX rates of developed and developing economies, in order to test the ability to forecast foreign exchange markets of different maturity. So, we consider USD/JPY, Philippine peso/ South African rand, New Zealand dollar/ Brasilian rand and Norwegian Krona/ Australian dollar FX rates, on daily and monthly rate forecasting. The accuracy of the autoregressive approach of the EEMD-SVR model in these bilateral rates leads to the rejection of the weak form market efficiency hypothesis and thus produces sustainable trading profits both on daily and monthly forecasting horizons.