Foercasting US/eur daily and monthly exchange rate with machine learning techniques

The extreme volatility of exchange rate time series has always been an active field of research. The main approaches used in the literature to forecast future rates are based on monetary exchange rate models, econometric modeling(such as ARCH, GARCH and ARIMA models) and machine learning techniques. We intend to combine signal processing to the forecasting abilities of machine learning methods by employing a hybrid Ensemble Empirical Mode Decomposition (EEMD) and Support Vector regression (SVR) model. SVR is a novel machine learning method that has proven empirically its abilities in forecasting exchange rates. On the other hand EEMD (Wu and Huang, 2005) is a data-driven, noise-assisted signal processing method that decomposes a given signal into its components, called Intrinsic Mode Functions (IMFs). EEMD has been implemented in a significant number of scientific areas, providing promising results in identifying physical processes through signal decomposition. Building on the energy approach of Rilling et al (2005), we select a mixture of IMFs as a non-linear, non-stationary trend.. The input variables consist of commodity prices, technical analysis indicators (Moving Averages, Relative Strong Index) and stock market indices, metal prices, basic macroeconomic indicators and past exchange rates of 8 currencies, spanning from 04/01/1999 to 31/10/2011. Multivariate Adaptive Regression Splines select the most informative variables among the plethora of the provided, which are then fed into an SVR model. The overall process is repeated twice, one for the trend and one for the fluctuation component of all time series. Then the forecasted trend and fluctuation are added and compared to real data. The above implementation marked superior forecasting abilities in predicting the US/EUR exchange rate compared to single SVR models both on monthly and daily forecasting horizons. Overall, the proposed model is a combination of effective techniques in forecasting time series, is data driven, relies on minimum initial assumptions and provides a structural aspect of the forecasting problem.