Forecasting foreign exchange rates using Support Vector Regression

Support Vector Regression (SVR) algorithms have received increasing interest in forecasting, promising nonlinear, non-parametric and data driven regression capabilities for time series prediction. But despite evidence on the nonlinear properties of foreign exchange markets, applications of SVR in price or return forecasting have demonstrated only mixed results. However, prior studies were limited to using only autoregressive time series inputs to SVR. This paper evaluates the efficacy of SVR to predict the Euro-US Dollar exchange rate using input vectors enhanced with explanatory variables on mean-reversion movements derived from Bollinger Bands technical indicators. Using a rigorous empirical out-of-sample evaluation of multiple rolling forecast origins, we assess the accuracy of different SVR input vectors, including upper and lower BB, binary trading signals of BB, and combinations of the above. As a result, a local SVR model using autoregressive lags in conjunction with BB bands and BB indicators, and recalibrated yearly, outperforms the random walk on directional and all other error metrics, showing some promise for an SVR application.