Interesting Forecast Research with SVR Network

Li, Hu, and Hirasawa made some interesting findings forecasting USD/JPY exchange rate moves in “Financial Time Series Prediction Using a Support Vector Regression Network”, or SVRN. Their Next-Day (t+1), Next-Week (t+5), and Next-2Week (t+10) directional predictions resulted with close to 80% accuracy. It rocks. For those unfamiliar with SVR, a detailed guide sits here.


SVRN test rundown

With the t+1, t+5, and t+10 USD/JPY exchange rates as output, input variables included historical USD/EUR exchange rates, the NIKKEI225, and oil prices. They essentially split the training data into three making up the “Transformation Layer”, whose outputs then became input variables for the “Prediction Layer” where the magic takes place.


What made this research interesting

Through the SVRN, the predictions displayed significantly lower SSE (Sum of Square Error) than those performed by SVR alone, especially with the longer time steps. This makes it much easier to create and support a practical trading strategy.


Did I mention they showed a close to 80% hit rate for all three forecasts? Keep in mind the SVRN derived forecast models remain INDEPENDENT. Just applying basic probability theory, one could theoretically wait until all three forecasts pointing to the same direction, bet on it and have a 98%+ hit rate (1- 0.2^3). Pretty cool huh!