Foreign Exchange Rate Prediction using Computational Intelligence Methods

This paper presents the application of six nonlinear ensemble architectures to forecasting the foreign exchange rates in the computational intelligence paradigm. Intelligent techniques such as Backpropagation neural network (BPNN), Wavelet neural network (WNN), Multivariate adaptive regression splines (MARS), Support vector regression (SVR), Dynamic evolving neuro-fuzzy inference system (DENFIS), Group Method of Data Handling (GMDH) and Genetic Programming (GP) constitute the ensembles. The data of exchange rates of US dollar (USD) with respect to Deutsche Mark (DEM), Japanese Yen (JPY) and British Pound (GBP) is used for testing the effectiveness of the ensembles. To account for the auto regressive nature of the time series problem, we considered lagged variables in the experimental design. All the techniques are compared with normalized root mean squared error (NRMSE) and directional statistics ( stat D ) as the performance measures. The results indicate that GMDH and GP based ensembles yielded the best results consistently over all the currencies. GP based ensembling emerged as the clear winner based on its consistency with respect to both stat D and NRMSE, but GMDH outperforms it in one of the currencies (DEM). Based on the numerical experiments conducted, it is inferred that using the correct sophisticated ensembling methods in the computational intelligence paradigm can enhance the results obtained by the extant techniques to forecast foreign exchange rates.