High frequency data: Making forecasts and looking for an optimal forecasting horizon

We illustrate the AutoRegessive/Generalised Conditionally Heterosscedastic (ARCH-GARCH) methodology on the developing a forecast model for exchange rates time series of the Czech crown (CZK) against the Slovak crown (SKK) and make comparisons the forecast accuracy with the class of Radial Basic Function Neural neural network RBF NN models. To illustrate the forecasting performance of these approaches the input/output function estimation based on RBF networks is presented. In a comparative study is shown that the RBF NN approach is able to model and predict high frequency data with reasonable accuracy and more efficient than statistical methods. In order to find the optimal forecasting horizon, we use the analysis of forecast errors and choose the values that give the smallest error variance. It is found that the error variance estimation process based on soft methods is simplified and less critical to the question whether the data is true crisp or white noise.