The sum-squared error measure which is commonly used in financial forecsting produces asymptiotically best estimators in the adequate model under the asumption of normally distributed noise with constant variance. In most practical applications the noise has a more complicated distribution, however, so that common regression networks yield suboptimal results. Imporved estimators may be derived in this case using density estimating neural networks, which are capable to embody more complex probability models. We will discuss appropratie distribution assumptions for the important cases of outliers and non-constant variances, and give interpretations of the new estimates in regression theory. The practical superiority of density-based estimators is shown in tory problems as well as in the task of forecasting the intraday volatiliity of the German stock index DAX.
展开▼
