In this paper a new neural network architecture, able to deal with uncertainty, is proposed. Capital market time series prediction is one of its main applications. The interest on this new modeling technique arises from the network's ability to draw a confidence interval on the forecasting task. By allowing a certain level of uncertainty, the networks' ability to generalize through the raining data is increased. Simulated results for a US Dollar/Swiss franc exchange rate series are presented to demosntrate the potential use of the proposed algorithm.
展开▼