Forecasting of time series with neural networks

During the East few years a number of new approaches to time series prediction have been developed. This paper presents some of the recent developments in artificial neural networks to tackle the problem of complex and noisy data, with an emphasis on the financial domain. The problem of time series prediction can be reduced to a three step approach: (1) One has to determine the relevant data and its representation (2) to implement an artificial neural network with optimal predictive powers and (3) to test the model. The selection of variables and their temporal extent as well as their representation and preprocessing are known to play a large role in the quality of the results. For practical applications the careful selection of the target time series is becoming increasingly important, in order to obtain a useful outcome. The overwhelming majority of models used in step two are based on the well-known backpropagation algorithm with a variety of topologies and transfer functions with an assortment of enhancements. The alternative is recurrent backpropagation used in non-feedforward nets. This approach requires more computational pourer but promises to out-perform the former for some time series.