Forecasting of the German stock index DAX with neural networks: Using daily data for experiments with input variable reduction and a modified error function

摘要

Using neural networks for the prediction of economic time series still involves many problems. Examples for using neural networks in financial market applications are de Groot (1993), Baun (1997) and Burgess (1996). In these studies neural networks were successfully applied. Intensive work has been done regarding data transformation and the selection of an appropriate topology for neural networks. By using daily data of the German stock index DAZ this study shows that: 1) Principal Component Analysis is not an appropriate technique for input variable reduction. 2) The Usage of a modified mean squared error as error function leads to significantly better results.