| Popis: |
Provides evidence about the reliability of neural networking models as applied to European industrial production time series. Takes 24 series for sectors of the UK, German and French economies from 1986, 1978 and 1985 respectively to 1995, and reserves the last two years' monthly data for post-sample tests. Separates nonlinear series, and subjects both subsamples to a specified one hidden layer feed forward neural network model, with one model for each forecast horizon. Compares with a linear autoregressive forecasting model, with different specifications for each forecasting horizon. Finds that the linear model produces smaller forecast root mean squared error than the neural network for up to 12 months ahead, but the neural network model is better at predicting the sign of the forecast value for up to 6 months ahead. Finds no evidence of greater neural network accuracy with nonlinear series. |
Full Text from ScienceDirect