| Abstract: | Although neural networks are increasingly used in a variety of disciplines there are few applications in political science. Approaches to electoral forecasting traditionally employ some form of linear regression modelling. By contrast, neural networks offer the opportunity to consider also the non-linear aspects of the process, promising a better performance, efficacy and flexibility. The initial development of this approach preceded the 2001 general election and models correctly predicted a Labour victory. The original data used for training and testing the network were based on the responses of two experts to a set of questions covering each general election held since 1835 up to 1997. To bring the model up to date, 2001 election data were added to the training set and two separate neural networks were trained using the views of our original two experts. To generate a forecast for the forthcoming general election, answers to the same questions about the performance of parties during the current parliament, obtained from a further 35 expert respondents, were offered to the neural networks. Both models, with slightly different probabilities, forecast another Labour victory. Modelling electoral forecasts using neural networks is at an early stage of development but the method is to be adapted to forecast party shares in local council elections. The greater frequency of such elections will offer better opportunities for training and testing the neural networks. |
