Abstract: | In this project I have explored the use of neural networks for the task of time series prediction. A multilayer feed-forward network was used and trained with standard back-propagation, quick-propagation and a new hybrid algorithm based on genetic algorithms and the perceptron learning algorithm. Some financial time series and the Mackey-Glass series were predicted with limited success. Fixed sized sliding input windows was used. The difficulty of choosing the right size input window was highlighted as was the difficulty of assessing the network's generalisation performance.
|