Academic

Forecasting the Financial Times Stock Exchange 100 Index using Neural Networks

Abstract

The ability of an investor to accurately forecast price movements in shares and share indices would potentially enable them to realise enormous wealth. Many analytical techniques exist to try to identify trends within these movements. Neural Networks is an area renowned for its capacity to recognize complex patterns in data which cannot otherwise be easily identified.

This project focuses on the ability of the Multilayer Perceptron, the Time Delay Neural Network and the Recurrent Neural Network at forecasting end of day closing price of the FTSE 100 Index.

Many different network topologies were examined, with a variety of different parameters and training levels, before the best network of each type was compared with a Box Jenkins ARIMA model.

The best performance was achieved using a Recurrent Network with 5 inputs, 15 hidden units, 0.1 memory depth, 0.8 learning rate and 0.5 momentum, when trained with 10000 epochs. During testing and validation of the networks, better results were obtained from MLP networks than TDNN networks. However during the final forecast test this was reversed.

This research concludes that the selection of appropriate network parameters is vital to the networks prediction capability as is the optimal amount of training.