City Research Online

Abstract

Financial markets are an important feature of modern economies, where trading decisions can be critical because of their significant impact on social and economic life. Various models and techniques have been applied to describe and predict financial time series in order to develop effective tools in financial prediction. In particular, neural networks have recently gained significant research interest in financial markets as well as in other domains. As financial time series data show a high degree of non-linearity, neural networks represent an attractive approach in this area.

This thesis introduces a novel neural network model, the FL-SMIA model, as well as several variations and extensions, namely the FL-SMIA*, D-FL-SMIA, MD-FLSMIA, MD-FL-SMIA-2, M FL-SMIA, and FL-SMIA-RBM. The FL-SMIA model is a model that uses the principles of the Functional Link Neural Network (FLNN) and the Self-organising Multilayer Neural Network using the Immune Algorithm (SMIA). The FL-SMIA model combines the higher-order inputs , i.e. the products of raw input features, with the self-organising hidden layer (SMIA) that dynamically grows and adapts to the input vectors.

Based on the promising results of the FL-SMIA network in initial experiments, variations and extensions have been developed using deeper architectures (D FLSMIA), mixed input representations (M-FL-SMIA), a combination of deep and mixed architectures (MD-FL-SMIA), and of the FL-SMIA with the Restricted Boltzmann Machine in the FL-SMIA-RBM. The proposed models have also been compared with other neural network architectures: FLNN, the Multilayer perceptron (MLP), and SMIA.

All networks have been evaluated for one day and five days ahead prediction using financial and statistical metrics, focusing on the Relative Profit (RP) and Annualised Volatility (AV). Data-sets of three different types have been used: exchange rates (USD/UKP, USD/EUR, JPY/USD), stock price indices (NASDAQ, DJIA), and commodity prices (OIL and GOLD).

In terms of average RP results for the one day ahead prediction, the FL-SMIA was slightly worse than the best model (FLNN) but FL-SMIA model reduced the investment risk by producing the lowest average AV value. We have also observed notable differences between data types.

For the five days ahead prediction, the M-FL-SMIA model has the highest average RP and the lowest average AV results. Correlation analysis on the residuals has shown differences in behaviour between FLNN model and FL-SMIA model, encouraging further extensions and variations.

Overall, the FL-SMIA model and its extensions will be useful for time series prediction because of their competitive performance and different behaviour to standard neural networks.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	Doctoral Theses School of Science & Technology > School of Science & Technology Doctoral Theses School of Science & Technology > Computer Science

Export

Downloads