Non-linear modelling of financial data using topologically evolved neural network committees

• Main Author: Fotinopoulos, Ioannis
• Published: Imperial College London 2008
• Subjects: 006.3
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.487972

Most of artificial neural network modelling methods are difficult to use as maximising or minimising an objective function in a non-linear context involves complex optimisation algorithms. Problems related to the efficiency of these algorithms are often mixed with the difficulty of the a priori estimation of a network's fixed topology for a specific problem making it even harder to appreciate the real power of neural networks. In this thesis, we propose a method that overcomes these issues by using genetic algorithms to optimise a network's weights and topology, simultaneously. The proposed method searches for virtually any kind of network whether it is a simple feed forward, recurrent, or even an adaptive network. When the data is high dimensional, modelling its often sophisticated behaviour is a very complex task that requires the optimisation of thousands of parameters. To enable optimisation techniques to overpass their limitations or failure, practitioners use methods to reduce the dimensionality of the data space. However, some of these methods are forced to make unrealistic assumptions when applied to non-linear data while others are very complex and require a priori knowledge of the intrinsic dimension of the system which is usually unknown and very difficult to estimate. The proposed method is non-linear and reduces the dimensionality of the input space without any information on the system's intrinsic dimension. This is achieved by first searching in a low dimensional space of simple networks, and gradually making them more complex as the search progresses by elaborating on existing solutions. The high dimensional space of the final solution is only encountered at the very end of the search. This increases the system's efficiency by guaranteeing that the network becomes no more complex than necessary. The modelling performance of the system is further improved by searching not only for one network as the ideal solution to a specific problem, but a combination of networks. These committces of networks are formed by combining a diverse selection of network species from a population of networks derived by the proposed method. This approach automatically exploits the strengths and weaknesses of each member of the committee while avoiding having all members giving the same bad judgements at the same time. In this thesis, the proposed method is used in the context of non-linear modelling of high-dimensional financial data. Experimental results are encouraging as both robustness and complexity are concerned.