Feed forward neural networks and genetic algorithms for automated financial time series modelling

摘要

This thesis presents an automated system for financial time series modelling. Formal and appliedudmethods are investigated for combining feed-forward Neural Networks and Genetic Algorithms (GAs) into audsingle adaptive/learning system for automated time series forecasting. Four important research contributionsudarise from this investigation: i) novel forms of GAs are introduced which are designed to counter theudrepresentational bias associated with the conventional Holland GA, ii) an experimental methodology forudvalidating neural network architecture design strategies is introduced, iii) a new method for network pruningudis introduced, and iv) an automated method for inferring network complexity for a given learning task isuddevised. These methods provide a general-purpose applied methodology for developing neural networkudapplications and are tested in the construction of an automated system for financial time series modelling.udTraditional economic theory has held that financial price series are random. The lack of a priori models onudwhich to base a computational solution for financial modelling provides one of the hardest tests of adaptiveudsystem technology. It is shown that the system developed in this thesis isolates a deterministic signal withinuda Gilt Futures prices series, to a confidences level of over 99%, yielding a prediction accuracy of over 60%udon a single run of 1000 out-of-sample experiments.udAn important research issue in the use of feed-forward neural networks is the problems associatedudwith parameterisation so as to ensure good generalisation. This thesis conducts a detailed examination ofudthis issue. A novel demonstration of a network's ability to act as a universal functional approximator forudfinite data sets is given. This supplies an explicit formula for setting a network's architecture and weights inudorder to map a finite data set to arbitrary precision. It is shown that a network's ability to generalise isudextremely sensitive to many parameter choices and that unless careful safeguards are included in theudexperimental procedure over-fitting can occur. This thesis concentrates on developing automated techniquesudso as to tackle these problems.udTechniques for using GAs to parameterise neural networks are examined. It is shown that theudrelationship between the fitness function, the GA operators and the choice of encoding are all instrumentaludin determining the likely success of the GA search. To address this issue a new style of GA is introducedudwhich uses multiple encodings in the course of a run. These are shown to out-perform the Holland GA on audrange of standard test functions. Despite this innovation it is argued that the direct use of GAs to neuraludnetwork parameterisation runs the risk of compounding the network sensitivity issue. Moreover, in theudabsence of a precise formulation of generalisation a less direct use of GAs to network parameterisation isudexamined. Specifically a technique, artficia1 network generation (ANG), is introduced in which a GA isudused to artificially generate test learning problems for neural networks that have known network solutions.udANG provides a means for directly testing i) a neural net architecture, ii) a neural net training process, andudiii) a neural net validation procedure, against generalisation. ANG is used to provide statistical evidence inudfavour of Occam's Razor as a neural network design principle. A new method for pruning and inferringudnetwork complexity for a given learning problem is introduced. Network Regression Pruning (NRP) is audnetwork pruning method that attempts to derive an optimal network architecture by starting from what isudconsidered an overly large network. NRP differs radically from conventional pruning methods in that itudattempts to hold a trained network's mapping fixed as pruning proceeds. NRP is shown to be extremelyudsuccessful at isolating optimal network architectures on a range of test problems generated using ANG.udFinally, NRP and techniques validated using ANG are combined to implement an Automated Neuraludnetwork Time series Analysis System (ANTAS). ANTAS is applied to the gilt futures price series The LongudGilt Futures Contract (LGFC).