Computational knowledge discovery techniques and their application to options market databases

Financial options are central to controlling investor's risk exposure. However, since 1987 parametric option pricing models have performed poorly in assessing risk levels. Also, electronic trading systems were introduced in this period, and these produce option price quotations at a rate of up to several times per second. There is a large and rapidly expanding amount of data to be analysed. A new generation of techniques for pattern recognition in large datasets has evolved, collectively termed 'computational knowledge discovery techniques' in this work. Preliminary evidence suggests that certain of these techniques are superior to parametric approaches in pricing options. Statistical confidence in models is of paramount importance in finance, hence there is a need for a systems framework for their effective deployment. In this thesis, a dedicated computational framework is developed, for the application of computational knowledge discovery techniques to options market databases. The framework incorporates practical procedures, methods, and algorithms, applicable to many different domains, to determine statistical significance and confidence for data mining models and predictions. To enable a fuller evaluation of the uncertainty of model predictions, these include a new method for estimating pointwise prediction errors, which is computationally efficient for large datasets, and robust to problems of regression and heteroskedasticity typical of options market data. A number of case study examples are used to demonstrate that computational knowledge discovery techniques can yield useful knowledge for the domain, when applied using the framework, its components, and appropriate statistical and diagnostic tests. They address an omission in the literature documenting the application of these techniques to option pricing, which reports few findings based on hypothesis testing. A contribution to the field of nonparametric density estimation is made, by an application of neural nets to the recovery of risk-neutral distributions from put option prices. The findings are also new contributions for finance. Finally, in a discussion of software implementation issues emerging technology trends are identified. Also, a case is made that future vertical data mining solutions for options market applications, should incorporate statistical analysis within the tool, and should provide access to values of partial derivatives of the models.

Name dc:type.qualificationname

phd

Level dc:type.qualificationlevel

doctoral

Grantor dc:publisher.institution

London Metropolitan University

Year dc:date.issued

2004

Author dc:creator

• Healey, Jerome V.

• 330 Economics

Dc Identifier Grantnumber

N/A

OAI identifier oai:identifier

oai:repository.londonmet.ac.uk:7260