Knowledge Driven Data Mining For Causal Relationships Between News And Financial Instruments

As news becomes more and more important in financial instruments trading algorithms, financial industry observers, investors and other analysts in financial markets are paying more attention to news. Some financial services companies have designed services which filter unrelated news information and documents and text information such as SEC or regulatory filings, and corporate web site information, which is another data source that financial investors can consider into their trading models. Besides financial services companies, some news publishers also provide similar services for their customers. However, after filtering, news served to financial brokerages and investors still need further human judgements for exploring the implications of news content and distinguishing significant from non-significant news, and for finding out the impact polar type of each kind of significant news. But these judgements are always limited by human information processing capability and cognitive biases. Thus, in order to support more objective decision making, by reducing human information processing limitations and cognitive biases, an ontology based framework, for investigating the relationships between news and financial instruments trading activities qualitatively and quantitatively, is proposed. This thesis contains two separate, but interrelated parts.The first part is an ontology, provided for demonstrating the domain knowledge about news in financial markets. The ontology model comprises two components. One is represented using OWL DL (which is a sub-language of Web Ontology Language), which provides a hierarchical framework for the domain knowledge, including primary classes of news, classes of financial markets participants, classes of financial instruments, and relations between these classes. This component is a specification of domain-specific vocabulary terms. The other component is a causal map, used to demonstrate impact of different classes of news events on financial instruments. It is of either a direct or an indirect"cause-effect"form, which can be written as rules using OWL rules language.The second part is an ontology based data mining framework designed to study the quantitative relationships between news and financial instruments trading activities. The framework is made of three components. The first is editing of the ontology from the first part with Protégésoftware tool. It is used to classify news and stocks into different groups according to the nature of businesses and stocks, and the news categories defined in the ontology model, when news and stock data come into the framework. The second part is an expert-defined rules reasoning system implemented in Jess Shell, a plug-in of the Protégétool. For a given financial instrument trading activity, it can indicate the possible significance of news, and generate a data mining model for the specific financial instrument. The third part is Bayesian network algorithm. Combined with the data mining model, this algorithm can specify the quantitative relationships between the possibly significant news and the given financial instrument trading activity.The major contributions of this research is that the ontology helps understand the knowledge about news in financial markets, helps build trading models based on news, and build systems for prediction of stock prices based on news. The ontology based data mining framework provides an ontology method for classifying news and financial instruments data, an expert reasoning system to integrate the background (domain) knowledge with current news, and a methodology to study the quantitative relationships between news and financial instruments activities.