My past Ph.D. research has resulted in many publications and, naturally, a dissertation, which was completed, published, and defended succesfully in 2014.
The flourishing data market and the financial resources pumped into research on the extraction of knowledge from data underline today's major stakes that are involved with the accurate extraction of knowledge and efficient usage thereof in financial decision making. Not only traditional and ubiquitous numerical data, but especially textual data such as news messages, are gradually receiving more attention. Such data often remain unused, due to their unstructured nature thwarting automatic processing. Thus, often large amounts of valuable knowledge remain undiscovered. Generally, this knowledge can be summarized in events, which are inextricably linked with financial markets. It has long been known that events, like acquisitions, product launches, natural disasters, or wars, could exert a notable influence on stock rates. Smart applications of detected events may therefore be beneficial for financial decision making. For instance, trading algorithms can be enriched with events, so as to better respond to new developments. Also, taking into account certain events may improve financial risk estimations. Because of the promising applications, but also the many inevitable challenges, my dissertation comprises a number of topics related to the semi-automatic detection of (financial) events in news text.
Taking into consideration the results of a detailed evaluation of currently existing knowledge-driven, data-driven, and hybrid extraction systems and methods focusing on events, my thesis presents a semi-automatic system for financial event extraction from news text. The system consists of various innovative, qualitatively competitive, and knowledge-driven components for natural language processing. Their interaction with a knowledge base fosters a feedback loop, so that newly detected events can be digested, allowing the incorporated knowledge to be used in future (extraction) processes.
In addition, two languages are proposed that can refine the aforementioned system. While the first language is focused on pattern-based event extraction from text, the second language is targeted toward the definition and execution of knowledge base updates, associated with the extracted events. After conducting a series of experiments focusing on pattern development times and result accuracies, it can be concluded that the extraction language offers, in contrast to many modern alternatives, a simple and flexible notation utilizing lexical, syntactical, and semantical elements, while maintaining expressivity. Moreover, an evolutionary approach for automatically generating patterns contributes to the practical employability of the language. The trigger-based knowledge base update language and various developed execution models are particularly suitable for event extraction applications. While modern languages are often less suited for fully automated updates, now an increased flexibility is offered for automatically executing predefined rules, providing the user with a wide range of options, e.g., immediate or deferred execution, update chaining, etcetera.
Last, two financial applications of events extracted from news text are presented in my dissertation. For both applications it is confirmed that the addition of events to prevailing computations or algorithms can yield more accurate results. In our analysis of an event-based automated trading application, rules are generated for trading stocks. The best performing rules do not only make use of numerical signals such as average historical stock rates, but also employ news-based event signals. Moreover, when cleaning stock data from disruptions caused by financial events, financial risk analyses yield more accurate results.
The reported results suggest that it is possible to semi-automatically and accurately detect events in news text in a knowledge-driven way, when making use of advanced extraction rules. Additional update rules and execution models accommodate a feedback loop to knowledge bases underlying event extraction systems. Events detected in news can be used as additional parameters in financial applications, thus yielding more accurate outcomes in the evaluated cases. Such advantageous applications can be of good use in (semi-)automated environments.
My Ph.D. work is linked to the Netherlands Organisation for Scientific Research (NWO) Physical Sciences (EW) Free Competition project 612.001.009: Financial Events Recognition in News for Algorithmic Trading (FERNAT). NWO funds thousands of top researchers at universities and institutes and steers the course of Dutch science by means of subsidies and research programmes. The NWO-EW Free Competition is intended for top innovative and risky fundamental research proposals that have a high scientific or practical urgency and that are significant for at least one of the disciplines of astronomy, computer science, or mathematics.
