Quantifying the risk of financial events using kernel methods and information retrieval

A financial event is any happening which dramatically changes the value of a firm. Examples of financial events are management fraud, bankruptcy, exceptional earnings announcements, restatements, and changes in corporate structure. This dissertation creates a method for timely detection of financial events using machine learning methods to create a discriminant function. As there are a myriad of possible causes for any financial event, the method created must be powerful. In order to increase the power of current methods of detection text related to the company is analyzed together with quantitative information on the company. The text variables are chosen based on an automatically created accounting ontology. The quantitative variables are mapped to a higher dimension which takes into account ratios and year-over-year changes. The mapping is achieved via a kernel. Support vector machines use the kernel to perform the learning task. The methodology is tested empirically on three datasets: management fraud, bankruptcy, and financial restatements. The results show that the methodology is competitive with the leading management fraud detection methods. The bankruptcy and restatement results show promise.