Evolving Multi-Variate Time-Series Patterns for the Discrimination of Fraudulent Financial Filings

ABSTRACT This paper considers an application of evolutionary computation (EC) to classification and pattern discovery. In particular we present a genetic algorithm (GA) utilized to discriminate cases of potential financial statement fraud. Of key interest to us is the ability to distinguish multidimensional patterns over time. The GA evolves strings over a pattern definition language to define class boundaries and to select classification features. The language defined allows for 1) the integration of data across time and across a number of variables 2) the integration of quantitative as well as qualitative data 3) the direct evolution by genetic algorithm and 4) easy interpretation by human experts. The data and method are described and results presented. Results offer a 63% true positive rate with a false positive rate of 5%. These results compare favorably with other published results on comparable data. Our technique captures behaviors not evident from traditional data analysis methods. The output from our system has the additional benefit of being easily understood and utilized by experts and practitioners in the field. This makes our approach more desirable than other black-box solutions. These techniques provide a foundation for multidimensional behavior analysis of data from a variety of domains including, financial, biological, manufacturing and clinical.