The Role of Imputation in Detecting Fraudulent Financial Reporting

Financial fraud detection plays a crucial role in the stability of institutions and the economy at large. Data mining methods have been used to detect/flag cases of fraud due to a large amount of data and possible concept drift. In the financial statement fraud detection domain, instances containing missing values are usually discarded from experiments and this may lead to a loss of crucial information. Imputation has been previously ruled out as an option to keep instances with missing values. This paper will examine the impact of imputation in financial statement fraud in two ways. Firstly, seven similarity measures are used to benchmark ground truth data against imputed datasets where seven imputation methods are used. Thereafter, the predictive performance of imputed datasets is compared to the original data classification using three cost-sensitive classifiers: Support Vector Machines, Näıve Bayes and Random Forest.