This paper proposes a tax auditing strategy capable of separately identifying tax evasion and unintentional misreporting using random yearly announcements. In the Netherlands, the tax authorities announce at the beginning of the taxpaying season one or several specific components in tax reports which will be heavily audited. Facing these unanticipated announcements, taxpayers who previously misreported their taxes adjust their declarations in revealing ways. In the paper I first present an incentive based theoretical model which classifies different types of behavior. Using these classifications, I then apply statistical learning methods (data mining) on an unprecedented administrative dataset collected from the Dutch tax authorities to extract individual probabilities of belonging to each type of misreporter. The paper further explains how these individual propensities can allow the tax authorities to discover patterns in misreporting and formulate optimal auditing strategies.