There are two strategies for implementing such methods. First, one can treat the elections believed to be accurate as a training set, estimate the features of the reference distribution from them, and compare the evaluated elections to those estimates. In that vein, Cantu and Saiegh 2011 use synthetic data to obtain the features of accurate elections. Alternatively, one can evaluate whether both the accurate elections and the elections under scrutiny can be fit adequately using the same numeral distribution (Medzihorsky 2015). The second option in effect means that the task is no longer to classify whether a set of returns has irregularities or not, but rather whether the distribution of the inspected returns is “close enough” to those believed to be accurate.