Optimizing Tax Administration Policies with Machine Learning

Tax authorities around the world are increasingly employing data mining and machine learning algorithms to predict individual behaviours. Although the traditional literature on optimal tax administration provides useful tools for ex-post evaluation of policies, it disregards the problem of which taxpayers to target. This study identifies and characterises a loss function that assigns a social cost to any prediction-based policy. We define such measure as the difference between the social welfare of a given policy and that of an ideal policy unaffected by prediction errors. We show how this loss function shares a relationship with the receiver operating characteristic curve, a standard statistical tool used to evaluate prediction performance. Subsequently, we apply our measure to predict inaccurate tax returns issued by self-employed and sole proprietorships in Italy. In our application, a random forest model provides the best prediction: we show how it can be interpreted using measures of variable importance developed in the machine learning literature.