Predicting the Deterrence Effect of Tax Audits. A Machine Learning Approach

We apply machine learning methods to the prediction of deterrence effects of tax audits. Based on tax declarations data, we predict the increase in future income declarations after being targeted by an audit. We find that flexible models, such as classification trees and ensemble methods based on them, outperform penalized linear models such as Lasso and ridge regression in predicting taxpayers more likely to increase their declarations after an audit. We show that despite the non-randomness of audits, their specific time structure and the distribution of changes in declared amounts suggest a causal interpretation of our results; that is, our approach detects a heterogeneity in the reaction to a tax audit, rather than just forecasting an unconditional future increase. We find that taxpayers identified by our model will on average increase their declared income by €14,461—the average among all audited taxpayers being €−205. Our approach allows the tax agency to yield significantly larger revenues by appropriately targeting tax audit.