Using Prediction Models to Design Tax Enforcement: Incentives vs Targeting

We study tax audit policies when the Tax Authority predicts true income using an inference model. When taxpayers are aware of model-based audit rules, using an inference model shapes both declaration incentives and the targeting of tax audits. The Tax Authority can achieve arbitrarily high tax collection rates if the model’s precision is sufficiently high. However, the targeting of audits yields minimal revenues as optimal reliance on the model focuses on enhancing the incentives to declare income in the first place. Prediction power is used to shape incentives rather than to direct audits. At the optimum, the predictions from the statistical model are used to screen larger true income taxpayers, tolerating evasion from taxpayers with lower incomes and high propensity to evade. We corroborate and extend our theoretical findings with numerical simulations calibrated on aggregate moments from administrative audit data. Enhanced model precision reduces tax evasion, particularly among higher incomes, thereby alleviating the inequality in effective tax rates induced by optimal enforcement. While plausible enhancements in model precision yield modest revenue gains, these gains are substantial compared to the audit budget increase required to achieve similar tax revenues without an inference model.