A Machine Learning Approach to Analyzing Corruption in Local Public Finances (with Sergio Galletta and Tommaso Giommoni)

This paper applies machine learning tools to detect local-government corruption using budget accounts data. In the context of Brazilian municipalities, we have gold-truth labels for corruption from a set of federally mandated audits (assigned randomly by lottery) for the years 2003 to 2010. Our tree-based gradient-boosted classifier can predict the presence of corruption in held-out test data, consistent with an expected link between corruption and budget composition. The trained model, when applied to new data, provides a synthetic measure of corruption which can be used for new empirical analysis. We confirm the empirical usefulness of this measure by replicating, and extending, some previous empirical evidence on corruption issues in Brazil. In particular, we exploit the longitudinal nature of our data to produce new evidence on the dynamic effects of audits on corruption, including spillover effects on neighboring municipalities.