Public procurement cartels : a large-sample testing of screens using machine learning

Abstract

Due to the high budgetary costs of public procurement cartels, it is crucial to measure and understand them. The literature developed screens that work well for selected cartel types and with high quality data, but it didn’t produce generalisable knowledge supporting policy and law enforcement on typically available datasets. We simultaneously measure multiple cartel behaviours on publicly available data of 73 cartels from 7 European countries covering 2004-2021. We apply machine learning methods, using diverse cartel screens characterising pricing and bidding behaviours in a predictive model. Combining many indicators in a random forest algorithm achieves 70-84% prediction accuracy, distinguishing behavioural traces of confirmed cartels from non-cartels across different cartel types and countries (accuracy is 97% when trained and tested on a single cartel case, typical of the literature). Most screens contribute to prediction in line with theory. These results could improve cartel detection and investigations and support pro-competition policies.