Who are you? Cartel detection using unlabeled data

Abstract

We propose a data-driven machine learning approach to flag bid-rigging cartels in the Brazilian road maintenance sector. First, we apply a clustering algorithm to group the tenders based on their attributes. Second, we use the labels created by the clustering algorithm as a target variable to predict them using a classifier. We rank the screens according to their relevance to decrease the number of false positive (detecting cartel when it does not exist) and false negative (not detecting cartel when it does exist) predictions. Our results shed light on the need to use a range of screens to recognize the vast profile of strategies practiced by bid-rigging cartels, such as misleading competitive dynamics, bid combination, and cover bidding behavior. Our method can improve cartels’ deterrence in different economic sectors, especially when labeled data are not available. In a controlled environment with a simulated labeled dataset, the overall average accuracy of the algorithm is 99.33%. In a real-world cartel case with a labeled dataset, the overall average accuracy is 80.25%. When applied to the road maintenance unlabeled dataset, our model identified a group containing 273 (31% of the total) suspicious tenders. We conclude by offering a policy prescription discussion for antitrust authorities.