A machine learning approach to detect collusion in public procurement with limited information

Public procurement spending reaches $9.5 trillion annually, yet bid-rigging remains a significant, undetected issue. Existing methods require detailed data often unavailable to competition authorities. Practitioners and researchers need effective and flexible tools with moderate data requirements to examine collusive behavior. This paper proposes a novel algorithm to detect collusion in auctions using readily available data about outcomes of public procurement processes. This method leverages theoretical findings about bidding behavior and machine learning algorithms. We demonstrate its effectiveness on data sets with known collusion cases (Italy, Japan, USA) and achieve superior results compared to traditional models. Further analyses on broader data sets encompassing Turkish and European contracts reveal a significant percentage (over 6% in Turkey, 4.5% in Europe) with high collusion probability. Collusion inflates procurement costs by 3-7%, highlighting the need for effective detection methods.