Machine Learning in Cartel Screening—The Case of Parallel Pricing in a Fuel Wholesale Market

Abstract

The detection and deterrence of collusive agreements among firms, such as price-fixing cartels, remain pivotal in maintaining market competition. This study investigates the application of machine learning methodologies in the behavioral screening process for detecting collusion, with a specific focus on parallel pricing behaviors in the wholesale fuel market. By employing unsupervised learning techniques, this research aims to identify patterns indicative of collusion—referred to as collusion markers—within time series data. This paper outlines a comprehensive screening research plan based on the CRISP-DM model, detailing phases from business understanding to monitoring. It emphasizes the significance of machine learning methods, including distance measures, motifs, discords, and semantic segmentation, in uncovering these patterns. A case study of the Polish wholesale fuel market illustrates the practical application of these techniques, demonstrating how anomalies and regime changes in price behavior can signal potential collusion. The findings suggest that unsupervised machine learning methods offer a robust alternative to traditional statistical and econometric tools, particularly due to their ability to process large and complex datasets without predefined models. This research concludes that these methods can significantly enhance the detection of collusive behaviors, providing valuable insights for antitrust authorities.