Won’t Get Fooled Again: A supervised machine learning approach for screening gasoline cartels

Abstract

We combine supervised machine learning techniques with statistical moments of the gasoline price distribution to detect cartels in the Brazilian retail market. Standard deviation, coefficient of variation, spread, skewness, and kurtosis are predictors that can help identify and predict anti-competitive market behavior. We evaluate each classifier and discuss the trade-offs related to false-positive (detect cartel when it does not exist) and false-negative (do not detect cartel when it does exist) predictions. The competition authority needs effective monitoring and often anticipating cartel movements. With this in mind, we test the algorithms’ performance in new datasets (ex-ante screening). Our results show that false-negative outcomes can critically increase when the main objective is to minimize false-positive predictions. The models’ overall average scoring rate for testing and predicting cartels in the same city is 96.22%. When we train the algorithms in one city and predict the cartel outcomes in other cities, on average, the overall scoring rate is equal to 73.75%. Our work suggests that machine learning classifiers have positive attributes and can provide valuable contributions to cartels’ deterrence. In addition, we offer a policy prescription discussion for antitrust authorities regarding the pros and cons of proactive tools for inhibiting collusive agreements in retail gasoline markets.