Fırat Bilgel

Tacit collusion is a market conduct that enables firms to raise prices, produce less output and obtain supra-normal profits, considered to be illegal in the United States, Canada and most of the EU countries due to antitrust laws. Some industries, such as steel, cement and banking, are particularly known to be prone to collusion, characterized by a high degree of concentration, similar cost structures, significant barriers to entry and product homogeneity. However, as current practices have shown through my field experience as a former consultant, detection of collusive behavior is a very difficult and occasionally a heuristic process because (1) the enforcement faces the difficulty of finding hard evidence, (2) the predictions of the enforcer/regulator or the courts may be imperfect and/or (3) the existing detection tools may be unable to isolate the confounding factors that indicate collusive behavior when there is actually none (type I error) or non-collusion when there actually is (type II error).

 

This paper aims to (1) introduce a powerful routine for detecting collusive behavior using counterfactuals; (2) assess the extent to which the imputed counterfactual is aligned with the claims of the competition authorities and the decision of the courts using data on landmark cases in the EU and the US and (3) identify whether collusive behavior can be well-predicted via machine learning tools for effective monitoring.

 

I propose the use of the synthetic control, to ex-post detect collusive practices. The idea is to reproduce the colluders' price trajectory before they enter into a collusive agreement by using the convex combination of firms in the same relevant market that are known to be unrelated to the collusion. Causal inference can be carried out by comparing the price level of the synthetic firms (non-colluders) against colluders in the collusion period. This allows us to (i) identify whether unfavorable price movements are the consequence of collusive behavior (through the imputation of the counterfactual and falsification tests), (ii) measure the intensity of collusion through time by the size of post-collusion gap between the observed and the counterfactual; and (iii) identify firms that are more likely to steer collusive behavior (through an exercise called, leave-one-out synthetic controls). Using data on cases in the EU and the US, a comparison of inferences drawn from the synthetic control and the actual court decisions can be made to assess the extent to which courts commit type I and type II errors. The identification of the cause and effect serves not only to design optimal law enforcement and punishment schemes, but also to carry out out-of-sample prediction of firm behavior through machine learning algorithms such as boosting regression to which effective monitoring of collusive behavior and facilitating practices hinge upon.

​

JEL Classification: C15; C54; K21; K42 L41; L51

Keywords: Synthetic control, collusive behavior, boosting regression