Estimating cartel damages using machine learning

This paper presents an alternative to the workhorse linear OLS regression model when predicting “but-for” prices in cartel damage estimation. By replicating the dataset from a prominent Vitamin C antitrust case of price-fixing, I show that a supervized machine learning algorithm achieves more accurate predictions of prices based on cross-validated out-of-sample testing than the OLS model applied by the court expert. The machine learning algorithm is therefore better suited in this case to predict prices for the counterfactual scenario that no cartel existed, and to calculate damages based on those predictions. I find that the machine learning algorithm p redicts damages that are 14% lower than those predicted by the court expert's OLS model. Given that millions of dollars are usually at stake in cartel litigation cases, it is recommended that machine learning algorithms should be in the toolbox of practitioners attempting to derive the most accurate estimate of cartel-related damages.