Transnational machine learning with screens for flagging bid‐rigging cartels

We investigate the transnational transferability of statistical screening methods originally developed using Swiss data for detecting bid‐rigging cartels in Japan. We find that combining screens with machine learning (either a random forest or an ensemble method consisting of six different algorithms) to classify collusive versus competitive tenders entails (depending on the model) correct classification rates of 88%–97% when training and testing the method on the Okinawa bid‐rigging cartel. As in Switzerland, bid rigging in Okinawa reduced the variance and increased the asymmetry in the distribution of bids. When training the models in data from one country to test their performance in the data from the other country, imbalance increases between the correct prediction of truly collusive and competitive tenders for all machine learners and classification rates go down substantially when using the random forest as machine learner, due to some screens for competitive Japanese tenders being similar to those for collusive Swiss tenders. Demeaning the screens reduces such distortions due to institutional differences across countries such that correct classification rates based on training in one and testing in the other country amount to 85% and to 90% when using the ensemble method as machine learner, which generally outperforms the random forest.