Where is the Limit? Assessing the Potential of Algorithm-Based Cartel Detection

Academic research on cartel detection has primary focused on algorithm-based screening of markets. However, only a few studies have assessed the extent to which competition authorities can generalize from a developed model to new markets. In our paper, we aim to fill this gap by investigating how close a market on which an algorithm is trained has to be to detect cartels in a new market. Our results confirm that when comparable training data are available, the machine-learning-based models are powerful tools to flag cartels. However, we show that the algorithms’ performances are limited when lacking comparable training data from the same industry, leading to the conclusion that practitioners should exercise a great deal of caution when choosing training data from different industries for cartel screening. In addition to our main contribution, we present a way of automated feature engineering and selection based on frequently used hand-crafted screens (descriptive statistics derived from firms’ prices) that are generally used in the recent cartel screening literature of algorithm-based cartel detection. Finally, to overcome the prerequisite of any pre-defined screen, we present the first-time application of various recurrent neural network architectures together with raw price data to flag potential cartels.