The use of descriptive statistics in detecting collusion in the Finnish asphalt industry

This thesis studies a method of cartel screening where descriptive statistics from the bid distribution in first-price sealed-bid auctions are used to detect collusive tenders. This screening method is introduced by Huber and Imhof (2019) who combine the bid distribution-based screens with a machine learning model to ex-post predict collusion from a Swiss road construction sector. In this thesis, the same screens and methodology are used to ex-post predict collusion from the Finnish asphalt industry. The results are then compared to the Swiss ones. Huber and Imhof (2019) introduce a total of 10 screens. These are based on two phenomena which are expected to happen under collusion: the convergence of bids and the emergence of a gap between the winning bid and the losing bids. The analysis conducted for the Finnish data shows that most of the screens exhibit statistically significant differences between collusive and competitive tenders. This would suggest that the Finnish cartel affected the distribution of the bids. Hence, the screens are potentially useful in detecting collusive tenders in the Finnish data. With the Finnish data, the predictive model, which uses as inputs a polynomial of the 10 screens introduced by Huber and Imhof (2019), was, on average, able to predict correctly 83.3 % of the tenders. However, this result was somewhat affected by the larger share of competitive tenders in the data (69 % of observations). This can be seen from the fact that for competitive tenders, the model was able to predict correctly 96.5 % of the tenders whereas for collusive tenders the prediction rate was only 53.2 %. The comparable prediction rate with the Swiss results was 74.9 % for the Finnish data and 83.8 % for the Swiss data. The results of this thesis show that the performance of the model was not as good as it was for the Swiss data. However, the model did have a clear improvement in the prediction rate compared to simply guessing the tenders. Whether the performance was good enough, is ultimately in the hands of the authorities planning to use to model. Lastly, this thesis suggests a novel way of interpreting the results of the predictive model: instead of looking at the predictions of individual tenders, one could search for structural breaks in collusion probabilities, and whether these breaks could be a sign of cartel behavior