Identifying Manipulations in Russian Procurement Contests: Machine Learning Approach

Sealed-bid scoring auction is a frequently used mechanism for determining the winner in public procurement. Due to the complexity and flexibility of the participant’s evaluation process, they are particularly vulnerable to different forms of manipulations. It provokes the appearance of corrupt participants and buyers. Exaggeration of non-price score, pre-agreed scoring rules and leaking bids are among the most popular ways to distort a contest's outcome. The purpose of this work is to estimate the proportion of corrupted auctions in Russia as well as to describe precise strategies of these manipulations using a machine learning approach. The main difficulty of this task is the absence of corruption labels, which makes supervised learning impossible. This problem is solved by positive-unlabeled classification method. This approach allows us to estimate a prior probability of corruption within the sample and then to predict a posterior probability of being corrupted for each participant. The model’s predictions are used for the further analysis of possible manipulation techniques and other patterns related to a higher corruption probability. We extract and analyze the data on 40,000 Russian procurement auctions between 2014 and 2019. According to the results, approximately 25.7% of auctions can be considered as suspicious. Generally, the manipulation method depends on the weights of the criteria in the total score of an auction, but the majority of suspicious ones is represented by non-price score exaggeration. We also find that corruption is more likely to be committed if the purchase object is relatively expensive and if the winner has met the procurer in earlier auctions.