[Paper Review] Detecting collusion in procurement auctions
The paper develops a machine learning model to predict signs of bidder collusion in Russian procurement auctions, achieving 91% accuracy on a 40-auction dataset split into 30/70 train/test, with model interpretability via the Shepley vector decomposition and validation through independent simulations.
The study aimed at detecting cartel collusion involved analyzing decisions of the Russian Federal Antimonopoly Service and data on auctions. As a result, a machine learning model was developed that predicts with 91% accuracy the signs of collusion between bidders based on their history after dividing 40 auctions into test and training samples in a 30/70 ratio. Decomposition of the model using the Shepley vector allowed the interpretation of the decision-making process. The behavior of honest companies in auctions was also studied, confirmed by independent simulation validation.
Motivation & Objective
- Motivate the detection of cartel collusion in procurement auctions.
- Develop a predictive model to identify signs of bidder collusion from historical auction data.
- Assess the model's interpretability and validate findings with simulation evidence.
Proposed method
- Construct a dataset of Russian procurement auctions and bidder histories.
- Split data into training (70%) and testing (30%) for evaluation.
- Train a machine learning model to predict collusion signs from historical decisions.
- Interpret model decisions using the Shepley vector decomposition.
- Validate honest-bidder behavior through independent simulations.
Experimental results
Research questions
- RQ1Can historical bidder decisions predict signs of collusion in procurement auctions?
- RQ2What is the predictive accuracy of a model trained on past auction data for detecting collusion?
- RQ3Does a decomposition method (Shepley vector) provide interpretable insights into the decision process?
- RQ4Do simulation studies corroborate the model’s findings about honest bidder behavior?
Key findings
- The model predicts collusion signs with 91% accuracy on the test split.
- A 30/70 train/test split was used across 40 auctions.
- Shepley vector decomposition enabled interpretation of the model’s decision-making process.
- Independent simulations confirmed the observed behavior of honest companies in auctions.
Better researchstarts right now
From reading papers to final review, dramatically reduce your research time.
No credit card · Free plan available
This review was created by AI and reviewed by human editors.