Hannes Wallimann, Silvio Sticher
9 min
As the field of industrial organization shifts from purely theoretical models to empirical analysis, students and practitioners face the challenge of needing both economic intuition and data science proficiency. This paper introduces a classroom game designed to bridge this gap, allowing students to experience the strategic dynamics of market collusion and the technical challenges of detecting it using machine learning.
The game is structured into three distinct parts, simulating a procurement market in the construction sector. In the first two parts, participants act as construction firms. In Part 1, they operate in an oligopoly without communication. In Part 2, they are permitted to coordinate, simulating illegal cartel behavior. In the final part, participants transition to the role of antitrust authority staff. They use machine learning algorithms—ranging from simple visual analysis to random forest models—to analyze bid data from the previous rounds and flag suspicious patterns. The game incentivizes accurate detection by rewarding correct classifications and penalizing false positives, which mirror the real-world reputational and legal risks faced by competition authorities.
This game-based learning approach addresses the multidisciplinary nature of modern economics. By breaking the activity into parts and integrating it with theoretical inputs, the authors aim to maintain long-term student engagement. The paper provides a flexible framework that can be adapted to different levels of data science expertise, ensuring that students gain practical experience in applying statistical markers and machine learning to real-world economic problems like bid-rigging.
We present a classroom game that integrates economics and data-science competencies. In the first two parts of the game, participants assume the roles of firms in a procurement market, where they must either adopt competitive behaviors or have the option to engage in collusion. Success in these parts hinges on their comprehension of market dynamics. In the third part of the game, participants transition to the role of competition-authority members. Drawing from recent literature on machine-learning-based cartel detection, they analyze the bids for patterns indicative of collusive (cartel) behavior. In this part of the game, success depends on data-science skills. We offer a detailed discussion on implementing the game, emphasizing considerations for accommodating diverging levels of preexisting knowledge in data science.
Alex: So a high bid need not mean coordination. It might reflect a genuine disadvantage in doing that particular job.
Sam: That’s the economic reasoning the setup asks students to use. The paper gives a railway-project example where a firm’s location and specialization determine additions to its base cost. Participants calculate their own costs before deciding how much to bid.
Alex: How do those decisions change across the game? I’m interested in what participants are allowed to coordinate, not just what the instructor hopes happens.
Sam: Initially, communication is prohibited. Firms bid using their costs and expectations about competitors. Then the bidding exercise repeats with communication allowed, giving firms the opportunity to coordinate. Across those phases, each group generates 32 tenders for later analysis.
Alex: Are participants rewarded simply for winning contracts, or for making money on them? That difference could change their bidding strategy.
Sam: The lowest bid wins, and the winner earns points equal to the bid minus its cost. Losing firms receive no points. Participants learn who won, but not the winning bid, so feedback supports strategic adjustment without revealing every pricing decision.
Alex: Then they switch sides. How do you stop someone recognizing their own bids and calling that detection?
Sam: The instructor pools and randomly rearranges the tenders. That makes participants’ own submissions harder to recognize, although the paper doesn’t claim recognition becomes impossible. Acting as competition-authority employees, they classify each tender as suspicious or non-suspicious.
Alex: What information goes into that classification? Raw prices alone don’t tell us whether bids are unusually close or unusually spread out.
Sam: The baseline proposal uses behavioral screens: statistics summarizing the bids within a tender. They capture features such as overall variation, the gap between the lowest bids, and that gap relative to differences among other bids. These summaries become inputs to a prediction model.
Alex: So rather than asking a model to understand the whole market, students give it economically motivated summaries of bidding behavior.
Sam: For supervised learning, meaning training on examples with known labels, the authors propose data from previously studied Swiss cartels. Students fit their models using that material, then classify the classroom tenders. The teaching exercise therefore connects existing research data with decisions students have experienced themselves.
Alex: But the classroom market isn’t the Swiss market. Does the document show that a model trained there transfers successfully to these student bids?
Sam: It doesn’t report that test. Nor does it report classroom classification accuracy. The external training data are part of the proposed procedure, not evidence that the procedure succeeds in this setting.
Alex: There’s another potential mismatch. Permission to communicate doesn’t guarantee collusion. How does the scoring define the correct answer?
Sam: The game treats tenders from the no-communication phase as non-suspicious and those from the communication-permitted phase as suspicious. That supplies convenient labels, but it does not establish collusion tender by tender. It’s a simplification readers should keep separate from real-world evidence.
Alex: And the task is screening, not proving an offence. Does the reward system preserve that distinction?
Sam: The story makes flagged tenders candidates for further investigation. If at least 75 percent of participants flag a tender, the game triggers that investigation. Points earned during the communication-permitted phase can then be revoked.
Alex: What discourages an investigator from flagging everything? They might catch more suspicious cases while accusing plenty of competitive firms.
Sam: A false positive, meaning a competitive tender incorrectly flagged, costs the participant five percent of their points. Eligibility to win also requires classification accuracy at or above the class median. The rules create a tradeoff between caution and overall accuracy.
Alex: That tradeoff sounds teachable even for students who cannot build a model. How much programming does the activity require?
Sam: The authors offer different routes. Advanced participants can choose algorithms and train models themselves. Those with basic knowledge can use prepared code and adjust the decision threshold: how strong the model’s suspicion must be before they flag a tender.
Alex: Raising that threshold could reduce false accusations, while missing more suspicious tenders. What’s the route for students without coding experience?
Sam: Visual inspection or simple numerical comparisons. The paper cites evidence that colluding firms often submit closely aligned bids, so students can examine bid distributions for such patterns. The authors also suggest allowing visual judgment to override model classifications, including in advanced classes.
Alex: We have a detailed mechanism, but not evidence of learning. What would be an overstatement of this document’s contribution?
Sam: Saying that the game improves learning, sustains motivation, or outperforms conventional instruction. The working paper explicitly precedes its first classroom implementation. Its contribution is a specified teaching design and advice for adapting it, not a measured educational effect.
Alex: Given that boundary, who should read the full document, and where should they begin?
Sam: Economics educators and interdisciplinary methods instructors should start with Section 4.2, on adapting the activity to prior data-science knowledge. Then read Section 3 for the bidding mechanics and scoring incentives. Those incentives matter because they shape both the market behavior and the later classifications.
Alex: And for researchers who aren’t planning a class, what’s the useful idea to retain?
Sam: Let learners produce the behavior before asking them to detect its traces. That is the proposal’s distinctive teaching move. Whether it improves learning remains an open question.
Alex: A teaching design to try, not a teaching effect already shown.