Bid rigging is particularly widespread in the construction sector. In such cases, construction companies coordinate their bids with one another during tender procedures. This restricts competition and results in customers paying unjustified price mark-ups.

Bid-rigging cartels have repeatedly been uncovered in Switzerland. The largest case to date took place in the canton of Graubünden, where the Competition Commission (COMCO) conducted several high-profile proceedings between 2017 and 2019. In the largest single case alone, twelve companies colluded on around 650 road construction projects with a total volume of at least CHF 190 million between 2004 and 2010. [1] Most recently, COMCO opened several investigations in French-speaking Switzerland; 22 construction companies in the canton of Neuchâtel are currently under investigation [2]. In November 2025, COMCO also opened an investigation in the neighboring canton of Jura; in February 2026, it extended the proceedings to a total of 20 construction companies, with more than 150 tenders issued between 2016 and 2025 potentially affected. [3]

Competition authorities are often alerted to bid rigging by whistleblowers. However, COMCO was also among the first competition authorities worldwide to use statistical methods to detect collusion based on bidding behavior. It is now well established in the academic literature that data generated during tender procedures can be systematically analyzed for indications of collusion, making bid rigging detectable at an early stage. This article provides an overview of the scientific literature on data-based detection of bid rigging and discusses how it can be applied in practice.

Thirty Years of Research: A Brief Overview

The academic literature on detecting bid rigging broadly distinguishes three methodological approaches.

Econometric Tests

Early contributions made use of the fact that cartel members often submit sham bids that differ systematically from the bids of competitive bidders and can be identified using econometric methods. Porter & Zona (1993) and Bajari & Ye (2003) show that such bids, owing to their strategic nature, often bear little relation to the company's cost structure — in contrast to competitive bids, which correlate strongly with firm costs.

Statistical Screening Markers

Furthermore, the literature has developed indicators that can be derived primarily from the statistical distribution of bid prices. Determining these indicators requires no information about the underlying cost structures of the companies. Instead, they exploit the fact that bid rigging is typically based on specific strategic behavior.

For example, companies may agree that not all possible bidders submit an offer (“bid suppression”). In such cases, a conspicuously low number of bids would be expected. Alternatively, sham bids may be submitted to support the bid of a predetermined winner (“cover bidding”). In such cases, the dispersion of bids is expected to be particularly low. The coefficient of variation of the bids can therefore be used as a marker. Another marker measures the normalized distance between the lowest and second-lowest bid, since strategic bidders typically build in a safety margin. In the case of bid rigging, this relative distance measure is therefore often particularly high. The literature has empirically documented exactly these patterns for bid rigging, including for Switzerland (Imhof, Karagök & Rutz, 2018).

Machine Learning

“Machine learning” methods represent the newest strand of research. Huber & Imhof (2019), for example, combine the aforementioned screening markers with “supervised learning” methods (LASSO regression and ensemble methods), enabling them to train models to identify individual groups of firms as possible bidding cartels. In further work, the bid interactions of two firms are visualized as scatter plots and examined for visual patterns of collective coordination using “convolutional neural networks” (Huber & Imhof, 2023). Most recently, “graph attention networks” have also been used, which analyze the entire network of firm interactions across many tenders to identify collusive patterns within it (Imhof, Viklund & Huber, 2025). Such approaches allow the identification of incomplete cartels as well as cartels that divide contract awards among themselves on a rotating basis (“bid rotation”).

What Distinguishes Practice from Theory

The literature thus offers a wide range of methodological approaches, which have also been validated using historical data from known bid-rigging cartels. In practice, however, many of these approaches reach their limits:

• Econometric tests are analytically elegant and grounded in auction-theoretic results, but in practice they often fall short on data availability. Procuring authorities rarely have access to the necessary firm-specific cost data — these regularly constitute business secrets.
• Statistical markers are simple to apply and require only relatively small datasets and limited prior information. This is why they are relatively widely used in practice. For example, Swiss Economics developed a tool for a canton that uses established markers such as the coefficient of variation and the relative distance measure to check tender data for indications of bid rigging. It has been successfully in use for several years. Experience shows, however, that calibrating and interpreting these markers requires specific market knowledge. What appears conspicuous in one market may correspond to normal competitive behavior in another. A further valid objection is that bidders could adjust their bidding behavior once the markers being used become known. However, such an adjustment simultaneously limits the scope for actual price-fixing, since collusion and statistical inconspicuousness are difficult to optimize for at the same time.
• With “machine learning” approaches the challenge is that models need to be trained on large datasets to deliver reliable results. The first promising approaches come from Brazil [4], Spain [5], and the United Kingdom [6], where competition authorities have access to large, centrally available public procurement datasets. In Switzerland, where a large share of public procurement is carried out by cantonal bodies and procurement data is therefore not centrally available, there are, to our knowledge, no comparable projects to date.

What all these approaches have in common is that data-based screening cannot provide conclusive proof of bid rigging. It can only identify markets, tenders, or bidder constellations that show an increased probability of collusion. It cannot replace the classic burden of proof in competition law proceedings. Data-based screening can, however, direct the attention of procuring authorities and competition authorities specifically to suspicious constellations and prompt further investigation.

Holistic Assessment Instead of a Single Method

However, the limitations mentioned do not fundamentally call data-based screening into question. Rather, they suggest that, particularly where data availability is limited, a mix of methods is advisable — one that specifically combines the strengths of different approaches and supplements them with qualitative analyses. Swiss Economics has successfully pursued such an integrated approach in practice.

Figure 1: Example screening process at Swiss Economics

Source: Swiss Economics

At the heart of Swiss Economics’ approach is proactive screening during the ongoing procurement process. To this end, tender data is continuously fed into a screening tool. Using the statistical screening markers mentioned above and others, this tool identifies conspicuous bids and recommends them for closer examination. The results are stored in a database and serve as a basis for further analyses.

In addition, periodic market analyses based on the accumulated screening data can provide added value. These make it possible to identify structural risk factors, such as conspicuous price trends in individual regions, changes in the number of bidders, high market concentration, or the growing prevalence of stable bidder constellations. This allows temporal and spatial patterns to be identified that point to the formation of a bidding cartel.

In addition, “machine learning”-based analyses can be carried out. These can detect conspicuous behavior in recurring bidder constellations that point to a role switch between winners and losers. Such coordination mechanisms are barely visible in classic individual screening or in aggregated market data. “Machine learning” methods make them tangible.

The combination of these three elements enables a comprehensive assessment of collusion risks in a tender market: conspicuous bids are identified promptly, market developments are tracked holistically, and modern “machine learning” methods are deployed specifically where they deliver the greatest added value.

Conclusion

Over the past thirty years, data-based bid-rigging screening has evolved from a predominantly academic field of research into an increasingly practical instrument. The methods are diverse and powerful, but none of them work universally. For procuring authorities in Switzerland, where decentralized procurement structures limit the availability of large, centrally managed datasets, low-threshold, robust approaches are especially useful: statistical markers that do not require cost data and already deliver good results with small datasets. These can be meaningfully complemented by periodic market analyses and, where the data allows, “machine learning” methods. Ultimately, the question is which combination of methods is feasible and effective in the specific institutional context. Swiss Economics helps procuring authorities answer exactly this question.

Sources

[1] See COMCO press release from 3.9.2019

[2] See COMCO press release from 14.03.2024

[3] See COMCO press release from 17.02.2026

[4] See Antitrust regulator uses AI to uncover cartel in outsourcing

[5] See Premio a BRAVA, el proyecto de IA de la CNMC contra el fraude en contratación pública - CNMC Blog

[6] See UK uses AI to tackle 'bid-rigging' collusion in public procurement contracts | TechCrunch

References

• Bajari, P. & Ye, L. (2003). Deciding Between Competition and Collusion. The Review of Economics and Statistics, 85(4), 971–989.

• Huber, M. & Imhof, D. (2019). Machine Learning with Screens for Detecting Bid-Rigging Cartels. International Journal of Industrial Organization, 65, 277–301.

• Huber, M. & Imhof, D. (2023). Flagging Cartel Participants with Deep Learning Based on Convolutional Neural Networks. International Journal of Industrial Organization, 89.

• Imhof, D., Karagök, Y. & Rutz, S. (2018). Screening for Bid Rigging – Does It Work? Journal of Competition Law & Economics, 14(2), 235–261.

• Imhof, D., Viklund, E.W. & Huber, M. (2025). Catching Bid-Rigging Cartels with Graph Attention Neural Networks. arXiv:2507.12369.

• Porter, R.H. & Zona, J.D. (1993). Detection of Bid Rigging in Procurement Auctions. Journal of Political Economy, 101(3), 518–538.

Do you have feedback or questions? Do not hesistate to get in contact directly by mail.