The use of machine learning in cartel detection
by Government Transparency Institute, DATACROS III partner
Public procurement engulfs roughly 13%-20% of global GDP (based on OECD1 and World Bank2 estimations). But due to entry costs, transparency, and frequent new tenders, the sector is highly vulnerable to collusion—illegal agreements between competing firms.
To ensure tenders fall their way, colluding firms coordinate prices, market shares, or timing – indeed, the United Nations Office on Drugs and Crime estimates that public procurement costs are inflated by 10% to 25% due to collusive behavior3. Given the size of the sector, even minor distortions can translate into major impacts and the waste of public resources.
Colluding firms – also known as cartels – are by nature covert and adaptive. Traditionally, detection was based on leniency policies, that is waiting for a whistleblower to come forward and spoil the party. But today, as administrative big data have become more available, investigations can be triggered by automated red flags.
In DATACROS III red-flags and anomaly indicators are used to signal to LEA and investigative journalists the potential involvement or infiltration of organized and financial crime actors – in a firm. Starting in 2019, DATACROS has been developing increasingly accurate and effective indicators, alongside practitioners and academics. In its third iteration, the Government Transparency Institute brings to the consortium a know-how and experience with public procurement tenders – with the aim of integrating both a European-wide coverage of tenders and a new set of indicators to aid DATACROS III users in evaluating them.
That being said, such red flags are not the final frontier in terms of analyzing public procurement. In a sense, they are limited, primarily capturing predefined risk patterns, overlooking the continuously evolving nature of cartels’ strategies. While previous research has focused largely on single-case studies or custom-built indicators, there remains a significant policy gap: the need for a set of tools that can be applied across countries and adapted to different types of markets. With the advent of machine learning, a path towards filling this gap is becoming clear.
Instead of following a provided set of rules, machine learning is a form of artificial intelligence where the computer studies data – looking for patterns in the information. Machine learning algorithms process large quantities of data and learn from a process of trial and error during what’s referred to as training. As a result, they can make predictions – like when a contract is the product of collusion.
A recent systematic mapping study reviewed 93 academic articles and found that machine learning classification models now dominate the research landscape for procurement fraud detection4. These tools work by analyzing past tenders for features commonly associated with collusion (such as irregular winning markups, bid rotation, or subcontracting patterns) and learning to distinguish between potentially collusive contracts and those that were not.
Recent research has found that using bid-level indicators alongside various machine learning models, trained on cartels’ strategies, can significantly improve accuracy in spotting cartels5. For example, bid-level indicators trained using proven cartel cases in Switzerland were tested on a cartel case in Japan, despite the stark differences between the two markets, cartel classification achieved accuracy rates of 88%–97%6. These recent advances suggest that machine learning can provide scalable, data-driven support for competition authorities seeking to find procurement cartels. Just as well, advancements are not limited to the scope of classification models; on the contrary, experimental and emerging modelling tools are also being explored, pioneering new, further ways to identify suspicious cartel behavior7-8.
Collusion practices are diverse – some cartels rotate winners, others mimic competition through cover bids. With machine learning models being trained, typically, on labeled examples, where the model has prior knowledge of what to look for (for public procurement, these are indicators that capture specific dimensions of collusion like the number of bids or price deviations) this variation poses a challenge. Further still, the data itself may be an obstacle. Many prior studies rely on small datasets, or on information typically not published by or available to governments—like losing bid prices, consortium memberships, or subcontracting details. These data gaps have limited the practical deployment of machine learning models in real-time procurement monitoring systems.
Therefore, we face an important question: can models that reliably detect collusion across cases, sectors, and countries be successfully developed and implemented?
In our study, we find an answer. We’ve built a universal machine learning model trained on cross-country cartel procurement data and demonstrated its potential for wide-scale, practical application.
Steps towards a Generalizable Procurement Collusion Detection Model
Our research set out to build a more general and practical machine learning model for detecting collusion in public procurement9. The models developed in this study work with standard, publicly available procurement data and deliver consistently reliable accuracy rates across countries, sectors, and cartel types.
We first collected large-scale procurement datasets from seven European countries—Bulgaria, France, Hungary, Latvia, Portugal, Spain, and Sweden—covering the 2004–2021 period. Then, we classified contracts as collusive or competitive based on 73 prosecuted cartel cases depending on whether the colluding companies won them before or after the start of an official investigation. We used the dates identified in court documents as the start and end points of cartel activity, assuming that companies stopped colluding once caught. This classification allowed us to examine how collusion indicators behave over different time periods—before, during, and after a given cartel’s operation (see Figure 1).
We create our training dataset in a couple of steps. First, we identify all contracts won by prosecuted companies (red symbols in Figure 1). Second, we distinguish between contracts won during the proven cartel period versus those won before or after. As a result, this also allows to distinguish different markets where colluding companies operate (Product A – circles; Product B – triangles). Once the contract grouping is done, we analyse the change in collusion indicator values based on the during (red filled) and after (grey filled) cartel period contracts. By comparing the behavior of prosecuted companies over time – i.e. contrasting rigged and presumably competitive tenders – our models learn to recognize the changes associated with cartel activity.
One common strategy used by cartels, such as cover-bidding, involves submitting artificially high losing bids to mask the real winner and avoid scrutiny, often leading to inflated price markups during the cartel’s operation that return to normal once the cartel is detected. This price markup, for example, can serve as one indicator among many, and when combined with other screens, it can help the model detect collusion effectively. To give our model this flexibility, we drew on a wide set of collusion indicators identified in the literature.
At the contract level, these included the number of bids per contract, whether subcontracting was used, participation in consortia, and the bid price relative to the reference price. We extended these with aggregated and lagged versions to capture persistent patterns. We also tracked company-level behavior over time, such as the number of markets and buyers each firm won contracts at, and incorporated deviations from Benford’s Law in bid pricing10. Finally, to ensure fair comparisons across cases, we controlled for product markets using standardized product classifications11 and approximated company size by the number of contracts awarded.
Learning from a diverse set of cartel cases – i.e., different countries and markets – enables the models to capture a wider range of collusion strategies. In fact, we found that predicting collusion status of individual contracts in a country significantly improves if our training data contains proven cartel cases from other countries and markets. Figure 2 shows how accuracy – that is correctly predicting rigged versus competitively won contracts – on a Hungarian testing sample improves as we incrementally add more cartel cases from different countries to the training data. Adding prosecuted cases from other countries improved accuracy – by 5.3 percentage points for Random Forest models and up to 14.4 percentage points for Boosting models.
Using ensemble models like Random Forest—a method that combines the results of many decision trees to improve predictions—we achieved prediction rates of 70–84% across diverse cartel cases. When tested on a single cartel case, prediction rates reached up to 95%, further confirming that these machine learning models can reliably detect patterns of collusive behavior.
Our tests show that no single indicator detects all types of collusion consistently. In some cartel strategies, the number of bidders may increase to mimic competition; in others, it may drop to a single firm. While price and bidding-pattern based indicators are powerful, they are incomplete on their own. However, the combination of multiple weak signals increases collusion detection substantially.
Using our model, we predicted collusion risks to 3.3 million contracts across the seven analysed countries. These predictions allow us to detect high-risk sectors and suppliers, which offer promising starting points for audits and in-depth investigations. For example, the likelihood of collusion and total spending can be calculated at the market level12. Figure 3 shows the relationship between these statistics across Hungary, Latvia, and Sweden13. Each point represents a product market, with the Occupational Clothing market highlighted in red as an example. Along the vertical axis, the figure shows that the average predicted likelihood of collusion ranges between 20% and 70%. For the clothing market, these values are highest in Hungary (67%) and lowest in Latvia (50%). Total market spending is shown on the horizontal axis, ranging from about 10 million EUR in Latvia to 25 million EUR in Sweden. This information can help competition authorities prioritize investigations into high-volume, high-risk markets.
Ways Forward
Machine learning tools for detecting collusion are designed to support human expertise. They act as early warning systems that help investigators focus their efforts more effectively. To remain useful as cartels adapt, these models should be continuously updated with new investigative findings. At the same time, our results highlight major policy challenges. Even in high-income EU countries, essential data, such as losing bid prices or subcontracting information, is often missing or inconsistently reported. Improving data quality would boost prediction accuracy and strengthen investigative capabilities. Nevertheless, this line of research proves that meaningful, large-scale risk estimation is possible even with imperfect data, offering a practical path forward for authorities seeking to detect procurement collusion.
[1] OECD, Implementing the OECD Recommendation on Public Procurement in OECD and Partner Countries. Available here.
[2] World Bank, Global Public Procurement Database: Share, Compare, Improve. Available here.
[3] UNODC, Guidebook on Anti-Corruption in Public Procurement and the Management of Public Finances (2013). Available here.
[4] Schneider dos Santos, E., Machado dos Santos, M., Castro, M., & Tyska Carvalho, J. (2025). Detection of fraud in public procurement using data-driven methods: a systematic mapping study. EPJ Data Science, 14(1), 52.
[5] Wallimann, H., Imhof, D., & Huber, M. (2023). A machine learning approach for flagging incomplete bid-rigging cartels. Computational Economics, 62(4), 1669-1720.
[6] Huber, M., Imhof, D., & Ishii, R. (2022). Transnational machine learning with screens for flagging bid-rigging cartels. Journal of the Royal Statistical Society Series A: Statistics in Society, 185(3), 1074-1114.
[7] Huber, M., & Imhof, D. (2023). Flagging cartel participants with deep learning based on convolutional neural networks. International Journal of Industrial Organization, 89, 102946.
[8] Harrington Jr, J. E., & Imhof, D. (2022). Cartel screening and machine learning. Stan. Computational Antitrust, 2, 133.
[9] Adam, I., Fazekas, M., Kazmina, Y., Teremy, Z., Tóth, B., Villamil, I. R., & Wachs, J. (2022). Public procurement cartels: A systematic testing of old and new screens. Governance Transparency Institute, Working Paper Series, 1. Available here.
[10] Benford’s Law, often used in fraud detection, measures how closely the distribution of first digits in bid prices follows a natural pattern. Significant deviation may signal artificial price manipulation.
[11] We use CPV, the Common Procurement Vocabulary, which is the EU’s standardized classification system for public procurement contracts.
[12] We aggregated tenders at the 3-digit CPV level
[13] We use a logarithmic scale for total spending because tender values vary across local currencies, even though the figure displays the nominal scale.
