In new research, Yoan Hermstrüwer and David Imhof analyze how AI can help antitrust authorities predict cartels by assessing international bidding data in countries with similar legal and market structures.
A cartel exposed in Switzerland. Another dismantled in Finland. And more bid-riggers prosecuted in Brazil and the United States. Each case generates valuable data on how colluding firms behave: the telltale patterns in their bids, the statistical fingerprints of coordination. But can lessons learned in Zurich help catch bid-riggers in Los Angeles?
The implications are significant, especially for government budgets. With public procurement representing roughly 13 percent of GDP in countries in the Organisation for Economic Co-operation and Development—and nearly 30 percent in developing economies—the damage from bid-rigging is staggering. Cartels typically inflate government contract prices in the range of 20 to 50 percent, draining taxpayer money meant for roads, schools, and hospitals. Enforcement agencies are increasingly turning to machine learning—and artificial intelligence, more generally—to assess patterns in data and fight back against collusion. An increasing number of antitrust authorities worldwide are now exploring algorithmic screening tools.
The promise is tantalizing. If AI models trained on the behavior of known cartels in one jurisdiction or industry could reliably flag collusion in different ones, resource-strapped agencies could pool their data and multiply their detection power. But the risk is equally clear. Models that fail to reliably generalize data could flood investigators with false alarms, or miss real cartels hiding behind unfamiliar legal and institutional rules.
In our new paper, “Cross-Jurisdictional Machine Predictions in Antitrust,” we provide systematic evidence on how cartel detection models can be used to predict collusion across borders and different industries. We term the quality of a model to apply accurately to different jurisdictions and industries “transferability.” Our findings reveal both the promises and limitations of such models.
First, cross-jurisdictional cartel detection is feasible, with accuracy exceeding 85 percent for institutionally similar markets. Second, novel bidder-level screens that track firms across multiple auctions substantially outperform traditional tender-level approaches that only treat offers in isolation. Third, the transferability of machine learning models breaks down when procurement rules, contract types, or market structures diverge significantly across jurisdictions. Specifically, models that are applied across different industries (e.g., from construction to milk markets) fail almost completely, performing worse than random guessing.
How We Measure Transferability
Detecting collusion is fundamentally a classification problem: given the statistical properties of a bid, the challenge is to predict whether it is collusive or competitive. We employ ensemble machine learning methods that combine multiple algorithms to improve predictive accuracy and robustness. Specifically, we train three distinct models—Random Forest, Gradient Boosting, and XGBoost—and aggregate their predictions using various techniques. Combining these models exploits their complementary strengths while mitigating individual weaknesses.
We compile bidding data from documented cartel cases across six jurisdictions: Switzerland, Finland, Sweden, Japan (Okinawa), Brazil, and the U.S. (California and Ohio). The dataset includes nearly 30,000 individual bids from public procurement auctions, primarily in road construction and civil engineering.
Our approach combines traditional screens—statistical indicators (e.g. the spread of bids in a tender) that flag suspicious bidding patterns, as recommended by the latest version of the OECD Guidelines for Fighting Bid Rigging in Public Procurement—with a novel screening method: bidder-level measures that track how individual firms behave across multiple procurement auctions over time. While existing methods treat each tender in isolation, our bidder-level screens capture the persistent coordination patterns that emerge when the same firms repeatedly collude.
We test four model specifications, each capturing how firms bid across different time horizons. The first uses only tender-level screens: measures like bid dispersion, skewness, and the spread between highest and lowest bids within a single procurement auction. The second incorporates moving averages, tracking each firm’s bidding patterns over its last five tenders. The third uses long-run bidder-level averages, computing each firm’s typical behavior across all observed auctions. The fourth combines medium-term and long-run measures. This design allows us to assess whether persistent behavioral patterns are more informative than snapshot observations.
To test cross-jurisdictional transferability, we use a leave-one-jurisdiction-out design: we train models on data from all jurisdictions except one, then predict collusion in the held-out jurisdiction. This approach simulates what would happen if an enforcement agency tried to use foreign cartel data to screen its own markets.
A critical challenge in applying machine learning to antitrust is the black box problem: complex models may achieve high accuracy, but they may not reveal why they flag certain bids as suspicious. We address this by using SHAP (SHapley Additive exPlanations), a technique grounded in cooperative game theory that decomposes each prediction into individual feature contributions. SHAP values reveal which statistical screens drive the model’s decisions, both globally (which features matter most across all predictions) and locally (why a specific bid was flagged). This transparency is essential in the antitrust context, where agencies must be able to explain and defend their screening decisions.
Key Findings
When procurement rules and contract types align, cross-jurisdictional detection works remarkably well. For Finland, our models correctly classify nearly 90 percent of bids as collusive or competitive—despite being trained entirely on foreign data. Sweden and Switzerland show similarly strong results, with accuracy above 75 percent. These jurisdictions share comparable tendering procedures, contract types, and market structures, all predominantly involving road construction.
The gains from tracking firms over time are substantial. In Finland, accuracy jumps from 76 percent using tender-level screens alone to nearly 90 percent with bidder-level measures. This improvement reflects a fundamental insight: collusion is not a single-tender event; it is a dynamic process rooted in repeated interaction. Firms that coordinate do so persistently, and our screens capture this pattern.
Japan presents a striking case of how specific procurement laws and auction design affects model performance. Japanese tender rules impose explicit upper and lower bounds on admissible bids, effectively compressing the range of observable prices. This truncation mechanically reduces bid variance—the very signal our screens rely on to distinguish collusion from competition. As a result, competitive Japanese bids look statistically similar to collusive bids elsewhere, generating a flood of false positives. Training models on Japanese data is therefore not helpful in generating accurate predictions of collusion in Finland, Sweden or Switzerland.
The starkest limitation emerges when models cross sectoral lines. Our Brazilian cartel data comes from oil infrastructure contracts rather than road construction. These projects are larger, more complex, and more heterogeneous—generating naturally higher bid dispersion even under competition. Models trained on construction data systematically fail to detect collusion in Brazil, misinterpreting its wide variation in bids as normal competitive variation. The lesson is blunt: for machine learning models, the product matters. Pavement is not oil—and pretending otherwise would be misleading.
The Ohio school milk cartel of the 1980s illustrates an even more fundamental barrier. This bid-rigging arrangement relied primarily on bid suppression and market allocation—strategies that reduce the number of bids rather than distort their distribution. With few bidders per tender and standardized products, the statistical fingerprints that construction-sector screens detect simply do not appear. The models perform worse than a coin flip.
Policy Lessons
Cross-border collaboration is viable—with caveats. Enforcement agencies can meaningfully leverage cartel data from other jurisdictions, but only when markets share similar institutional frameworks. For European road construction markets, our results suggest that pooling data across jurisdictions could substantially enhance detection capacity. The OECD’s push for international cooperation in cartel screening finds empirical support here.
Institutional awareness is essential. Procurement rules are not neutral—they shape the very data that detection models analyze. Features like bid price bounds, qualification requirements, or contract bundling practices can systematically alter statistical patterns in ways that confound algorithmic screening. Authorities deploying AI tools must understand how their local rules interact with model assumptions.
The failure of cross-sectoral transfer is perhaps our most cautionary finding. A model that performs almost perfectly in construction may be useless—or worse, actively misleading—in healthcare procurement or food supply contracts. Different industries feature different cost structures, different competitive dynamics, and different collusion mechanisms. One-size-fits-all solutions do not exist.
Models are risk indicators, not verdicts. Even our best-performing specifications generate some false positives and miss some true cartels. These tools should flag suspicious patterns for human review, not trigger automatic enforcement actions. The goal is to help authorities allocate scarce investigative resources more efficiently, not to replace prosecutorial or, more generally, human judgment.
As procurement systems digitize and generate ever richer audit trails, the scope for algorithmic cartel screening will only expand. Our research suggests this is a promising frontier—but one that requires careful attention to institutional context. Machine learning can help catch cartels across borders, but only if we understand the legal rules and practices applied in the respective markets.
Author’s Disclosure: The author reports no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
