A case study on Danish public procurement contracts, 2016–2022
A new study from Utrecht University tests whether machine learning models can identify corruption risk in public procurement before a contract is awarded. The research applies twelve algorithms to 20,777 Danish procurement contracts covering the period 2016 to 2022. Its central finding is that the right models, operating on well-prepared data that includes information about beneficial ownership, substantially outperform random audit selection. The wrong models, by contrast, add nothing.
The study forms part of the BridgeGap project, funded by the European Commission under the Horizon Europe programme. It is the second major deliverable from the project’s work package on artificial intelligence, building on an earlier comparative assessment of data suitability across EU member states and accession countries.
The Problem:
Supervisory authorities responsible for public procurement cannot audit every contract. The practical challenge is prioritisation: which tenders merit closer scrutiny? Traditional red-flag systems check individual indicators in isolation. Machine learning models can instead learn from combinations of variables simultaneously, detecting patterns that no single rule would surface.
Corruption detection faces a structural difficulty. The study examines whether correcting for this imbalance improves detection, and how much it matters in practice. Artificial intelligence can meaningfully improve the detection of corruption risk in government contracts — but only when applied carefully, with clean data and the right algorithms.
The study paper, authored by Andrea Longobucco and Joras Ferwerda, has the goal: to determine whether AI tools could flag high-risk tenders before a contract is awarded — and whether linking procurement records with beneficial ownership data makes those tools substantially more effective.
The study also examines what information matters most. Its main hypothesis is that linking procurement data to beneficial ownership information — data that identifies the natural persons who ultimately control the winning companies — should improve detection beyond what procurement data alone can achieve.
The Case Study: Denmark
The choice of Denmark as a test case is deliberate and somewhat counterintuitive. Denmark currently ranks first in the world on Transparency International’s Corruption Perceptions Index, widely regarded as the least corrupt country globally.
That status makes it a controlled testing environment. Corruption is sufficiently rare that anomalies in procurement patterns are more likely to reflect genuine irregularities than systemic distortions. Where corruption is entrenched, anomaly detection logic tends to invert: clean cases become the outliers.
Denmark is not, however, immune. The Atea case (2009–2014) involved systematic bribery of public officials through gifts and travel funded by IT procurement contracts. The FES defence case saw public officials create shadow accounts with contractors to extract private benefits from routine maintenance contracts; the Western High Court delivered a conviction in September 2024. Transparency International Denmark’s 2025 report notes that known cases, taken together, suggest corruption is more widespread than commonly assumed, with many incidents never reaching public attention.
The Council of Europe’s Group of States against Corruption (GRECO) found Denmark in insufficient compliance with anti-corruption recommendations in 2025, citing the absence of a coherent national preventive framework. Denmark continues to treat corruption primarily as a matter for criminal law after the fact.
Denmark also holds a public beneficial ownership register, the Central Business Register (CVR), which has been open to the public since May 2017. This register links companies to the natural persons who ultimately own or control them, and it is machine-readable. That combination — a mature procurement data infrastructure and an accessible ownership register — makes Denmark one of very few countries where this type of analysis is currently feasible.
“Denmark’s corruption paradox“, as researchers describe it, makes it an ideal laboratory: clean enough for anomaly detection to work well, but not so clean as to make the exercise meaningless.
Data and Methods:
The dataset originates from a linked database of Danish public procurement contracts and beneficial ownership structures, covering 41,312 awarded contracts from 2016 to 2022. After quality filtering and variable selection, the working sample was reduced to 20,777 observations. The study is explicit that this reduction was deliberate: variables with very high rates of missing data were excluded rather than imputed, because filling missing values with estimated figures would produce an unrealistically clean environment and overstate the practical performance of any detection model.
The analysis uses only information available before contract award. Final contract price is included as a proxy for bid price and estimated contract value, both of which had missing rates close to fifty percent in the raw data. Sixteen variables were selected for analysis, covering procurement procedure type, buyer characteristics, tender structure, and beneficial ownership attributes. After preprocessing and encoding, this expanded to seventy-six features.
Beneficial ownership data has been publicly available in Denmark since May 2017 through the Central Business Register, and includes information on ownership shares, the number of companies controlled by the same individual, and how frequently that ownership information changes. Crucially, companies bidding for public contracts are not required to disclose their beneficial owners during the tender process — the link can only be established retrospectively, or proactively by oversight bodies with access to the register.
The researchers deliberately avoided imputing missing values — roughly 16% of cells in the analysis variables were incomplete — to simulate the real conditions that a supervisory authority would face in practice. This decision reduced performance compared to an artificially clean dataset, but produced more honest results.
Two families of models were tested. Six unsupervised models identify anomalous tenders without requiring labelled examples of past corruption. These include density-based methods (Local Outlier Factor and DBSCAN), a distance-based method (k-nearest neighbours), a tree-based method (Isolation Forest), a kernel-based method (one-class support vector machine), and a neural network reconstruction method (Autoencoder). Unsupervised models were evaluated against a single, directly observed indicator: single bidding, defined as a tender that received only one valid offer.
Six supervised models were trained on the Corruption Risk Index (CRI), a composite of five procurement signals: single bidding, short submission periods, irregular decision timelines, non-competitive procedure types, and high buyer–supplier concentration. Contracts scoring above 0.5 on this index — approximately the top seven percent — were classified as high risk. The supervised models tested were Distributed Random Forest, Gradient Boosting Machine, Deep Learning (feed-forward neural network), Regularised Logistic Regression, Naive Bayes, and Support Vector Machine with a Gaussian kernel.
To test the impact of class imbalance, the supervised models were trained on both the original imbalanced dataset and three rebalanced versions using SMOTE-NC oversampling at minority-class proportions of twenty, thirty, and forty percent. To test the value of beneficial ownership data, a parallel set of models was trained on procurement data alone, with all ownership variables removed.
Results:
Among unsupervised models, Isolation Forest is the most effective at tight audit thresholds. Focusing on the top one percent of anomaly scores, it identifies single-bid tenders at more than three and a half times the rate of random selection. Density-based methods add almost nothing beyond chance. Autoencoders become competitive as the audit threshold widens to five percent, at which point they maintain reasonable precision while covering more of the risk population.
Among supervised models, Distributed Random Forest (DRF) is the standout performer. On the imbalanced training set with beneficial ownership information, it achieves a balanced accuracy of 0.931 and correctly identifies nearly nine in ten high-risk tenders while keeping the false-alarm rate low. Gradient Boosting Machine ranks second. Deep Learning performs reasonably well but does not improve on the tree-based methods in this context. The Support Vector Machine is effectively useless: it flags almost every tender as high risk, achieving near-perfect recall but negligible precision, and performs no better than random classification on any balanced metric.
Oversampling via SMOTE-NC does not improve the best models in terms of overall performance. Its main effect is a shift in the trade-off between precision and recall. At a twenty-to-eighty oversampling ratio, DRF becomes more precise: fewer false alarms, but more high-risk contracts missed. This configuration is more appropriate when audit capacity is limited and each false positive is costly. The study also shows that DRF on the original imbalanced data tends to underestimate predicted risk probabilities relative to observed outcomes. This systematic underestimation is less problematic than overestimation for audit purposes, since only the highest-scoring tenders would be selected for review.
Removing beneficial ownership variables from the supervised models reduces balanced accuracy by an average of six percentage points. This difference was positive in every one of a thousand bootstrap replications. Ownership variables — particularly ownership share, the number of contracts held by the same beneficial owner, and company age — contribute signal that procurement data alone cannot provide. Variable selection analysis using stability selection confirms this: these ownership variables are consistently chosen by Random Forest models across repeated subsamples, indicating robust predictive importance. Their contribution is largely nonlinear, which is why it appears more clearly in tree-based models than in linear specifications such as Lasso regression.
Among unsupervised methods, results varied sharply. Density-based algorithms barely outperformed random selection of contracts for audit. By contrast, the Isolation Forest algorithm — which detects anomalies by measuring how quickly individual observations can be isolated from others in randomly constructed decision trees — identified single-bid contracts at more than three-and-a-half times the rate of random auditing when focusing on the top 1% of anomaly scores. Autoencoders, neural networks that learn to reconstruct normal patterns and flag records they cannot explain, performed comparably at broader thresholds.
A central concern in the study was class imbalance: only 7% of contracts were labelled high-risk, raising the possibility that models would simply learn to predict “low risk” for everything and still achieve 93% accuracy.
The authors are candid about the study’s limits. The models detect patterns correlated with corruption risk, not corruption itself. The CRI, their primary label, is a composite risk indicator constructed from process characteristics — not a record of confirmed wrongdoing. A tender flagged by the model may be perfectly legitimate; an unflagged one may be corrupt. The models should be understood as triage tools, not verdicts.
The dataset, at around 20,000 observations, is insufficient for deep learning methods to demonstrate any advantage over tree-based models. The use of final contract price as a proxy for bid price — necessitated by high rates of missing values in the preferred variables — introduces a minor violation of the study’s own ex-ante principle, since final prices are technically known only after award. The authors note this would not be an issue in a well-maintained real-world data system.
Most structurally: the countries where these tools would deliver the greatest public benefit — those with weaker institutions and higher baseline corruption — are precisely the countries least likely to have the transparent, reliable, accessible beneficial ownership registers that make the models work. Denmark’s CVR is a best-case data environment. Replicating this approach in contexts where BO registers are incomplete, unverified, or non-public would produce substantially weaker results.
Conditions for Effective Use of AI in Corruption Detection:
The study draws three conditions under which machine learning tools can function as effective audit-prioritisation instruments.
Data quality is the first and most decisive condition. The gap between a well-maintained procurement dataset and a poorly maintained one is larger than the gap between any two algorithms. Missing values, input errors, and unstructured records all degrade model performance. Artificial intelligence learns from what it is given; it cannot recover signal that was never recorded.
Beneficial ownership linkage is the second condition. Procurement data alone leaves measurable predictive signal inaccessible. Open, machine-readable, company-linkable ownership registers are a practical prerequisite for high-performance detection. Countries where beneficial ownership data is unavailable or unreliable face a fundamental constraint that no algorithmic improvement can substitute for.
Anomaly detection works only where corruption is genuinely atypical. In procurement environments where irregular practices are the norm, the statistical logic of anomaly detection inverts. Denmark functions as a viable test environment precisely because single bidding and non-competitive procedures are exceptions rather than rules. The same models would not transfer directly to high-corruption settings without reconceptualisation.
Model choice must match institutional context. High precision with lower recall suits supervisory bodies where false positives are costly and investigative capacity is constrained. High recall suits bodies with broader investigative mandates and the resources to follow up more leads. There is no universally optimal model; the choice is a policy decision as much as a technical one. The study also notes that fairness and explainability — whether models treat different categories of suppliers consistently, and whether their outputs can be explained to courts and audited bodies — remain open research questions and are the subject of forthcoming BridgeGap deliverables.
The evidence from this study is specific: the right machine learning models, operating on well-prepared data enriched with beneficial ownership information, produce audit prioritisation tools that are substantially better than random selection. The best supervised model correctly identifies close to nine in ten high-risk tenders.
These models rank tenders by risk. They do not determine guilt. They are triage tools, and their practical value depends entirely on what happens next: whether flagged cases are investigated by people with the knowledge, legal authority, and institutional backing to act on what they find.
The study also confronts its own limits. The dataset covers one country with unusually good data infrastructure. The sample is 20,777 contracts, which is meaningful but not large enough to support more demanding deep learning architectures. The labels used for supervised training are themselves proxies for corruption, not confirmed instances of it. And the country-specific nature of the analysis means that models cannot be imported directly into different institutional and market environments without revalidation.
The study is one component of a broader research programme examining AI tools for anti-corruption across EU member states. Forthcoming deliverables will address fairness — whether AI screening disproportionately flags certain types of companies or sectors — and explainability, including the use of SHAP and LIME methods to translate model outputs into actionable audit rules. The researchers also recommend extending the analysis to countries with higher corruption levels to test whether the same indicators and algorithms generalise across institutional environments.
For policymakers, the study’s practical message is measured: AI tools can meaningfully improve the efficiency of procurement audits, but naïve deployment of off-the-shelf algorithms on raw data is likely to produce little of value. The returns come from careful data preparation, thoughtful model selection matched to the agency’s operational constraints, and — critically — investment in the data infrastructure, particularly beneficial ownership registers, that makes meaningful detection possible in the first place.
The practical message for supervisory bodies is narrow but concrete. Investing in data quality and in open beneficial ownership registers is more impactful than investing in algorithm sophistication. The choice of algorithm matters, but it matters less than the data it runs on. And any institution considering these tools needs to decide in advance what kind of errors it is least able to tolerate — because the choice of model embeds that decision whether it is made explicitly or not.
Here you can find the full deliverable for a more extensive insight: https://corruptiondata.eu/wp-content/uploads/2026/02/BridgeGap-D6-2-The-Effectiveness-of-AI-tools-to-detect-corruption_V1.0.pdf
Source: Longobucco, A. & Ferwerda, J. (2025). Paper on the Effectiveness of AI Tools to Detect Corruption: A Case Study on Danish Public Procurements. BridgeGap Deliverable 6.2, Utrecht University. Funded by the European Commission Horizon Europe, Grant Agreement 101132483.