The Single Market of the EU – one of the largest economies in the world, accounting for around 15% of global GDP – aims to ensure the free movement of people, goods, services, and capital across member states and to prohibit any discrimination among economic operators. While, on paper, a large economic area functioning as a single national market appears to be a win-win strategy, the reality is less reassuring: internal barriers still persist more than 30 years after the launch of the Single Market (Draghi 2025). This incompleteness is also evident through the lens of state aid regulation. For instance, during the Covid-19 pandemic, harmonising support programmes across EU member states became crucial to avoid market distortions (Peitz and Motta 2020). At the same time, these programmes generated additionality mainly for micro and small firms (Canzian et al. 2025) with non-trivial targeting issues (Rodano et al. 2022).
Even in normal times, the enforcement of an effective regulation of state aid seems rather fragile. The general rule is that the EU Treaty generally prohibits state aid unless exceptionally justified by some market failure (Article 107). Nevertheless, concerns remain regarding the effective enforcement of state aid regulations since member states may provide unauthorised aid (for example, to promote national champions). Moreover, there may be a lack of accountability if member states do not comply with the transparency regulation that requires that information on the main state aid awards must be in the public domain. Finally, it is possible for a firm operating within the Single Market to benefit from ‘foreign’ subsidies – meaning financial support granted by a government or other public entity outside the EU. This could occur, for example, if the undertaking is ultimately owned or controlled by a non-EU government (European Commission 2020). These situations can harm competition and market efficiency by enabling less efficient firms to expand and gain market share at the expense of more efficient, unsubsidised rivals. Yet, despite the policy significance of these issues, no systematic evidence currently exists.
Detecting hidden state aid (in normal times)
In a recent paper (Barone and Letta 2025), we propose an original machine learning (ML) approach to identify potential cases of hidden state aid recipients. We merge data from the Transparency Award Module (TAM) database, maintained by the European Commission and collecting information on state aid recipients, with financial statement data from the Bureau van Dijk Orbis database to train three ML classification models capable of accurately predicting state aid recipients based on financial accounting information. The final dataset refers to the pre-Covid-19 period and comprises over 11 million observations and includes 187 predictor variables. Suspect cases are defined as firms that, according to multiple algorithmic predictions, exhibit financial characteristics consistent with having received grants in a given year but do not appear in the Commission’s database of authorised state aid recipients (more precisely, those pinpointed as ‘false positives’ by the majority of our ML models). These cases are considered ‘suspect’ because, in a given year, the firms may: (i) have received aid that is incompatible with the Single Market; (ii) have received compatible aid that was not disclosed as required by the transparency regulation; or (iii) have benefited from aid granted by non-EU states, which falls outside EU rules but may still pose competitive concerns.
We employ three distinct ML algorithms, carefully selected based on three key criteria: transparency and interpretability, robustness to pervasive missing data, and the ability to capture potential nonlinearities in the data. The first model is a simple and easily interpretable classification tree, which addresses the ‘black box’ criticism and ensures transparency and accountability for policymakers. The second is a more complex, though less interpretable, tree-based bagging algorithm. Since both of these models require training on artificially rebalanced datasets to manage the highly unbalanced nature of the outcome variable, we also use a third approach: eXtreme Gradient Boosting (XGBoost). This state-of-the-art ensemble algorithm offers strong predictive performance and can effectively handle highly unbalanced outcomes without the need for data rebalancing.
How many potential hidden state aid occurrences are there?
Table 1 presents the prediction performance of our ML models. For the classification tree (Panel A), which selects industry, baseline fixed assets, baseline operating revenue turnover, and baseline financial revenues as key predictors – combined in a non-linear fashion – the model achieves high overall accuracy: out of more than 3.6 million firm-year observations in the testing sample, over 90% are correctly classified as either true positives or true negatives. The remaining observations fall into the two misclassification categories: false negatives and false positives (9.5% of all observations). The model’s sensitivity – defined as the share of actual positives correctly identified – is also moderately high, at 72%.
The Bagging model (Panel B) identifies industry, baseline fixed assets, and baseline financial expenses as its top three predictors. This algorithm outperforms the classification tree, achieving an accuracy of over 91% and, more importantly, a sensitivity of more than 81% in detecting true positives. The number of false positives amounts to nearly 310,000 firm-year observations, corresponding to 8.5% of the testing set.
Finally, XGBoost (Panel C; top three predictors: baseline total assets, baseline fixed assets, change in total assets) also delivers outstanding predictive performance. It almost matches the other models in terms of overall accuracy (exceeding 90%) and achieves the highest sensitivity among all approaches, at 82.6%. The model flags approximately 360,000 firm-year observations as false positives, which represents 9.9% of the testing sample.
From a policy perspective, the most relevant category is the false positives. These are firms not listed as state aid recipients in the TAM data but predicted by each model to be recipients. This group includes two distinct components: pure prediction error and potential underreporting. While it is not possible to precisely disentangle these components, we can reasonably assert that the higher the model’s sensitivity, specificity (i.e. the ability to identify true negatives), and accuracy, the smaller the share attributable to pure prediction error – and thus, the greater the likelihood that a false positive reflects a genuine case of underreporting. Since our outcome variable is highly unbalanced (fewer than 0.1% of observations are actual state aid recipients), false positives are numerous relative to true positives, but few in comparison to true negatives.
We then proceed to combine the predictions of the different learners to identify suspect observations. Specifically, we take observations in the false positives cell of all learners and consider an observation suspect if it is a false positive according to the majority (i.e. two out of three) of our ML models. This way, about 326,000 firm-year observations are suspect, corresponding to almost 9% of the sample – the magnitude of hidden aid is far from negligible. According to a straightforward back-of-the-envelope calculation, if we conservatively multiply only half of the suspected share (4.5%) by the number of nonfinancial private sector firms (over 23.4 million firms in 2020), then by the median subsidy amount (1 million euros), we arrive at a number that is above 4% of aggregate turnover.
Table 1 Performance of ML models
What can we say about suspect cases?
An ex-post descriptive analysis of suspect cases reveals interesting patterns. In some countries (Cyprus, Lithuania), the incidence of suspect recipients is twice the average value or higher, while units from Italy, Malta, Spain, and Slovenia show a lower share of pooled false positives. The heterogeneity is even larger across industries: on average, firms in the mining and manufacturing sectors are much more likely to be predicted as hidden recipients. In the case of the energy and water-related sectors, the incidence is even larger. Firm size is also highly correlated to the suspect recipient status: the incidence is high for firms in the upper total asset quartile, and close to zero for the others.
Finally, we study the correlates of the country-specific estimates of hidden recipients (Figure 1). It turns out that suspect cases are more likely in countries with larger public ownership of firms (consistent with the idea that member states support national champions; Panel A), with higher levels of inward foreign direct investment (FDI) from outside the EU (consistent with the idea that countries outside the EU might be interested in supporting companies in which they have invested), and with higher values of other measures of the accuracy gap in reporting to the EU (Panels C and D).
Figure 1 Suspect cases and cross-country correlations
Notes: Scatter plots coming from country-level regressions of the residualised share of suspect cases on (i) an indicator of public ownership, computed as the simple average of the OECD economy-wide product market regulation indicators relative to “Scope of state-owned enterprises”, “Direct control over enterprises”, “Governance of state-owned enterprises”, measured in 2015; these variables are on a 0-6 scale, with higher values indicating a larger role of the state in the economy (Panel A); (ii) the stock of the inward net position of FDI from non-EU countries (taken from UNCTAD), measured in 2015, as a percentage of gross fixed capital formation (Panel B); (iii) the χ2 statistics of the Benford’s law applied to the SA amount taken from TAM (Panel C); (iv) the number of infringements of competition and internal market rules over the 1987-2024 period sourced from the European Commission (Panel D).
How can policy move beyond classification?
Once suspect labels are identified, policymakers can follow up with more targeted investigations using soft information, alternative data sources, micro-level financial analysis, or on-site inspections. Given the high cost of such efforts, it is crucial to narrow the pool of cases. Our ML tools help by flagging false positives with the highest predicted likelihood of hidden aid, making the process scalable and resource-efficient. Notably, our approach also aligns with the EU’s new foreign subsidies regulation (effective since January 2023), offering a practical first step for prioritising firms for further audit and investigation.
References
Barone, G and M Letta (2025), “Unlevel playing field? Machine learning meets state aid regulation”, International Journal of Industrial Organization 103175.
Canzian G, E Crivellaro, T Duso, A R Ferrara, A Sasso and S Verzillo (2025), “State aid in times of crisis: Lessons from COVID-19 support for firms in Italy and Spain”, VoxEU.org, 13 June.
European Commission (2020), “White paper on levelling the playing field as regards foreign subsidies”.
Draghi, M (2025), “Forget the US — Europe has successfully put tariffs on itself”, Financial Times.
Peitz, M and M Motta (2020), “EU state aid policies in the time of COVID-19”, VoxEU.org, 18 April.
Rodano G, E Sette and M Pelosi (2022), “Zombie firms and the take-up of support measures during Covid-19”, VoxEU.org, 4 May.