Can deep learning tell which suppliers the antimonopoly regulator will sanction for bid rigging next?
We scored every active supplier on Prozorro from its screens, its last 128 bids and its place in a graph of co-bidding and shared contacts, then checked the ranking against AMCU decisions made in the 21 months after the cutoff. No model saw that period while it was being built.
The answers
Each question was fixed in a locked protocol before the test year was scored.
Bid-rigging sanctions per year
Firms named in AMCU decisions. Later years are what the models predict.
Learn on the past, test on the future
Each origin uses three years of history (grey) and labels the next two (coloured).
Tuned trees rank best; the graph network is the best deep model
Ranking quality (PR-AUC, 95% interval resampling whole cartels)
Coloured: the models the protocol compares. Grey: other benchmarks and ablations.
Is the complexity worth it?
Paired difference in test PR-AUC. A line clear of zero means a real gain after Holm correction.
Cumulative gains
Share of future-sanctioned firms found as more firms are screened.
Where each model works
Test PR-AUC lift over the base rate inside each subgroup. Darker is better; the number is the lift.
All metrics
Precision at k = share of the top k firms later sanctioned. Recall at x% = share of all future cases inside the top x%.
Size explains much of the ranking, but not all of it; geography barely matters
The two questions
Bigger firms really are sanctioned more often
Share of firms sanctioned in 2025–26, by size decile (number of bids in the previous three years).
But the top-1% list leans much harder on size
Where LightGBM's 408 top-1% flags come from, against where the 911 real cases came from.
Inside each size band, the models still rank
PR-AUC lift over the band's own base rate. Above 1× means the model separates firms of the same size.
Oblasts: more or fewer sanctions than their size mix predicts
Observed ÷ expected cases, where expected uses each firm's size decile. Hover for counts and model lift. Open the map →
What the evidence looks like
Share of firms later sanctioned in each band of one signal (dashed: overall base rate). Hover for the model's mean percentile in the band.
Audit simulator
Choose how many firms the regulator can examine. Each model ranks all active suppliers at the end of 2024; we count how many were then sanctioned in 2025–26.
Hits at this budget
Hits as the budget grows
Watchlist
Where the risk sits, and where it sells
Colour an oblast by what you want to compare, switch on connections, and click any oblast for its profile, where its suppliers sell and its highest-risk firms.
Boundaries: geoBoundaries, OpenStreetMap contributors (ODbL). Crimea and Sevastopol have no supplier data; Donetsk, Luhansk and Kherson are thin because of the war.
Did the model see these cartels coming?
Cartels AMCU sanctioned in 2025–26. Each dot is a lot in 2022–24 where at least two members bid against each other, before the model scored them. Cards show each member's risk percentile at the 2022 and 2024 origins.
More deep learning did not beat tuned trees
Each experiment was designed before it ran, kept the same leakage rules and scored its test period once.
Data, models and protocol
CartelNet: three encoders fused into one risk score
What LightGBM relies on
Mean absolute SHAP value, top 12 features.
What CartelNet relies on
Drop in test PR-AUC when one family of evidence is shuffled.
Hyper-parameter search
Validation PR-AUC of every random-search trial.
A data artefact we removed
Before 2023 the bid timestamp is a last-modified time, so "bid within 10 minutes of a rival" was dropped as a feature.