STT's Machine-Learning Risk Model for Public-Procurement Fraud and Corruption

Lithuania's Special Investigation Service (STT) worked with the OECD and the Government Transparency Institute, funded by the European Commission's Technical Support Instrument, to build a machine-learning risk-assessment tool for public procurement. The team integrated eleven administrative datasets — covering criminal and administrative offences, company registries, beneficial ownership, blacklisting and EU-funded project data — into a single analytical database, restructuring procurement records from tender-level to lot-level to capture bidding dynamics.

The goal was to help STT analysts prioritise investigative attention on the procurement contracts most likely to involve fraud or corruption, while keeping every prediction explainable so it guides investigator judgement rather than replacing it.

Because only a small share of contracts are linked to confirmed offences, the team used a Hellinger Distance Stratified Random Forest, a method designed for Positive-Unlabelled learning and severe class imbalance. Every prediction is paired with SHAP (SHapley Additive exPlanations) values in an interactive dashboard, so analysts can trace which factors — bidding patterns, ownership concentration, network position — drove a given contract's risk score.

In back-testing, focusing investigative attention on the top 25% of contracts by predicted risk captured roughly 60% of contracts linked to proven or suspected criminal offences, versus the 25% expected under random selection, and the top 10% captured about five times more known cases than chance.

The OECD's published case study is explicit that this is a proof-of-concept, not an operational system: it recommends an initial phase of annual updates backed by a stable data pipeline before wider deployment, and stresses that the tool is meant to guide investigator judgement, not replace it.