Mexico CityMexico’s sanctioned-supplier list trained a fraud-detection model.
Training on the government’s banned-firm list solves machine learning’s key problem: no confirmed “clean” examples exist.
Shared directors and bid-clustering were the top predictors; researchers say the method works on any national debarment dataset.
Sources: Nature
THE LABEL PROBLEM
Supervised learning needs labeled examples of both classes. Fraud datasets have a structural asymmetry: caught fraud is labeled, uncaught fraud sits silently in the 'clean' pile. A model trained naively learns to detect what auditors already catch — not what they miss.
WHY DEBARMENT LISTS WORK
A sanctioned-supplier registry is a rare gift: it is a government-curated list of firms that have been adjudicated as fraudulent. Most countries maintain one — the World Bank's debarred-firms list, the US System for Award Management exclusions, the EU's EDES database — and they share enough structural features that a method built on one transfers to others.
THE TELLTALE SIGNATURES
Two patterns dominate procurement fraud globally. Shared directors across nominally-independent bidders signal collusive bid-rigging — the same person controls multiple 'competitors'. Bid-clustering — prices suspiciously close to each other or to the reserve price — signals coordinated bidding where firms take turns winning.
THE SCALE OF THE LEAK
Public procurement is roughly 13–15% of global GDP. The OECD estimates 10–30% of procurement spend is lost to corruption and inefficiency in developing economies. For Mexico, that translates to tens of billions of dollars annually — a fiscal hole larger than most social programs.
THE MEXICAN CONTEXT
Mexico's procurement scandals have toppled administrations. The Estafa Maestra exposed in 2017 routed ~$400 million through shell universities. The 2014 Casa Blanca affair tied the presidency to a contractor's housing gift. Detection has historically been journalistic, not algorithmic — which is why training a model on the resulting debarment list is, in effect, learning from investigative reporting.
WHY IT GENERALIZES
Corruption mechanisms are structurally similar across jurisdictions because the underlying game — extracting rents from a competitive-bidding requirement — has the same solutions everywhere. A model that learns 'shared directors plus bid-clustering' in Mexico will flag the same pattern in Brazil, Nigeria, or Indonesia. The auditors change; the geometry of fraud does not.