Machine learning methods are increasingly used to screen for bid rigging in public procurement. Prior research shows that supervised models can substantially improve on traditional screening rules when a sufficiently rich archive of confirmed cartel cases is available, but it also suggests that predictive performance may deteriorate when models are transferred across jurisdictions with different procurement environments. In practice, these constraints may arise jointly: confirmed domestic cartel cases may be scarce, and models trained elsewhere may not transport reliably to a different institutional setting. This work studies that joint problem directly. Using a multi-jurisdiction dataset of 40,960 bids from confirmed cartel and competitive tenders, we first evaluate cross- jurisdiction transfer by training models on all jurisdictions except the target one and testing them on the excluded jurisdiction. We then estimate within-country learning curves under extreme label scarcity by varying the absolute number of labeled tenders available for training. The results point to a consistent pattern. Cross-jurisdiction transfer often deteriorates materially, reinforcing earlier evidence that collusive bidding patterns are not fully portable across procurement environments. In our benchmark, these losses appear especially marked once performance is evaluated relative to a local oracle on a prevalence-adjusted scale. At the same time, supervised gradient boosting recovers a substantial share of attainable local performance with relatively small labeled sets in some jurisdictions, although the rate of recovery is highly heterogeneous. In this benchmark, the interaction between label scarcity and institutional heterogeneity appears more constraining than label scarcity on its own. These findings suggest that, in settings like ours, even a modest domestic labeled archive can be valuable when authorities seek to build jurisdiction-specific screening tools.
Díaz-Gil et al. (2026) studied this question.