Explainable Prediction of Supplier Sanction Risk in Public Procurement: Comparing Random Forest and XGBoost under Forward Temporal Validation

Abstract

Administrative sanctions against suppliers to the Peruvian state are imposed after the events that give rise to them, which limits their usefulness for guiding anticipatory oversight. This study evaluates the ability of Random Forest and XGBoost to predict debarments or fines in the following year based on the contractual and sanction trajectory of suppliers. A panel of 210,434 supplier-year observations from 94,359 suppliers awarded contracts between 2018 and 2025 was built, and the sanction history was reconstructed from archived versions of administrative registries. Using forward temporal validation over 2022–2024, both algorithms were compared with logistic regression and a model based only on prior sanctions. Random Forest achieved the best performance, with a ROC-AUC of 0.792 and a capture of 42.7% of sanctioned suppliers when the 10% with the highest risk were reviewed, compared with 0.789 and 41.5% for XGBoost. The difference between the two was significant only in PR-AUC. Annual performance ranged from 0.766 to 0.808 and reached 0.769 among suppliers without prior sanctions, who accounted for 75.2% of sanctioned suppliers. The temporal audit of sanctions and penalties barely changed performance, whereas using the current registry instead of the reconstructed one reduced the ROC-AUC by 0.027. With Random Forest, 12.5 suppliers had to be reviewed for each sanctioned supplier identified, compared with 53.2 under random selection. This study demonstrates that the reconstruction of the sanction history and the application of Random Forest and XGBoost make it possible to anticipate the risk of administrative sanction and to direct oversight toward the suppliers with the highest risk.