Predicting Early Corruption Risk in Public Procurement: Ecuador's Case Study

Public procurement accounts for nearly a third of Ecuador's public spending and is a critical vector for corruption. In response to this challenge, this work presents an earlywarning framework that predicts corruption risk in Ecuador's Reverse Electronic Auction (REA) processes. Specifically, we extend the Kapak transparency platform by constructing a stagebased dataset (2020-2023) of 2,000 REA contracts, integrating 12 binary red-flag indicators and 20 structural features per stage. Corruption risk is operationalized as a four-level label derived from a composite indicator built on quartile thresholds. To address this predictive task, we formulate prediction as a multi-class classification task and benchmark Support Vector Machines, Multilayer Perceptrons, Random Forests, and Gradient Boosting across three decision points: (i) Stage 1: Questions-Answers-Clarifications, (ii) Stage 2: Pending Award, and (iii) Stage 3: Contract Execution. Across all stages, Gradient Boosting consistently ranks first, achieving macro-F1/AUC-ROC scores of 0.60/0.84 in Stage 1, 0.61/0.85 in Stage 2, and 0.79/0.94 in Stage 3, with performance gains confirmed by Wilcoxon signed-rank tests (threshold < 0.05). Notably, high-risk contracts are the easiest to discern (AUC up to 0.91), while medium-risk levels remain challenging. Furthermore, feature-importance analysis reveals that clarifications with restrictive language, late contract publication, and post-award document changes are among the strongest predictors. Overall, the results demonstrate that meaningful corruption-risk signals are detectable well before contract execution, enabling oversight bodies to allocate investigative resources proactively and enhance procurement integrity in emerging economies.