ARTIFICIAL INTELLIGENCE-BASED PROCUREMENT RISK SCREENING AND SINGLE-BID COMPETITION IN NIGERIA

Abstract

This study assesses whether artificial-intelligence methods add practical value when used to screen single-bid competition risk in Nigerian federal procurement. The evidence comes from Bureau of Public Procurement records published in Open Contracting Data Standard format. The underlying file contains 98,866 compiled contracting processes, of which 9,124 satisfy the study's analytical requirements: a positive tenderer count and an interpretable tender-closing date between 2017 and 2024. A single-bid process is treated as a signal of weak competition that may justify closer examination; it is not treated as evidence of corruption. To keep the information sequence intact, each buyer-history variable is calculated only from procurement processes that closed before the process being scored. Random forest and logistic regression are estimated with observations from 2017-2022 and evaluated on a later 2023-2024 holdout, together with simpler benchmarks based on buyer history and procurement method. Random forest produces a ROC AUC of 0.761, average precision of 0.633 and a Brier score of 0.190. Logistic regression produces the same AUC of 0.761, average precision of 0.619 and a Brier score of 0.187. A buyer-cluster bootstrap based on 1,000 replications provides no statistically meaningful evidence that the two AUCs differ. Both multivariable models exceed the buyer-history-only benchmark, whose AUC is 0.705, while logistic regression is more closely calibrated. The findings show that structured procurement records can support the prioritisation of audit attention, but they do not justify assuming that greater model complexity automatically produces better decisions. AI-assisted procurement assurance should therefore be built around transparent comparison models, strict control over when information becomes available, regular calibration review and professional judgement.

DOI: https://doi.org/10.5281/zenodo.23123223