Predicting Public Procurement Irregularity: An Application of Neural Networks

ABSTRACT Using the data describing the characteristics of contractors provided by the Comptroller General of the Union, Brazil (CGU), this paper mainly implements two artificial neural networks, traditional neural network (TNN) and deep neural network (DNN), to develop prediction models of public procurement irregularities designed for the initial screening of contractors. This is the first application of DNN in the context of government auditing. To examine the effectiveness of DNN, the authors compare its predictive performance to TNN and two other algorithms (logistic regression and discriminant function analysis) and find that DNN significantly outperforms TNN and other algorithms in terms of accuracy, precision, F-scores, AUC, and other metrics, as suggested by the high Z-scores of the Z-tests. Although TNN has a higher recall than DNN, the difference of recall between TNN and DNN is insignificant. Logistic regression and discriminant function analysis achieve the highest recall scores, but their Z-scores are much lower than those of other metrics. Therefore, DNN generally performs more accurately than other approaches and meets the requirement of the CGU for an early alarm system.