A neural network approach for predicting corruption in public procurement

We apply topic modelling and neural network algorithms to a sample of more than 70,000 public procurement tenders from 33 European countries between 2016 and 2018. Rather than identifying a binary indicator of possible corruption, we establish a more precise red-flag indicator with four different levels. We initially identify some selection criteria that are more present in tenders with a low number of received offers, a common proxy of corruption. Our model then detects different corruption risk profiles depending on the selection criteria reported in the procurement announcement. Tenders awarded based mainly on price criteria present a higher risk of low competition. Consequently, non-price evaluation criteria (technical quality, environmental issues, etc.) are useful indicators for preventing corruption.