Data Science for the Assessment of Operation Errors and their Impact

Abstract

The vast amount financed annually by the European Union for community funds requires rigorous monitoring of the operations allocated to funds. Nowadays, this monitoring is achieved through auditing, which is laborious, costly, and naturally done solely upon a sample of operations, not guaranteeing the conclusiveness of the results. This study explores the feasibility of applying data-science techniques to predict the error of the operations, in complement or, ultimately, replacement of the auditing procedure. We tested two estimation methods based on the data from three funds over two auditing years. We compared them with a benchmark estimation approach in terms of the ability and precision to predict the total error amount and conclusiveness. We experimented with the framework of this study with 1-stage and 2-stage approaches, the last adding a classification step to filter only the operations with error to the prediction stage. Our findings confirm that the assessment of operation error benefits from using a model compared to the traditional estimation method, especially when using the model-assisted estimation. Furthermore, a 1-stage prediction pipeline produces satisfactory results, which are not improved with an additional modeling step.