Using Bayesian Dyalisis and Tetrads To Detect The Persistent Characteristics of The Fraud

In this paper, we propose a methodology combining Bayesian and big data tools designed to optimize the investigation of fraud. This methodology is called Bayesian dialysis. We address three issues: a) Is it possible to capitalize on the evidence provided by data indicating fraud without a parametric model and using an interpretable approach? b) If so, would it be the best solution in any case? c) What is the effect size of all unobservable, even unknown, varia-bles? We prove the viability of a new method using as an exemplary case the selection for VAT control in the Spanish Tax Agency (Agencia Estatal de Ad-ministración Tributaria —AEAT). The new method improves fraudster detec-tion precision by 12,29%, which is increased from an average of 82.28% to 94.36%. We also use 2018-2019 corporate tax data to test the scope of this ap-proach. Finally, based on the concept of tetrads, we propose a method to quan-tify the effect of unknown latent variables on models analysis.