Detecting value-added tax evasion by business entities of Kazakhstan

Abstract

This paper presents a statistics-based method for detecting value-added tax evasion by Kazakhstani legal entities. Starting from features selection we perform an initial exploratory data analysis using Kohonen self-organizing maps; this allows us to make basic assumptions on the nature of tax compliant companies. Then we select a statistical model and propose an algorithm to estimate its parameters in unsupervisedmanner. Statistical approach appears to benefit the task of detecting tax evasion: our model outperforms the scoring model used by the State Revenue Committee of the Republic of Kazakhstan demonstrating significantly closer association between scores and audit results.