Predictive Analytics for Controlling Tax Evasion Using Deep Learning: The Case of Ethiopia Customs Commission Taxpayers.

Abstract

Tax evasion is an illegal evasion by a person or a company. This is the intentional misrepresentation of a tax return in order to reduce your liability in tax. Tax evasion represents one of the big dilemmas of governments and tax authorities around the world, but most especially in developing countries. Tax authorities control tax evasion through a system based on Tax audits. Tax audit selection is based on audit selection criteria, auditor experience, and a non-adaptive randomization selection manual system. The process therefore faces subjective judgments, limited data utilization, and susceptibility to corruption all these ultimately cause tax evasion. Herein, we present predictive analytics for controlling tax evasion using deep learning. In this study, raw data is collected from three sources: Customs Commission (138,644 data), domestic (15,070 data) and post-clearance audit history (2,686 data) over 5 years in the Ethiopia Customs Commission taxpayers. Using domain knowledge we integrated and pre-processed data. The final pre-processed dataset is the 15070 dataset; then we split the dataset to 80% training dataset(12056), 10%test datasets(1507), and 10% validation datasets(1507).The data have been trained using deep neural networks, AutoEncoder, and feed-forward neural networks then we used two hyperparameter tuning techniques random search CV and Bayesian optimize; and transfer learning to accurate predictions. We have predicted five consecutive annual selections for a tax audit and three highest tax evasion conditions. We Model then evaluated its mean bias, mean squared error, root mean squared error, R-squared, explained variance score, and training and validation loss comparing two hyperparameter tuning technique. We concluded that the random search CV hyperparameter tuning technique better by reducing error. We also determined using random search CV the feed-forward neural network the best model, with a mean bias of - 0.0009, mean squared error of 0.0000, and root mean squared error of 0.0038. The best performance yielded an R-squared of 99.98%, explained variance core of 99.98%, and accurate prediction of tax audit selection and three top tax evasion cases over 5 consecutive years. Predictive analysis of each variable is done sequentially but feature researcher uses any parallel processing system to minimize analysis time and integrated it with behavioural model.