Bayesian Vector Auto-Regression Method as an Alternative Technique for Forecasting South African Tax Revenue

Abstract

Tax revenue forecasts are important for tax authorities as they contribute to the budget and strategic planning of any country. For this reason, various tax types need to be forecast for a specific fiscal year, using models that are statistically sound and have a smaller margin of error. This study models and forecasts South Africa’s major tax revenues, i.e. Corporate Income Tax (CIT), Personal Income Tax (PIT), Value-Added Tax (VAT) and Total Tax Revenue (TTR) using the Bayesian Vector Auto-regression (BVAR), Auto-regressive Moving Average (ARIMA), and State Space exponential smoothing (Error, Trend, Seasonal [ETS]) models with quarterly data from 1998 to 2012. The forecasts of the three models based on the Root mean square error (RMSE) were from the out-of-sample period 2012Q2 to 2015Q1. The results show the accuracy of the BVAR method for forecasting major tax revenues. The ETS appears to be a good method for TTR forecasting, as it outperformed the BVAR method. The paper recommends that the BVAR method may be added to existing techniques being used to forecast tax revenues in South Africa, as it gives a minimum forecast error.