Forecasting the South African tax-to-GDP ratio series utilizing seasonal autoregressive integrated moving average and artificial neural networks models

Abstract

South Africa has experienced successful tax collections but has not achieved the governance outcomes desired to establish an efficient fiscal contract compared to most developed countries. The objective of this study is to compare the performance of the conventional Seasonal Autoregressive Integrated Moving Average (SARIMA) model with that of the recently developed machine learning approach, the Artificial Neural Network (ANN) model in forecasting the South African Tax-To-GDP ratio. The focus is on accurately forecasting the South African tax-to-GDP ratio using the historical series from January 2008 to November 2024. The sampled period is characterized by random, irregular, and seasonal fluctuations, which are critical for accurate forecasting in the context of macroeconomic policy. The lower mean absolute percentage error indicates that the machine learning model outperformed the conventional time series model in terms of accuracy and reliability when forecasting South Africa’s tax-to-GDP ratio. The findings further show that there will be slight growth in the tax-to-GDP ratio during the financial year of 2025, with a sharp decline forecasted between the end of 2025 and the beginning of 2026. These results add to the growing literature on the application of machine learning methods to economic forecasting. For policy considerations, this study suggests that South Africa's policy to expand its tax base, enhance tax administration efficiency, diversify revenue sources, and promote sustainable economic growth to minimize tax distortions and maintain macroeconomic stability during economic downturns.