Article Preview
TopIntroduction
The South African public policy budget continues to rise with a spending growth of about 49% over the last decade; moving from ZAR 33764 billion ($2597 billion) in 2010/2011 to ZAR 1.79 trillion ($0.16 trillion) in 2018/2019. Included in this expenditure category are health care and social protection, which represent respectively 12% and 14% of the total government spending in 2018 (SSA, 2015 and 2019) and are expected to escalate with aging population. South African population is estimated at 57.3 million in 2018, 8.3% of whom are aged over 60 years, with a predicted increased in the share of elderly to 9.1% in 2022. While the old age grant has grown to about 30% of the social spending in 2019, the country faces a hiking chronic disease burden as more South Africans live beyond the age of 60 years (Solanki et al., 2019). In addition to the Covid-19 outbreak, which has triggered the health spending, 40% of older population in South Africa are poor and depend entirely on the government for their basic survival. The uncertainty about the future increase requires accurate forecasting of such expenditures, in order to assist government policy makers with their planning, assessment, and budget allocation in support of funding decisions. The present study addresses this specific objective by introducing a machine learning model to forecast public expenditures based on the historical trend in population aging; with the ultimate goal of providing a new support framework for the budgetary policy decision.
A sound governmental fiscal planning and management requires accurate expenditures predictions. This is particularly crucial for developing countries characterized by persistent budget imbalances. In addition, modifications in fiscal expenditures due to unanticipated changes in socio-economic and demographic conditions entail delayed effects, which necessitate accurate forward-looking policy strategies and/or actions. To address such uncertainties, the economic and forecasting literature (Lassila et al., 2014; Kudrna et al., 2015) advocates incorporation of both demographic and macroeconomic factors in predicting future government expenditures.
Existing literature offers a variety of approaches for forecasting government expenditures using time series-based forecasting. Recently, researchers have proposed Artificial Neural Networks (ANN) as an alternative approach to time series-based forecasting (Zang et al., 1998; Zang, 2003; Ghiassi et al., 2006; Ghiassi et al., 2008; Ramyar & Kianfar, 2019 among others). For example, Ghiassi et al. (2006) introduce a dynamic architecture for artificial neural network (DAN2) to forecast electric consumption for Taiwan, and Ghiassi et al. (2008) follow the same methodology to forecast urban water demand for the city of San Jose, California. They find that DAN2 delivers exceptional fit and forecasts using time series datasets. Similarly, Basaran et al. (2010) implement a feed forward neural network architecture to forecast public expenditures in Turkey and obtain high prediction precision. In the same vein, Ramyar & Kianfar (2019) compare the forecast ability of a multilayer perceptron neural network to the vector autoregressive model in forecasting oil prices and conclude that machine learning improves prediction accuracy.
This study aims to develop a model based on a traditional approach (ARIMAX) and compare and contrast its performance against a model based on DAN2 to predict government expenditures in South Africa. The accurate forecast values from these models can be used to assist policy makers to better manage and/or keep fiscal imbalances at sustainable levels. Moreover, with the increasing pace of elderly cohort expected to trigger fiscal pressures, this study addresses the inclusion of demographic changes and analyses the role of population aging in such forecast models.