Forecasting the South African Unemployment Rate: A Comparative Analysis of ARIMAX and LSTM Models

This study looks into the predictive performance of linear econometric and deep learning methodologies for the South African unemployment rate quarterly data. In this paper, the Autoregressive Integrated Moving Average with exogenous variables (ARIMAX) model was compared to the Long Short-Term Memory (LSTM) network using unemployment rate quarterly data. Exploratory Data Analysis (EDA) suggested that the unemployment rate series is non-stationary, with structural breaks around 2020 and time-varying volatility. Stationarity tests established the need for differencing, whereas diagnostic tests revealed the presence of autocorrelation and ARCH effects in the raw data. The ARIMAX model added labour market covariates, and the differenced Not Economically Active (NEA) variable was statistically significant, whereas Discouraged workers were not. Although the ARIMAX model provided a good in-sample fit, residual diagnostics showed deviations from normality. Out-of-sample forecast study revealed moderate predictive accuracy, with relatively substantial forecast errors and increasing prediction intervals over time. In contrast, the LSTM model showed significant learning capacity, with early convergence and well-behaved residuals that meet both independence and homoskedasticity criteria. The model achieved significantly lower forecast errors, with RMSE, MAE, and MAPE values much lower than those of the ARIMAX model. Comparative forecast analysis using Diebold-Mariano (DM) test and model confidence Set (MCS) method and bootstrap confidence intervals consistently demonstrated the statistical superiority of the LSTM model. The findings give strong evidence that the LSTM model outperformed the ARIMAX model for projecting South African unemployment rate. The findings emphasise the importance of nonlinear modelling approaches in capturing complex labour market dynamics while also demonstrating the limitations of classic linear models. These findings also emphasise the importance of using nonlinear machine learning algorithms in macroeconomic forecasting.