Abstract
This study models perceived safety in South Africa through a comparative evaluation of statistical and machine learning (ML) approaches. Using 215,301 observations from Statistics South Africa, the study examines perceptions of safety when walking alone during the day and after dark, distinguishing between Safe and Unsafe perceptions. Stationarity tests indicate that all series are stationary in levels, while the AIC identified VARMA (3,1) as the selected baseline specification. To assess whether ML methods provide additional predictive value, the VARMA model was integrated with RF and LightGBM algorithms using VARMA residuals, resulting in VARMA–RF and VARMA–LightGBM hybrid models. The models are evaluated using MSE and RMSE. The results indicate that the hybrid models do not improve predictive performance relative to the standalone VARMA model. Across the two perceived-safety series, the VARMA specification produces lower MSE and RMSE values than both the VARMA–RF and VARMA–LightGBM models. These findings suggest that the temporal dependence captured by VARMA provides substantial predictive information, while the incorporation of ML algorithms does not necessarily yield additional forecasting gains in this application. The findings highlight the importance of empirically evaluating model complexity and predictive performance rather than assuming that hybrid statistical and ML approaches will outperform conventional time-series models. The study provides empirical evidence on the comparative performance of statistical and ML approaches for modelling perception-based public safety measures in the South African context.