Forecasting passenger rail ridership under systemic disruption: A regime-based comparison of statistical and machine-learning models

Accurate ridership forecasting is essential for public transport planning and resilience, especially during disruptions. The Gautrain, South Africa's rapid rail system, experienced significant ridership declines during the COVID-19 pandemic, highlighting the need for forecasting models that can adapt to sudden external shocks. This study evaluates traditional time-series models including ARIMA, SARIMA, and Holt-Winters against machine learning methods like LSTM, CNN, and an Ensemble model in forecasting Gautrain ridership. Three datasets were used: the full timeline, a stable pre-COVID period, and the disruptive COVID-19 period. Model accuracy was assessed using Mean Absolute Percentage Error (MAPE). Results show that forecast performance is highly context dependent. The Ensemble model was most robust on the full dataset (MAPE = 13.04%) and showed strong adaptability, reducing error by 39% during the disruption. The Holt-Winters Multiplicative model performed best in the stable pre-COVID period (MAPE = 13.89%), while SARIMA achieved the highest accuracy during the disruption (MAPE = 8.85%). LSTM maintained consistent accuracy across all periods, whereas CNN accuracy declined slightly. Residual analysis revealed increased unexplained variance during COVID-19, reflecting external influences. Overall, traditional models excel in stable conditions, while ensemble approaches are more resilient in volatile environments, suggesting that planners should adopt context-specific forecasting strategies.