Abstract
Accurate ridership forecasting is integral for effective public transport planning and resilience, particularly during periods of systemic disruption. The Gautrain experienced significant ridership declines during the COVID-19 pandemic due to lockdown restrictions, shifts to remote working, and reduced commuter mobility. These disruptions highlighted the system’s vulnerability to sudden external shocks and underscored the importance of developing forecasting models capable of adapting to unprecedented conditions.
This study assessed the performance of traditional time-series models (ARIMA, ARIMA-GARCH, SARIMA, and Holt-Winters methods: additive and multiplicative) against machine learning methods (LSTM, CNN, and ensemble) for forecasting Gautrain rapid rail ridership. The models were trained and tested on three datasets: the full dataset, a stable pre-COVID-19 period, and a disruptive COVID-19 period. Using the mean absolute percentage error (MAPE) as the primary accuracy metric for its intuitive interpretation, the results show a context-dependent performance: the ensemble model was most robust on the full dataset (MAPE =13.04%), the Holt-Winters multiplicative model performed best in the stable pre-COVID period (MAPE =13.89%), and the SARIMA model was most accurate during the disruption (MAPE =8.85%). The LSTM model demonstrated stable performance, maintaining a comparable accuracy across all periods. Notably, the ensemble model demonstrated a remarkable 39% reduction in error during the disruption, reducing its MAPE from 16.01% to 9.95%, whereas the CNN model's performance worsened slightly.
While the models demonstrated strong performance, residual analysis during the disruption phase showed high unexplained variance (±100,000 to ±200,000), indicating the significant effect of exogenous factors that were not captured by the models. These findings demonstrate the context-dependent superiority of simpler statistical models and underscore the necessity of incorporating exogenous variables (e.g. economic indicators, population, and factors from other transport modes) to improve forecast robustness during unprecedented events.
From a management perspective, the findings highlight the crucial need for ridership forecasting strategies adaptable to different contexts. The study recommends a dual approach: using traditional models, such as the Holt-Winters or SARIMA, for periods with known stable or disruptive patterns and employing the ensemble model for overall forecasting or when future conditions are uncertain. Ultimately, this research provides Gautrain planners with practical, evidence-based guidance to enhance forecasting reliability and strengthen strategic resilience.