Abstract
Objectives: On-demand Shared Mobility Systems (ODSMS) operate in stochastic and dynamic environments. Yet many simulation studies model core operational processes deterministically, failing to capture or predict real-world performance accurately. This study determines whether jointly modeling and calibrating four critical and interdependent sources of uncertainty–link travel time variability, post-match booking cancellations from riders, rider no-shows, and stop service duration variability– improves the predictive accuracy of ODSMS simulation and improves the overprediction bias of critical KPIs relative to deterministic and partially stochastic baselines.
Methods: We implement the joint stochastic modeling framework of operational uncertainties in FleetPy with components to simulate each of the four operational uncertainties. The three model systems–the full stochastic model with four operational uncertainties, a partially stochastic model with only link travel time variability, and a deterministic baseline model with none of the four operational uncertainties–are calibrated and validated using real-world operational data from the Silicon Valley (SV) Hopper microtransit system.
Findings: We find that the fully stochastic model significantly improves out-of-sample (validation) predictive accuracy for vehicle productivity, reducing the root-mean-square error by 48% compared to the deterministic and 29% compared to the partially stochastic model. It reduces the optimism bias in vehicle productivity produced by the deterministic model from about 13% to 3% and evening-peak wait-time underestimation from about 17% to 11%, with comparable accuracy on offer creation and morning-peak wait time.
