Assessing Bayesian Forecasting for Ride-Hail Demand in San Francisco

Thesis Abstract

This study addresses the challenge of accurately forecasting ride-hail demand in San Francisco amid spatial-temporal variability, network effects, and dynamic pricing, by evaluating the effectiveness of Bayesian forecasting approaches relative to traditional time-series models. The aim is to develop a robust Bayesian framework that integrates spatial heterogeneity, temporal non-stationarity, and exogenous covariates to improve short- to medium-term demand predictions for operational decision-making in ride-hail platforms. Specific objectives include (1) to compare predictive performance of Bayesian hierarchical models against ARIMA and Prophet benchmarks across multiple boroughs and times of day; (2) to quantify the value of incorporating spatial priors, covariates such as weather, events, and traffic conditions, and surge pricing indicators on forecast accuracy; (3) to assess calibration and uncertainty quantification through probabilistic forecasting metrics; (4) to examine the transferability of the models to near-real-time forecasting using streaming data; and (5) to provide actionable implications for dispatch optimization and fleet sizing under uncertainty. The research adopts a quantitative, case-study design focused on the San Francisco ride-hail market, utilizing a population of trip records from a major platform over a 12-month period. A stratified random sample of 5,000 trip-origin–destination pairs is drawn, with oversampling in high-demand zones (e.g., downtown, Mission District, SoMa) to ensure adequate representation of spatial heterogeneity. Data sources comprise trip-level records (timestamps, origin-destination coordinates, estimated duration and distance, fare, and surge indicators), platform-provided hourly demand aggregates, and external covariates including weather (precipitation, temperature), major events calendars, public transit disruptions, and real-time traffic indices. Instruments include API-fed weather and traffic feeds and event calendars to construct exogenous covariates, supplemented by official San Francisco open data for validation of spatial delineations. Methodologically, the study implements a suite of Bayesian models, beginning with a Bayesian hierarchical time-series framework (Gaussian process plus dynamic linear model) to capture spatial-temporal dependencies, augmented with a log-Gaussian Cox process for demand intensity. Competing models include a baseline ARIMA with exogenous variables (ARIMAX) and Facebook Prophet, both equipped with covariates for weather, events, and surge pricing. Model estimation employs Markov chain Monte Carlo techniques with robust convergence diagnostics and prior elicitation grounded in literature on traffic demand and Bayesian spatial statistics. Model comparison relies on predictive accuracy measures (WAPE, RMSE, CRPS), calibration checks, and reliability diagrams, with out-of-sample validation over a rolling 28-day window. Sensitivity analyses assess the impact of prior specification, spatial granularity (neighborhood vs. grid cell), and sample size on forecast performance. Ethical considerations address data privacy and the anonymization of user-level information, with adherence to institutional review board requirements. Expected findings indicate that Bayesian hierarchical models with spatial priors and covariates yield superior predictive accuracy and well-calibrated probabilistic forecasts across most time horizons (15 minutes to 24 hours ahead) relative to ARIMAX and Prophet baselines. The integration of event indicators and traffic conditions is anticipated to substantially reduce forecast errors during peak hours and major city events, while the spatially varying coefficients reveal heterogeneous demand elasticities across neighborhoods. The probabilistic forecasts are expected to provide narrower credible intervals during stable periods and wider intervals under exogenous shocks, thereby enabling more resilient dispatch and fleet-rebalancing decisions. The study anticipates demonstrating the practical value of Bayesian predictive uncertainty for fleet sizing, surge management, and real-time decision support at San Francisco’s ride-hail operations. Contributions to knowledge encompass (i) a methodologically rigorous evaluation of Bayesian forecasting for urban ride-hail demand with explicit spatial-temporal modeling; (ii) empirical evidence on the gains from integrating exogenous covariates and spatial structure into predictive distributions; (iii) a transferable modeling framework and implementation blueprint for city-scale ride-hail platforms seeking to optimize operations under uncertainty. The conclusion emphasizes that Bayesian methods offer meaningful improvements in forecast accuracy and uncertainty quantification, suggesting policy and managerial recommendations such as adopting probabilistic forecasts for dispatch optimization, investing in real-time data integration of weather and event signals, and maintaining adaptive priors that reflect evolving urban mobility patterns. Further research directions include extending the framework to multi-modal competition scenarios and evaluating long-horizon forecasts under scenario-based planning.