Phase 2 of this project builds on Phase 1 foundational work to develop causal inference frameworks for event-driven travel behavior, and informed by those causal insights, forecasting models for the 2028 Los Angeles Olympic and Paralympic Games. This research can directly inform the work of the White House Task Force on the 2028 Summer Olympics (established by Executive Order 14328) and broader innovations in transportation planning for large-scale events.

In Phase 1 (January–June 2026), we are leveraging large-scale smartphone mobility data from approximately 100,000 anonymized mobile devices to characterize travel behavior during historical major events at approximately 20 venue sites where Olympic and Paralympic Games will be held across Los Angeles County. Analyzing mobility data linked to a Points of Interest database, we are producing descriptive analyses for two key populations: (1) event visitors, characterized by place of origin, lodging location, attendance and post-event lingering duration, and secondary activities; and (2) local residents, for whom disruptions to routine travel are analyzed before, during, and after events. From these analyses we are constructing origin-destination (O-D) matrices disaggregated by visitor/resident status, time, and event type.

Phase 2 advances this work to develop a causal inference-informed forecasting pipeline through two interconnected research thrusts. In the first stage, we will use quasi-experimental methods (synthetic control and difference-in differences) to estimate the causal effect of each historical event on mobility outcomes, isolating what changes in mobility were caused by the event vs. confounders including baseline trends and seasonal variation. In the second stage, an invariant prediction framework [1, 2, 3] will identify which event features reliably predict those causal effects across approximately 20 venues and diverse event types;those features then become the inputs to the forecasting model. This design is motivated by a key insight from recent work in causal transfer learning and causal representation learning: prediction performance improves when models rely on features whose relationship to the outcome remains invariant, i.e., stable across different environments, rather than based on correlational patterns that may not generalize. This causal-then-forecast approach is especially well suited to the proposed Olympic Games planning task for three reasons: there is large heterogeneity across training “environments” characterized by different venues, event scales, types, seasons, and audiences simultaneously; observations per environment are limited by modest sampling rates (1-3%) and a finite number of historical events, so restricting to causally stable features can provide performance-improving regularization; and the practical purpose is prospective planning for unprecedented scenarios (e.g., simultaneous events across multiple venues), where causally grounded models can more reliably extrapolate.