From Crowdsourced Data to Policy Design: Monitoring and Forecasting Homeless Tents

Homelessness in the United States remains a persistent and growing crisis, further complicated by a shortage of high-quality, high-resolution data. The existing Point-in-Time (PIT) count data has several limitations, as it relies on single-night counts with inconsistent coverage and cannot capture the dynamic and seasonal nature of unsheltered homelessness. To address this gap, we integrate multiple crowdsourced data sources---including 311 Service Call records, Mapillary street-view images, and OpenStreetMap (OSM) amenities---into a spatiotemporal variational Gaussian Process (ST-VGP) model for monitoring and predicting daily homeless tent trends. Using San Francisco as a case study, we leverage quarterly tent counts from the city’s Department of Emergency Management to calibrate and validate the model over the period from January 2016 to May 2024. Our results demonstrate that crowdsourced data provide fine-grained insights into the spatial and temporal dynamics of homelessness, thereby complementing the PIT counts. We find that tents dispersed from downtown to other parts of the city following aggressive crackdown policies in 2018, highlighting the unintended consequences of short-term enforcement measures. Moreover, urban amenities and structures such as banks, restaurants, bridges, and highway ramps are significantly associated with tent locations. These findings underscore the transformative potential of crowdsourced data for designing sustainable, data-driven homelessness interventions. By offering a cost-efficient and adaptive framework, our approach enables policymakers and service providers to allocate resources proactively, evaluate policy effectiveness, and move toward systemic solutions for ending homelessness.