An updated hierarchical Bayesian framework for earthquake fatality estimation

We present a framework for accurately forecasting a class of nondecreasing temporal count processes with a finite limit. Specifically, we apply this framework to rapidly estimate total earthquake-induced fatalities. We model trends in successive reported fatality counts by public media within the first hours and days after a major earthquake, using an updated Bayesian framework to forecast the likely final count. This work extends previous research with several key improvements: (1) Addressing past earthquake fatality trends that deviated significantly from the previous framework's assumptions; (2) Leveraging reported fatality data from past earthquakes to improve projections for new events; and (3) Optimizing and scaling the observed data's influence to enhance estimates and associated uncertainty. We demonstrate the predictive capabilities of this high- dimensional framework by comparing its performance to its predecessor across several past deadly earthquakes, including the 2023 M 7.8 Türkiye/Syria earthquake sequence. We show that the proposed high-dimensional updating framework improves forecasting and significantly reduces overall uncertainties compared to PAGER's initial estimates. MSC 2010 subject classifications: Primary 60K35, 60K35; secondary 60K35. • The manuscript presents a novel framework for accurately forecasting earthquake-induced fatalities using prior loss models and reported fatality data from news media outlets. • The rapid assessment of likely impact within the first hours and days after a major earthquake is critical for post-earthquake response. • The proposed framework can be adapted for forecasting impacts from other disasters using ground truth reports and helps to enhance our ability to effectively respond to natural disasters worldwide.