An efficient Bayesian framework for updating PAGER loss estimates

Abstract

We introduce a Bayesian framework for incorporating time-varying noisy reported data on damage and loss information to update near real-time loss estimates/alerts for the U.S. Geological Survey’s Prompt Assessment of Global Earthquakes for Response (PAGER) system. Initial loss estimation by PAGER immediately following an earthquake includes several uncertainties. Historically, the PAGER’s alerting on fatality and economic losses has not incorporated location-specific reported data on physical damage or casualties for a given earthquake. The proposed framework provides the ability to include early reports on fatalities at any given time and improve the overall impact forecast for the earthquake. The reported data on fatalities or damage are generally incomplete and noisy, especially in the early hours of the disaster. To address these challenges, we develop a recursive Bayesian updating framework that takes into account the loss projection model and the measurement and model uncertainties. The framework is applied to loss data for three example earthquakes, and the results show that the proposed updating improves the loss estimates and alert level to the correct level within the first day of the earthquake.