SCEC Award Number 15169 View PDF
Proposal Category Collaborative Proposal (Integration and Theory)
Proposal Title Ensemble Earthquake Forecasting Models for the 2010 Darfield, New Zealand, Earthquake Sequence
Investigator(s)
Name Organization
Maximilian Werner University of Bristol (United Kingdom) Matthew Gerstenberger GNS Science (New Zealand) Maria Liukis University of Southern California Warner Marzocchi Istituto Nazionale di Geofisica e Vulcanologia (Italy)
Other Participants
SCEC Priorities 2b, 2d, 2e SCEC Groups WGCEP, EFP, CSEP
Report Due Date 03/15/2016 Date Report Submitted 11/14/2016
Project Abstract
 The project’s objective was to investigate the ability of ensemble models to forecast the 2010 Canter-bury, New Zealand, earthquake sequence. Ensemble modeling refers to methodologies for merging forecasts from multiple models according to different criteria. Ensemble models offer two main ad-vantages. First, no single model has to be selected a priori for making decisions. Second, merged fore-casts may perform better than individual forecasts. Ensemble models may therefore contribute to Opera-tional Earthquake Forecasting (OEF) systems. The Darfield earthquake sequence offers a unique oppor-tunity for studying the performance of ensembles. The Collaboratory for the Study of Earthquake Pre-dictability (CSEP) had previously assembled fourteen individual forecast models for the sequence. We developed four different ensemble models for the 1-day and 1-month forecasting experiments, using both real-time and best-available datasets. The weights of models in our ensembles are based on a measure of past performance. The weights reveal temporal fluctuations that show model performance evolves substantially over time. We ranked the ensembles against individual models using the likelihood metric. All ensemble models perform almost as well as the best individual model. This suggests that en-semble models provide a robust OEF approach for merging forecast without selecting one model a pri-ori. However, the current approaches do not merge information in a complimentary manner to improve on the single best forecast (which is, however, only known a posteriori). Two ensemble models are now implemented within the California testing region and are running live.
Intellectual Merit Our results contribute to SCEC’s goals of improving the science of Operational Earthquake Forecasting (OEF). This project advanced our understanding of how forecasts can be merged for optimal performance. Bayesian Model Averaging (BMA) techniques are well known, but many other ensembles are possible and indeed perform comparatively well. Ensemble models also provide new tools for visualizing the performance of models over time.
Broader Impacts The INGV uses an ensemble model of CSEP-tested individual models to provide real-time information about the time-dependence of seismic hazards to the Italian Civil Protection Agency. Other government agencies, including the USGS, are making plans to deploy OEF systems. Our results show that the tested ensembles perform well: they are robust and suitable for OEF systems. The project has strengthened collaborations between CSEP nodes at SCEC, INGV, GFZ Potsdam and GNS Science.
Exemplary Figure Figure 1: Evolution of weights in the 1-month ensemble model forecasts over the 20 forecast horizons of the 2010-12 Canter-bury earthquake sequence. Weights are calculated from the cumulative performance of individual models; performance met-rics vary between ensemble models. Left panel: best-available data is used to generate forecasts and ensembles. Right pan-el: real-time data is used. The BMA and gSMA show strong sensitivity to past performance and interesting fluctuations. The SMA and PGSMA ensembles are more robust and conservative. [from Taroni, Marzocchi, Werner & Zechar, 2015, in prep.]