SCEC Award Number 13026 View PDF
Proposal Category Travel Only Proposal (SCEC Annual Meeting)
Proposal Title CSEP Participation, Model development, and Testing and Optimization of Hybrid Models
Investigator(s)
Name Organization
Matthew Gerstenberger GNS Science (New Zealand) David Rhoades GNS Science (New Zealand)
Other Participants
SCEC Priorities 2 SCEC Groups CSEP, EFP, Seismology
Report Due Date 03/15/2014 Date Report Submitted N/A
Project Abstract
 The major work undertaken in this project concerned the testing and optimization of the operational hybrid earthquake forecasting model for Canterbury for the next 50 years – the Expert Elicitation (EE) model, and an assessment of information gains that could be obtained by forming multiplicative hybrids of models in the Regional Earthquake Likelihood Model (RELM) five-year experiment in California. In the Canterbury study, the EE model was evaluated against its individual components over 26 years in the whole New Zealand CSEP test region with time lags up to 25 years. The main result was that the EE hybrid was more informative than most of the individual models for most of the time, even when looking ahead for up to 25 years. In the California study, multiplicative hybrids were constructed involving the best individual model from the RELM experiment (the Helmstetter et al. HKJ model) as a baseline with each of the other models in turn as conjugate models. Many two-model hybrids were found to have an appreciable information gain (log probability gain) per earthquake relative to the best individual model. In contrast, no additive hybrids of the same models gave an appreciable information gain over the best individual model.
Intellectual Merit In this project we have applied rigorous statistical procedures to the investigation of hybrid earthquake forecasting models in New Zealand and California. These studies consistently demonstrate the superiority of hybrid models, based on a range of different ideas or data inputs, over individual models, based on a single idea and data input. The retrospective testing described here will be supported by independent prospective testing of the hybrid models in the CSEP testing centers in New Zealand and California.
Broader Impacts The challenge of continually increasing the information value of earthquake forecasting models can be met by learning how to integrate information from a variety of data and modelling inputs into hybrid models. The studies described take a step in that direction. We aim to provide systematic methods of assimilating new elements–data or modelling inputs–into statistical forecasting models, by fitting a few extra parameters for each new element. This is similar to the way in which multiple regression analysis can be used to explain much more of the variation of a response variable than any individual explanatory variable could on its own. We anticipate that these methods can be applied in the CSEP testing centers around the world to derive hybrid models which are more informative than the individual models submitted to the centers. The method applied in the RELM study can also be used to integrate models of different types (e.g. likelihood models, alarm-based models) and with different updating intervals into hybrid earthquake likelihood models.
Exemplary Figure Figure 3. Map of earthquake rates, relative to reference (RTR), in the best three-model hybrids from the RELM experiment for (a) the whole of California: a hybrid of Helmstetter et al. HKJ, Bird & Liu Neokinema, and Holliday et al. PI; and (b) southern California: a hybrid of Helmstetter et al. HKJ, Shen et al. Geodetic, and Holliday et al. PI. In the reference model, one earthquake per year is expected to exceed any magnitude m in an area of 10m km2. This figure is from Rhoades et al. (BSSA submitted), [4] in reference list.