SCEC Award Number 13184 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Short-Term Earthquake Predictability: Improving Stochastic Models and CSEP Experiments
Investigator(s)
Name Organization
Maximilian Werner Princeton University
Other Participants Dr. Agnes Helmstetter
SCEC Priorities 2b, 2e, 4e SCEC Groups CSEP, EFP, Seismology
Report Due Date 03/15/2014 Date Report Submitted N/A
Project Abstract
 The objective of this project was to establish the predictive skills of seismicity forecasting models in the immediate and short term. The ultimate goal is a better understanding and description of triggered seismicity as well the development of improved models for Operational Earthquake Forecasting (OEF). We designed modular Epidemic Type Aftershock Sequence (ETAS) models with a variety of spatial aftershock kernels (isotropic and anisotropic) as well as different Omori-Utsu law parameterizations. We estimated parameters of these models using Californian data between 1984 and 2012 and compared the information gains per earthquake and the Akaike Information Criterion (AIC), which penalizes for additional parameters. We find that the spatial power-law kernel by Zhuang performs better than other spatial kernels by Helmstetter, Kagan, Ogata and Werner, despite the additional parameter. We also find that estimating anisotropic kernels obtained by smoothing early aftershock locations improves the likelihood score. Including small M>2 earthquakes for forecasts of M>3.95 targets increases the predictive skills of all models. Using a c value in the Omori-Utsu law that scales with magnitude is not warranted by our analysis. We have also evaluated the information gain as a function of forecast horizon from 30 minutes to 24 hours, showing substantial increases with smaller intervals.
Intellectual Merit We have quantified how the short and immediate term predictability of earthquakes depends on forecasting model choices such as the spatial aftershock distribution, the Omori-Utsu law formulation, and the learning catalog magnitude threshold. The literature contains many different ETAS model formulations that are each tested on different data sets with different model choices; here, we compare a large selection of these side-by-side in a consistent manner to determine optimal choices.
Broader Impacts Possible benefits to society may include improved operational earthquake forecasts based on rigorous evaluation and calibration of statistical seismicity models. ETAS models are a popular choice for OEF (see, e.g., UCERF-ETAS and the USGS global aftershock advisories); investigating optimal model choices may help increase their reliability and accuracy, which in turn may lead to better hazard and risk estimates and governance.
Exemplary Figure Figure 5: Influence of updating interval on information gain per earthquake for target earthquakes $M\geq 3.95$ in California between 1992 and 2012 for different learning catalog magnitude thresholds (colors). Anisotropic aftershock kernels are estimated by smoothing early aftershocks and provide a small but consistent increase. (Werner et al., 2015, in preparation)