第56回統計地震学セミナー / The 56th Statistical Seismology Seminar
- Date&Time
- 2016年6月7日
(火)
/ 7 June, 2016 (Tuesday) 16:00 – 17:00
- Place
- 統計数理研究所 セミナー室6 (A508)
/ Seminar room 6 (A508) @ The Institute of Statistical Mathematics
- Speaker
- Dr. Anne Strader (GFZ German Research Centre for Geosciences)
- Title
- Evaluation of Current CSEP Testing Methods: Case Studies for Japan and California
- Abstract
- The Collaboratory for the Study of Earthquake Predictability (CSEP) was developed to rigorously test earthquake forecasts retrospectively and prospectively through reproducible, completely transparent experiments within a controlled environment (Zechar et al., 2010). Forecasts are individually evaluated using a set of likelihood-based consistency tests, which measure the consistency between the number, spatial and magnitude distribution of the observed and forecasted seismicity during the testing period (Schorlemmer et al., 2007; Zechar et al., 2010). Additionally, the classical paired t-test and non-parametric w-test are used to directly compare two forecasts' performances at target earthquake locations. These tests rely on a hypothesis testing framework, resulting in a final decision (to reject or not reject a forecast), rather than quantifying the model's lack-of-fit or localized performance. Residual methods are employed by the CSEP to discern spatial variation in model performance compared to the observed seismicity distribution and other models, but are not currently incorporated into decision-making processes. To illustrate what can be learned from commonly utilized current CSEP tests, we present two case studies. The first is a retrospective evaluation of a rate-and-state forecast for the Japan CSEP testing classes, where spatiotemporal seismicity rate fluctuations are inverted for Coulomb stress changes. Although the model underestimates the number of earthquakes following the M9.0 Tohoku mainshock, it displays positive information gain over baseline ETAS seismicity rates (Ogata, 2011) within the rupture region. The second forecasting experiment is a continued prospective evaluation of the time-independent California earthquake forecasts tested in the Regional Earthquake Likelihood Model (RELM) experiment, from 2011-2016. Additionally, we test two models developed by the United States Geological Survey (USGS): the time-dependent Uniform California Earthquake Rupture Forecast (UCERF2) and time-independent National Seismic Hazard Mapping Project (NSHMP) models. To reduce bias from expert-based decision making utilized in current testing methods, we introduce the framework of a Dynamic Risk Quantification (DRQ) platform, that will be developed to combine and optimize ensemble forecasts and hazard models using a data-driven approach, and updated as new data become available.
