CSEP Minutes 02-20-2019

CSEP Working Group Home Page

Participants: D. Jackson, W. Savran, D. Rhoades, A. Michael, M. Werner and W. Marzocchi

• Use histograms as opposed to ECDF using some type of cumulative statistic.
• The K-S statistic provides biased values bc we are likely always going to be seeing largest differences at the greatest magnitudes, move to cumulative statistics.
• Could also start the test above magnitudes of interest with a significance of earthquake hazard.
• Imagine a hierarchy of tests that starts at a high-level and becomes more granular.
• Look into the discrepancy in the ECDF.
• Conditional magnitude test based on the observed number of events, this would involve some type of sampling based of the combined distribution.
• Spatial distribution is very important, could be important more so than the number-distribution of the forecast; from USGS perspective.
• What is interesting about UCERF3-ETAS:
  • Spatial distribution is most important and the differentiator between UCERF3-ETAS and regular ETAS.
  • Time-frame matters; most catalogs are computed for long times such that most catalogs are circular.
  • The value of the model might be seen from the long-term behavior of the model.
• We could bin the models spatially and perform the N-test. We’d want to know how the models vary at the different between look-ahead times.
• Need effective ways of communicating CSEP tests with the public. Tests should be accompanied with meaningful figures. See figures from the USGS wrt spatial forecasts.
• Think about aggregate testing, for example, within Italy, people are interested in final outcomes. A powerful way is to use some sort of tool that David mentioned of plotting the time varying results.
• Look at the weather forecasting community.