1. K. Z. Nanjo (ERI, Univ. Tokyo)
Workshop on "Earthquake Forecast Systems Based on Seismicity of Japan:
Toward Constructing Base-line Models of Earthquake Forecasting“
1-2 Nov. 2010
CSEP earthquake forecasts based on the RI algorithm for the Japanese experiment
K. Z. Nanjo (ERI, Univ. Tokyo)

2. RI (Relative Intensity of Seismicity) Model
Hypothesis
Future large earthquakes likely occur at the locations where seismicity was active in the past
Model construction
Spatial variation of the a-value of the GR law
Smoothing seismicity to estimate a- value
Constant b-value=0.9
Image
Smoothing radius of 10 km
Color bar: Log10(forecast number of events with M≥4)
■：M≥4 events in the forecast period
Nanjo (2010 submitted to EPS)

3. Comparison among RI, PI & HSHM in CA
NSHM
National Seismic Hazard Map
★: M≥5 earthquakes that have occurred in
Colors：Alarm levels
Zechar and Jordan (2008)

4. Molchan Diagram Square: RI Triangle: PI Reference: NSHM
Nanjo (2010 submitted to EPS)

5. RI models submitted to the experiment
4 different smoothing version in each of the 9 categories
Testing region
Testing class
1 day
3 months
1 year
3 years
Total
All Japan 5 9 12 35
Mainland 2 11 7 29
Kanto 4 8 27 25 31 24 91
0.05x0.05°
d=0-100km
0.1x0.1°
d=0-30km
0.1x0.1°
d=0-100km
Nanjo (2010 submitted to EPS)

6. Highlighted parameter tuning (1)
Completeness magnitude, MC, is used to determine ML
ML: the minimum magnitude, above which all events are used to generate forecasts
RI is based on the GR a-value, which is desirably determined by using as many events as possible
All Japan: ML=3.0
Kanto: ML=2.5
Mainland: ML=2.0
Nanjo (2010 submitted to EPS)

7. Highlighted parameter tuning (2)
Constant b=0.9 is used for all the testing regions
b=0.9 is known to be a typical value for Japanese seismocity (Ishibe and Shimazaki, 2009)
Confirm this value, using data in
Nanjo (2010 submitted to EPS)

8. Highlighted parameter tuning (3)
Non-declustering (original) is used
Compare between non-declustering and declustering
Log of probability gain per earthquake (LG): LG=(LLnd-LLd)/Nt
Declustering does not always improve forecast capability
Nt: total # of target earthquakes
LLnd: nondeclustering
LLd: declustering
LLnd>LLd
LLd>LLnd
Nanjo (2010 submitted to EPS)

9. Retrospective forecast: All Japan
3yrs: 2006/1/ /2/1
1yr: 2008/1/ /2/1
3m: 2008/11/1-2009/2/1
Images show general agreement between forecast and observation
s= km
3yrs
1yrs 3m
Nanjo (2010 submitted to EPS)

10. Prospective forecasts
First 3months (Tsuruoka et al., 2010)
2009/11/ /01/31
M≥4.0
All Japan: 115 events
Mainland: 15 events
Kanto: 14 events

11. All Japan
R-test
Tsuruoka et al. (2010)

12. Mainland
R-test
Tsuruoka et al. (2010)

13. Kanto
R-test
Tsuruoka et al. (2010)

14. Discussion and conclusion
A brief overview of the RI models and its results
Large smoothing is not so good
Simple way to construct the RI model
But, it can show comparable performance to more sophisticated models, based on the first 3-month experiment (Tsuruoka et al., 2010)
My proposal
A base-line model that can be used as a reference in future predictability research ideally has to be better than the RI models
RI model must not be an ideal reference
To find the candidate(s), the experiment must be iterated.