Experiment in prediction by RTP algorithm
An overview, December 28, 2006

 

 # 
Region/
target 
earthquakes
Period of 
alarm
Prediction 
was put on 
record on
Target earthquake 
Prediction
outcome
Probability of a 
random success
 
Probability of a 
random success
with R1=2.5R
1
Japan
MJMA>=7.0
Mar 27, 2003 - 
-  Jan 27, 2004
July 1, 2003
Sep 25, 2003, 
Mw=8.3
within the alarm
Correct 
0.25
(0.34)
2
California
MANSS>=6.4
May 5,  2003 -
-  Feb 27, 2004 
June 24, 2003
Dec 22, 2004, 
M=6.5 
within the alarm
Correct 
0.05
(0.07)
 
3
Southern 
California
MANSS>=6.4
Oct 29, 2003 -
-  Sep 05, 2004
May 12, 2004
 
False alarm
0.08
(0.10)
4
Honshu,
Japan
Mw>=7.2
Feb 8, 2004 -
-  Nov 8, 2004
June 1, 2004
Sep 5, 2004, 
Mw=7.4 
outside the region;  
127 km outside alarm
Near miss
(correct with 
R1=2.5R)
0.07
(0.11)
5
Northern 
Dinarides
Mw>=5.5
Feb 29, 2004 -
-  Nov 29, 2004
May 12, 2004
Jul 12, 2004, 
Mw=5.2, ML=5.7
within the alarm
Near miss
(correct for  
ML>=5.5 )
0.07
estimate was made 
for ML>=5.5
(0.08)
estimate 
for ML>=5.5
6,
6a
6b
6c
Southern 
California
MANSS>=6.4
Nov 14, 2004 -
- Dec 24, 2006
Nov 16, 2004,
Oct 5, 2005
Mar 17, 2006
Mar 30, 2006
 
False alarm 
0.19
0.30
7
Oregon 
off coast
MANSS>=6.4
Nov 16, 2004 -
-  Aug 16, 2005
Jan 29, 2005
Jun 15, 2005, 
Mw=7.2
60 km outside alarm
Near miss
(correct with 
R1=2.5R)
0.01
0.03
8,
8a
Central Italy 
M>=5.5
Jan 1, 2005
- Oct 1, 2005
- Feb 6, 2006 
Jan 29, 2005,
Oct 1, 2005
 
False alarm
0.09 
0.10 
9
Honshu, 
Japan
Mw>=7.2
June 14, 2005 -
-  Mar 14, 2006
Oct 1, 2005 
Aug 16, 2005, 
Mw=7.2 
within the alarm 
Due to technical 
delay of data, the alarm 
was determined after the  
earthquake Aug 16, 2005
 
0.05
0.14
10,
10a
Hokkaido-S. Kurils
Mw>=7.2
May 11, 2006 -
- Feb 11, 2007
- June 30, 2007
May 22, 2006
Oct 9, 2006
Nov 15, 2006
Mw=8.3 
within the alarm
Correct
0.25
0.30

 

These results lead to the following conclusions.

1. The binary score is: three successes, six false alarms, one failure to predict. This score includes neither current alarms nor the target earthquake near coast of Honshu, Japan, August 16, 2005, Mw=7.2. This earthquake is not admissible for scoring due to the technical delay of data which made impossible advance determination of alarm. This statistics is not yet sufficient for evaluating significance of RTP predictions.

2. Regarding further development of RTP algorithm: Note, that among six false alarms are three near misses: in cases ## 4 and 7 (see the Table) a target earthquake occurred near the alarm area; in one more case (#3) a rare strong earthquake has occurred, but its magnitude was by 0.3 units below the target. In cases ## 4 and 7, we would obtain correct predictions after increasing by factor 2.5 parameter R of the algorithm. The area of alarm would then expand and cover the nearly missed earthquakes. At the same time, the probabilities of random success would remain quite acceptable (compare two last columns of the Table; values obtained a posteriori are in brackets). Motivated by that retrospective observation we launched on October 1, 2005 a parallel test of the algorithm changing as indicated above only the R value. We will call the original test A, and the new test - B.


Fig. 1. RTP predictions in California, Western Nevada and Oregon.

Notations. Contours outline the area of alarms. Their color shows the outcome of prediction: Green - correct (confirmed) alarm; blue - false alarm; light green - "near miss". Circles show epicenters forming a precursory chain.  Stars show strong earthquakes targeted for prediction, within or near alarm.  Predictions, from North to South, correspond to ## 7, 2, 3, 6 in the Table.

                      


Fig. 2. RTP predictions in and near Japan.

See notations in Figure 1. Panels on top left, top right, bottom left, and bottom right correspond to lines ## 1, 4, 9 and 10 in the Table. An area surrounding the target epicenter in the top right panel (thin contour) was absent in the original prediction. The corresponding epicenter forming the precursory chain is located outside the formal region boundaries. The total shape of the alarm area is obtained in case of absence of this limitation.       
 

                   


Fig. 3. RTP prediction in Central Apennines, Alps, Northern Dinarides and Po Valley.

See notations in Figure 1. Predictions, from North to South, correspond to ## 5 and 8 in the Table.


Last updated: 2007-01-03 14:10:03