Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Bivariate Count Processes for Earthquake Frequency Mathieu Boudreault & Arthur Charpentier Université du Québec à Montréal
[email protected] http ://freakonometrics.blog.free.fr/
Séminaire GeoTop, January 2012
1
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Figure : Time and distance distribution (to 6,000 km) of large (56.5) in days
4
Number of earthquakes (magnitude >4) per 15 sec., average before=100
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
220
Number of earthquakes before and after a major one, magnitude of the main event, small events more than 4 ● ●
200
Main event, magnitude > 6.5 Main event, magnitude > 6.8 Main event, magnitude > 7.1 Main event, magnitude > 7.4
●
●
180
● ● ●
80
100
120
140
160
●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ●● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ●● ● ●●● ●● ● ● ● ●●●● ●● ●● ● ●● ● ● ●● ● ●● ●● ● ● ●● ●● ●● ● ● ● ●●●● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ●● ● ●●● ●● ●●● ●● ● ●● ●● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ●●● ● ● ●● ● ●●● ● ● ●● ●●● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ●● ● ● ● ●● ●●● ● ● ●● ●● ●● ● ● ● ●●● ● ● ● ●●● ● ●● ● ●●● ●●● ● ● ●● ● ●● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ●●●● ● ●●●● ● ● ● ● ●●● ● ●● ●●● ● ● ● ● ●● ● ● ●●● ● ● ●● ● ● ●● ●●● ● ● ●●●●● ● ●● ●● ●●●● ● ● ● ●●● ● ● ● ● ●● ●● ● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ●● ●● ●●●● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ●●●●●●●●● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●●●● ● ●●● ● ●● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●●● ● ● ● ● ● ●● ●● ● ●●● ●●● ●● ●● ●●● ● ● ●● ● ● ●●●●●● ● ● ●●● ●● ●●● ●● ● ● ● ● ● ●● ●● ● ● ● ● ●●● ● ● ● ● ● ●● ●● ●● ●●● ●●●●● ● ●● ● ●● ●●● ●● ● ● ● ●● ● ● ●●●● ● ●● ●●● ● ● ●●● ●●● ●●● ● ●● ● ●● ●●● ● ● ●● ● ● ● ● ●● ●● ● ● ●● ●● ● ●● ●● ● ●● ● ● ● ● ● ●●●● ●● ● ● ●●● ● ●● ●● ● ●●● ● ● ● ● ●● ●●● ● ● ● ● ●●● ● ●● ● ●● ● ●●● ● ●● ●● ● ●● ● ● ●●● ● ● ●● ●● ● ● ●●● ● ●● ● ● ●● ●● ● ●● ●● ●● ● ● ● ● ● ●● ●● ●●●●●● ● ● ●● ● ●● ● ●● ●●● ●●●● ●● ●● ● ●●● ●● ●●● ●●● ● ●● ●● ●● ● ●● ● ● ●● ● ●●● ●● ●● ● ● ● ● ● ● ●●● ●●● ●●● ●● ● ● ●● ● ●● ●●●● ● ● ● ● ● ● ● ●● ● ●● ●● ● ●●●● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ●●● ● ● ●●●●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ●● ●●● ●●● ● ●● ● ●●●● ●● ● ●●● ● ● ● ● ●● ● ●● ● ●● ●● ● ● ●● ●● ● ● ● ●●● ● ● ● ● ●●● ●● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ●● ●● ●● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ●●●●● ● ● ●● ● ●● ● ● ●● ●●●● ● ● ● ● ● ●●●● ● ●● ● ●● ●● ●● ● ● ● ●● ● ● ● ● ●●●● ●● ● ● ● ● ● ● ●● ●●● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●● ● ●● ●● ● ●●● ● ● ● ● ●●●● ● ●● ● ●●● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ●● ●● ● ●● ● ●●● ●● ●● ● ●● ●● ● ●●● ●● ●● ●●● ● ● ●● ● ● ●● ● ●● ● ● ●●● ● ●● ● ● ● ● ● ●● ● ●●● ●●● ● ●● ● ●●●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●● ● ●● ● ● ●● ● ●●● ● ● ● ● ● ● ●●● ●● ●● ● ● ●● ● ●●● ● ●● ● ●●●● ● ●● ●● ● ● ● ●● ●●●●● ●● ● ● ● ●●● ●● ●● ●● ●● ● ● ●● ● ●● ● ● ● ● ● ● ●● ●● ● ●●● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ●● ● ●●●● ●●●●●● ● ● ●●● ● ●● ● ● ● ●● ● ●● ●● ● ● ●● ● ●● ● ● ● ● ● ●●● ● ● ● ●● ● ●● ●●●●●● ●● ●● ● ● ●● ●● ● ●●● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ● ●● ●●●● ● ● ● ● ● ●● ●●●● ● ●●● ●● ● ●●● ● ● ●●● ● ●● ●● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ● ●●● ● ●● ● ●● ●● ●●●● ●● ●● ● ● ● ● ● ● ● ● ● ●●● ●●● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ●● ●● ● ● ●●●● ●● ● ● ● ●● ● ● ●●●●● ● ● ●● ●●● ● ● ●●● ●● ● ●● ● ● ● ● ● ●●●●●●● ● ● ●● ● ● ● ● ●● ● ● ●● ● ●● ● ● ●●● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ●● ●● ● ●● ●●●● ●● ●●● ● ●●● ● ● ●●●● ● ●● ●● ●● ●● ●●● ●● ● ● ● ● ● ● ● ●● ●●●● ●● ●● ● ● ● ● ●● ● ● ● ●● ●● ●●●●●● ● ● ●●●● ● ●● ● ●● ●● ●● ●●● ● ● ● ● ●● ● ●●● ● ● ● ●●● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ●● ● ● ● ● ● ●● ●● ● ●● ● ●● ● ●● ●● ●● ● ● ●● ● ● ●●●● ● ●●● ● ● ● ●● ●●● ● ● ●●● ●● ●● ● ●● ●● ●● ● ● ●● ● ● ●● ● ●●● ●● ●● ●●●●● ● ● ● ●●● ●●● ● ●●●● ● ● ● ●● ● ● ●● ● ● ●●● ● ●●● ● ●●● ●● ●●●●●● ● ●●● ● ●● ● ● ● ●● ●● ● ●●● ● ● ● ● ● ●● ●● ●●● ● ●●● ●● ●●●●● ●● ● ●●● ● ● ●● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
−15
−10
−5
0
5
10
15
Time before and after a major eathquake (magnitude >6.5) in days
5
1000
●
● ●
800
Same techtonic plate as major one Different techtonic plate as major one ● ●
●
200
400
600
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ●● ●● ● ● ● ● ●●● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●●● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ●● ● ●●● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●●● ● ● ●●●●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●●● ●●●●●● ●● ● ● ● ● ●● ● ●●● ● ●● ● ●● ●● ●●●● ● ● ●● ● ● ● ●●●●● ● ● ● ● ● ● ●
0
Number of earthquakes (magnitude >2) per 15 sec., average before=100
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
−15
−10
−5
0
5
10
15
Time before and after a major eathquake (magnitude >6.5) in days
6
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Shapefiles from http://www.colorado.edu/geography/foote/maps/assign/hotspots/hotspots.html
7
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
We look at seismic risks with the eyes of actuaries and statisticians...
8
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
9
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Agenda • Motivation (Parsons & Velasco (2011)) • Modeling dynamics ◦ AR(1) : Gaussian autoregressive processes (as a starting point) ◦ VAR(1) : multiple AR(1) processes, possible correlated ◦ INAR(1) : autoregressive processes for counting variates ◦ MINAR(1) : multiple counting processes • Application to earthquakes frequency ◦ counting earthquakes on tectonic plates ◦ causality between different tectonic plates ◦ counting earthquakes with different magnitudes
10
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Part 1
Modeling dynamics of counts
11
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
(ANSS) http://www.ncedc.org/cnss/catalogsearch.html
350
Number of earthquakes (Magnitude ≥ 5) per month, worldwide
●
300
●
●
● ● ●
●
●
250
● ●
● ● ●
●
● ●
200
●
150
●
●
●
●
●
● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ●● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ●●● ● ●● ●●●● ●● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ●●● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ●● ● ● ●●● ● ● ● ● ● ● ●
● ●
●
● ●
●
● ●
●
50
100
● ●
●
●
1970
1980
1990
2000
2010
Number of earthquakes (Magn.>5) worldwide per month
12
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
(ANSS) http://www.ncedc.org/cnss/catalogsearch.html Number of earthquakes (Magnitude ≥ 5) per month, in western U.S.
15
●
●
●
10
●
● ●
● ● ●
5
● ●
●
● ●
●
●
● ●●● ●●
0
● ● ●●●●●●●●● ●● ● ● ● ● ●●●
●●
● ●●
●●●● ● ●● ●●
● ● ●●● ●
●
● ●
●●
●
● ●
●●● ● ●●● ● ● ●●● ● ●
●
● ● ●
●●
●
●● ●
●
●● ● ●
● ● ● ●● ●●
● ●
●
●● ●●●● ●● ● ● ● ●● ● ● ●● ● ●● ●●● ● ●● ● ● ● ● ● ●● ●● ●●●● ● ● ● ●● ●●●●●● ● ● ●●●●● ● ● ●●● ● ●
●● ● ● ●
●
●●●
●
●●●●
● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ●●●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●
1950
1960
1970
1980
1990
2000
2010
Number of earthquakes (Magn.>5) in West−US per month
13
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
(Gaussian) Auto Regressive processes AR(1) Definition A time series (Xt )t∈N with values in R is called an AR(1) process if Xt = φ0 +φ1 Xt−1 + εt
(1)
for all t, for realvalued parameters φ0 and φ1 , and some i.i.d. random variables εt with values in R. It is common to assume that εt are independent variables, with a Gaussian distribution N (0, σ 2 ), with density 2 1 ε ϕ(ε) = √ exp − 2 , ε ∈ R. 2σ 2πσ Note that we assume also that εt is independent of X t−1 , i.e. past observations X0 , X1 , · · · , Xt−1 . Thus, (εt )t∈N is called the innovation process.
14
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Example : Xt = φ1 Xt−1 + εt with εt ∼ N (0, 1), i.i.d., and φ = 0.6
● ●
● ● ●
● ●
●
2
●
●
0
●
−2
●
● ●● ● ●
● ●
● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●
●
●
−4
● ●●
0
50
100
150
200
250
300
15
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Example : Xt = φ1 Xt−1 + εt with εt ∼ N (0, 1), i.i.d., and φ = 0.6
● ●
● ● ●
● ●
●
2
●
●
0
●
−2
●
● ●● ● ●
● ●
● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●
●
●
−4
● ●●
0
50
100
150
200
250
300
16
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
0.4 0.2 0.0
ACF
0.6
0.8
1.0
Example : Xt = φ1 Xt−1 + εt : autocorrelation ρ(h) = corr(Xt , Xt−h ) = φh1
0
5
10
15
20
Lag
17
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Definition A time series (Xt )t∈N is said to be (weakly) stationary if • E(Xt ) is independent of t ( =: µ) • cov(Xt , Xt−h ) is independent of t (=: γ(h)), called autocovariance function
Remark As a consequence, var(Xt ) = E([Xt − E(Xt )]2 ) is independent of t (=: γ(0)). Define the autocorrelation function ρ(·) as ρ(h) := corr(Xt , Xt−h ) = p
cov(Xt , Xt−h ) var(Xt )var(Xt−h )
=
γ(h) , ∀h ∈ N. γ(0)
Proposition (Xt )t∈N is a stationary AR(1) time series if and only if φ1 ∈ (−1, 1). Remark If φ1 = 1, (Xt )t∈N is called a random walk. Proposition If (Xt )t∈N is a stationary AR(1) time series, ρ(h) = φh1 ,
∀h ∈ N.
18
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
From univariate to multivariate models Density of the Gaussian distribution
Univariate gaussian distribution N (0, σ 2 ) 2 x 1 exp − 2 , for all x ∈ R ϕ(x) = √ 2σ 2πσ
0.20 0.15 0.10 0.05 0.00 3 2 1 0 −1 1
−2
2
3
Multivariate gaussian distribution N (0, Σ) x0 Σ−1 x 1 , ϕ(x) = p exp − d 2 (2π)  det Σ for all x ∈ Rd .
0 −1
−3
−2 −3
X = AZ where AA0 = Σ and Z ∼ N (0, I) (geometric interpretation)
19
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Vector (Gaussian) AutoRegressive processes V AR(1) Definition A time series (X t = (X1,t , · · · , Xd,t ))t∈N with values in Rd is called a VAR(1) process if X1,t = φ1,1 X1,t−1 + φ1,2 X2,t−1 + · · · + φ1,d Xd,t−1 + ε1,t X = φ X 2,t 2,1 1,t−1 + φ2,2 X2,t−1 + · · · + φ2,d Xd,t−1 + ε2,t (2) ··· X = φ X d,t d,1 1,t−1 + φd,2 X2,t−1 + · · · + φd,d Xd,t−1 + εd,t or equivalently φ X 1,t 1,1 X2,t φ2,1 .. = .. . . Xd,t  {z } Xt
φd,1

φ1,2
···
φ2,2 .. .
···
φd,2 · · · {z Φ
φ1,d
ε X 1,t−1 1,t φ2,d X2,t−1 ε2,t .. .. + .. . . . φd,d Xd,t−1 εd,t }  {z }  {z } X t−1
εt
20
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
for all t, for some realvalued d × d matrix Φ, and some i.i.d. random vectors εt with values in Rd . It is common to assume that εt are independent variables, with a Gaussian distribution N (0, Σ), with density 0 −1 εΣ ε 1 exp − , ∀ε ∈ Rd . ϕ(ε) = p 2 (2π)d  det Σ Thus, independent means time independent, but can be dependent componentwise. Note that we assume also that εt is independent of X t−1 , i.e. past observations X 0 , X 1 , · · · , X t−1 . Thus, (εt )t∈N is called the innovation process. Definition A time series (X t )t∈N is said to be (weakly) stationary if • E(X t ) is independent of t (=: µ) • cov(X t , X t−h ) is independent of t (=: γ(h)), called autocovariance matrix 21
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Remark As a consequence, var(X t ) = E([X t − E(X t )]0 [X t − E(X t )]) is independent of t (=: γ(0)). Define finally the autocorrelation matrix, q ρ(h) = ∆−1 γ(h)∆−1 , where ∆ = diag γi,i (0) . Proposition (X t )t∈N is a stationary AR(1) time series if and only if the d eignvalues of Φ should have a norm lower than 1. Proposition If (X t )t∈N is a stationary AR(1) time series, ρ(h) = Φh , h ∈ N.
22
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Statistical inference for AR(1) time series Consider a series of observations X1 , · · · , Xn . The likelihood is the joint distribution of the vectors X = (X1 , · · · , Xn ), which is not the product of marginal distribution, since consecutive observations are not independent (cov(Xt , Xt−h ) = φh ). Nevertheless L(φ, σ; (X0 , X )) =
n Y
πφ,σ (Xt Xt−1 )
t=1
where πφ,σ (·Xt−1 ) is a Gaussian density. Maximum likelihood estimators are bσ (φ, b) ∈ argmax log L(φ, σ; (X0 , X ))
23
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Poisson distribution  and process  for counts N as a Poisson distribution is P(N = k) = e
k −λ λ
k!
where k ∈ N.
If N ∼ P(λ), then E(N ) = λ. (Nt )t≥0 is an homogeneous Poisson process, with parameter λ ∈ R+ if • on time frame [t, t + h], (Nt+h − Nt ) ∼ P(λ · h) • on [t1 , t2 ] and [t3 , t4 ] counts are independent, if 0 ≤ t1 < t2 < t3 < t4 , (Nt2 − Nt1 ) ⊥⊥ (Nt4 − Nt3 )
24
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Poisson processes and counting models Earthquake count models are mostly based upon the Poisson process (see Utsu (1969), Gardner & Knopoff (1974), Lomnitz (1974), Kagan & Jackson (1991)), Cox process (selfexciting, cluster or branching processes, stressrelease models (see Rathbun (2004) for a review), or Hidden Markov Models (HMM) (see Zucchini & MacDonald (2009) and Orfanogiannaki et al. (2010)). See also VereJones (2010) for a summary of statistical and stochastic models in seismology. Recently, Shearer & Starkb (2012) and Beroza (2012) rejected homogeneous Poisson model,
25
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Thinning operator ◦ Steutel & van Harn (1979) defined a thinning operator as follows Definition Define operator ◦ as p ◦N = Y1 + · · · + YN if N 6= 0, and 0 otherwise, where N is a random variable with values in N, p ∈ [0, 1], and Y1 , Y2 , · · · are i.i.d. Bernoulli variables, independent of N , with P(Yi = 1) = p = 1 − P(Yi = 0). Thus p ◦ N is a compound sum of i.i.d. Bernoulli variables. Hence, given N , p ◦ N has a binomial distribution B(N, p). L
Note that p ◦ (q ◦ N ) = [pq] ◦ N for all p, q ∈ [0, 1]. Further E (p ◦ N ) = pE(N ) and var (p ◦ N ) = p2 var(N ) + p(1 − p)E(N ).
26
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
(Poisson) Integer AutoRegressive processes IN AR(1) Based on that thinning operator, AlOsh & Alzaid (1987) and McKenzie (1985) defined the integer autoregressive process of order 1 : Definition A time series (Xt )t∈N with values in R is called an INAR(1) process if Xt = p ◦ Xt−1 + εt ,
(3)
where (εt ) is a sequence of i.i.d. integer valued random variables, i.e. Xt−1
Xt =
X
Yi + εt , where Yi0 s are i.i.d. B(p).
i=1
Such a process can be related to GaltonWatson processes with immigration, or physical branching model.
27
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Xt+1 =
Xt X
Yi + εt+1 , where Yi0 s are i.i.d. B(p)
i=1
28
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
E(εt ) pE(εt ) + var(εt ) Proposition E (Xt ) = , var (Xt ) = γ(0) = and 1−p 1 − p2 γ(h) = cov(Xt , Xt−h ) = ph . It is common to assume that εt are independent variables, with a Poisson distribution P(λ), with probability function P(εt = k) = e
k −λ λ
k!
, k ∈ N.
Proposition If (εt ) are Poisson random variables, then (Nt ) will also be a sequence of Poisson random variables. Note that we assume also that εt is independent of X t−1 , i.e. past observations X0 , X1 , · · · , Xt−1 . Thus, (εt )t∈N is called the innovation process. Proposition (Xt )t∈N is a stationary INAR(1) time series if and only if p ∈ [0, 1). Proposition If (Xt )t∈N is a stationary INAR(1) time series, (Xt )t∈N is an homogeneous Markov chain. 29
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
π(xt , xt−1 )
=
P(Xt = xt Xt−1 = xt−1 ) =
xt X
P
xt−1
!
X
Yi = xt − k · P(ε = k) .  {z } {z } Poisson
i=1
k=0

Binomial
30
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Inference of Integer AutoRegressive processes IN AR(1) Consider a Poisson INAR(1) process, then the likelihood is " n # X0 Y λ λ L(p, λ; X0 , X ) = ft (Xt ) · exp − X0 X ! (1 − p) 1−p 0 t=1 where min{Xt ,Xt−1 }
ft (y) = exp(−λ)
X i=0
λy−i Yt−1 i p (1 − p)Yt−1 −y , for t = 1, · · · , n. (y − i)! i
Maximum likelihood estimators are b ∈ argmax log L(p, λ; (X0 , X )) (b p, λ)
31
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate Integer Autoregressive processes M IN AR(1) Let Nt := (N1,t , · · · , Nd,t ), denote a multivariate vector of counts. Definition Let P := [pi,j ] be a d × d matrix with entries in [0, 1]. If N = (N1 , · · · , Nd ) is a random vector with values in Nd , then P ◦ N is a ddimensional random vector, with ith component [P ◦ N ]i =
d X
pi,j ◦ Nj ,
j=1
for all i = 1, · · · , d, where all counting variates Y in pi,j ◦ Nj ’s are assumed to be independent. L
Note that P ◦ (Q ◦ N ) = [P Q] ◦ N . Further, E (P ◦ N ) = P E(N ), and E ((P ◦ N )(P ◦ N )0 ) = P E(N N 0 )P 0 + ∆, with ∆ := diag(V E(N )) where V is the d × d matrix with entries pi,j (1 − pi,j ). 32
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Definition A time series (X t ) with values in Nd is called a dvariate MINAR(1) process if X t = P ◦ X t−1 + εt
(4)
for all t, for some d × d matrix P with entries in [0, 1], and some i.i.d. random vectors εt with values in Nd . (X t ) is a Markov chain with states in Nd with transition probabilities π(xt , xt−1 ) = P(X t = xt X t−1 = xt−1 )
(5)
satisfying π(xt , xt−1 ) =
xt X
P(P ◦ xt−1 = xt − k) · P(ε = k).
k=0
33
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Parameter inference for M IN AR(1) Proposition Let (X t ) be a dvariate MINAR(1) process satisfying stationary conditions, as well as technical assumptions (called C1C6 in Franke & Subba b of θ = (P , Λ) Rao (1993)), then the conditional maximum likelihood estimate θ is asymptotically normal, √ L b − θ) → n(θ N (0, Σ−1 (θ)), as n → ∞. Further, L b 0 ) − log L(N , θN 0 )] → 2[log L(N , θN χ2 (d2 + dim(λ)), as n → ∞.
34
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality with BIN AR(1) (X1,t ) and (X2,t ) are instantaneously related if ε is a noncorrelated noise,g g g ggggggggggg
X p p1,2 X ε ε λ 1,t = 1,1 ◦ 1,t−1 + 1,t , with var 1,t = 1 X2,t p2,1 p2,2 X2,t−1 ε2,t ε2,t ϕ  {z }  {z }  {z }  {z } Xt
P
X t−1
ϕ λ2
εt
35
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality with BIN AR(1) 1. (X1 ) and (X2 ) are instantaneously related if ε is a noncorrelated noise, g g g g gggggggggg
X p p1,2 X ε ε λ 1,t = 1,1 ◦ 1,t−1 + 1,t , with var 1,t = 1 X2,t p2,1 p2,2 X2,t−1 ε2,t ε2,t ?  {z }  {z }  {z }  {z } Xt
P
X t−1
? λ2
εt
36
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality with BIN AR(1) 2. (X1 ) and (X2 ) are independent, (X1 )⊥(X2 ) if P is diagonal, i.e. p1,2 = p2,1 = 0, and ε1 and ε2 are independent,
X p 0 X ε ε λ 1,t = 1,1 ◦ 1,t−1 + 1,t , with var 1,t = 1 X2,t 0 p2,2 X2,t−1 ε2,t ε2,t 0  {z }  {z }  {z }  {z } Xt
P
X t−1
0 λ2
εt
37
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality with BIN AR(1) 3. (N1 ) causes (N2 ) but (N2 ) does not cause (X1 ), (X1 )→(X2 ), if P is a lower triangle matrix, i.e. p2,1 6= 0 while p1,2 = 0,
X p 0 X ε ε λ 1,t = 1,1 ◦ 1,t−1 + 1,t , with var 1,t = 1 X2,t ? p2,2 X2,t−1 ε2,t ε2,t ϕ  {z }  {z }  {z }  {z } Xt
P
X t−1
ϕ λ2
εt
38
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality with BIN AR(1) 4. (N2 ) causes (N1 ) but (N1,t ) does not cause (N2 ), (N1 )←(N2,t ), if P is a upper triangle matrix, i.e. p1,2 6= 0 while p2,1 = 0,
X p ? X ε ε λ 1,t = 1,1 ◦ 1,t−1 + 1,t , with var 1,t = 1 X2,t 0 p2,2 X2,t−1 ε2,t ε2,t ϕ  {z }  {z }  {z }  {z } Xt
P
X t−1
ϕ λ2
εt
39
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality with BIN AR(1) 5. (N1 ) causes (N2 ) and conversely, i.e. a feedback effect (N1 )↔(N2 ), if P is a full matrix, i.e. p1,2 , p2,1 6= 0
X p ? X ε ε λ 1,t = 1,1 ◦ 1,t−1 + 1,t , with var 1,t = 1 X2,t ? p2,2 X2,t−1 ε2,t ε2,t ϕ  {z }  {z }  {z }  {z } Xt
P
X t−1
ϕ λ2
εt
40
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Bivariate Poisson BIN AR(1) A classical distribution for εt is the bivariate Poisson distribution, with one common shock, i.e. ε = M + M 1,t 1,t 0,t ε2,t = M2,t + M0,t where M1,t , M2,t and M0,t are independent Poisson variates, with parameters λ1 − ϕ, λ2 − ϕ and ϕ, respectively. In that case, εt = (ε1,t , ε2,t ) has joint probability function 1 ,k2 } k2 min{k k1 X (λ2 − ϕ) k1 k2 ϕ −[λ1 +λ2 −ϕ] (λ1 − ϕ) i! e k1 ! k2 ! i i [λ1 − ϕ][λ2 − ϕ] i=0 with λ1 , λ2 > 0, ϕ ∈ [0, min{λ1 , λ2 }]. λ1 λ1 λ= and Λ = λ2 ϕ
ϕ λ2
41
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Bivariate Poisson BIN AR(1) and Granger causality For instantaneous causality, we test H0 : ϕ = 0 against H1 : ϕ 6= 0 b denote the conditional maximum likelihood estimate of Proposition Let λ λ = (λ1 , λ2 , ϕ) in the nonconstrained MINAR(1) model, and λ⊥ denote the conditional maximum likelihood estimate of λ⊥ = (λ1 , λ2 , 0) in the constrained model (when innovation has independent margins), then under suitable conditions, ⊥ L b b 2[log L(N , λN 0 ) − log L(N , λ N 0 )] → χ2 (1), as n → ∞, under H0 .
42
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Bivariate Poisson BIN AR(1) and Granger causality For lagged causality, we test H0 : P ∈ P against H1 : P ∈ / P, where P is a set of constrained shaped matrix, e.g. P is the set of d × d diagonal matrices for lagged independence, or a set of block triangular matrices for lagged causality. b denote the conditional maximum likelihood estimate of P in Proposition Let P c b the nonconstrained MINAR(1) model, and P denote the conditional maximum likelihood estimate of P in the constrained model, then under suitable conditions, c L b b 2[log L(N , P N 0 ) − log L(N , P N 0 )] → χ2 (d2 − dim(P)), as n → ∞, under H0 .
Example Testing (N1,t )←(N2,t ) is testing whether p1,2 = 0, or not.
43
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Autocorrelation of M IN AR(1) processes Proposition Consider a MINAR(1) process with representation X t = P ◦ X t−1 + εt , where (εt ) is the innovation process, with λ := E(εt ) and Λ := var(εt ). Let µ := E(X t ) and γ(h) := cov(X t , X t−h ). Then µ = [I − P ]−1 λ and for all h ∈ Z, γ(h) = P h γ(0) with γ(0) solution of γ(0) = P γ(0)P 0 + (∆ + Λ).
44
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Part 2
Application to earthquakes
45
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models ?
Shapefiles from http://www.colorado.edu/geography/foote/maps/assign/hotspots/hotspots.html
46
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
The dataset, and stationarity issues We work with 16 (17) tectonic plates, – Japan is at the limit of 4 tectonic plates (Pacific, Okhotsk, Philippine and Amur), – California is at the limit of the Pacific, North American and Juan de Fuca plates. Data were extracted from the Advanced National Seismic System database (ANSS) http://www.ncedc.org/cnss/catalogsearch.html – 19652011 for magnitude M > 5 earthquakes (70,000 events) ; – 19922011 for M > 6 earthquakes (3,000 events) ; – To count the number of earthquakes, used time ranges of 3, 6, 12, 24, 36 and 48 hours ; – Approximately 8,500 to 135,000 periods of observation ;
47
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models : comparing dynamics p X 1,t = 1,1 X2,t p2,1
ε ε X p1,2 λ ◦ 1,t−1 + 1,t with var 1,t = 1 ε2,t ϕ ε2,t X2,t−1 p2,2
ϕ λ2
Complete model, with full dependence
48
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models : comparing dynamics p X 1,t = 1,1 X2,t 0
λ ε ε X 0 ◦ 1,t−1 + 1,t with var 1,t = 1 ϕ ε2,t ε2,t X2,t−1 p2,2
ϕ λ2
Partial model, with diagonal thinning matrix, nocrossed lag correlation
49
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models : comparing dynamics p X 1,t = 1,1 X2,t 0
λ ε ε X 0 ◦ 1,t−1 + 1,t with var 1,t = 1 0 ε2,t ε2,t X2,t−1 p2,2
0 λ2
Two independent INAR processes
50
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models : comparing dynamics p X 1,t = 1,1 X2,t 0
λ ε ε X 0 ◦ 1,t−1 + 1,t with var 1,t = 1 0 ε2,t ε2,t X2,t−1 p2,2
0 λ2
Two independent INAR processes
51
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models : comparing dynamics X 0 1,t = 0 X2,t
X ε ε λ 0 ◦ 1,t−1 + 1,t with var 1,t = 1 0 X2,t−1 ε2,t ε2,t ϕ
ϕ λ2
Two (possibly dependent) Poisson processes
52
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models : tectonic plates interactions – For all pairs of tectonic plates, at all frequencies, autoregression in time is important (very high statistical significance) ; – Long sequence of zeros, then mainshocks and aftershocks ; – Rate of aftershocks decreases exponentially over time (Omori’s law) ; – For 713% of pairs of tectonic plates, diagonal BINAR has significant better fit than independent INARs ; – Contribution of dependence in noise ; – Spatial contagion of order 0 (within h hours) ; – Contiguous tectonic plates ; – For 79% of pairs of tectonic plates, proposed BINAR has significant better fit than diagonal BINAR ; – Contribution of spatial contagion of order 1 (in time interval [h, 2h]) ; – Contiguous tectonic plates ; – for approximately 90%, there is no significant spatial contagion for M > 5 earthquakes 53
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality N1 → N2 or N1 ← N2 1. North American Plate, 2. Eurasian Plate,3. Okhotsk Plate, 4. Pacific Plate (East), 5. Pacific Plate (West), 6. Amur Plate, 7. IndoAustralian Plate, 8. African Plate, 9. IndoChinese Plate, 10. Arabian Plate, 11. Philippine Plate, 12. Coca Plate, 13. Caribbean Plate, 14. Somali Plate, 15. South American Plate, 16. Nasca Plate, 17. Antarctic Plate
Granger Causality test, 3 hours
1
3
2
5
4
6
8
7
Granger Causality test, 6 hours
9 10 11 12 13 14 15 16 17
1
●
2
●
3
●
4
●
5
●
6
●
7
●
8
●
9
●
10
●
11
●
12
●
13
●
14
●
15
●
16 17
● ●
1
●
3
2
5
4
6
8
7
9 10 11 12 13 14 15 16 17
1 2
●
3
●
4
●
5
●
6
●
7
●
8
●
9
●
10
●
11
●
12
●
13
●
14
●
15
●
16 17
●
●
● ●
54
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality N1 → N2 or N1 ← N2 1. North American Plate, 2. Eurasian Plate,3. Okhotsk Plate, 4. Pacific Plate (East), 5. Pacific Plate (West), 6. Amur Plate, 7. IndoAustralian Plate, 8. African Plate, 9. IndoChinese Plate, 10. Arabian Plate, 11. Philippine Plate, 12. Coca Plate, 13. Caribbean Plate, 14. Somali Plate, 15. South American Plate, 16. Nasca Plate, 17. Antarctic Plate
Granger Causality test, 12 hours
1
3
2
5
4
6
8
7
Granger Causality test, 24 hours
9 10 11 12 13 14 15 16 17
1
●
2
●
3
●
4
●
5
●
6
●
7
●
8
●
9
●
10
●
11
●
12
●
13
●
14
●
15
●
16 17
● ●
1
●
3
2
5
4
6
8
7
9 10 11 12 13 14 15 16 17
1 2
●
3
●
4
●
5
●
6
●
7
●
8
●
9
●
10
●
11
●
12
●
13
●
14
●
15
●
16 17
●
●
● ●
55
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Granger causality N1 → N2 or N1 ← N2 1. North American Plate, 2. Eurasian Plate,3. Okhotsk Plate, 4. Pacific Plate (East), 5. Pacific Plate (West), 6. Amur Plate, 7. IndoAustralian Plate, 8. African Plate, 9. IndoChinese Plate, 10. Arabian Plate, 11. Philippine Plate, 12. Coca Plate, 13. Caribbean Plate, 14. Somali Plate, 15. South American Plate, 16. Nasca Plate, 17. Antarctic Plate
Granger Causality test, 36 hours
1
3
2
5
4
6
8
7
Granger Causality test, 48 hours
9 10 11 12 13 14 15 16 17
1
●
2
●
3
●
4
●
5
●
6
●
7
●
8
●
9
●
10
●
11
●
12
●
13
●
14
●
15
●
16 17
● ●
1
●
3
2
5
4
6
8
7
9 10 11 12 13 14 15 16 17
1 2
●
3
●
4
●
5
●
6
●
7
●
8
●
9
●
10
●
11
●
12
●
13
●
14
●
15
●
16 17
●
●
● ●
56
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Multivariate models : frequency versus magnitude X1,t =
X i=1
1(Ti ∈ [t, t + 1), Mi ≤ s) and X2,t =
X
1(Ti ∈ [t, t + 1), Mi > s)
i=1
Here we work on two sets of data : mediumsize earthquakes (M ∈ (5, 6)) and largesize earthquakes (M > 6). – Investigate direction of relationship (which one causes the other, or both) ; – Pairs of tectonic plates : – Unidirectional causality : most common for contiguous plates (North American causes West Pacific, Okhotsk causes Amur) ; – Bidirectional causality : Okhotsk and West Pacific, South American and Nasca for example ; – Foreshocks and aftershocks : – Aftershocks much more significant than foreshocks (as expected) ; – Foreshocks announce arrival of largersize earthquakes ; – Foreshocks significant for Okhotsk, West Pacific, IndoAustralian, IndoChinese, Philippine, South American ; 57
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Risk management issues – Interested in computing P
T (N + N ) ≥ n F0 for various values of T 1,t 2,t t=1
P
(time horizons) and n (tail risk measure) ; – Total number of earthquakes on a set of two tectonic plates ; – 100 000 simulated paths of diagonal and proposed BINAR models ; – Use estimated parameters of both models ; – Pair : Okhotsk and West Pacific ; – Scenario : on a 12hour period, 23 earthquakes on Okhotsk and 46 earthquakes on West Pacific (second half of March 10th, 2011) ;
58
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Diagonal model n / days
1 day
3 days
7 days
14 days
5
0.9680
0.9869
0.9978
0.9999
10
0.5650
0.7207
0.8972
0.9884
15
0.1027
0.2270
0.4978
0.8548
20
0.0067
0.0277
0.1308
0.4997
Proposed model n / days
1 day
3 days
7 days
14 days
5
0.9946
0.9977
0.9997
1.0000
10
0.8344
0.9064
0.9712
0.9970
15
0.3638
0.5288
0.7548
0.9479
20
0.0671
0.1573
0.3616
0.7256
59
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Some references AlOsh, M.A. & A.A. Alzaid (1987) "Firstorder integervalued autoregressive process", Journal of Time Series Analysis 8, 261275. Beroza, G.C. (2012) How many great earthquakes should we expect ? PNAS, 109, 651652. Dion, J.P., G. Gauthier & A. Latour (1995), "Branching processes with immigration and integervalued time series", Serdica Mathematical Journal 21, 123136. Du, J.G. & Y. Li (1991), "The integervalued autoregressive (INAR(p)) model", Journal of Time Series Analysis 12, 129142. Ferland, R.A., Latour, A. & Oraichi, D. (2006), Integervalued GARCH process. Journal of Time Series Analysis 27, 923942. Fokianos, K. (2011). Count time series models. to appear in Handbook of Time Series Analysis. 60
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Franke, J. & Subba Rao, T. (1993). Multivariae firstorder integervalued autoregressions. Forschung Universität Kaiserslautern, 95. Gourieroux, C. & J. Jasiak (2004) "Heterogeneous INAR(1) model with application to car insurance", Insurance : Mathematics & Economics 34, 177192. Gardner, J.K. & I. Knopoff (1974) "Is the sequence of earthquakes in Southern California, with aftershocks removed, Poissonean ?" Bulletin of the Seismological Society of America 64, 13631367. Joe, H. (1996) Time series models with univariate margins in the convolutionclosed infinitely divisible class. Journal of Time Series Analysis 104, 117133. Johnson, N.L, Kotz, S. & Balakrishnan, N. (1997). Discrete Multivariate Distributions. Wiley Interscience. Kagan, Y.Y. & D.D. Jackson (1991) Longterm earthquake clustering, Geophysical Journal International 104, 117133. 61
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Kocherlakota, S. & Kocherlakota K. (1992). Bivariate Discrete Distributions. CRC Press. Latour, A. (1997) "The multivariate GINAR(p) process", Advances in Applied Probability 29, 228248. Latour, A. (1998) Existence and stochastic structure of a nonnegative integervalued autoregressive process, Journal of Time Series Analysis 19, 439455. Lomnitz, C. (1974) "Global Tectonic and Earthquake Risk", Elsevier. Lütkepohl, H. (2005). New Introduction to Multiple Time Series Analysis. Springer Verlag. Mahamunulu, D. M. (1967). A note on regression in the multivariate Poisson distribution. Journal of the American Statistical Association, 62, 25 1258. McKenzie, E. (1985) Some simple models for discrete variate time series, Water Resources Bulletin 21, 645650. 62
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Orfanogiannaki, K., D. Karlis & G.A. Papadopoulos (2010) "Identifying Seismicity Levels via Poisson Hidden Markov Models", Pure and Applied Geophysics 167, 919931. Parsons, T., & A.A. Velasco (2011) "Absence of remotely triggered large earthquakes beyond the mainshock region", Nature Geoscience, March 27th, 2011. Pedeli, X. & Karlis, D. (2011) "A bivariate Poisson INAR(1) model with application", Statistical Modelling 11, 325349. Pedeli, X. & Karlis, D. (2011). "On estimation for the bivariate Poisson INAR process", To appear in Communications in Statistics, Simulation and Computation. Rathbun, S.L. (2004), “Seismological modeling”, Encyclopedia of Environmetrics. Rosenblatt, M. (1971). Markov processes, Structure and Asymptotic Behavior. SpringerVerlag. 63
Arthur CHARPENTIER & Mathieu BOUDREAULT, Bivariate count processes for earthquake frequency
Shearer, P.M. & Stark, P.B (2012). Global risk of big earthquakes has not recently increased. PNAS, 109, 717721. Steutel, F. & K. van Harn (1979) "Discrete analogues of selfdecomposability and stability", Annals of Probability 7, 893899. Utsu, T. (1969) "Aftershocks and earthquake statistics (I)  Some parameters which characterize an aftershock sequence and their interrelations", Journal of the Faculty of Science of Hokkaido University, 121195. VereJones, D. (2010), "Foundations of Statistical Seismology", Pure and Applied Geophysics 167, 645653. Weiβ, C.H. (2008) "Thinning operations for modeling time series of counts : a survey", Advances in Statistical Analysis 92, 319341. Zucchini, W. & I.L. MacDonald (2009) "Hidden Markov Models for Time Series : An Introduction Using R", 2nd edition, Chapman & Hall
The article can be downloaded from http ://arxiv.org/abs/1112.0929 64