Professional Documents
Culture Documents
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Introduction
Basic problem.
Applications . .
Examples . . . .
Brief history . .
Resources. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Basic Notions
Principles of Stability. . . . . . . . . . . . . . . . . . . . . .
Probability framework for block maxima distribution.
Classical limit laws . . . . . . . . . . . . . . . . . . . . . . .
Extremal types theorem . . . . . . . . . . . . . . . . . . . .
Using the limit law . . . . . . . . . . . . . . . . . . . . . . .
Generalized extreme value distribution . . . . . . . . . .
Quantiles and return levels . . . . . . . . . . . . . . . . . .
Max-stability . . . . . . . . . . . . . . . . . . . . . . . . . . .
Outline proof of GEV as limit law . . . . . . . . . . . . .
Domains of attraction . . . . . . . . . . . . . . . . . . . . .
Convergence and approximation . . . . . . . . . . . . . .
Minima . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Interest and nuisance parameters . . . . . . . . . . . . . .
Modelling with the GEV. . . . . . . . . . . . . . . . . . . .
Port Pirie Annual Maxima data . . . . . . . . . . . . . . .
Bayesian Extremes. . . . . . . . . . . . . . . . . . . . . . . .
MCMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MetropolisHastings . . . . . . . . . . . . . . . . . . . . . .
Port Pirie: MCMC analysis . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
. 4
. 5
. 6
15
16
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
17
18
19
20
21
25
26
27
28
29
30
31
32
33
34
36
38
39
42
44
45
47
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Modelling Issues
D(un ) condition . . . . . . . . . . . . . .
Short-term dependence: Example . .
Moving maxima . . . . . . . . . . . . . .
Extremal index . . . . . . . . . . . . . . .
D. . . . . . . . . . . . . . . . . . . . . . . .
Modelling. . . . . . . . . . . . . . . . . . .
Wooster example . . . . . . . . . . . . .
Return levels . . . . . . . . . . . . . . . .
Rain example . . . . . . . . . . . . . . . .
Non-stationarity . . . . . . . . . . . . . .
Estimation . . . . . . . . . . . . . . . . . .
Model reduction . . . . . . . . . . . . . .
Other extreme value models . . . . . .
Examples . . . . . . . . . . . . . . . . . . .
Semiparametric . . . . . . . . . . . . . . .
Example: Swiss winter temperatures
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Multivariate Extremes
Multivariate extremes . . . .
Componentwise maxima . . .
Special cases . . . . . . . . . .
Parametric models. . . . . . .
Alternatives . . . . . . . . . . .
Inference . . . . . . . . . . . . .
Structure variables. . . . . . .
Sea levels . . . . . . . . . . . . .
Point process . . . . . . . . . .
Comments . . . . . . . . . . . .
Consistency of results . . . .
Poisson likelihood . . . . . . .
Exchange rate data . . . . . .
Oceanographic example . . .
Higher-dimensional models .
Asymptotic dependence . . .
Bivariate normal variables . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
51
52
53
57
61
62
63
66
67
69
70
71
72
73
75
76
78
80
83
84
85
86
88
91
92
93
95
96
105
107
111
112
113
116
117
118
119
120
122
126
129
130
131
133
137
140
141
142
Spatial Extremes
143
Spatiotemporal extremes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Geostatistics
Geostatistics . . . . . . . . . . . . . .
Temperature data . . . . . . . . . .
Approaches . . . . . . . . . . . . . .
Latent variable approach . . . . .
HeffernanTawn (2004) . . . . . .
Max-stable processes . . . . . . . .
Extremal coefficient . . . . . . . . .
Madogram . . . . . . . . . . . . . . .
Spectral representations I . . . . .
Spectral representations II . . . .
Pairwise likelihood. . . . . . . . . .
Example: Temperature data . . .
Swiss summer temperature data
Fit of marginal model . . . . . . .
Estimated correlations . . . . . . .
Fit of correlation curve. . . . . . .
Pairwise fits . . . . . . . . . . . . . .
Simulated fields . . . . . . . . . . .
......................
......................
......................
Discussion . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
145
145
146
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
Overview
1. Basic notions: maximum analysis, and threshold models
2. Alternative charactizationsimproved inferences
3. Modelling issues
4. Multivariate extremes
5. Spatial models for extremes
Statistics of Extremes
Introduction
slide 3
Basic problem
iid
estimates are often required beyond the largest observed data value;
standard density estimation techniques fit well where the data have greatest density, but can be
severely biased in estimating tail probabilities.
Usual lack of physical or empirical basis for extrapolation leads to the extreme value paradigm:
Base tail models on asymptotically-motivated distributions.
Statistics of Extremes
Applications
Historically there have been two main application areas of extreme value theory:
Environmental:
sea levels
wind speeds
pollution concentrations
river flow
...
Reliability Modelling:
4.2
4.0
3.6
3.8
Sea-Level (metres)
4.4
4.6
1930
1940
1950
1960
1970
1980
Year
15
10
0
Frequency
20
25
0.0
0.5
1.0
1.5
2.0
2.5
Breaking Strength
1.6
1.2
1.4
Sea-level (meters)
1.8
1900
1920
1940
1960
1980
Year
1.6
1.2
1.4
Sea-level (meters)
1.8
-1
SOI
240
238
232
234
236
242
244
246
1975
1980
1985
1990
Year
140
120
60
80
100
Sea-level (cm)
160
180
200
1930
1940
1950
1960
1970
1980
Year
40
0
20
60
80
1920
1930
1940
1950
1960
Year
0
-20
-40
-60
20
Examples: Nonstationarity
1983
1984
1985
1986
1987
1988
Year
Index
-0.02
-0.06
6000
-0.04
8000
Index
0.0
10000
0.02
0.04
12000
1996
1997
1998
1999
2000
1996
Year
1997
1998
1999
2000
Year
Dow Jones Index. (Left Panel: Raw. Right Panel: After transformation to induce approximate
stationarity)
Statistics of Extremes
65
60
55
50
45
40
50
60
70
80
Brief history
1940s: Asymptotic theory unified and extended by Gnedenko and von Mises
1950s: Use of asymptotic distributions for statistical modelling by Gumbel and Jenkinson
Statistics of Extremes
Resources
Books
Galambos (1987) The Asymptotic Theory of Extreme Order Statistics, Krieger
Resnick (1987) Extreme Values, Regular Variation, and Point Processes, SV
Leadbetter, Lindgren and Rootzen (1983) Extremes and Related Properties of Random Sequences and Processes,
SV
Statistics of Extremes
Basic Notions
slide 17
Principles of Stability
Fundamental to all characterizations of extreme value processes is the concept of stability.
For example, we might propose one model for the annual maximum of a process, and another for
the 5-year maximum. Since the 5-year maximum will be the maximum of 5 annual maxima, the
models should be mutually consistent.
Similarly, a model for exceedances over a high threshold should remain valid (in a precise sense)
for exceedances of higher threshold.
Statistics of Extremes
iid
= Pr{X1 x} Pr{Xn x}
= F (x)n .
Statistics of Extremes
Statistics of Extremes
10
II : G(x)
III : G(x)
0,
x 0,
exp(x ),
x > 0, > 0;
Conversely, each of these Gs may appear as a limit for the distribution of (Mn bn )/an , and does so when G itself is
the distribution of X.
Statistics of Extremes
10
0
x
10
0.3
PDF
0.2
0.1
0.0
0.0
0.0
0.1
0.1
PDF
0.2
PDF
0.2
0.3
0.3
0.4
0.4
Gumbel
10
0
x
10
10
0
x
10
The Frechet (Type II) is bounded below, and the negative Weibull (Type III) is bounded above.
Statistics of Extremes
11
1 x
e
n
exp{ exp(x)}.
CDF
0.0
0.0
0.2
0.2
0.4
0.4
CDF
0.6
0.6
0.8
0.8
1.0
1.0
6
x
10
12
2
x
Distributions of maxima and renormalised maxima of n = 1, 7, 30, 365, 3650 standard exponential variables (from left to right), with Gumbel distribution (heavy).
Statistics of Extremes
an = (2 log n)0.5 ,
1.0
0.8
0.6
0.0
0.2
0.4
CDF
0.6
0.4
0.0
0.2
CDF
0.8
1.0
-4
-2
-4
-2
Distributions of maxima and renormalised maxima of n = 1, 7, 30, 365, 3650 standard normal variables (from left to right), with Gumbel distribution (heavy).
Statistics of Extremes
12
That is, the family of extreme value distributions may be fitted directly to a series of observations
of Mn .
Statistics of Extremes
This family encompasses all three of the previous extreme value limit families:
(
)
x 1/
G(x) = exp 1 +
,
+
defined on {x : 1 + (x )/ > 0}.
Statistics of Extremes
13
where G(xp ) = 1 p.
In extreme value terminology, xp is the return level associated with the return period 1/p.
15
Quantile
10
Shape=0.2
Shape=0
Shape=-0.2
-2
Log y
Statistics of Extremes
Max-stability
Some insight into these results is obtained using the concept of max-stability.
Definition 3 A distribution G is said to be max-stable if
Gk (ak x + bk ) = G(x),
k = 1, 2, . . . ,
In other words, taking powers of G results only in a change of location and scale.
The connection with extremes is that a distribution is max-stable if and only if it is a GEV
distribution.
Statistics of Extremes
14
so that G and G1/k are identical apart from scaling coefficients. Thus, G is max-stable and therefore GEV.
Statistics of Extremes
Domains of attraction
For a given distribution function F , is it easy to determine suitable sequences an and bn and to know what limit
G will occur?
For sufficiently smooth distributions, defining the reciprocal hazard function
r(x) =
1 F (x)
.
f (x)
and letting
bn = F 1 (1 n1 ), an = r(bn ), = lim r (x)
x
the limit distribution of (Mn bn )/an is
1/
exp{(1 + x)+
} if 6= 0
and
exp(ex )
if = 0.
Statistics of Extremes
15
There has been a good deal of work on the speed of convergence of Mn to the limiting regime, which
depends on the underlying distribution F for example, convergence is slow for maxima of n Gaussian
variables.
From a statistical viewpoint, this is not so useful: we use the GEV as an approximate distribution for
sample maxima for finite (small?) n, so the key question is whether the GEV fits the available dataassess
this empirically.
Direct use of the GEV rather than the three types separately allows for flexible modelling, and ducks the
question of which type is most appropriatethe data decide.
Testing for fit of one type or another is usually unhelpful, because setting (say) = 0 can give
unrealistically precise inferences. Often uncertainty in extremes is (appropriately) large, and it can be
misleading to constrain inferences artificially.
Statistics of Extremes
Minima
All the ideas apply equally to minima, because
min(X1 , . . . , Xn ) = max(X1 , . . . , Xn ).
Our general discussion is for maxima, and we make this transformation without comment when we
model minima.
Statistics of Extremes
Inference
Given observed annual maxima X1 , . . . , Xk , aim now is to make inferences on the GEV parameters
(, , ) . Possibilities include:
incorporation of prior information via Bayes theorem also uses likelihood as key ingredient.
Statistics of Extremes
16
Likelihood
Definition 4 Let y be a data set, assumed to be the realisation of a random variable Y f (y; ), where is
unknown. Then the likelihood and log likelihood are
L() = L(; y) = fY (y; ),
() = log L(),
Statistics of Extremes
Likelihood approximations
For large n, and if the data were generated from f (y; 0 , then
b 1 ,
b Np 0 , J()
b 1 .
so a standard error for br is the square root of the rth diagonal element of J()
b we just have to maximise () and to obtain the Hessian matrix at the
To obtain b and J(),
maximumpurely numerically, if a routine to compute () is available;
W (0 ) 2p .
We base tests and/or confidence intervals on these results, even when the sample is small (e.g. n = 20 or
less). Often the approximations are adequate for practical work.
For details, see for example Chapter 4 of Davison (2003) Statistical Models, CUP.
Statistics of Extremes
17
b ) (, ),
(,
Wp () 2q ,
for large n.
We often base confidence intervals and tests for 0 on this, particularly if the profile log likelihood for ,
b ) is asymmetricgenerally the case for quantiles, Value-At-Risk, and similar quantities.
p () = (b ) = (,
Statistics of Extremes
Non-regular models
With extremal models there is a potential difficulty the endpoint of the distribution (if finite) is a function
of the parameters, so usual asymptotic results may not hold.
Smith (1985, Biometrika) established that the limiting behaviour of the MLE depends on the value of shape
parameter :
when > 0.5, maximum likelihood estimators obey standard theory above;
when 1 < < 1/2, the MLE satisfies the score equation;
when < 1, the MLE does not satisfy the score equation.
When < 1/2, the MLE for the endpoint converges to the endpoint faster than when > 1/2 (good)
at a convergence rate that depends on (bad), and other parameters converge at the usual rate (OK), so
there is no simple theory.
In most environmental problems 1 < < 1 (often 0) so maximum likelihood works fine.
If not, Bayesian methods (which do not depend on these regularity conditions) may be preferable.
Statistics of Extremes
18
1+
o
xi 1/
xi
Calculation of standard errors from inverse of observed information matrix (also obtained numerically).
Diagnostic checks: probability plots, quantile plots, return level plots.
Comparison of competing nested models through deviance (likelihood ratio statistic), and of non-nested
b or other similar criteria.
models by minimising the Akaike information criterion AIC = 2{dim() },
Calculation of confidence intervals for return levels.
Statistics of Extremes
4.2
4.0
3.6
3.8
Sea-Level (metres)
4.4
4.6
1930
1940
1950
1960
1970
1980
Year
portpirie.fit<- gev.fit(portpirie$SeaLevel)
$conv
[1] 0
$nllh
[1] -4.339058
$mle
[1] 3.87474692 0.19804120 -0.05008773
$se
[1] 0.02793211 0.02024610 0.09825633
Statistics of Extremes
19
4.4
4.2
Empirical
3.8
4.0
0.6
0.4
0.0
3.6
0.2
Model
0.8
4.6
1.0
Probability Plot
0.2
0.4
0.6
0.8
1.0
3.6
3.8
4.0
4.2
Empirical
Model
Density Plot
4.4
4.6
0.0
0.5
1.0
f(z)
4.5
4.0
Return Level
1.5
5.0
0.0
0.1
1.0
10.0
100.0
1000.0
3.6
3.8
4.0
Return Period
4.2
4.4
4.6
Statistics of Extremes
2
1
0
Profile Log-likelihood
-0.3
-0.2
-0.1
0.0
0.1
0.2
0.3
Shape Parameter
2
0
-2
Profile Log-likelihood
4.5
5.0
5.5
6.0
Return Level
Statistics of Extremes
20
Bayesian Extremes
Markov chain Monte Carlo (MCMC), and other stochastic computation algorithms, have enabled
Bayesian techniques to be applied to extreme value problems.
Advantages include
Statistics of Extremes
Bayesian inference
Key idea: everything is treated as random variable, so probability calculus is applied to unknown parameters
and potential data
Use Bayes theorem to compute
Pr( unknowns | data ) =
Statistics of Extremes
MCMC
Idea is that when direct simulation from a density f is difficult, the problem can be tackled by
setting up a Markov chain whose limit density is f .
Consequently, simulation of x1 , x2 , . . . from the chain yields a series with the property that the
marginal density of xj for large enough j is approximately f .
In other words, for large enough N , xN , xN +1 , . . . can be regarded as a dependent series with
marginal density f . Hence, for example, empirical moments of this series yield approximations of
the moments of f .
Statistics of Extremes
21
MetropolisHastings
To generate a Markov chain {xt } with limit density f :
1. Set t = 0 and choose x(0) .
3. Calculate
f (x )q(x(t) | x )
= (x(t) , x ) = min 1,
f (x(t) )q(x | x(t) )
4. Define
(t+1)
x
x(t)
with probability
with probability 1 .
Comments on MetropolisHastings
The algorithm remains valid when f is only proportional to a target density function. This is why
the algorithm is so important in Bayesian applications.
These conditions facilitate the use of fairly short series of simulated data for inference, but use of
convergence diagnostics is essential, as there is usually no theoretical guarantee of convergence
Statistics of Extremes
22
0.6
scale
0.2
0.4
4.6
4.2
3.8
location
5.0
0.8
MCMC chains for GEV parameters and (by transformation) 100-year return level in analysis of Port Pirie annual
maxima. Analysis based on vague, though proper, priors.
1000
3000
5000
1000
1000
3000
5000
6.5
6.0
5.5
5.0
4.5
0.2
shape
0.2
0.6
3000
Index
Index
5000
Index
1000
3000
5000
Index
Statistics of Extremes
b = 3.87,
b = 0.208, b = 0.045,
23
3.85
3.95
20
5 10
4.05
0.15
0.20
0.25
0.30
Shape
0.2
0.4
1.0
0.0
6
5
4
3
2
1
0.0
2.0
scale
location
0.4
20
15
10
5
3.75
Scale
Location
shape
return level
It states that when (normalized) maxima have a limit, that limit will be a member of the GEV
family. It does not (by itself) guarantee the existence of a limit, and there exist wide classes of
models for which no limit exists.
One interesting case is the Poisson distribution. If {Xi } is a sequence of Poisson variables with
mean , a sequence of constants {In } can be found such that
lim Pr max = In or In + 1 = 1
n
1in
and this degeneracy implies that no limit distribution for normalized maxima exists.
The discreteness in the Poisson distribution causes the oscillation between In and In + 1. It is not
the cause of the limit degeneracy, which is actually a consequence of the Poisson tail decay.
Statistics of Extremes
24
slide 51
Improved Inferences
The annual maxima method can be inefficient if other data are available. Alternative methods include:
15
1970
1975
1980
1985
Time
Statistics of Extremes
15
1970
1975
1980
1985
Time
Statistics of Extremes
25
15
1970
1975
1980
1985
Time
Statistics of Extremes
Point process
For any suitable set A, we simply count the number of points in A, and for some (random) n we
can write the process as
n
X
Xj
P=
j=1
where the Xj are the positions of the points, and x puts unit mass at x.
Statistics of Extremes
Poisson process
Definition 5 Let X Rd , and let a function (A) 0 be defined for any measurable A X . A
Poisson process P on X with intensity measure satisfies
the number of points of P in A, denoted N (A), has the Poisson distribution with mean (A);
(x1 , . . . , xd ) =
d (A)
x1 xd
26
iid
Form a 2-dimensional point process {(i, Xi ); i = 1, . . . , n} and characterize the behaviour of this
process in regions of the form A = [t1 , t2 ] [u, ).
More formally, suppose (Mn bn )/an converges to the distribution
o
n
1/
.
G(x) = exp (1 + x)+
Statistics of Extremes
Poisson limit
Poisson limit when rescaling samples of sizes 10, 100, 1000, 10,000.
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
t
Statistics of Extremes
0.8
1.0
-2
(X-b)/a
-4
-6
-8
(X-b)/a
-8
-6
-4
-8
-8
-6
-6
(X-b)/a
-2
-4
-4
(X-b)/a
-2
-2
= Pr{Mn x}
o
n
1/
,
exp (1 + x)+
1/
.
January 2008 slide 60
27
Statistical application
For statistical application, it is convenient to:
Assume the limiting process is a reasonable approximation for finite n above a high threshold, u.
To absorb the unknown scaling coefficients into the intensity function and so work directly with
the original series.
To rescale the intensity so that the annual maximum has the GEV distribution with parameters
(, , ) .
This leads to modelling the series {(i, Xi ); i = 1, . . . , n} as a Poisson process above u with intensity
function
1/
where ny is the number of years of observation, and the time limits are rescaled to 0 t1 < t2 1.
Statistics of Extremes
Likelihood
L(Av ; , , )
= exp{(Av )}
d(ti , xi )
i=1
NAv
n
1/ o Y
1/1
1
= exp ny 1 + v
,
1 + xi +
i=1
where x1 , . . . , xNAv is an enumeration of the NAv points that exceed the threshold v.
The parameters are the same as for the maximum model.
Statistics of Extremes
15
1970
1975
1980
1985
Time
Model is now applied to a time series of hourly rainfall aggregates measured in mm with a threshold
of u = 5mm.
Statistics of Extremes
28
scale
1.86637
Standard Errors
loc
scale
0.35380 0.23673
shape
0.06696
shape
0.05379
Statistics of Extremes
0.0
0.2
10
Empirical
15
Model
0.4
0.6
0.8
20
1.0
Probability Plot
0.0
0.2
0.4
0.6
Empirical
0.8
1.0
10 12
Model
14
16
18
0.0
0.1
10
0.2
Density
0.3 0.4
Return Level
15
20
0.5
25
0.6
Density Plot
10 12
Quantile
14
16
0.05
Statistics of Extremes
0.20
1.00
5.00
Return Period
20.00
29
Threshold methods
Let Xn,i
= (Xi bn )/an , for i = 1, . . . , n. Then, by the Poisson limit,
{(0, 1) (u + x, )}
{(0, 1) (u, )}
1/
x
.
=
1+
+ (u ) +
Absorbing the unknown scaling coefficients leads to the survivor function of the Generalized Pareto
Distribution (GPD):
Pr{Xn,i
> u + x | Xn,i
> u}
x 1/
Pr{Xn,i > u + x | Xn,i > u} = 1 +
,
+
x > 0,
where = + (u ).
Taking the limit 0 gives the exponential distribution as a special case.
Statistics of Extremes
Alternative inference
An equivalent inference to the point process method is to apply the GPD model to the exceedances of
a threshold u. For the rainfall data we obtain:
(rain.gpdfit <- fpot(esk.rain,threshold=5,npp=365*24))
Deviance: 1058.954
Threshold: 5
Number Above: 356
Proportion Above: 0.0024
Estimates
scale
shape
1.52239 0.06702
Standard Errors
scale
shape
0.11488 0.05383
Statistics of Extremes
30
Interpretation
There are now just two parameters, since the model conditions on the exceedance (one parameter
is lost corresponding to the crossing rate of u).
The estimate of is (almost!) identical to the point process analysis, while the change of
parameterization leads to a different value of .
+ u
,
1
we obtain a simple diagnostic for threshold selection. The mean exceedance above u should be
linear in u at levels for which the model is valid.
Suggests looking for linearity in a plot of the empirical mean residual life.
Statistics of Extremes
0.0
0.5
1.0
Mean Excess
1.5
2.0
2.5
3.0
3.5
6
8
Threshold
10
12
Threshold selection
0.2
14
Scale
3
4
Location
16 18 20
Shape
0.0 0.1 0.2 0.3
22
Bias-variance trade-off: threshold too lowbias because of the model asymptotics being invalid; threshold too
highvariance is large due to few data points.
An alternative approach is to fit the Poisson process model at many thresholds and look for parameter stability:
3 4 5
Threshold
3 4 5
Threshold
3 4 5
Threshold
Statistics of Extremes
31
!1/ r
(r)
Y 1
Mn
L = exp 1 +
i=1
(i)
Mn
1+
!1/1
This is the likelihood contribution for a single block (year). The full likelihood is obtained by
multiplying such terms across the blocks.
Statistics of Extremes
12.6929941
-0.1148273
se:
[1] 1.34704941 0.52516203 0.01901015
Statistics of Extremes
32
Modelling Issues
slide 73
Modelling issues
Extreme value data usually show:
other forms of nonstationarity (for example, the deterministic effect of tides on sealevels).
For temporal dependence there is a sufficiently wide-ranging theory which can be invoked. Other
aspects have to be handled at the modelling stage.
We now discuss how to deal with
short-term dependence;
Statistics of Extremes
D(un ) condition
The usual (weak) condition which is adopted to eliminate the effect of long-range dependence is
Leadbetters D(un ) condition:
For all i1 < . . . < ip < j1 < . . . < jq with j1 ip > l
|Pr{Xi1 un , . . . , Xip un , Xj1 un , . . . , Xjq un }
Pr{Xi1 un , . . . , Xip un }Pr{Xj1 un , . . . , Xjq un }| (n, l),
where (n, ln )0 for some sequence ln = o(n). This implies that rare events that are sufficiently
separated are independent.
Theorem 6 If D(un ) is satisfied with un = an x + bn , and if
Pr{Mn an x + bn }G(x),
then G is a GEV distribution.
Conclusion: Same models apply for processes with (restricted) long-range dependence.
Statistics of Extremes
33
x
3
20
40
60
80
100
Exponential example
Marginal distribution:
Pr{Xi < x}
Let X1 , X2 , . . . be an independent series with the same marginal distribution, and Mn = max{X1 , . . . , Xn }.
Then
1 x 2n
e
Pr{Mn log(2n) < x} = [1 exp{x log(2n)}]2n = 1 2n
exp ex = G2 (x),
while, for Mn = max{X1 , . . . , Xn },
Pr{Mn log(2n) < x}
Statistics of Extremes
34
Let Y0 , Y1 , Y2 , . . . F , with
n
o
1
F (y) = exp (a+1)y
,
y > 0,
Xi = max{aYi1 , Yi },
5.0
50.0
i = 1, . . . , n.
0.1
0.5
x(i)
5.0
0.1
0.5
x(i)
50.0
X0 = Y0 ,
10
20
30
40
50
10
20
40
50
30
40
50
50.0
5.0
0.1
0.5
x(i)
5.0
0.1
0.5
x(i)
30
i
50.0
10
20
30
40
50
10
20
Statistics of Extremes
Now, let X1 , X2 , . . . be a series of independent variables having a marginal standard Frechet distribution, and
define Mn = max{X1 , . . . , Xn }. Then,
n
It follows that
Statistics of Extremes
35
Extremal index
The previous examples illustrate the following general result.
Theorem 7 Let {Xi } be a stationary process and {Xi } be independent variables with the same marginal
distribution. Set Mn = max{X1 , . . . , Xn } and Mn = max{X1 , . . . , Xn }. Under suitable regularity conditions,
Pr {(Mn bn )/an z} G1 (z)
as n for normalizing sequences {an > 0} and {bn }, where G1 is a non-degenerate distribution function, if
and only if
Pr {(Mn bn )/an z} G2 (z),
where
G2 (z) = G1 (z)
for a constant called the extremal index that satisfies 0 < 1.
Thus if G1 is GEV, then so is G2 , with the same a strong robustness result.
Statistics of Extremes
Extremal index
Loosely, is the probability that a high threshold exceedance is the final element in a cluster of
exceedances.
Thus extremes occur in clusters whose (limiting) mean cluster size is 1/.
In fact the distribution of a cluster maximum is the same as the marginal distribution of an
exceedance, so there is no bias in considering only cluster maxima, if we can identify clusters . . .
Statistics of Extremes
36
Consequences
If we can estimate the tail of marginal distribution F (e.g. by fitting to block maxima), then
Pr(Mn x) F (x)n G(x),
where G is GEV with parameters , , . The marginal quantiles are approximately
F 1 (p) G1 (pn ) > G1 (pn ),
so may be much larger than would be the case with = 1.
A similar argument shows that ignoring can lead to over-estimating a return level estimated using G.
Statistics of Extremes
D (un ) condition
lim sup n
n
X
j=2
If a stationary process satisfies both the D(un ) and D (un ) conditions, then its extremal index is
= 1.
D (un ) holds for many well-known processes, for example Gaussian autoregressive processes. This
raises the issue of how best to model the extremes of such processes, since at any realistic level,
dependence is likely to be strong, while the asymptotic limit is independence.
Statistics of Extremes
Modelling techniques
The above discussion tells us that under the weak condition D(un ) and some mild additional
conditions, the maxima of stationary series can be modelling using the GEV, and that exceedances
can be modelling using the GPD. This is a strong robustness result, but it implies that extremes will
occur in clusters of correlated observations.
Possible modelling strategies using exceedances and the GPD are:
to ignore the dependence on the basis that the marginal model is valid, but to inflate standard
errors to account for reduction in independent information;
Statistics of Extremes
37
10
-30
-20
-10
Degrees Below 0 F.
10
0
-10
-30
-20
Degrees Below 0 F.
20
20
10
20
30
40
Day Index
10
20
30
40
Day Index
Simple estimates of the extremal index are based on empirical means of clusters. Here the runs method is used:
a cluster is deemed to have terminated when there are r consecutive observations below the threshold. Left:
r = 1 gives 4 clusters. Right: r = 3 gives 2 clusters.
Statistics of Extremes
i
(mu ) 1 ,
where and are the parameters of the threshold excess generalized Pareto distribution, u is the
probability of an exceedance of u, and is the extremal index.
nu
nc
bu =
and b =
.
n
nu
Statistics of Extremes
nc
b
b
x
b100
b
u = 10
r=2
r=4
31
20
11.8 (3.0)
14.2 (5.2)
0.29 (0.19) 0.38 (0.30)
27.7 (12.0)
26.6 (14.4)
0.42
0.27
u = 20
r=2
r=4
43
29
17.4 (3.6)
19.0 (4.9)
0.36 (0.15) 0.41 (0.19)
26.2 (9.3)
25.7 (9.9)
0.24
0.16
38
scale
1.86637
Standard Errors
loc
scale
0.35380 0.23673
shape
0.06696
shape
0.05379
Statistics of Extremes
shape
0.04178
Standard Errors
loc
scale
shape
0.3519 0.2406 0.0632
Statistics of Extremes
39
0.2
0.3
0.4
Extremal Index
0.5
0.6
0.7
0.8
For finite threshold u, increases with u, suggesting that very extreme hourly rainfall totals occur
singly.
3
4
Threshold
Statistics of Extremes
Non-stationarity
General results not broad enough for application hence, model trends, seasonality and covariate
effects by parametric or nonparametric models for the usual extreme value model parameters.
Some possibilities for parametric modelling:
(t) = + t;
(t) = exp( + t);
1 , t t0 ,
(t) =
2 , t > t0 ;
(t) = + y(t).
Statistics of Extremes
Parameter estimation
m
Y
t=1
Standard likelihood techniques also yield standard errors, confidence intervals etc.
Statistics of Extremes
40
Model reduction
Statistics of Extremes
Model diagnostics
b
Zt GEV(b
(t),
b(t), (t)),
Zt
b(t)
1
b
log 1 + (t)
,
Zt =
b
b(t)
(t)
each have the standard Gumbel distribution, with probability distribution function
Pr{Zt z} = exp(ez ),
z R.
Possible diagnostics:
probability plot: i/(m + 1), exp( exp(
z(i) )); i = 1, . . . , m
quantile plot: log [ log{i/(m + 1)}] , z(i) ; i = 1, . . . , m
Statistics of Extremes
Similar techniques are applicable for the threshold exceedance and point process models, but
threshold selection is likely to be a more sensitive issue.
Time-varying thresholds may also be appropriate, though there is little guidance on how to make
such a choice.
Statistics of Extremes
Log-likelihood
43.6
49.9
50.6
50.7
Statistics of Extremes
41
1.6
1.2
1.4
Sea-level (metres)
1.8
1900
1920
1940
1960
1980
Year
Fitted trend in location parameter for Fremantle annual maximum sea levels.
Statistics of Extremes
0.0
0.2
0.4
Model
Empirical
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
-1
Empirical
Model
Probability and quantile plots for nonstationary GEV analysis of Fremantle annual maximum sea levels.
Statistics of Extremes
Log-likelihood
54.5
Linear
51.8
Quadratic
48.4
239.3
(0.9)
(242.9, 0.311)
(1.4, 0.101)
(247.0, 1.395, 0.049)
(2.3, 0.420, 0.018)
b
3.63
(0.64)
2.72
(0.49)
2.28
(0.45)
b
0.469
(0.141)
0.201
(0.172)
0.182
(0.232)
42
240
238
232
234
236
242
244
246
1975
1980
1985
1990
Year
Fitted models for location parameter in womens 1500 metre race times. Note quadratic model would
lead to slower races in recent and future events.
Statistics of Extremes
(t) = 0 + 1 e2 t .
240
238
232
234
236
242
244
246
has log-likelihood 49.5. Not so good as quadratic model, though comparison via likelihood ratio test
is invalid as models are not nested. Better behaviour for large t suggests a preferable model though.
1975
1980
1985
1990
Year
Statistics of Extremes
43
-20
-60
-40
0
-20
-40
20
Winter
100
200
300
400
100
200
Day Index
Day Index
Summer
Autumn
300
400
300
400
-30
-70
-50
-50
-60
-70
-40
-10
100
200
300
400
Day Index
100
200
Day Index
One approach to handle seasonality is to model seasons separately. This is not likely to be sufficient
here, at least with only 4 seasons.
Statistics of Extremes
0
-20
-40
-60
20
1983
1984
1985
1986
1987
1988
Year
p
3
5
7
9
9
10
143.6
126.7
143.6
145.9
145.1
143.9
44
Semiparametric regression
We may relax parametric assumptions, or mix them with more flexible forms. For example, we may want to
model temperatures at sites in the Alps as dependent on altitude a(x), where x is location, plus a smooth
function of time t, giving
(x, t) = 0 + 1 a(x) + g(t), or (x, t) = 0 + 1 a(x) + g{t, a(x)}.
Statistics of Extremes
Formulation
Often take splines, which have optimality properties, and which correspond to prediction of stochastic
processes.
Usually penalise the smoothing parameters = (1 , . . . , p )T , for example, maximising a log likelihood of
form
(, ) + T K,
where smoothing parameter > 0 controls smoothness of result (equivalent to effective degrees of
freedom).
Statistics of Extremes
45
Grachen
1980
1990
Sils Maria
Arosa
5
10
15
20
0
Chateau dOex
La Brevine
Guttannen
Oberiberg
Chaumont
Andermatt
Fribourg
Haidenhaus
Heiden
Einsredeln
Vattis
Elm
5
10
15
20
0
5
10
15
20
0
Rheinfelden
Grono
Altstatten
Thun
Ebnat Kappel
Meiringen
5
10
15
20
1970
1980
1990
1970
1980
1990
1970
1980
1990
Year
0.4
0.3
0.2
lambda
0.0
0.1
0.2
0.0
0.1
lambda
0.3
0.4
20
40
60
80
8
6
2
sigma
6
sigma
4
2
20
40
60
80
0.10
xi
0.20
0.25
0.25
0.20
0.15
0.15
0.10
0.05
0.05
Day
xi
Day
500
1000
1500
Altitude
Statistics of Extremes
46
30
30
30
15
20
25
10
25
25
10
15
20
20
15
10
Arosa
Vattis
Rheinfelden
Statistics of Extremes
Care needed with choice of model: with GPD, change of threshold u 7 u changes scale
parameter:
u 7 u = u + (u u),
so, for example, the formulation
(u , ) = (exp{g(t)}, h(x))
at threshold u will become
(u , ) = exp{g(t)} + h(x)(u u ), h(x) ,
Better to fit on GEV scale, or using point process formulation, for which parametrization is
invariant.
Statistics of Extremes
47
Multivariate Extremes
slide 111
Multivariate extremes
Lack of data means the precision of extreme value estimates is often poor.
The only way to overcome this is to incorporate additional information, suggesting the use of
multivariate models.
Questions include:
Statistics of Extremes
Componentwise maxima
iid
and
then
M n = (Mx,n , My,n )
is the vector of componentwise maxima.
The issue is partly resolved by recognizing that {Xi } and {Yi } considered separately are sequences
of independent, univariate random variables, to which the earlier theory may be applied.
Statistics of Extremes
Marginal standardization
Representations are especially simple if we assume that both Xi and Yi have the standard Frechet
distribution, with distribution function
F (z) = exp(1/z),
Then defining
M n
M n
z > 0.
i=1,...,n
i=1,...,n
48
Let M n = (Mx,n
, My,n
) be componentwise maxima of independent vectors with standard Frechet marginal
distributions. Then if
d
Pr{Mx,n
x, My,n
y} G(x, y),
max
0
x > 0, y > 0
w 1w
, y
x
dH(w),
max
0
w 1w
, y
x
h(w)dw.
Statistics of Extremes
Special cases
x > 0, y > 0.
x > 0, y > 0,
which is the distribution function of variables that are marginally standard Frechet, but which are
perfectly dependent: X = Y with probability 1.
Statistics of Extremes
Parametric models
For modelling purposes, it is usual to specify a parametric family for H or h that encompasses a
wide range of dependence types over the parametric domain.
In this case
o
n
,
G(x, y) = exp x1/ + y 1/
x > 0, y > 0.
Statistics of Extremes
49
Alternative models
A limitation of the logistic model is its symmetry. Asymmetric alternatives include the
bilogistic model
h(w) = 12 (1 )(1 w)1 w2 (1 u)u1 {(1 u) + u}1
on 0 < w < 1, where 0 < < 1 and 0 < < 1, and u = u(w, , ) is the solution of
(1 )(1 w)(1 u) (1 )wu = 0;
0 < w < 1,
Structure variables
2.
multivariate analysis of the full process, followed by marginalization to the structure variable
of interest.
Statistics of Extremes
50
Statistics of Extremes
4.4
4.2
4.0
3.8
3.6
4.6
1.3
1.4
1.5
1.6
1.7
1.8
1.9
51
2.0
1.0
1.5
2.5
3.0
10
50
100
500
1000
Return levels of structure variable Z = max{Mx , (My 2.5)} using both univariate and multivariate
methods.
Statistics of Extremes
2.0
1.6
1.8
2.2
2.4
10
50
100
500
1000
Z = max{Mx , (My 2.5)} in logistic model analysis of Fremantle and Port Pirie annual maximum
sea-level series with = 0, 0.25, 0.5, 0.75, 1 ( = 0 is lowest curve).
Statistics of Extremes
52
1.6
1.2
1.4
1.8
2.0
10
50
100
500
1000
Z = min{Mx , (My 2.5)} in logistic model analysis of Fremantle and Port Pirie annual maximum
sea-level series with = 0, 0.25, 0.5, 0.75, 1 ( = 0 is highest curve).
Statistics of Extremes
As in the univariate case, there are threshold exceedance and point process characterizations of
extremes that enable a greater use of available information.
The point process representation, in particular, also provides some insight into the measure
function H that appears in the componentwise limit law.
Statistics of Extremes
Pr{Mx,n
x, My,n
y} G(x, y).
53
Nn N,
and w =
x
x+y ,
max
w 1w
x, y
dH(w)}
Statistics of Extremes
Comments
The intensity function of the limit process factorizes across r and w. Loosely, the relative
magnitude of extreme events in the two variables is independent of the magnitude itself.
In practiceas usualthe limit process is taken to be exact for extreme enough events: in this
case for events which are sufficiently far from (0, 0).
Statistics of Extremes
Consistency of results
Pr{Mx,n
x, My,n
y} = Pr{Nn (A) = 0},
where
A = {(0, ) (0, )}\{(0, x) (0, y)}.
So,
Pr{Mx,n
x, My,n
y} Pr{N (A) = 0} = exp{(A)},
where
(A) =
=
= 2
2
A
1
dr
dH(w)
r2
Z
w=0
Z 1
w=0
r=min{x/w,y/(1w)}
max
w 1w
x, y
dr
dH(w)
r2
dH(w).
Statistics of Extremes
54
dr
dH(w) = 2
r2
r=r0
dr
r2
dH(w) = 2/r0 ,
w=0
NA
Y
i=1
NA
Y
h(wi ),
i=1
where wi = x(i) /(x(i) + y(i) ) for the NA points (x(i) , y(i) ) falling in A.
Statistics of Extremes
Comments
Simultaneous estimation of marginal and dependence features is possibly advantageous, but again
results in a considerably more complicated likelihood.
Statistics of Extremes
0.0
-0.02
Log-daily return
0.02
1997
1998
1999
2000
2001
2000
2001
0.0
-0.02
Log-daily return
0.02
Year
1997
1998
1999
Year
Log-daily returns of exchange rates. Top panel: UK sterling/US dollar exchange rate. Bottom panel:
UK sterling/Canadian dollar exchange rate.
Statistics of Extremes
55
0.01
0.0
-0.01
-0.02
0.02
-0.02
-0.01
0.0
0.01
0.02
Sterling v. US Dollars
100
10
1
1000
10
100
1000
56
1.5
1.0
0.0
0.5
Relative Frequency
2.0
0.0
0.2
0.4
0.6
0.8
1.0
Fitted logistic model density h in point process analysis of exchange rate data. (b
= 0.434 (0.025))
Statistics of Extremes
0.2
-0.2
0.0
Surge (m)
0.4
0.6
0.8
10
57
1.0
0.0
0.5
Relative Frequency
1.5
2.0
0.0
0.2
0.4
0.6
0.8
1.0
227.2
230.2
238.2
0.659
0.704
0.852
0.603
0.502
Details of various fits to oceanographic data give support for asymmetric dependence structure.
Statistics of Extremes
Higher-dimensional models
d
X
wj = 1}
j=1
subject to
Z
wj dH(w) = 1,
j = 1, . . . , d
Sd
2.
3.
Statistics of Extremes
58
Asymptotic dependence
A limitation of multivariate extreme value models (and point process equivalents) is that apart from
the degenerate case of independence, all such models are asymptotically dependent; the intuition is
that because these are asymptotic (limiting) models, the extent of dependence cannot vary with the
rareness of the event, whereas this happens in practice.
Two variables X and Y with the same distribution are said to be asymptotically independent if
lim Pr{Y > z | X > z} = 0,
Loosely, two variables are asymptotically independent if an extreme value in one has zero
probability of happening on the occurrence of an extreme in the other variable.
Statistics of Extremes
The class of asymptotically independent variables is non-trivial, and includes, for example, all
bivariate normal variables with positive correlation.
The limit is independence, but this provides a poor approximation to dependence at finite
levels.
Much recent work has looked at inference for asymptotic independent models: Ledford and Tawn,
Heffernan, Coles and Tawn, Heffernan and Tawn.
Statistics of Extremes
59
Spatial Extremes
slide 143
Spatiotemporal extremes
Many environmental extremal problems are spatial or temporal in nature, or both:
heatwaves
rainfall
avalanches
forest fires
storms
sea levels
Likely to be more important in the future, prediction (on different scales) needed for long-range
planning, short-range evacuation, . . .
Statistics of Extremes
Geostatistics
slide 145
Geostatistics
and use MetropolisHastings algorithm for inference on S(x), predictions of future Y , etc.
(Diggle, Tawn, Moyeed, 1998), lots of Bayesians
Statistics of Extremes
60
Statistics of Extremes
Jungfraujoch (3580 m)
Santis (2490 m)
!
Gd!St!Bernard (2472 m)
!
!
Arosa (1840 m)
!
!
Davos!Dorf (1590 m)
!
Montana (1508 m)
!
Engelberg (1035 m)
Bern!Liebefeld (565 m)
!
Zurich!MeteoSchweiz (556 m)
!
Neuchatel (485 m)
!
Oeschberg!Koppigen (483 m)
!
!
Montreux!Clarens (405 m)
!
!
Locarno!Monti (366 m)
!
!
!
Basel!Binningen (316 m)
!
Lugano (273 m)
2001
2002
2003
2004
Statistics of Extremes
2005
61
Geostatistics of extremes
Basic setup:
where r(x, t) is population at risk if Y (x, t) exceeds some level ydanger at time t
Statistics of Extremes
Approaches
Four (?) main approaches:
copulas
HeffernanTawn
Statistics of Extremes
Conditional on underlying process S(x), observations Y (x), for x X follow an extremal distribution
Examples:
Y (x)
Y (x)
|
|
ind
Examples: Casson and Coles (1999, Extremes); Cooley et al. (2007, JASA)
Could use copulas to transform margins to Gaussian, then fit geostatistical modelsGaussian
anamorphesis
Advantages: computationally feasible for large-scale problems using standard simulation techniques
(MetropolisHastings algorithm, . . .), possibility of estimating probabilities for complex events
Disadvantages: all extremal dependencies are incorporated through S(x); marginal distributions are not
extremal; difficult to incorporate full range of possible extremal dependencies
Statistics of Extremes
62
HeffernanTawn (2004)
Use decomposition
Pr(Y C) =
where C =
d Cd
D
X
d=1
Use GPD for Yd and nonparametric linear model for conditional probability, and simulate to estimate
Pr(Y C).
Advantages: can deal with large D, extends to near-independence models
Disadvantages: models on different Cd need not be coherent, uniqueness of representation for conditional
probability unclear, inference messy
Statistics of Extremes
Max-stable processes
General setup:
Z(x) Frechet(1), for x X , so if the Zm () are independent and identically distributed, then
m1 max{Z1 (x), . . . , Zm (x)} Frechet(1)
where D = |D|, SD is the unit simplex in RD , and H is a measure on SD with all its marginal
expectations equal to 1; we have 1 D D.
Statistics of Extremes
Extremal coefficient
z > 0,
Not a complete summary of joint distribution, but basis of useful diagnostics, as can estimate D quite
easily.
Rainfall example (sketch!)
Statistics of Extremes
63
Madogram
Exploratory tool (analogous to variogram), designed for dealing with many extremes on a map
(x + h)} F {Y (x)}|] =
1 (h) 1
,
2 (h) + 1
max
w 1w
,
y1 y2
dHh (w).
Statistics of Extremes
Spectral representations I
de Haan (1984):
Z(x) = max Uk f (x Tk ),
k
where (Uk , Tk ) are the points of a Poisson process on R+ [0, 1] with measure du/u2 (dt),
where is a finite measure on [0, 1] and fk are non-negative L1 functions.
Exploited by Smith (1990), Coles (1993), Coles and Walshaw (1994), de Haan and Pereira (2006)
with various choices of f (normal, t, Laplace, circular)
Statistics of Extremes
64
Spectral representations II
Schlather (2002):
Let V (x) be a stationary process on Rd with = E max{0, V (x)} < , and let be a Poisson process on
(0, ) with intensity 1 ds/s2 . If the Vs (x) are independent copies of V (x), then
Z(x) = max sVs (x),
s
1
2
1
1
+
z1
z2
1/2 !
{(h) + 1}z1 z2
1+ 12
(z1 + z2 )2
Corresponding extremal coefficient (h) = 1 + 21/2 {1 (h)}1/2 has natural bounds (Schlather and
Tawn, 2003)
Can only represent positive dependencebut most likely in practice
Statistics of Extremes
Pairwise likelihood
Data Z1 , . . . , Zn are from (unavailable) full joint density f (z1 , . . . , zn ; ), yielding log likelihood
() = log f (z1 , . . . , zn ; ).
If lower-order marginal densities are available, construct composite likelihood by taking product of
(non-independent!) densities.
log f (zi , zj ; ),
i>j
iid
j
If is identifiable from the pairwise marginal densities, and if Z 1 , . . . , Z n f (z; ), with Z j = (Z1j , . . . , ZD
), then
under mild regularity conditions the maximum pairwise likelihood estimator is consistent, and satisfies
n
o
1 K()J(
)
1 as n .
N , J()
Statistics of Extremes
65
Statistics of Extremes
d ,
d ,
d = 1, . . . , 17, j = 1, . . . , 46,
iid
use likelihood estimation to obtain bij for each pair of sites (model checking)
0 1 1, 2 , 3 > 0,
simulate max-stable random fields Z (x) from fitted model, then transform back to real scale Y (x) and
assess . . .
Statistics of Extremes
66
50.0
Station Engelberg
0.5
Station GdStBernard
2.0
10.0
10.0 50.0
0.5
1.0
5.0
50.0
0.5
50.0
50.0
10.0 50.0
Station Montana
2.0
0.5
2.0
10.0 50.0
Station Lugano
5.0
0.2
Station MontreuxClarens
1.0
5.0
50.0
Station Neuchatel
0.2
1.0
5.0
50.0
Station OeschbergKoppigen
10.0 50.0
Station DavosDorf
50.0
2.0
10.0 50.0
0.5
2.0
10.0 50.0
0.2 1.0 5.0
10.0 50.0
5.0
0.2
2.0
50.0
2.0
1.0
50.0
Station LocarnoMonti
0.5
0.2
5.0
0.5
0.2 1.0 5.0
1.0
2.0
50.0
10.0 50.0
5.0
0.2
10.0 50.0
5.0
0.5
1.0
50.0
0.5
1.0
2.0
10.0 50.0
2.0
0.5
0.2
5.0
2.0
10.0 50.0
2.0
0.2
Station BernLiebefeld
1.0
Station BaselBinningen
0.5
0.5
2.0
10.0 50.0
50.0
0.5
0.2
0.5
10.0 50.0
0.5
2.0
10.0 50.0
0.5
Station Arosa
2.0
10.0 50.0
50.0
10.0 50.0
10.0
2.0
2.0
0.5
0.5
0.5
2.0
10.0 50.0
2.0
0.5
0.5
2.0
10.0 50.0
50.0
0.5
Station Jungfraujoch
2.0
10.0 50.0
0.5
Station Satis
5.0
50.0
Station ZrichMeteoSchweiz
Statistics of Extremes
Estimated correlations
1.0
0.5
0.0
1.0
0.5
0.5
0.0
0.5
1.0
1.0
Individual points: MLEs of correlations 2 SE (grey bars) as a function of distance (103 km) between pairs of
points
Solid line: fitted correlation estimated using pairwise likelihood
Left: temperature data; right: simulated data
0.0
0.2
0.4
0.6
0.8
1.0
0.0
Statistics of Extremes
0.2
0.4
0.6
0.8
1.0
67
1
0
2
Sample Quantiles
Normal QQ Plot
Theoretical Quantiles
Statistics of Extremes
68
69
Statistics of Extremes
y (x, 2007) dx
Simulated fields
5.0
50.0
5.0
50.0
50.0
0.1
0.5
50.0
0.5
5.0
50.0
5.0
0.5
5.0
50.0
5.0
0.5
5.0
50.0
pair(1,13) altitudes=(1840,405)
0.1
0.5
5.0
50.0
pair(1,16) altitudes=(1840,2490)
0.1
0.5
5.0
50.0
0.1
0.5
5.0
50.0
0.5
5.0
0.5
0.5
pair(1,9) altitudes=(1840,3580)
0.1
0.1
pair(1,12) altitudes=(1840,1508)
0.1
0.5
pair(1,15) altitudes=(1840,483)
0.1
0.1
50.0
5.0
50.0
pair(1,5) altitudes=(1840,985)
0.1
50.0
5.0
0.5
5.0
0.1
0.1
0.5
50.0
0.5
5.0
0.5
50.0
pair(1,8) altitudes=(1840,2472)
0.1
0.1
pair(1,11) altitudes=(1840,273)
0.1
0.5
5.0
0.5
pair(1,14) altitudes=(1840,485)
0.1
0.1
50.0
5.0
0.1
50.0
5.0
5.0
5.0
50.0
5.0
50.0
0.5
0.5
50.0
0.5
0.5
5.0
pair(1,10) altitudes=(1840,366)
0.1
pair(1,4) altitudes=(1840,565)
0.1
0.5
5.0
0.1
pair(1,7) altitudes=(1840,1035)
0.1
50.0
50.0
5.0
50.0
50.0
50.0
5.0
50.0
0.5
0.1
0.5
5.0
50.0
5.0
pair(1,3) altitudes=(1840,316)
0.1
0.5
5.0
0.1
0.5
5.0
50.0
0.5
0.5
0.1
0.1
0.5
5.0
0.1
0.5
50.0
50.0
5.0
pair(1,17) altitudes=(1840,556)
0.1
0.5
5.0
50.0
Pairwise fits
50.0
pair(1,6) altitudes=(1840,1590)
0.1
0.1
Statistics of Extremes
pair(1,2) altitudes=(1840,496)
0.1
10
10
45.5
45.5
47.5
47.5
47.0
47.0
46.5
46.5
46.0
46.0
48.0
48.0
Statistics of Extremes
10
45.5
45.5
47.5
47.5
47.0
47.0
46.5
46.5
46.0
46.0
48.0
48.0
Statistics of Extremes
10
45.5
45.5
47.5
47.5
47.0
47.0
46.5
46.5
46.0
46.0
48.0
48.0
Statistics of Extremes
Discussion
Schlather representation is very general, can envisage for other types of data:
rainfall time series at a single site, using addition of a random set (Mehdi)
rainfall in space and time?
snowfall in space (Juliette?)
etc.
Modelling issues:
What processes are sensible? Gaussian random fields? Levy processes?
How to extend to near-independence cases?
How to build in covariates? Smoothing? Pre-whitening or not?
How best to perform estimation and testing?
Bayesian inference?
How to build in physical knowledge of underlying processes?
Statistics of Extremes