A Journey Into the Emotions of Software Developers
Santiágo Beguería: Covariate-dependent modeling of extreme events by non-stationary Peaks Over Threshold analysis
1. Covariate-dependent modeling of
extreme events by nonstationary
Peaks Over Threshold analysis
A review and a case study in Spain
Santiago Beguería
santiago.begueria@csic.es
Estación Experimental de Aula Dei,
Consejo Superior de Investigaciones Cientícas (EEADCSIC), Zaragoza, España
Liberec, February 17th 2012
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 1 / 52
2. Motivation, I
We all love the extreme value theory (EVT) applied to the analysis of
climatic hazard: it's elegant, simple, and provides useful and
understandable results in terms of magnitude / frequency curves.
The stationarity assumption, though, was an important limitation of
the EVT.
Relatively recent extensions of the EVT allow for non-stationary
analysis (Coles, 2001), and an increasing number of authors are
exploring their possibilities for the analysis of climatic hazard.
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 2 / 52
3. Motivation, I
We all love the extreme value theory (EVT) applied to the analysis of
climatic hazard: it's elegant, simple, and provides useful and
understandable results in terms of magnitude / frequency curves.
The stationarity assumption, though, was an important limitation of
the EVT.
Relatively recent extensions of the EVT allow for non-stationary
analysis (Coles, 2001), and an increasing number of authors are
exploring their possibilities for the analysis of climatic hazard.
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 2 / 52
4. Motivation, I
We all love the extreme value theory (EVT) applied to the analysis of
climatic hazard: it's elegant, simple, and provides useful and
understandable results in terms of magnitude / frequency curves.
The stationarity assumption, though, was an important limitation of
the EVT.
Relatively recent extensions of the EVT allow for non-stationary
analysis (Coles, 2001), and an increasing number of authors are
exploring their possibilities for the analysis of climatic hazard.
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 2 / 52
5. Motivation, II
Most studies up to now focused on identifying temporal trends in the
occurrence of extreme events, i.e. making time a covariate.
In the last few years other covariates with an expected inuence on
the occurrence of extreme events are being used, too high
relevance for the statistical downscaling of reanalysis or model data,
which typically cannot be used for local impact studio.
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 3 / 52
6. Motivation, II
Most studies up to now focused on identifying temporal trends in the
occurrence of extreme events, i.e. making time a covariate.
In the last few years other covariates with an expected inuence on
the occurrence of extreme events are being used, too high
relevance for the statistical downscaling of reanalysis or model data,
which typically cannot be used for local impact studio.
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 3 / 52
7. Summary
In this talk I will present ongoing research on the relationship between
extreme precipitation and teleconnection indices in Spain, using
non-stationary EVT techniques. The talk is organized as follows:
A review of non-stationary Peaks Over Threshold analysis
Case study: relationship between teleconnection indices and extreme
rainfall events in Spain
Conclusions and future work
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 4 / 52
8. Summary
In this talk I will present ongoing research on the relationship between
extreme precipitation and teleconnection indices in Spain, using
non-stationary EVT techniques. The talk is organized as follows:
A review of non-stationary Peaks Over Threshold analysis
Case study: relationship between teleconnection indices and extreme
rainfall events in Spain
Conclusions and future work
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 4 / 52
9. Summary
In this talk I will present ongoing research on the relationship between
extreme precipitation and teleconnection indices in Spain, using
non-stationary EVT techniques. The talk is organized as follows:
A review of non-stationary Peaks Over Threshold analysis
Case study: relationship between teleconnection indices and extreme
rainfall events in Spain
Conclusions and future work
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 4 / 52
10. Summary
In this talk I will present ongoing research on the relationship between
extreme precipitation and teleconnection indices in Spain, using
non-stationary EVT techniques. The talk is organized as follows:
A review of non-stationary Peaks Over Threshold analysis
Case study: relationship between teleconnection indices and extreme
rainfall events in Spain
Conclusions and future work
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 4 / 52
11. Short review of NSPOT analysis, I
200
150
Intensity (mm day-1)
100
50
0
1950 1960 1970 1980 1990 2000 2010
Time (n = 266)
Peaks-over-threshold (POT) sampling: take only exccedances over a threshold, x − x0
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 5 / 52
12. Short review of NSPOT analysis, II
200
150
Intensity (mm day-1)
100
50
0
1950 1960 1970 1980 1990 2000 2010
Time (n = 266)
Peaks-over-threshold (POT) sampling: take only exccedances over a threshold, x − x0
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 6 / 52
13. Short review of NSPOT analysis, III
Stationary POT: assuming independent inter-arrival times, the POT data
follows a Generalized-Pareto distribution.
Probability of exceedance:
−1/κ
x − x0
P (X x |X x0 ) = 1 − λ 1 + κ (1)
α
Quantile corresponding to a return period T:
κ
α 1
XT = x0 + 1− (2)
κ λT
(beware of alternative conventions: x0 = u , α = σ, κ = ξ )
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 7 / 52
14. Short review of NSPOT analysis, IV
Approaches for assessing non-stationarity in POT modeling:
Split-sample approach (Li et al., 2005)
Moving kernel (Hall and Tajvidi, 2000)
Non-stationary POT (NSPOT) modeling
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 8 / 52
15. Short review of NSPOT analysis, GPar distribution
Return period curves,
V
2500
NAO−
2000
1500
daily EI
NAO+
1000
500
0
1 2 5 10 20 50 100
return period (years)
Split-sample approach: independent models for positive and negative phases of NAO (Angulo et al., 2011).
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 9 / 52
16. Short review of NSPOT analysis, VI
2110 ´
S. BEGUERIA et al.
X287 X164
0.00 0.02 0.04 0.06 0.08 0.10
0.00 0.05 0.10 0.15 0.20 0.25
1930 1950 1970 1990 2010 1930 1950 1970 1990 2010
X9198 X2234
0.0 0.1 0.2 0.3 0.4 0.5 0.6
0.5
0.4
0.3
0.2
0.1
0.0
1930 1950 1970 1990 2010 1930 1950 1970 1990 2010
Figure 6. Model uncertainty: time variation of the 10-year return period event intensity standard error (mm by a nonparametric NSPOT d−1 )
model (dots), a linear NSPOT model (solid line), a high order polynomial NSPOT model (short-dashed line) and a stationary POT model
Moving kernel approach: time variability in the stations. Location of the stations a moving Figure 1. of the previous 20 years of
(long-dashed line) in four P10 quantile, based on is shown in window
data (Beguería et al., 2011).
by more than 50% (e.g. series X164). The time evolution and the most parsimonious (i.e. with the lowest polyno-
of q10 depended on the evolution
(santiago.begueria@csic.es) of both α (scale) and mial order) was chosen. Following this approach, statis- 10 / 52
NSPOT with covariates Liberec, 17/02/2012
17. Short review of NSPOT analysis, VII
Stationary POT:
−1/κ
x − x0
P (X x |X x0 ) = 1 − λ 1 + κ
α
Non-stationary POT:
−1/κ(c )
x − x0 (c )
P (X x |X x0 , C ) = 1 − λ(c ) 1 + κ(c ) (3)
α(c )
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 11 / 52
18. Short review of NSPOT analysis, VIII
Some examples of NSPOT analysis of climatic variables:
Time dependence of T and P (Smith, 1999)
Nogaj et al. (2006) time trends of T extremes over the NA region
Laurent and Parey (2007), Parey et al. (2007), T extremes in France
Méndez et al. (2006), trends and seasonality of POT wave height
Yiou et al. (2006) trends of POT discharge in the Czech Republic
Abaurrea et al. (2007) trends of POT T in the IP
Acero et al. (2011), Beguería et al. (2011), trends in POT P, IP
Friederichs (2010), Kallache et al. (2011), downscaling of POT P based on
reanalysis / GCM data
Tramblay et al. (2012), covariation between POT P extremes and
atmospheric covariates, SE France
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 12 / 52
19. Teleconnections aecting precipitation in IP, I
The North Atlantic Oscillation (NAO).
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 13 / 52
20. Teleconnections aecting precipitation in IP, II
The North Atlantic Oscillation (NAO).
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 14 / 52
21. Teleconnections aecting precipitation in IP, III
The Mediterranean Oscillation (MO, Palutikof 2003).
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 15 / 52
22. Teleconnections aecting precipitation in IP, IV
The Mediterranean Oscillation (MO, Palutikof 2003).
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 16 / 52
23. Teleconnections aecting precipitation in IP, V
The Western Mediterranean Oscillation (WEMO, Martín-Vide and López-Bustins 2006).
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 17 / 52
24. Teleconnections aecting precipitation in IP, VI
The Western Mediterranean Oscillation (WEMO, Martín-Vide and López-Bustins 2006).
(santiago.begueria@csic.es) NSPOT with covariates Liberec, 17/02/2012 18 / 52