Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory...
Transcript of Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory...
![Page 1: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/1.jpg)
Statistical Extreme Value Theory (EVT)
Part I
Eric Gilleland Research Applications Laboratory
21 July 2014, Trieste, Italy
![Page 2: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/2.jpg)
What is extreme?
• The upper quartile of temperatures during the year?
• The maximum temperature during the year? • The highest stream flow recording during the
day? • The lowest stream flow recordings accumulated
over a three-day period for the year? • Winning the lottery?
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 3: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/3.jpg)
What is extreme?
• When a variable exceeds some high threshold?
• When the year’s maximum of a variable is very different from the usual maximum value?
• When an unusual event takes place, regardless of whether or not it is catastrophic?
• Only when an event causes catastrophes?
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 4: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/4.jpg)
What is extreme?
• The United States wins the World Cup for Soccer?
• Any team wins the World Cup? • The Denver Broncos win the Super Bowl?
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 5: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/5.jpg)
What is extreme?
Suppose an “event” of interest, E, has a probability, p, of occurring and a probability 1 – p of not occurring.
Then, the number of events, N, that occur in n trials follows a binomial distribution. That is, the probability of having k “successes” in n trials is governed by the probability distribution:
Pr N = k{ } = nk
⎛⎝⎜
⎞⎠⎟pk 1− p( )n−k
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 6: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/6.jpg)
What is extreme?
Now, suppose p is very small. That is, E has a very low probability of occurring.
Simon Denis Poisson introduced the probability distribution, named after him, obtained as the limit of the binomial distribution when p tends to zero at a rate that is fast enough so that the expected number of events is constant.
Pr N = k{ } = nk
⎛⎝⎜
⎞⎠⎟pk 1− p( )n−k
That is, as n goes to infinity, the product np remains fixed.
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 7: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/7.jpg)
What is extreme?
That is, for p “small,” N has an approximate Poisson distribution with intensity parameter nλ.
Pr N = 0{ } = e−nλ , Pr N > 0{ } = 1− e−nλ
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 8: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/8.jpg)
What is extreme?
We’re looking for events with a very low probability of occurrence (e.g., if n ≥ 20 and p ≤ 0.05 or n ≥ 100 and np ≤ 10). Does not mean that the event is impactful! Does not mean that an impactful event must be governed by EVT!
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 9: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/9.jpg)
Inference about the maximum
• If I have a sample of size 10, how many maxima do I have?
• If I have a sample of size 1000? • Is it possible, in the future, to see a value
larger than what I’ve seen before? • How can I infer about such a value? • What form of distribution arises for the
maximum?
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 10: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/10.jpg)
Extreme Value Theory Background
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Maxima of samples of size 30from standard normal distribution
Density
1 2 3 4
0.0
0.2
0.4
0.6
0.8
sampling experiment
![Page 11: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/11.jpg)
Extreme Value Theory Background: Max Stability
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
max{ x1, …, x2n } = max{ max{ x1, …, xn }, max{ xn+1, …, x2n }}.
0 20 40 60 80 100
−4−3
−2−1
01
2
0 20 40 60 80 100
−4−3
−2−1
01
2 ●
0 20 40 60 80 100
−4−3
−2−1
01
2 ●
●
![Page 12: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/12.jpg)
Extreme Value Theory Background: Max Stability
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
In other words, the cumulative distribution function (cdf), say G, must satisfy G2(x) = G(ax + b) where a > 0 and b are constants.
![Page 13: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/13.jpg)
Extreme Value Theory Background: Extremal Types Theorem
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Let X1,…,Xn be independent and identically distributed, and define Mn = max{X1,…,Xn }.
Suppose there exist constants an > 0 and bn such that Pr{ (Mn – bn) / an ≤ x } G(x) as n ∞, where G is a non-degenerate cdf.
![Page 14: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/14.jpg)
Extreme Value Theory Background: Extremal Types Theorem
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Then, G must be a generalized extreme value (GEV) cdf. That is,
G x;µ,σ ,ξ( ) = exp − 1+ ξ x − µσ
⎡⎣⎢
⎤⎦⎥
−1/ξ⎧⎨⎪
⎩⎪
⎫⎬⎪
⎭⎪
defined where the term inside the [ ] and σ are positive.
![Page 15: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/15.jpg)
Extreme Value Theory Background: Extremal Types Theorem
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
G x;µ,σ ,ξ( ) = exp − 1+ ξ x − µσ
⎡⎣⎢
⎤⎦⎥
−1/ξ⎧⎨⎪
⎩⎪
⎫⎬⎪
⎭⎪
Note that this is of the same form as the Poisson distribution!
![Page 16: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/16.jpg)
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Extreme Value Theory Background: Extremal Types Theorem
Three parameters: • µ is the location parameter • σ > 0 the scale parameter • ξ is the shape parameter
![Page 17: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/17.jpg)
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Extreme Value Theory Background: Extremal Types Theorem
Three types of Extreme Value df’s
Weibull; bounded upper tail
Gumbel; light tail
Fréchet; heavy tail
![Page 18: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/18.jpg)
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Extreme Value Theory Background: Extremal Types Theorem
Denny, M.W., 2008, J. Experim. Biol., 211:3836–3849.
Predicted Speed Limits Thoroughbreds (Kentucky Derby)
≈ 38 mph
Greyhounds (English Derby) ≈ 38 mph
Men (100 m distance) ≈ 24 mph
Women (100 m distance) ≈ 22 mph
Women (marathon distance) ≈ 12 mph
Women (marathon distance using a different model)
≈ 11.45 mph
Paula Radcliffe, 11.6 mph world marathon record London Marathon, 13 April 2003
![Page 19: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/19.jpg)
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Extreme Value Theory Background: Extremal Types Theorem
Weibull type: temperature, wind speed, sea level negative shape parameter bounded upper tail at:
µ − σξ
![Page 20: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/20.jpg)
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Extreme Value Theory Background: Extremal Types Theorem
Gumbel type: “Domain of attraction” for many common distributions (e.g., normal, exponential, gamma) limit as shape parameter approaches zero. “light” upper tail
![Page 21: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/21.jpg)
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Extreme Value Theory Background: Extremal Types Theorem
Fréchet type: precipitation, stream flow, economic damage positive shape parameter “heavy” upper tail infinite k-th order moment if k ≥ 1 / ξ (e.g., infinite variance if ξ ≥ ½)
![Page 22: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/22.jpg)
Extreme Value Analysis • Fit directly to block maxima, with relatively
long blocks § annual maximum of daily precipitation
amount § highest temperature over a given year § annual peak stream flow
• Advantages § Do not necessarily need to explicitly model
annual and diurnal cycles § Do not necessarily need to explicitly model
temporal dependence UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
![Page 23: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/23.jpg)
Extreme Value Analysis
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Parameter estimation • Maximum Likelihood Estimation (MLE) • L-moments (other moment-based
estimators) • Bayesian estimation • various fast estimators (e.g., Hill estimator
for shape parameter)
![Page 24: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/24.jpg)
Extreme Value Analysis
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
MLE Given observed block maxima Z1 = z1,…,Zm = zm ,
minimize the negative log-likelihood (-ln L(z1,…, zm; µ, σ, ξ)) of observing the sample with respect to the three parameters.
![Page 25: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/25.jpg)
Extreme Value Analysis
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
MLE Allows for employing the likelihood-ratio test to test one model against another (nested) model. Model 1: -ln L(z1,…, zm; µ, σ, ξ = 0) Model 2: -ln L(z1,…, zm; µ, σ, ξ) If ξ = 0, then V = 2 * (Model 2 – Model 1) has approximate χ2 distribution with 1 degree of freedom for large m.
![Page 26: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/26.jpg)
Extreme Value Analysis
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
1950 1960 1970 1980 1990
1520
25
Year
Max
imum
Spr
ing
Tem
pera
ture
(C)
Sept, Iles, Québec
![Page 27: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/27.jpg)
Extreme Value Analysis
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Sept, Iles, Québec
●
●●
●●●●●●●●●●●●●
●●●●●●●●
●●●●●●●
●●●●●●●●●●●
●●●
●
●●● ●
●
14 16 18 20 22 24 26 28
1520
25
Model Quantiles
Empi
rical
Qua
ntile
s
●●
●
●●●● ●●●●●
●●●●●● ●●
●●●●●●● ●●
●●
●●●●●●●
●●●●
● ●●
●
●●
●
●
●
15 20 25
1520
2530
TMX1 Empirical Quantiles
Qua
ntile
s fro
m M
odel
Sim
ulat
ed D
ata
1−1 lineregression line95% confidence bands
10 15 20 25 30
0.00
0.04
0.08
0.12
N = 51 Bandwidth = 1.448
Den
sity
EmpiricalModeled
2 5 10 50 200 1000
2025
3035
40
Return Period (years)
Ret
urn
Leve
l
●●●
●●●●●●●●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●
●●●●●●
●● ● ●
●
fevd(x = TMX1, data = SEPTsp)
![Page 28: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/28.jpg)
Extreme Value Analysis
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Sept, Iles, Québec
95% lower CI
Estimate 95% upper CI
µ 17.22 18.20 19.18
σ 2.42 3.13 3.84
ξ -0.37 -0.14 0.09
100-year return level
24.72 °C 28.81 °C 32.90 °C
![Page 29: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/29.jpg)
Extreme Value Theory: Return Levels
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
Assume stationarity (i.e. unchanging climate) Return period / Return Level Seek xp such that G(xp) = 1 – p, where 1 / p is the return period. That is, xp = G-1(1 – p; µ, σ, ξ), 0 < p < 1 Easily found for the GEV cdf. Example, p = 0.01 corresponds to 100-year return period (assuming annual blocks).
![Page 30: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/30.jpg)
Extreme Value Theory: Return Levels
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.
xp
![Page 31: Statistical Extreme Value Theory (EVT) Part Iericg/EVT1.pdf · Statistical Extreme Value Theory (EVT) Part I Eric Gilleland Research Applications Laboratory 21 July 2014, Trieste,](https://reader034.fdocuments.in/reader034/viewer/2022042121/5e9ab5e4c4b2d755dc50a20f/html5/thumbnails/31.jpg)
Conclusion of Part I: Questions?
UCAR Confidential and Proprietary. © 2014, University Corporation for Atmospheric Research. All rights reserved.