Extreme Value Analysis Extreme Value Analysis
Nov 8, 2012
Extremes
Extreme Value Analysis
Motivation
I
Quantification of extremes is important for environmental disaster planning purposes - flood, wind, mudslides, fire, tornado, temperatures, drought, etc
I
Financial applications: insurance, risk analysis, stock gain/loss, etc
I
Human health: heat waves, ozone/pollutant levels, contamination levels, etc
Extremes
Extreme Value Analysis
Extreme Value Distribution
*Simulation from a χ23 Extremes
Extreme Value Analysis
Extreme Value Distribution
χ23
GEV(10.8, 0.8, 1)
Extremes
Extreme Value Analysis
Extreme Value Distribution
Extremes
Extreme Value Analysis
Extreme Value Analysis
Why do we need EVA?
What if the distributional model is correctly specified?
But what if the data are normally distributed?
Extremes
Extreme Value Analysis
Motivating Example
8000 7000 6000 5000
dowjones[, 2]
9000
10000
11000
12000
Daily Dow Jones, 1996−2000
0
200
400
600
800
1000
Index
Extremes
Extreme Value Analysis
1200
Motivating Example Log transformation for normality (djd), Differenced to remove temporal correlation:
Density
30
0.00
20
−0.02
10
−0.04
0
−0.06
diff(log(dowjones[, 2]))
0.02
40
0.04
50
density.default(x = djd)
0
200
400
600
800
1000
1200
Index
−0.08
−0.06
−0.04 N = 1303
Log Difference
−0.02
0.00
Density Extremes
Extreme Value Analysis
0.02
Bandwidth = 0.001849
0.04
Motivating Example
But what if the data are normally distributed? ●
●
0.04
●
● ●
● ●● ●
0.00 −0.02
Sample Quantiles
0.02
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●
−0.04
● ●
−0.06
●
●
●
●
−2
0
2
Standard Normal Quantiles
djd ∼ N(µ, σ) reasonable assumption for > 99% of the data.
Extremes
Extreme Value Analysis
Motivating Example
djd ∼ N(µ, σ) estimated by: x¯ = 0.0006697099, and σ = 0.01068724 P(drop of 249 points or more ) = P(djd ≤ -0.02843519) = P(drop of 617 points or more ) = P(djd ≤ -0.07454905) = And yet both of these were observed. −249.43 as the 1st quantile; −617.78 as the minimum difference (maximum drop). Extreme value analysis yields much more realistic (block) quantiles (return levels) and probabilities for extreme events.
Extremes
Extreme Value Analysis
Interpretable Quantity
A common measure of extreme events is the n-year return level.1 The n-year return level, yn , is the level so extreme it is expected to occur once every n time-units. Time units are usually natural groupings of the measured quantity - years, months, days, hours. Note: the n-year (block) return level corresponds to the (1 − n1 ) quantile of the predictive distribution.
1
Cooley, D., Nychka, D., Naveau, P., ”Bayesian Spatial Modeling of Extreme Precipitation Return Levels” Extremes
Extreme Value Analysis
Extreme Value Distribution
20 15 10
Temp
25
30
How do we decide what is extreme?
0
50
100
150 Months
Extremes
Extreme Value Analysis
200
20 15 10
Temp
25
30
Extreme Value Distribution
0
50
100
150 Months
Extremes
Extreme Value Analysis
200
Theory Given a sequence of iid RV’s X with common distribution F, consider the maxima Mn = max{X1 , ..., Xn }. The distribution of Mn can be derived exactly: P{Mn ≤ y } = {F (y )}n (Coles, 2001). Extremal Types Theorem: If there exists sequences of normalizing constants {an > 0} and {bn } such that P{(Mn − bn )/an ≤ y } → G (y ) where G is non-degenerate distribution function, then G belongs to the class of Generalized Extreme value distributions. Note: In practice, an and bn are not estimated.
Extremes
Extreme Value Analysis
Theory
The Generalized Extreme Value (GEV) distribution is defined by −1/ξ y −µ G(y) = exp − 1 + ξ ψ +
The Generalized Extreme Value (GEV) distribution is defined by location µ, scale ψ, and shape ξ. Parameter Interpretation:
Extremes
Extreme Value Analysis
Theory GEV: −1/ξ y −µ Pr{Y ≤ y } = exp − 1 + ξ ψ +
Three Types: I
Gumbel
I
Fr´echet
I
Weibull
Extremes
Extreme Value Analysis
Theory Parameters: Location µ, scale ψ, and shape ξ. Main Advantage: Direct parameterization for n-year Return Level, yn directly obtained w/ estimated GEV parameters 1 yn − µ −1/ξ = 1+ξ ψ n leads to: yn =
( ξ µ + ψ n ξ−1
if ξ 6= 0,
µ + ψ log n
if ξ = 0.
Main Disadvantage: Loss of information - biased return levels
Extremes
Extreme Value Analysis
Motivating Example Dow Jones Example:
50
density.default(x = djd) ●
0.04
0.04
●
● ●
0.00
0.02
20
−0.02
Density
Sample Quantiles
30
0.00 −0.02
● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●●
−0.04
●
●
−0.06
10
−0.04
●
−0.06
●
●
0
diff(log(dowjones[, 2]))
0.02
40
● ●● ●
0
200
400
600
800
1000
1200
Index
Log Difference
●
−0.08
−0.06
−0.04
−0.02
0.00
0.02
0.04
N = 1303 Bandwidth = 0.001849
Density
Extremes
Extreme Value Analysis
−2
0
2
Standard Normal Quantiles
QQ Norm Plot
●
GEV Analysis of Dow Jones Monthly Blocks: ∼22 days/month. 61 months (blocks). Minimums Quantile Plot ● ●●
●
●
Empirical
0.6
●●●● ● ●●● ● ●●
0.4
Model
0.8
●● ●● ● ●● ● ● ●●● ●●●● ● ●● ● ●● ● ● ● ●
0.0
0.2
●● ●● ● ●● ●● ● ● ●● ● ● ●● ●●●
0.2
0.4
0.6
0.8
1.0
●
● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●
0.01
0.02
● ●●
0.03
●
0.04
Empirical
Model
Density Plot
0.06
0.08
40
0.05
30
●
f(z)
0.06
●
20
●
0.04
●
1e−01
1e+00
1e+01
10
●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●
0
0.00
0.02
Return Level
●
●
Return Level Plot
0.10
0.0
0.01 0.02 0.03 0.04 0.05 0.06 0.07
1.0
Probability Plot
1e+02
1e+03
Return Period
● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●
0.00
0.02
●
●
0.04
0.06
●
●
0.08
z
library(ismev); library(extrRemes) djd.gevMIN u} = 1 − 1 + ξ σxu +
Advantage of GPD: I
Richer source of extreme data Disadvantages of GPD:
I
Parameters tied to threshold value, u, ie sclae param: σGPD = ψGEV + ξ(u − µ)
I
Extremes tend to occur in clusters - violates independence assumptions
Extremes
Extreme Value Analysis
POT Analysis of Dow Jones
0.02 0.00
0.01
Mean Excess
0.03
How to chose threshold?
0.00
0.01
0.02
0.03
0.04
0.05
0.06
u
Mean Residual Life Plot mrl.plot(djd.neg, umin=0, umax=max(djd.neg)-0.01) Extremes
Extreme Value Analysis
Threshold: Theory
Peaks Over Thresholds (POT) If the threshold is chosen to be “high enough” I
the parameters will stabilize
I
parameter estimations are equivalent to the GEV parameters, after a linear transformation of the scale parameter, (and allowing for sample estimation effects).
I
ξGPD = ξGEV
I
return value estimates are robust and asymptotically equivalent to GEV return value estimates
Bias-Variance trade-off:
Extremes
Extreme Value Analysis
POT Analysis of Dow Jones Thresholds: 0.015, 0.022
0.8
0.07
1.0
0.07
0.8
0.06
●
● ●
●
1.0
0.02
●
●
●
●● ●
0.03
● ●
0.0
●
0.04
0.05
0.06
●
0.0
0.2
● ●
●
0.03
0.2
● ● ● ● ● ●
●
0.04
●
● ●● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
●
●
Empirical
0.6
Model
● ● ● ●
0.4
●
0.05
0.06 Empirical
0.03 0.6
●
●
● ●
0.4
0.6
0.8
● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
1.0
●● ●
0.03
0.04
0.05
Empirical
Model
Empirical
Model
Return Level Plot
Density Plot
Return Level Plot
Density Plot
150
1.0
50
0.5
f(x)
100
Return level
100 f(x)
0.1
50
0.2
Return level
0.3
150
0.4
0.06
1.5
0.0
● ●●●
0.4
● ●
●
●
0.02
0.2
● ●● ● ● ●● ●● ● ●● ● ●●● ●●● ●●● ● ●●●
0.2
Quantile Plot ● ● ●
● ●
●● ●●● ●● ●● ●● ● ●●● ●●
0.0
Probability Plot ●
●
0.05
0.8 0.6
●●● ● ●●●●● ●●● ●● ● ●●
0.4
Model
Quantile Plot ●●● ● ● ● ●● ●● ●● ●● ●● ●●
0.04
1.0
Probability Plot
●
0.0
0 1e+03
Return period (years)
0.01
0.03
0.05
0.07
x
●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
1e−01
● ●
●
1e+01
0
●
1e+01
0.0
● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
1e−01
1e+03
Return period (years)
0.02
0.04
0.06 x
djd.gpdMIN15 0 for each i (if these constraints are violated, L is automatically set to 0).
Extremes
Extreme Value Analysis
Point Process
I
The basic method of estimation is therefore to choose the parameters (µ, ψ, ξ) to maximize the negative log-likelihood. This is performed by numerical nonlinear optimization.
I
In practice it is convenient to replace (µ, ψ, ξ) by (θ1 , θ2 , θ3 ) where θ1 = µ, θ2 = log ψ, θ3 = ξ (defining θ2 to be log ψ rather than ψ itself makes the algorithm more numerically stable, and has the advantage that we don’t have to build the constraint ψ > 0 explicitly into the optimization procedure).
Extremes
Extreme Value Analysis
Point Process
Provided the model fits the data: I
PP approach should produce equivalent parameter values as the POT approach
I
Parameters are independent of the threshold (adjusting for estimation error)
I
Ideal threshold determined by considering where the parameter values stabilize
Extremes
Extreme Value Analysis
PP Analysis of Dow Jones
Quantile Plot
0.0
0.2
0.4
0.6
0.8
1.0
empirical
●
0.06
● ●
0.04
empirical
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
0.02
model
0.0 0.2 0.4 0.6 0.8 1.0
Probability plot
● ● ● ● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
0.02
0.03
0.04
0.05
0.06
model
djd.ppMIN 0, y > 0
Extremes
Extreme Value Analysis
Bivariate Extremes - Block Maxima
1
Z V (x, y ) = 2
max( 0
w 1−w , dH(w ) y y
where H is a distribution function on [0, 1] satisfying the mean constraint: 1
Z
wdH(w ) = 1/2 0
If H is differentiable with density h, then Z V (x, y ) = 2
1
max( 0
w 1−w , h(w )dw y y
This defines the class of bivariate extreme value distributions
Extremes
Extreme Value Analysis
Bivariate Extremes - Block Maxima
For any constant a > 0, V (a−1 x, y −1 ) = aV (x, y ) Thus, G n (x, y ) = (exp{−V (x, y )})n = exp{−nV (x, y )} = G (n−1 x, n−1 y )
So if (X , Y ) has distribution function G , then Mn also does, with rescaling n−1
Extremes
Extreme Value Analysis
Bivariate Extremes - Block Maxima Theory was derived using standard unit Fr´echet. Note that this can be generalized to the complete class of bivariate extreme value distributions by generalizing the marginal distributions:
x − µx σx
1 ξx
y − µy σy
1 ξy
x˜ = 1 + ξx and y˜ = 1 + ξy with distribution function:
G (x, y ) = exp{−V (˜ x , y˜ )}
Extremes
Extreme Value Analysis
Bivariate Extremes - Block Maxima H does not have to be differentiable. For example, when H places mass 0.5 on w = 0 and w = 1, V (x, y ) = x −1 + y −1 and the corresponding bivariate extreme value distribution is: G (x, y ) = exp{−(x −1 + y −1 )} This corresponds to independent variables. When H places mass 1 on w = 0.5: G (x, y ) = exp{−(max(x −1 , y −1 )} This corresponds to dependent variables.
Extremes
Extreme Value Analysis
Bivariate Extremes - Block Maxima Families for which the mean is parameter-free and the V integral is tractable define dependence classes. I
Logistic Family:
I
Bi-logistic Family:
I
Dirichlet model:
Extremes
Extreme Value Analysis
References
I
Coles, 2001. An Introduction to Statistical Modeling of Extremes.
I
Cooley, D. 2012 course notes from ENVR short course on uni- and multi-variate extremes
I
R-packages: ismev, evd, evir, SpatialExtremes, extRemes
I
Richard Smith spatemp notes http://www.stat.unc.edu/faculty/rs/s321/spatemp.pdf
I
http://cran.r-project.org/web/packages/SpatioTemporal/ vignettes/Tutorial.pdf
Spatial Extremes
Extremes
Extreme Value Analysis