NCL: Basic Extreme Value Statistics

NCL has a small number of basic extreme value (EV) and recurrence statistical functions. However, NCL is not R or S+ or Matlab or IDL or Excel or Python's SciPy. These tools contain many more EV related functions. That said: none of these tools is NCL either!

The Gilleland et al reference below provides an overview of selected EV software.

Extreme value theory (EVT) is a branch of statistics dealing with the extreme deviations from the median of probability distributions. There exists a well elaborated statistical theory for extreme values. It applies to (almost) all (univariate) extremal problems. From EVT, extremes from a very large domain of stochastic processes follow one of the three distribution types: Gumbel, Frechet/Pareto, or Weibull.

The generalized extreme value (GEV) distribution is a family of continuous probability distributions developed within EVT. The GEV combines three distributions into a single framework. The distributions are:

• Type I: Gumbel
• Type II:Frechet; (aka: Pareto; Fat-tailed Pareto)
• Type III:Weibull

The GEV allows for a continuous range of possible shapes. The shape parameter, S, governs the tail behavior of the distribution. The sub-families defined by, S ~ 0, S > 0 and S < 0 correspond, respectively, to the Gumbel, Frechet and Weibull families. Note the differences in the ranges of interest for the three extreme value distributions: Gumbel is unlimited, Frechet has a lower limit, while the reversed Weibull has an upper limit.

The GEV facilitates making decisions on which distribution is appropriate. The GEV distribution is often used as an approximation to model the minima or maxima of long (finite) sequences of random variables. In general, the GEV distribution provides better fit than the individual Gumbel, Frehet, and Weibull models. For example, in most hydrological applications, the distribution fitting is via the GEV as this avoids imposing the assumption that the distribution does not have a lower bound (as required by the Frechet distribution).

From Wikipedia:

The maximum value (or last order statistic) in a sample of a random variable following an exponential distribution approaches the Gumbel distribution closer with increasing sample size.[4] In hydrology, therefore, the Gumbel distribution is used to analyze such variables as monthly and annual maximum values of daily rainfall and river discharge volumes,[5] and also to describe droughts.

Most commonly, extreme value analyses focus on high extremes: maxima. It should be noted that dealing with minima follows the same approaches except for a sign reversal:

                               min[x(i)] = -max[-x(i)]

From Wikipedia:

Two approaches exist for practical extreme value analysis. The first method relies on deriving block maxima (minima) series as a preliminary step. In many situations it is customary and convenient to extract the annual maxima (minima), generating an "Annual Maxima Series" (AMS). The second method relies on extracting, from a continuous record, the peak values reached for any period during which values exceed a certain threshold (falls below a certain threshold). This method is generally referred to as the "Peak Over Threshold" [1] method (POT) and can lead to several or no values being extracted in any given year

Example 6 (below) offers a sample block maxima approach based on extracting the maximum rate of river flow ineach year.

An estimate of the likelihood of an event, such as a flood or heat wave, is commonly called the return period. However, other terms are used, including: repeat interval, recurrence interval, exceedance probability, expected frequency and return interval. These terms are synonymous.

General References:

   Alexandersson, H. et al (2001): Extreme Value Analysis in Nordic countries: 
                                   Pilot Studies of minimum temperatue and maximum daily precipitation
                                   and a review of methods in use 

   Coles, S. (2001): An Introduction to Statistical Modeling of Extreme Values. 
   Springer ISBN 978-1-4471-3675-0

   Coles, S. and A. Davison (2008):  Statistical Modelling of Extreme Values

   Gilleland, E. et al (2013): A software review for extreme value analysis 
   Extremes 16:103-119
   DOI 10.1007/s10687-012-0155-0

   Gilleland, E. and Katz, R.W. (2014): extRemes 2.0:  An Extreme Value Analysis Package R

   Gong, S. (2012): Estimation of hot and cold spells with extreme value theory. 
   U.U.D.M. Project Report 2012:19 (Uppsala Universitet). 

   Katz, R. (???): Introduction to Statistical Theory of Extreme Values 

   Katz, R. et al (2002): 
Statistics of Extremes in Hydrology. Advances in Water Resources: 25: 1287–1304.

   NASA: Generalized Extreme Value Distribution and Calculation of Return Value

   Rieder, H.E. (2014):  Extreme Value Theory: A primer.  
   Lamont Doherty Earth Observatory.


   Wikipedia: (a) Extreme Value Theory
              (b) Generalized Extreme Value Distribution
   
   Wilks, D. (2006): Statistical Methods in the Atmospheric Sciences. Academic Press.  

NCL (6.4/5.0) currently has several GEV statistical functions. Currently, the only ones documented are:

• extval_mlegev
• extval_mlegam
• extval_mlegev
• extval_gev
• extval_gumbel
• extval_frechet
• extval_pareto
• extval_weibull
• extval_recurrence_table
• extval_return_period
• extval_return_prob
• others may be added

Several of the above functions use the following parameters:

• shape - affects the distribution shape [ ;-) ]
• scale - stretches or shrinks the distribution
• center, location - shifts the distribution

The extval_mlegev and extval_mlegam provide maximum-liklihood estimates of these parameters for the GEV and Gamma distributions, respectively. The method of moments can readily be used to derive parameter estimates for other extreme value distributions. However, the moment method ahs some issues (biases).

     pdfcdf = extval_gev(x, shape, scale, center, 0)   ; [/ PDF, CDF /]

    pdfcdf = extval_gumbel(x, scale, center, 0)   ; [/ PDF, CDF /]

    pdfcdf = extval_frechet(x, shape, scale, center, 0)   ; [/ PDF, CDF /]

    pdfcdf = extval_weibull(x, shape, scale, center, 0)   ; [/ PDF, CDF /]

    pdfcdf = extval_pareto(x, shape, scale, center, ptype, 0)   ; [/ PDF, CDF /]