Extreme Value Theory for Time Series using Peak-Over-Threshold method

Gianluca Rosso
GradStat – RSS, the Royal Statistical Society, London
Correspondent Researcher – SIS, Società Italiana di Statistica, Roma
gianluca.rosso@sis-statistica.org

ABSTRACT: this brief paper summarize the chances offered by the Peak-Over-Threshold method, related with
analysis of extremes. Identification of appropriate Value at Risk can be solved by fitting data with a Generalized
Pareto Distribution. Also an estimation of value for the Expected Shortfall can be useful, and the application of
these few concepts are valid for the most wide range of risk analysis, from the financial application to the
operational risk assessment, through the analysis for climate time series; resolving the problem of borderline data.

In Extreme Value Theory (EVT) the method defined as

Peak-Over-Threshold (POT) is common used. It is also
commonplace to calculate its Value At Risk (VaR) and
Expected Shortfall (ES) fitting data with a Generalized
Pareto Distribution (GPD). This method is well known in
financial risk analysis, but it is in general applicable to all
kind of risk analysis (for example in climate time series; see
Rick Katz, Extreme Value Analysis for Climate Time
Series, Institute for Mathematics Applied to Geosciences
National Center for Atmospheric Research Boulder, CO
USA).
We know that the Normal Distribution is assumed as a
standard for the most of analysis. The Normal Distribution
has two parameters: mean and standard deviation.

= ( ,   ) Fig. 1
(1)
where When it is not easy to assume that the owned data could

1
fitting a Normal Distribution, the way is to use a LogNormal

 =  

Distribution. This function transform not-normal distribution
into normal. Its shape is not symmetrical, therefore its
 inflection point and mean are not corresponding:

Inflection point =   
(2)


1
  = (  − )

(6)


1
 =   
(3)

Is assumed that the distribution in symmetrical, and that the 
mean divide the distribution into two symmetrical parts. The (7)

1
top of the distribution (corresponding with the mean) is
 = (   − )


defined by

 

 √
Inflection point = (8)
(4)
The height of the density at any value x is given by

1
The height of the density at any value x is given by
 !  
1     ( ) =     

>0
( ) =       √2
√2 (9)
(5)

Extreme Value Theory for Time Series using Peak-Over-Threshold method - Gianluca Rosso (2015)
1
 97:∝ (=) = =(∝)
(17)

97:∝ (=) = inf A ∈ ℝ ∶ D(= > ) ≤ 1−∝F

= inf A ∈ ℝ ∶ GH ( ) ≥ ∝F ∝ ∈ (0,1)

(18)

This means that the 97:∝ is given by the smallest value

such that the actual loss X exceeds with the probability
Fig. 2 1−∝ , at most.

more between the given ∝ and ∞ , but (probabilistically

Note that in this case we have the Worst Case that lies not

Looking at he figure above, we can define speaking) between ∝ and 1.



0 −  ∶ % ( )& ( )*+& ,-.. ((,)

'

0

 − / ∶ % ( )& 1 )*+& ,-.. (1,)



/ − ∞ ∶ % ( )& 4-5.+ 67. (46)

0

(10, 11, 12) Fig. 3

The value of ∝ is decided in according with the attended

-
So, our Worst Case will be

regulators. For example a very precuationally value for ∝

results of the analysis, or in according with rules defined by
46 = 100 (1−∝)%

Once decided the value for ∝, also the value for 97:∝ will
could be at 99.5%. (19)

be strictly linked to. The density function between 0 and ∝ and (<∝ will be
represents the VaR

1
97: = (, + 1, (<∝ (=) = % 97:K (=) &L
1−∝
(13) ∝

97:∝ = ∝
(20)

(14) The model will be defined by the function

( ) = (  ,  , … ,  |O)
= (  |O) ∙ (  |O) ∙ … ∙ (  |O)
The Expected Shortfall (ES) is

3
1
(<∝ = % ( )&
∝
(21)
∝
(15) where O are the parameters of the model.
As said above
and therefore is


46 = ∝ (<∝ ( ) = Q (  | O)
(16) 
(22)
So, as 97:∝ = ∝ it is also true that
in log-normality is

Extreme Value Theory for Time Series using Peak-Over-Threshold method - Gianluca Rosso (2015)
2
 Gc (S) ≈ UV,W (S)


( ) =  (  | O)
(28)

 it is that
(23)
G( ) = e1 − G(T)f ⋅ UV,W ( − T) + G(T) -5 > T
Now, we can consider that one of the best way to analyze
the peak of our time series is the POT method. About the (29)
method we suggest to refer to the very large literature
where
written during last years. We focus now to the analysis via

−h
Let = , R = 1, … ,
our data set (or sample), and let u the G(T) =
GPD and the possible way to estimate VaR and ES.

value of the threshold, so that (30)

S= −T given N the number of observations and m the number of

(24) exceedances.
So the estimation is


]k ( − T) Vm

h
i
G( ) = 1 − j1 + l

^k
(31)

We know that observations are denoted by , so for and

97:/ is

= 97:/
(32)


h ]k e97:
i/ − Tf Wo
Fig. 4

n = 1 − j1 + l
+

^k
The GPD is defined by
(33)

[ ] V ^k
Vm
Y1 − \1 + _ R ] ≠ 0 i/ = T +
97: pq (1 − n)r − 1s
UV,W ( ) = ^ ]k h
Z
Y 1 −  ) \− _ R ] = 0
X ^
(34)
(25)
About the Worst Case, and reguarding what said above, it is
possibile to define the estimation for ES
and are valid

^ > 0 7& ≥ 0 → ] ≥ 0 46 = ∝ (<∝

(35)
^
0 ≤ ≤ − → ] < 0 1

i∝ ^k − ]k
97:
] t∝ = iu &S =
(< % 97: +
1−∝ 1 − ]k 1 − ]k
∝
(26)

where ] is the shape parameter, and ^ is the scale parameter.

(36)

So, we need to put our exceedances in relation with the GPD where

G( + T) − G(T)
Gc ( ) = D(= − T ≤ | = > T) = ]k , ^k
1 − G(T)
= estimated GPD parameters
N = number of observations
(27) m = number of exceedances
u = value of the threshold
since

Extreme Value Theory for Time Series using Peak-Over-Threshold method - Gianluca Rosso (2015)
3
Brief conclusions.
The POT method assumes that once we have placed a
threshold, all the items over this level are peaks. But as I
wrote in a paper about the clustering (Rosso, 2014, Outliers
Emphasis on Cluster Analysis - The use of squared
Euclidean distance and fuzzy clustering to detect outliers in
a dataset, arXiv:1403.5417), why items very close between
them must be attributed to a cluster instead to another one?
The conclusion of that paper was that we must have a
method that runs more close with the fuzzy thinking of our
brain. The solution was: fuzzy clustering, in first instance.
After that, possibilistic clustering could be a substantial help
to manage data with a more flexible vision.
Take a look at Figure 4, again, here down as Figure 5. We
should consider the item number 1 a peak, and the item
although they are so close in terms of value of
.
number 2 should be considered as a standard data. This is
+ in the Tail but the Poison Depends on the Dose.
Fig. 5
It will be absolutely more easy to consider a plot as in
Figure 6.
Fig. 6
Surely in this situation we have no dubt about peaks. We
have a large blue zone without data, virtually an extremely
wide threshold. But in plot in Figure 5 we must mitigate the
euclidean distance between the threshold line and each
single item, or event. We need an average value, but thought
Extreme Value Theory for Time Series using Peak-Over-Threshold method - Gianluca Rosso (2015)
in term of risk because we must remember that our analysis
is for the value at risk. So, the applications of the rules
described above can conduct to an assessment numerically
substainable, and surely consistent in its reliability, despite
borderline cases.
______________________________
REFERENCES.
– McNeil, A. (1997). Estimating the tails of loss severity
distributions using extreme value theory. Astin Bulletin,
27:117-137.
– McNeil, A. (1999). Extreme value theory for risk
managers. In Internal Modelling and CAD II, pages 93-
113. Risk Waters Group, London.
– McNeil, A. and Frey, R. (2000). Estimation of tail-
related risk measures for heteroscedastic financial time
series: an extreme value approach. J. Empirical Finance,
7:271-300.
– McNeil, A., Frey, R., and Embrechts, P. (2005).
Quantitative Risk Management: Concepts, Techniques
and Tools. Princeton University Press, Princeton.
– Mikosch, T. (2003). Modeling dependence and tails of
financial time series. In Finkenstadt, B. and Rootz_en,
H., editors, Extreme Values in Finance,
Telecommunications, and the Environment. Chapman &
Hall, London.
– Rydell, S. (2013). The use of extreme value theory and
time series analysis to estimate risk measures for extreme
events, Umea University.
– Jobst, A. A. (2007). Operational Risk—The Sting is Still
International Monetary Fund.
– Katz, R. (2008). Extreme value analysis for climate time
series. Institute for Mathematics Applied to Geosciences
National Center for Atmospheric Research Boulder, CO
USA .
– Katz, R.W. Brush, G.S. Parlange, M.B. (2005). Statistical
of extremes: modelling ecological disturbances. by the
Ecological Society of America. Ecology, 86(5), 2005, pp.
1124–1134.
– Towler, E. Rajagopalan, B. Gilleland, E. Summers, R.S.
Yates, D. Katz, R.W. (2010). Modeling hydrologic and
water quality extremes in a changing climate: A
statistical approach based on extreme value theory.
WATER RESOURCES RESEARCH, VOL. 46,
W11504, doi:10.1029/2009WR008876.
– Rootzen, H. Katz, R.W. (2013). Design Life Level:
Quantifying risk in a changing climate. WATER
RESOURCES RESEARCH, VOL. 49, 1–9,
doi:10.1002/wrcr.20425.
– Katz, R.W. Grotjahn, R. (2014). Statistical methods for
relating temperature extremes to Large-Scale
Meteorological Patterns. US CLIVAR VARIATIONS •
Winter 2014 • Vol. 12, No. 1.
– Cheng, L. AghaKouchak, A. Gilleland, E. Katz, R.W.
(2014). Non-stationary extreme value analysis in a
changing climate. Climatic Change (2014) 127:353–369,
DOI 10.1007/s10584-014-1254-5.
4