International Journal of Statistics and Applied Mathematics 2018; 3(5): 20-27
ISSN: 2456-1452
Maths 2018; 3(5): 20-27
© 2018 Stats & Maths
www.mathsjournal.com
Received: 03-07-2018
Accepted: 04-08-2018
Olanrewaju O Rasaki
Department of Statistics,
University of Ibadan, Ibadan,
Nigeria
Oseni Ezekiel
Department of Banking and
Finance, University of Lagos,
Lagos, Nigeria
Adekola Lanrewaju Olumide
Department of Physical Sciences,
Bells University of Technology,
Ota, Nigeria
Oyinloye Adedeji Adigun
Department of Statistics,
University of Ibadan, Ibadan,
Nigeria
On skew generalized extreme value-ARMA model: An
application to average monthly temperature (19012016) in Nigeria
Olanrewaju O Rasaki, Oseni Ezekiel, Adekola Lanrewaju Olumide and
Oyinloye Adedeji Adigun
Abstract
This study describes the approach for modeling extreme and lengthy time-varying series of an
Autoregressive Moving Average of order ( p, q) via a Skew Generalized Extreme Value distribution as
the white noise. This approach establishes the procedure for parameters’ estimation and their standard
errors for the SGEV-ARMA ( p, q) model via the iterative Fisher information scores derived from the
Maximum Likelihood Estimation for a chosen optimal degree of flexibility (bandwidth) " " . The study
was applied to a lengthy series of average monthly temperature (report in oC) of Lagos, Nigeria from
January 1901 to December 2016 with 1381 data points. It was noted that SGEV-ARMA (3,3) recorded a
subjacent model performance error via the evaluated indexes of AIC, BIC and HQIC (103.02, 141.35 &
124.50) respectively compare to an intensive error performance in the white noise Gaussian-ARMA (3,
3) with (108, 144.4 & 129.26) respectively. In addition, the forecast error indexes with the SGEV
subjected white noise were miniaturized compared to the Gaussian white noise.
Keywords: Autoregressive moving average, bandwidth, maximum likelihood estimation, skew
generalized extreme value, temperature, and white noise
Correspondence
Olanrewaju O Rasaki
Department of Statistics,
University of Ibadan, Ibadan,
Nigeria
1. Introduction
The origin of extreme value theory started its course by Gnedenko (1943) [10] when it was used
to study the maxima series of Gaussian subjected variables under general hypothesis of
limiting distribution called Generalized Extreme Value (GEV) distribution for series of
extreme (s), lengthy series, lengthy observations, ecological observations, climate observations
etc. Its course was extended when an unusual or usual event takes place regardless of whether
or not it is catastrophic or when an event causes catastrophes (Farago and Katz, 1990 [6];
Faranda et al. 2012 [7].). However, its development could be traced back to the work done by
Bernoulli in 1975[3]. Kotz and Nadarajah (2000) [11] and its first application was made by Fuller
in 1914 [8]. It is based on large deviations from the median of probability distributions such
that the theory assesses the type of probability distribution generated by processes.
Rieders (2014) [17] affirmed that limiting distributions (which are distinct from the normal
distribution) are the Extreme Value Distributions (EVD) for maximum, minimum or extreme
lengthy contaminated series or collection of observations of either dependent or covariates
random variable (s). It is widely used in modeling phenomena in disciplines, such as structural
hydrology, meteorology, engineering, finance, earth sciences, traffic prediction and risk
management. Estimates of extreme precipitation are consistent in forecasting planning
infrastructures such as dams flood frequency etc. Engineers often need such statistics for the
design of structures for flood protection using Areal Reduction Factors (ARFs) to convert
quantiles for point rainfall to the corresponding quantiles of areal rainfall. ARFs have been
derived empirically by estimating the areal rainfall as a function of point rainfall
measurements e.g. Natural Environment Research Council (NERC) (1975) [13]; Bell (1976) [2];
or by statistical modeling (Bacchi and Ranzi, 1996) [1].
~20~
International Journal of Statistics and Applied Mathematics
Three approaches had been in existence for the practical applications extreme value- the first method evolves deriving block
maxima (that is maxima/minima) series as a preliminary step. The approach relies partly on the results of the Fisher–Tippett–
Gnedenko theorem, leading to the Generalized Extreme Value Distribution (GEVD) being selected for fitting. The second method
relies on extracting "Peak over Threshold" (POT) that is peak values above or below a certain threshold from a continuous record
while the third approach tries to strike the balance between the block maxima and peak over threshold approaches via r th-largest
order statistics approach (Ragulina and Reitan, 2017) [16].
Chavez-Demoulin and Davison (2012) [5] superposed monthly maximum river flow at the station Muota-Ingenbohl in Switzerland,
for the years 1923-2008 by fitting a nonparametric GEV with time dependent location parameter where " " plays the same role
as the bandwidth in local likelihood estimation to the random effects model. In addition, Laurini & Pauli (2009) [12], applied it
using Bayesian computational tools. Ning and Bloomfield (2017) [4] water flow dataset is collected from French Broad River at
Asheville in North Carolina. The datasets contain annual maximum water flow level from 1941 to 2009 and used the Dependent
Generalized Extreme Value (DGEV) model as the white noise of the Autoregressive model to explain the water flow phenomena.
Hence, this article subjected the Skew Generalized Extreme Value (SGEV) distribution as the stochastic error distribution of
Autoregressive Moving Average (ARMA) as a deterministic time varying model for trend or non-trend effect and used the
iterative Fisher information scores via the maximum likelihood method of estimation in estimating their standard errors for a
chosen optima degree of flexibility (bandwidth) " " .
2. Specification of the skew generalized extreme value- autoregressive moving average (SGEV-ARMA)
A stochastic process of yt is said to follow an Autoregressive Moving Average (ARMA) of order ( p, q) if it satisfies
yt 0 i yt i i t i
p
q
i 1
i 0
1
that is independently and identically distributed with mean zero and variance . Replacing and forgoing the standard Gaussian
2
of zero mean and variance
E ( yt )
2 of t for Skew Generalized Extreme Value (SGEV) with mean and variance
2
1 1 (1 ) ; 2 v ( yt )
(1 2 ) 2 (1 ) Where, is the Gamma
2
2
1
function with identity ( )
x
e x , for > 0 the location parameter of GEV; the scale parameter; the
1 x
shape parameter; the degree of flexibility (bandwidth) measure
0
With
0 1 in equation (1), yt will be stable and equals
p
q
1
L
y
i t i
i i t
0
i 0
i 1
2
So, equation (1) becomes,
yt 0 i yt i i t i
p
q
i 1
i 0
3
SGEV ARMA( p, q)
That is, SGEV-ARMA ( p, q) . Ribereau et al. (2011) [15] defined the Probability Density Function (PDF) of a random variable
( yt ) for Skew Generalized Extreme Value (SGEV) distribution to be
f ( yt ) ( 1) g ( yt )G ( yt )
4
distribution respectively, where must be strictly greater than -1.
Where g ( yt ) and G( yt ) are the PDF and Cumulative Distribution Function (CDF) of the Generalized Extreme Value (GEV)
G( yt ) e
1
1
yt
5
~21~
International Journal of Statistics and Applied Mathematics
g ( , , ; yt )
1
1
yt
1
1
1
yt
exp 1
yt
1
So, f ( , ; yt )
1
Such that, 1
yt
1
1
6
1
y
exp ( 1) 1 t
7
0 ; 0 ; y
i yt i i t i
3. Parameter Estimation
Given that yt 0
p
q
i 1
i 0
p
q
y
i t i
i t i
1
i 1
i 0
f ( , , , ; y )
1
i
i
t
The log-likelihood function,
( E ( yt ), v( yt )) with order ( p, q) so,
follows SGEV
1
p
q
1
y
i t i
i t i
i 1
i 0
exp ( 1) 1
1
1
8
L() log L(i ,i , , ; yt ) n log( 1) n log 1
i yt i i t i
i 0
log (1 ) i 1
i 1
p
n
q
i yt i i t i
n
i 0
1 (1 ) i 1
i 1
p
q
1
9
Where
i ,i , ,; yt
p
q
yt2ii i t i
1
i 1
i 0
1
1
1 (1 )
q
p 2
i 1
y i t i
n t i i
L ()
1 i 1
i 0
p
q
i
i 1
i yt i i t i
1 n
i 1
i 0
1
log
(1
)
i 1
n
q
p 2
y
i t i
n t i i
i 0
(1 ) i 1
i 1
p
q
i yt i i t i
n
i 0
log (1 ) i 1
i 1
q
p 2
y
i t i
n t i i
i 1
i 0
1
i 1
~22~
1
10
1
11
International Journal of Statistics and Applied Mathematics
p y q 2
i
n i t i
i 0
(1 ) i 1
p y ( y ) q 2
i 1
t i i t i i
L ()
1 n i 1
i 0
p
q
i
i 1
i yt i i t i
n
i 1
i 0
log (1 )
i 1
t i
t i
L()
n 1 n
1
1 log
i 1
1 1
1
1
1
1
12
13
1
q
1 ( 1) n p
L() 1
2 log
y
y
(
)
t i i t i i t i (i t i )
3
i 1 i 1
i 0
p
q
2
p y q
i 1
yt ii i t i
i t i
n t i i
L ()
i 1
2
i 0
i 0
(1 ) i 1
p
q
2
i
i 1
n
i yt i i t i
log (1 ) i 1
i 0
i 1
14
n
p
q
i t i
y
i t i
n
i
i
1
0
log (1 )
i 1
q
p 3
y
n t i i i t i
i 0
i 1
i 1
q
p 3
y
i t i
t i i
1 n
i 1
i 0
1 (1 )
i 1
1
p
q
2
p y q
i 1
yt ii i t i
i t i
n t i i
L ()
i 1
2
i 0
i 0
(1 ) i 1
p
q
2
i
i 1
n
i yt i i t i
log (1 ) i 1
i 0
i 1
n
p
q
y
i t i
i t i
n
i
i
1
0
log (1 )
i 1
q
p
3
y
n t i i i t i
i 0
i 1
i 1
~23~
15
International Journal of Statistics and Applied Mathematics
q
p
y
i3 t i
i
i
n
1
i 0
1 (1 ) i 1
i 1
1
16
n
1
1
2
2
L() n
1
1
i 1
1
2
2
2
1 n
1
1
log
i 1
q
1 ( 1) n p
L()
1
1
2
log
y
y
3
(
)
t i i t i i t i (i t i )
2
3
3
4
i 1 i 1
i 0
17
18
The Hessian matrix of the parameter space,
2 L ( )
2
i
2 L ( )
i2
2
H () L ()
2 L ( )
2
2 L ( )
2
19
H () is a block diagonal matrix and a square matrix of n by n dimension depending on the order of the SGEVARMA, that is the order of ( p, q) .
Such that the
The Fisher Scoring algorithm,
m 1
L()
/
E H () / , , ,
i i
i ,i , ,
1
m
m1 m I n( m ) S ( m )
1
(m)
20
S ( m ) are the Fisher information and Score matrixes respectively to be evaluated by via Newton-Raphson
iterative procedure for a chosen value of the degree of flexibility measures ( ) that must be strictly greater than one.
Where I n
and
3.1 Model identification criteria
Model selection via information criteria is being defined as
criteria(m) (Maximized likelihood) + f (n, m)
21
However, Tsay (2016) [18]defined the information criteria in terms of the Akaike’s Information Criteria (AIC), Bayesian’s
Information Criteria (BIC) and Hannan and Quinn Information Criteria (HQIC). The criteria defined will be extended to by
substituting the maximum likelihood of the SGEV-ARMA to the residual variance needed.
4. Analysis and discussion of results
The average monthly series of temperature of Lagos, Nigeria from January 1901 to December 2016 was used. Observations are
reported in Degree Celsius (oC), recorded on monthly bases starting from the inception of meteorological section under the
ministry of environment. The models’ estimation and Exploratory Data Analysis utilized One thousand three hundred and eighty
one (1381) data points of the average monthly temperature. All the data points maintained the same unit of measurement of oC.
~24~
International Journal of Statistics and Applied Mathematics
4.1 Preliminary and descriptive analysis
Fig 1: Time Plot of the Temperature Series from 1901 to 2016.
From the visualization above, it is noted that the average monthly temperature ranges approximately from 21 to 32 oC throughout
the stipulated period. It can also be deduced that the recorded oCs maintained a near steady and constant trend around the 90 to
100 percentiles of the range. The minimum oC was recorded within the first quarter of the year 1920 with the trend around 10 to
20 percentiles not really steady and constant.
Table 1: Descriptive statistics and stationary test of the temperature report.
Mean
Min.
Max.
Standard Deviation
skewness
kurtosis
Augmented Dickey-Fuller Test
Phillips-Perron Unit Root Test
KPSS Test
Box-Pierce test
Cox-Stuart trend test
Estimates
26.91
21.90
31.57
1.8075
3.3803
-0.6504
-16.666
-14.747
0.5154
592.44
370
P-values
------------------------0.01
0.01
0.0382
0.0023
0.0086
Table.1 represents the descriptive measurements for the meteorological temperature. The oC clustered around 26.91 with a
collaborated maximum oC of 21.90 and minimum oC of 31.57. The recorded meteorological oCs was affected by extreme values
(possibly lower outlier) via an indication by the estimated value (3.3803) of the skewness that is strictly greater than three. The
effect of the estimated kurtosis as shown in figure. 2 led to the unusual peakedness or flatness of the graph of the frequency
distribution of the meteorological data, especially with respect to the concentration of values near the mean as compared with the
normal distribution. The Augmented Dickey- Fuller, Phillips-Perron Unit Root and Kwiatkowski-Phillips-Schmidt-Shin (KPSS)
tests with P-values (0.01, 0.01 and 0.0382) respectively suggested and indicated the probability of the meteorological data having
a unit root; being non-stationary are 0.01, 0.01 and 0.0382 respectively, so the tests tell that there is a very high probability that
the data is stationary. Similarly, the Box-Pierce and Cox-Stuart trend tests for invertible and trend effect of the data with P-values
(0.0023 and 0.0086) respectively betokens the invertible of the Moving Average embedded coupled with constant trend.
Fig 2: Kernel density plot of the temperature series.
~25~
International Journal of Statistics and Applied Mathematics
The kernel density plot confirmed the unusual emerged of the fat-tailed of the frequency distribution of the weather report
meteorological distributed. It is nothing but merged mesokurtic distributions.
4.2 Model estimation and analysis
Table 2: Minimum lag (order) selection and white noise test.
MA0
5460.5
4682.3
4030.6
MA1
4567.7
4322.5
4031.8
MA2
4177.5
4142.6
3994.4
MA3
4106.1
4105.8
3965.4
AR3
4031.4
4015.6
3995.5
3961.0
Check
Non-White-Noise
AR0
AR1
AR2
lag
1
2
3
4
5
LB
594
605
746
929
989
P-value
0.0000
0.0000
0.0000
0.0000
0.0000
6
1006
0.0000
Table 2 presents the Minimum and optima lag selection for the weather series. An alternate technique from the exponentiation
decaying of the Partial Autoregressive Correlation Function in describing the appropriate order for ARMA model is the minimum
selection of the optima order selection of the Autoregressive (AR) cross sectioning with the Moving Average (MA). The ideal
order for exponentiated decaying series of the weather series was at (3, 3). Furthermore, table. 2 unveiled the non-white noise test
for the series from lag one to lag six, the hypotheses being the stated as “Non-white noise series” were accepted from lag one to
lag six where it cut-off.
Table 3: Coefficients of the Gaussian-ARMA (3, 3) and SGEV-ARMA (3, 3)
1
2
3
1
2
3
Gaussian-ARMA (3, 3)
Estimates
SD
Z-ratio
1
2
3
1
2
3
P-Value
0.36058
0.1386
9.8184
0.0000
-0.85391
0.1653
-5.1664
0.0099
0.1849
0.10051
1.8401
0.0078
0.3275
0.1360
2.4065
0.0621
-0.0138
0.04247
-0.3270
0.0000
0.2685
0.0369
7.2724
0.0000
Log-likelihood = -46.98, AIC = 108, BIC= 144.4, HQIC=129.26.
RMSE=19.7837; MAPE=23.147; MPE=24.147; MAE=20.9082;
R-.squared=88.8334.
SGEV-ARMA (3, 3)
Estimates
SD
Z-ratio
P-Value
0.7933
0.0624
12.5027
0.0000
-0.5297
0.0089
15.4093
0.0036
0.9442
0.0305
-2.6028
0.0041
0.5209
0.0007
-10.002
0.0000
-0.2821
0.0620
-8.3091
0.0000
0.4893
0.1209
9.8943
0.0000
Log-likelihood = -50.34, AIC = 103.02, BIC= 141.35, HQIC=124.50.
RMSE=17.0088; MAPE=21.818; MPE=21.346; MAE=18.2028; Rsquared= 89.4605.
yt 0.7933yt 1 0.5297yt 2 0.9442yt 3 0.5209 t 1 0.2821 t 2
0.4893 t 3 SGEV 26.004,1.0392 ;
yt 0.3606 yt 1 0.85391yt 2 0.1849yt 3 0.3275 t 1 0.0138 t 2
0.2685 t 3 GAUSSIAN 26.91,1.8075
Table. 3 present the Skew Generalized Extreme Value- Autoregressive Moving Average; SGEV-ARMA (3, 3) in comparison with
the conventional Gaussian-ARMA (3, 3) in terms of parameterization and indexes. Firstly, is to be noted that the no differencing
in any form was subjected to the series, since it was stationary at raw. The crucial step in an appropriate model performance is the
determination of optimal model via the models with subjacent performance criteria; the AIC, BIC and HQIC for the Gaussian
white-noise were (108, 144.4 and 129.26) respectively at optima selected of (3, 3) compared to SGEV white-noise with lesser
AIC, BIC and HQIC (103.02, 141.35, 124.50) respectively, that best described and captured lesser stochastic error. In addition,
the mean of the two models differs by 0.906, connoting a smaller magnitude in each of the models’ clustering around the mean oC,
but the variation (1.8075) in the SGEV-ARMA (3, 3) model was dinky compared to the Gaussian-ARMA (3, 3). Furthermore, the
evaluation of the computed forecast error indexes of the two models of the weather series report were estimated. The Residual
Mean Squared Error (RSME), Mean Absolute Percentage Error (MAPE), Mean Percentage Error (MPE) and Mean Absolute
Error (MAE) were estimated as (19.7837, 23.147, 24.147, 20.9082, and 88.8334) for the Gaussian white noise ARMA (3, 3)
compared with a miniaturized SGE-ARMA (3, 3) forecast error indexes of the same RSME, MAPE, MPE, and MAE as (17.0088,
21.818, 21.346, 18.2028). The error forecast of the SGEV-ARMA is subjacent to Gaussian-ARMA, suggesting a more robust
evaluation and performance of the incorporated non-white-noise of the SGEV.
~26~
International Journal of Statistics and Applied Mathematics
5. Conclusions
In conclusion, the recorded weather report series in oCs maintained a near steady and constant trend around the 90 to 100
percentiles of the range and minimum oC was recorded within the first quarter of the year 1920 with the trend around 10 to 20
percentiles not really steady and constant. The series was favored with a miniaturized chance of 0.01 of non- stationary. In
addition, the variation absolution via skewness and kurtosis; and model performance was superincumbent in the SGEV whitenoise compression in matchmaking with the conventional Gaussian error term subjected to the ARMA.
6. Reference
1. Bacchi B, Ranzi R. On the derivation of the areal reduction factor of storms. Atmospheric Research. 1996; 42:123-135.
2. Bell FC. The areal reduction factors in rainfall frequency estimation. Institute of Hydrology, Wallingford, 1976, 35.
3. Bernoulli N. De Usu Artis conjectandi in lure. Doctoral Thesis, Basel. In: Die Werke von Jakob Bernoulli. 1975; 3(1):287326.
4. Ning B, Bloomfield P. Bayesian inference for generalized extreme value distribution with Gaussian copula dependence.
Cornell University Library 2017.
ArXiv: 1703.00968v1 [stat. ME].
5. Chavez-Demoulin V, Davison AC. Modeling Time Series Extremes. REVSTAT-Statistical Journal, 2012; 10(1):109-133.
6. Faragó T, Katz RW. Extremes and design values in climatology. World Meteorological Organization WMO-TD. 1990;
386(14):46.
7. Faranda D, Lucarini V, Turchetti G, Vaienti S. Generalized extreme value distribution parameters as dynamical indicators of
stability. International Journal of Bifurcation and Chaos. 2012; 22(11):1793- 6551.
doi.org/10.1142/S0218127412502768.
8. Fuller WE. Flood flows. ASCE Trans. 1914; 77:567-617.
9. Galina R, Trond R. Generalized extreme value shape parameter and its nature for extreme precipitation using long time series
and the Bayesian approach. Hydrological Sciences Journal, 2017; 62(6):863-879.
doi:10.1080/02626667.2016.1260134. 2150-3435.
10. Gnedenko B. Sur la distribution limite du terme maximum d’une série aléatoire. The Annals of Mathematics. 1943; 44:423453.
https://www.jstor.org/stable/1968974
11. Kotz S, Nadarajah S. Extreme Value Distributions: Theory and Applications. London: Imperial College Press, 2000.
12. Laurini F, Pauli F. Smoothing sample extremes: The mixed model approach. Computational Statistics and Data Analysis.
2009; 53:3842-3854.
13. Natural Environment Research Council (NERC). Flood studies report. NERC, London, 1975:2.
14. Padoan SA, Wand MP. Mixed model-based additive models for sample extremes. Statistical Probability Letter. 2008;
78:2850-2858.
15. Ribereau P, Masiello E, Naveau P. Skew generalized extreme value distribution: probability weighted moments estimation
and application to block maxima procedure. Communication in Statistics-Theory and Methods, Taylor & Francis, 2014, 1-25.
https://hal.archives-ouvertes.fr/hal-01018877
16. Ragulina G, Reitan T. Generalized extreme value shape parameter and its nature for extreme precipitation using long time
series and the Bayesian approach. Hydrological Sciences Journal, 2017; 4(1):2150-3435.
doi: 10.1080/02626667.2016.1260134.
17. Rieder HE. Extreme Value Theory: A primer. Lamont-Doherty Earth Observatory, 2014.
18. Tsay RS. Lecture Notes of Bus 41202 (Spring) Analysis of Financial Time Series, 2016.
~27~