QUANTIFICATION OF RISK IN NORWEGIAN STOCKS VIA THE

NORMAL INVERSE GAUSSIAN DISTRIBUTION

ERIK BLVIKEN AND FRED ESPEN BENTH

Abstra t. The quanti ation of risk in Norwegian sto ks via the normal inverse
Gaussian distribution is studied. On seven sto ks quoted on the Norwegian Sto k
Ex hange and one quoted in New York we estimate Value-at-Risk on dierent time
horizons for the tted Gaussian distribution, the normal inverse Gaussian distribu-
tion and a non-parametri model. Results are ompared and dis ussed.

1. Introdu tion
The family of normal inverse Gaussian distributions, onstru ted more than two
de ades ago by Barndor-Nielsen [2℄, has the remarkable property of being losed
under onvolutions, yet at the same time able to portray sto hasti phenomena that
have heavy tails or are strongly skewed. Unlike the gamma familiy it is not onned
to the positive half axis. The model is an obvious andidate for nan ial data, where
experien e di tates that the Gaussian familiy often underestimates random variation,
even after passing to logarithms of, say pri es of sto ks and other se urities. With
the normal inverse Gaussian distributions the nan ial analyst has at its disposal a
model that an exibly be adapted to many dierent shapes while the distribution of
sums of independent random variables are still trivial to ompute.
Appli ations in nan e have been reported in several papers, for example Barndor-
Nielsen [3℄, Rydberg [14℄ and Prause [11℄. The purpose of this arti le is to investigate
this model as a tool to evaluate the un ertainty in future pri es of sto k listed on the
Norwegian Sto k Ex hange in Oslo. For omparison we shall also onsider a Norwe-
gian se urity quoted in New York. Chie y among the questions to be addressed is
how well the normal inverse Gaussian family ts share pri es at one of the smaller
nan ial ommunities of the world where trading is relatively low and where the ow
of information may dier in nature from what it is at larger nan ial entra? How
mu h bias is introdu ed by postulating that the day to day u tuations follow a nor-
mal inverse Gaussian distribution, whi h an not possibly be orre t? How mu h data
are needed to estimate the parameters of normal inverse Gaussian models suÆ iently
a urate for Value-at-risk evaluations?
Date : February 16, 2000.
1
2 BLVIKEN AND BENTH
Our framework are perfe t markets where the hanges from one day to the next
are independent of what has happended before. This means that we are imposing
Levy pro esses or, in dis rete time, random walk models. A brief dis ussion of these
models is given in Se tion 2.1 below. Note the usefulness of the onvolution prop-
erty in this ontext. If Value-at-Risk are sought over longer time horizons, we an
simply plug in the appopriate member of the normal inverse Gaussian familiy, with
parameters depending on the time interval in question. Other ways of evaluating
Value-at-Risk will have to perform some sort of numeri al integration to a hieve the
same thing. It is outside our brief to ompare the normal inverse Gaussian approa h
to other proposals in the literature, for example the ideas in Embre hts, Kluppelberg
and Mikos h [9℄, who employ Extreme Value Theory and generalised Pareto distri-
butions. In the next se tion the normal inverse Gaussian family of models and its
statisti al inferen e will be introdu ed in a tutorial manner. Value-at-Risk evaluation
will be onsidered in Se tion 3, and eight dierent Norwegian se urities studied in
Se tion 4.

2. Mathemati al and statisti al modelling

2.1. Levy pro esses. A Levy pro ess Lt is a time- ontinuous sto hasti pro ess with
independent and stationary in rements; for every t > s  0, Lt Ls is independent
of Ls and its distribution depends only on the time in rement t s and not on t
or s. Classi al examples are Brownian motion and the Poisson pro ess. Ex ept for
Brownian motion, the paths of a Levy pro ess have jumps at random time points, and
that is pre isely what goes on in a market of se urities if we take the view that the
value is the pri e agreed last. But the Levy model also imposes serious restri tions.
The sizes of the jumps, whether positive or negative for example, have no relation to
the pri e level rea hed, and the ontinuous time viewpoint for es the distribution of
the in rements Lt Ls to belong to the innitely divisible ones; e.g the distribution
must for every n be the n-fold onvolution of some other distribution. There are not
too many models with this property around, but the normal inverse Gaussian family
is one of them.
Arguably there is no lower limit as to how lose in time two dierent tradings an be
ompleted, but in pra ti e it is ommon to pass to a dis rete viewpoint and tra k the
sto k pri e at some sequen e of equidistant points in time. If a very ne resolution is
used (say se onds!) the distribution of the in rement Lt Ls will have a huge atom
at the origin ('no trading'), but when we deal in terms of days and liquid shares,
a purely ontinuous distribution, without atoms, may be in order. Experien e also
di tates that it is reasonable to relate the sto k pri e St to the Levy pro ess Lt trough


(2.1) St = S0 exp Lt ;
 QUANTIFICATION OF RISK IN NORWEGIAN STOCKS 3
known as a Geometri Levy pro ess. The ratio of pri es at t and t
now be omes
(2.2) ln(St=St
) = Lt Lt
;
and we pass in a dis rete time to an ordinary random walk based on independent
in rements Xt = Lt Lt
, the distribution of whi h being our modelling tool. In
this paper the distribution of Xt will be assumed to belong to a spe i parametri
form.
2.2. The normal inverse Gaussian distribution. A random variable X follows a
normal, inverse Gaussian distribution with parameter ve tor (; Æ; ;  ), in symboli
notation X  NIG(; Æ; ;  ), if its probability density fun tion is
nig(x; ; Æ; ;  ) = Æ 1nig x Æ  ;0; 1; ; 


(2.3)
where p
 p
K1 ( 1 + z 2 )
(2.4) nig(z;0; 1; ;  ) =  exp   + z p1 + z2 :
2 2

Here K1 is a modied Bessel fun tion of the third kind with index 1; see Se tion 2.3
below. Note that  and Æ are ordinary parameters of lo ation and s ale whereas  and
 determines the shape of the density; i.e. if X  NIG(; Æ; ;  ), then Z = (X )=Æ
is the standardized version NIG(0; 1; ;  ). We have hosen to parametrize the family
as suggested in Barndor-Nielsen [3℄, but dierent from many other papers, notably
Rydberg [14℄, where  and  have a slightly dierent meaning. It is in our view more
transparent to let the shape of the density be determined by exa tly two of the four
parameters. The onditions for a viable density is Æ > 0,  > 0 and j j= < 1.
The mean, varian e, skewness and kurtosis of X are
EX  =  + Æ (1 (= =)2 )1=2
 
Var X = Æ2 1 (1 (= 1
)2)3=2
SkewX  = 3 1=4 (1 (= =)2 )1=4
=2 1 + 4(=)
  2
Kurt X = 3 (1 (=)2)1=2
1

Here, SkewX  = E(X E[X ℄)3= Var[X ℄
3=2 and KurtX  = E(X E[X ℄)4= Var[X ℄
2
3. Note that
Skew[X ℄2 = 3(=)2 ;
Kurt[X ℄ 1 + 4(=)2
so that
Kurt[X ℄ > 0 and jSkew[X ℄j  53 Kurt[X ℄;
4 BLVIKEN AND BENTH
(sin e j j= < 1). There is thus a bound on the skewness relative the kurtosis, and
the tails of the distribution is always heavier than those of the normal (whi h has
zero kurtosis). In fa t, (see e.g. [3℄),
nig(x; ; ; ; Æ)  kjxj 3=2 exp (=Æ)jxj + (=Æ)x
; when jxj ! 1 ;
where k is a positive onstant. Heavy tails is hardly any pra ti al restri tion in em-
piri al nan e. The distribution onverges to the normal as  ! 1 (if Æ / 1=2)
whereas the Cau hy distribution appears as the opposite limit as  ! 0. The param-
eter  determines the degree of skewness. For symmetri al densities  = 0.
An important property of the the family of normal inverse Gaussian distributions
is its behaviour under onvolutions. Suppose X1 and X2 are independent and dis-
tributed as NIG(1; Æ1; 1; 1) and NIG(2; Æ2; 2; 2) respe tively. Then, X1 + X2 
NIG(1 + 2; Æ1 + Æ2; 1 + 2; 1 + 2), provided 1=Æ1 = 2=Æ2. In parti ular, any
sum of independent and identi ally NIG-distributed random variables is also NIG-
distributed. By appealing to the entral limit theorem the result underpins the earlier
laim that the normal distribution will eventually appear, as  grows. If logreturns
follow a random walk, as suggested in Se tion 2.1, see (2.1) in parti ular, one would
expe t logreturn data to behave more like normal ones over longer time spans. For
example, Eberlein and Keller [6℄ on lude that the normal distribution is appropriate
for monthly logreturn data, but reje t this distribution for daily or weekly logreturn
data1 .
The normal inverse Gaussian distribution an be generalised with a fth parame-
ter to the so- alled generalized inverse Gaussian distributions; see e.g. Rydberg [14℄.
Eberlein and Keller [6℄ used a subfamily alled the hyperboli distributions to study
logreturn data from the German sto k market (see also [7℄ and [12℄).
2.3. Fitting normal inverse Gaussian models.   It is easy to see from the pre eed-
ing dis ussion that the expressions for E X , Var X , Skew[X ℄ and Kurt[X ℄ an be
inverted to yield a unique set
of2 parameters (; Æ; ;  ). For example, the ratio =
is determined from Skew[X ℄ =Kurt[X ℄, and the others then follow readily. This
means that the so- alled method of moments is available to estimate the parameters
from a given sample x1; : : : ; xn of normal inverse Gaussian distributed random vari-
ables. The mean, varian e, skewness and kurtosis are then repla ed by their sample
versions and the four equations are solved for ; ;  and Æ. Su h equations are un-
wieldy in many other situations in statisti s, but here they are straightforward, and
the whole operation is numeri ally so simple that the Bessel fun tion does not have
to be evaluated at all.
1Theirstudy was performed on German sto ks. See the referen es in [6℄ for empiri al investiga-
tions of US sto ks
 QUANTIFICATION OF RISK IN NORWEGIAN STOCKS 5
There is, however, reason to be suspi ious of the quality of su h estimates based
on higher order sample moments, and there is, of ourse, no guarantee that these
sample moments satisfy the restri tions laid down by the normal inverse Gaussian
family at all, so that the moment equations have a tually a solution. Likelihood
estimation seems a better, albeit more ompli ated alternative. This method was
suggested in Blsild and Srensen [5℄ and Rydberg [14℄. The former used the method
of steepest des ent to maximize the likelihood surfa e whereas the latter found a zero
of the dierentiated likelihood fun tions. The numeri al literature (for example Gill,
Murray and Wright [10℄) usually re ommends other approa hes. Prominent are are
the two lasses of optimization methods known as the variable metri s and the on-
jugated gradients. We have tested both on the data in the next se tion. Both worked
well when properly implemented, but the variable metri method (BFGS version)
onverged faster and is our re ommendation.
Either of these approa hes require the evaluation of the log-likelihood fun tion
n
X
(2.5) `(; ; ; Æ ) = log nig(xi; ; Æ; ;  )

i =1
and its partial derivatives with respe t to the parameters. The evaluation of the Bessel
fun tion K1(z) and is derivative is then required. There exist ountless representations
of these fun tions; see Se tion 9.6 of Abramowitz and Stegun [1℄, for example,
Z 1
p
(2.6) K1 (z ) = z e zt t2 1 dt
1
from whi h it an be proved that
(2.7) K10 (z ) = z 1 K1 (z ) z 2 K0 (z )
where K0(z) = R11 e zt (t2 1) 1=2 dt. Subroutines for the evaluation of K0 and K1
are available in the Numeri al Re ipe library, see Press, Teukolsky, Vetterling and
Flannery [13℄.
3. Value-at-Risk
We dene the on ept of Value-at-Risk a ording to JP Morgans RiskMetri s: The
logreturn from the sto k St at time t is Lt = ln(St=S0). If p is a spe ied level risk
toleran e, the Value-at-Risk for St at time t is
(3.1) VaRp(St) = ELt qp(t)
where qp(t) is the p-quantile of Lt. In the standard sto k pri e model, Lt = t + Bt
for a Brownian motion Bt with drift and volatility . In this ase we re over the
well-known formula for Value-at-Risk
p
(3.2) VaRp(St) = p t
6 BLVIKEN AND BENTH
where p is the p-quantile of a standard normal distribution. It is of interest to om-
pare the Value-at-Risk estimate in the standard model with that of a normal inverse
Gaussian model. We expe t the two estimates to onverge for large time horizons. To
explain this, let L1 be normal inverse Gaussian distributeed with parameters ; ; 
and Æ. Then by the onvolution property the logreturn at time t is distributed as
Lt  nig(x; t; t; t; Æt), and thus
   Æ 
E Lt =  + p2  2
t
  Æ 22
Var Lt = (p
 2  2 )3

t
When t in reases, we know that Lt tends to a normal distribution. In fa t for large
t, we approximately have
 Æ
Æ 22 
Lt  N  + p 2
t; p
t
 2 ( 2  2)3
Hen e, if we are able to t perfe tly both a normal inverse Gaussian and a Gaussian
model to the same set of data, the orresponding Value-at-Risk estimates will be
nearly identi al for large time horizon. However, on shorter time s ales there will be
a signi ant dieren e between the two Value-at-Risk estimates.
4. Empiri al study

4.1. Summary des ription and t. We have used eight Norwegian sto ks as ba-
sis for an empiri al evaluation of the normal inverse Gaussian model. Seven of them
(Agresso, Bergesen, Den norske Bank, Kreditkassen, Merkantildata, Norsk Hydro and
Petroleum Geo-Servi es) are quoted on the Oslo Sto k Ex hange while the eight (also
Norsk Hydro, but now denoted Norsk Hydro (NYSE) ) was taken from the notations
in New York. The returns from the former group of seven papers were end-of the-day
pri es from O tober 16, 1997 until November 17, 1999, that is 506 dierent values.
The data from New York was longer, lasting from January 2, 1990 until De ember 31,
1998, a total of 2274 notations, all at losing time. When dealing with in rements,
all series be ome one unit shorter. All analyses below are in terms of log-returns and
assume the sto k pri es to follow a random walk.
The rst part of the analysis is a summary des ription of the data and an evalu-
ation of the t. The estimates of the parameters are shown in Table 1 with estimates
of their standard error obtained by the bootstrap te hnique explained in 4.2 below.
We have in Figure 1 onverted the estimates of the shape parameters  and  to an
alternative, more pi torial form sometimes used. Introdu e
p 
 = 1 + 2  2 1=2;  = :


(4.1) 
 QUANTIFICATION OF RISK IN NORWEGIAN STOCKS 7
Here (; ) 0  jj <  < 1 (sin e j j < ), and the line  = 0 divides the triangle in
two parts a ording to the orientations of skewness, distributions skewed to the right
(left) being on the right (left) hand side. The verti al middle line = 0 (where  = 0)
represents symmetri distributions, the normal being approa herd as  ! 0 and the
Cau hy as  ! 1. The general pi ture is that of a phenomenon with onsiderably
higher variability than the normal ( being a good deal greater than 0) whereas a
possible skewness in the original in rements evidently has been removed by passing
to logarithms. The variability in  among the various papers re e t an underlying
volatility that seems to dier between some of the shares.
This omes out in Figure 2, as well, where the estimated normal inverse Gauss-
ian densities have been plotted against non-parametri estimates (obtained from the
kernel method with a Gaussian kernel; see Silverman [15℄). Note that the s ale of the
verti al axis has been made logarithmi , so that the normal density would ome out
as a parabola. The t of the tted normal inverse Gaussian distributions is ex ellent
everywhere. The dis repan ies at the extreme tails an be explained by larger random
error of the non-parametri estimate in these regions.
4.2. Value-at-Risk. Using normal inverse Gaussian models to evaluate Value-at-
Risk raises the obje tion that the estimated model is a ompromise t over the whole
range of variation whereas value at risk deals with the left tail. However, with time
horizons longer than one time unit, for example 5 or 10 days, as below, one does need
the whole distribution to ompute the ve-fold or ten-fold onvolution. The question
is how mu h bias the normal inverse Gaussian assumption will bring and how large
this ee t is ompared to the random variation. This, in turn, depends on how long
the histori al re ords are; i.e. how far ba k experien e is deemed to be of relevan e
for the un ertainty of the future. The answer is, of ourse, also in uen ed by the
per entile in question.
Consider Value-at-Risk at 1% and ompare evaluations based on the normal familly,
the normal inverse Gaussian family and a purely non-parametri approa h. The
latter uses, for ea h sto k, the empiri al distribution fun tion F^ of the observered
in rements of logreturns as the estimate of the underlying distribution fun tion F .
The1 pre ise denition of F^ is as the distribution fun tion assigning probability mass
n to ea h observered dieren e xj . To ompute Value-at-risk t days ahead under this
distribution, t in rements were drawn with equal probabilities and with repla ement
from the set fx1; : : : ; xng. Their sum is then a realisation of the random walk over
t days under F^ . When this operation is repeated a large number N times, the 1%
smallest is an approximation to the lower 1% per entile. The results reported are
based on N = 100000 times repetitions, whi h is large enough for the impa t of
the Monte Carlo error to be negligible. The method was used even for t = 1 as
a onvenient devi e to ir umvent that the exa t 1% per etile of F^ is not pre isely
dened.
8 BLVIKEN AND BENTH
The un ertainty of all the Value-at-Risk estimates was assessed by non-parametri
bootstrapping; see Eron and Tibshriani [8℄. Arti al data are then sampled from
the estimate F^ of the true distribution fun tion F and the omputations leading
from the real data x1; : : : ; xn to the estimate of the Value-at-Risk is opied exa tly on
the Monte Carlo data. When the latter operation is repeated, we an estimate the
standard error of the Monte Carlo estimates and obtain a pi ture of the un ertianty
of the original estimate. The method has the advantage that no assumption on the
underlying distribution is imposed. With the Model Carlo estimate the approa h
implies a nested sampling s heme, but sampling from F^ is so fast that the omputer
time is still minor. The number of Monte Carlo samples was 50.
The Value-at-Risk estimates are shown in Table 2. Note that the normal ones are
onsistently lower than the two others, onsistent with the tails in Figure 1 being
heavier than the normal. There is no onsistent pattern of dieren e between the
two other estimates, and this supports the normal inverse Gaussian model as suitable
for Value-at-Risk evaluations. Note, however, the mu h higher variablity of the non-
parametri estimates. We have not investigated the reason for the un ertainty of the
latter varying so little with the time horizon t, but it seems that the underlying exa t
fun tion must grow very slowly.
A nal aspe t is how long histori al re ords must be for the estimation of the param-
eters of a normal inverse Gaussian model to be suÆ iently a urate for the Value-at-
Risk estimate to be a reliable guide of future un ertainty. To obtain a partial answer
we divided the available data into periods about a half year ea h ( orresponding to
around 125 logreturns in rements). That gave four dierent set of notations for ea h
of the seven sto ks from the Oslo Sto k Ex hange and twenty for the one from New
York. The orresponding Value-at-Risk estimates are plotted jointly in Figure 3 for
a period of up to a month. The dotted urves are based on all available data; e.g two
years for the Norwegian notations.
The results were not quite as expe ted. In all of the seven Norwegian ases the Value-
at-Risk assessments based on half year data are onsistently lower than the urves
based on all data. There is no bug! The likelihood estimates of the parameters of
the normal inverse Gaussian distributions were arefully he ked to represent (lo al)
maxima on the log likelihood surfa e with zero derivatives(in some ases the iteration
onverged to an innte  orresponding to a Gaussian model). We also tried to nd
other lo al maxima by starting the iteration elsewhere, but without su ess. In other
words, this rather pe uliar phenomenon is real and is aused by huge upward bias in
the estimates of  for half-year data, whi h undervalues the un ertainty of the sto k.
Note the ompletely dierent behaviour of the notations in New York. It is tempting
to spe ulate that the ould be some systemati dieren es between the two sites, but
the results need to be ba ked up by Monte Carlo studies based on arti ial data.
 QUANTIFICATION OF RISK IN NORWEGIAN STOCKS 9
Sto k  Æ  =
Agresso -0.0028 (0.003) 0.033 (0.010) 0.64 (0.6) 0.047 (0.09)
Bergesen -0.0067 (0.003) 0.025 (0.004) 1.25 (0.6) 0.210 (0.09)
Den norske Bank 0.0005 (0.001) 0.019 (0.002) 0.55 (0.2) -0.028 (0.09)
Kreditkassen 0.0001 (0.001) 0.020 (0.002) 0.49 (0.2) 0.028 (0.09)
Merkantildata -0.0031 (0.003) 0.034 (0.005) 0.85 (0.3) 0.018 (0.09)
Norsk Hydro -0.0039 (0.009) 0.034 (0.025) 2.81 (9.0) 0.096 (0.12)
Norsk Hydro (NYSE) -0.0006 (0.003) 0.015 (0.001) 0.86 (0.1) 0.047 (0.05)
Petroleum Geo servi es -0.0048 (0.004) 0.041 (0.007) 1.07 (0.5) 0.083 (0.08)
Table 1. Estimated parameters in the normal inverse Gaussian distri-
bution for eight Norwegian sto k; estimated standard error in paren-
thesis
1.0

4
0.8

3
1
7 5
8 2
0.6

6
xi
0.4
0.2
0.0

-1.0 -0.5 0.0 0.5 1.0

chi

Figure 1. The estimated normal inverse Gaussian distributions plot-

ted in the shape triangle. The numbers orresponds to the ordering of
the sto k in Table 1.
5. Con lusion
The normal inverse Gaussian distribution is a very exible family of distributions
enjoying the onvolution property. With only four parameters it aptures stylized
fa ts like heavy tails and skewness observed in the marginals of nan ial time series
data. By using sophisti ated numeri al optimization te hniques, likelihood estimation
of parameters in the normal inverse Gaussian distribution an be al ulated eÆ iently
and reliable, making it admissible for nan ial appli ations. We have demonstrated
for seven sto ks quoted on the Norwegian Sto k Ex hange and one quoted in New
York that the normal inverse Gaussian distribution ts logreturn data nearly perfe tly
both in the tails and in the enter, learly outperforming the normal distribution.
Value-at-Risk with 1% risk toleran e was estimated for all eight sto ks using the
normal inverse Gaussian, normal and a non-parametri model. The omparison of the
10 BLVIKEN AND BENTH
Agresso Bergesen
4

4
•
• ••••••••
••• •• •• ••
2

2
• • • • ••••
• ••• • •••
• •
••• •• •• •• ••• • ••
0

0
••
• ••• • •••• • •• • • •••
• • • • • •
• •• •• • • •
-2

-2
• • •
• •
•
-4

-4
-6

-6
-0.2 -0.1 0.0 0.1 0.2 -0.2 -0.1 0.0 0.1 0.2

Den norske Bank Kreditkassen

4

4
• •• • • •• •
•• ••
•• • •• • • • •
2

2
• •• • •• • • •
• •
•••• •••• • • •
0

0
• • •
• • •• •
• • • • • • • • • •
• • • •
-2

-2
•• • • • •
• • • •
• • • •
-4

-4
• •
•
-6

-6
-0.2 -0.1 0.0 0.1 0.2 -0.2 -0.1 0.0 0.1 0.2

Merkantildata Norsk Hydro

4

4
•••• ••••••••••••
•• ••• ••
2

•• 2 •• ••
•• •• ••• •••••
••• • • • • ••
0

• • • •
• •••• • •••• • •• •
• • • • • • • • • •
-2

-2

• • • •
• • •
-4

-4

• • •
• •
•
-6

-6

•
-0.2 -0.1 0.0 0.1 0.2 -0.2 -0.1 0.0 0.1 0.2

Norsk Hydro (NYSE) Petroleum Geo-Services

4

•••
•• •• • •••••••
•• • ••
2

•• •• •••
•• • ••• • ••
•• ••• •••
•
0

• •• • ••
• •• •
•• • • ••
••••• • • • •
-2

-2

• •• • • •
••••• •• •
•
-4

-4

• • •
•
-6

-6

•
•
-0.2 -0.1 0.0 0.1 0.2 -0.2 -0.1 0.0 0.1 0.2

The tted normal inverse Gaussian densities (solid lines)

Figure 2.
against non-parametri estimates (dots). Logarithmi s ale on verti al
axix.
three models was performed on a time horizon ranging from 1 day to nearly a month.
The normal distribution gave onsistently too optimisti Value-at-Risk estimates for
all sto ks in question ompared to the non-parametri estimate. This is expe ted in
view of the poor tting of the tails. The normal inverse Gaussian distribution, on the
other hand, seems to be loser to the non-parametri Value-at-Risk estimate. We also
al ulated the standard error based on bootstrapping te hniques. The estimates from
the non-parametri model showed high errors, both relative to its estimated Value-at-
Risk and to the two other statisti al models. This learly indi ates the un ertainty
(and danger!) onne ted to Value-at-Risk estimation for per entiles far out in the
tails.
By going through the same analysis on data divided into periods about half a year
ea h, we found a huge upward bias in the estimate of  for the seven Norwegian sto ks.
 QUANTIFICATION OF RISK IN NORWEGIAN STOCKS 11
1 day 5 days 10 days
Sto k VaR 1% VaR
d VaR
d VaR
d
nig 0.113 (0.013) 0.226 (0.021) 0.311 (0.024)
Agresso normal 0.095 (0.005) 0.208 (0.014) 0.290 (0.025)
non-par. 0.104 (0.040) 0.211 (0.041) 0.291 (0.041)
nig 0.056 (0.004) 0.119 (0.007) 0.168 (0.009)
Bergesen normal 0.054 (0.003) 0.116 (0.009) 0.160 (0.014)
non-par. 0.060 (0.023) 0.124 (0.023) 0.173 (0.123)
Den nig 0.074 (0.005) 0.147 (0.009) 0.200 (0.011)
Norske normal 0.060 (0.003) 0.133 (0.008) 0.188 (0.014)
Bank non-par. 0.073 (0.022) 0.139 (0.023) 0.192 (0.023)
Kreditt- nig 0.081 (0.008) 0.160 (0.014) 0.219 (0.017)
kassen normal 0.067 (0.004) 0.152 (0.012) 0.216 (0.020)
non-par. 0.089 (0.029) 0.161 (0.030) 0.219 (0.032)
Merkantil- nig 0.093 (0.009) 0.192 (0.013) 0.268 (0.017)
data normal 0.087 (0.005) 0.197 (0.014) 0.283 (0.023)
non-par. 0.081 (0.037) 0.204 (0.037) 0.277 (0.037)
Norsk nig 0.049 (0.004) 0.106 (0.005) 0.149 (0.007)
Hydro normal 0.047 (0.002) 0.103 (0.007) 0.144 (0.012)
non-par. 0.045 (0.023) 0.106 (0.022) 0.140 (0.022)
Norsk nig 0.044 (0.002) 0.089 (0.003) 0.124 (0.004)
Hydro normal 0.039 (0.001) 0.087 (0.003) 0.123 (0.005)
NYSE non-par. 0.043 (0.014) 0.093 (0.015) 0.127 (0.016)
Petrolem nig 0.101 (0.007) 0.209 (0.012) 0.291 (0.015)
Geo normal 0.091 (0.005) 0.199 (0.014) 0.277 (0.024)
servi es non-par. 0.105 (0.041) 0.210 (0.039) 0.294 (0.039)
Table 2. Estimates of 1% Value-at-Risk from tted inverse normal
Gaussian and ordinary Gaussian models and from a non-parametri
pro edure; estimated standard error in parenthesis.

This led to signi antly lower Value-at-Risk estimates than the orresponding Value-
at-Risk based on the omplete dataset. Interestingly, the same phenomenon was not
observed for the New York sto k. Here, as one would a priori expe t, Value-at-Risk
based on subsets of data enveloped the estimate from the full data series.

Referen es

[1℄ M. Abramowitz and I.A. Stegun (1972): Handbook of Mathemati al Fun tions. Dover Publ.
In . New York.
[2℄ O. E. Barndor-Nielsen, Exponentially de reasing distributions for the logarithm of parti le
size, Pro . Roy. So . London A 353 (1997), 401-419.
 12 BLVIKEN AND BENTH
Agresso Bergesen
• • • •
0.10 0.20 0.30 0.40

• • • •
• •

0.05 0.10 0.15 0.20

• • • •
• • • •
• • • •
• • • •
• • •
• •
•
• •
• •
• •
• •
5 10 15 20 5 10 15 20

Den norske Bank Kreditkassen

0.30
• • • •
0.10 0.15 0.20 0.25

• • • •
• • • •
• • • •
• • • •
• •

0.20
• • • •
• • •
•
• •
• •
• •

0.10
• •
• •
5 10 15 20 5 10 15 20

Merkantildata Norsk Hydro

0.05 0.10 0.15 0.20

• • • •
• • • •
• • • •
0.30

• • • •
• • • •
• • • •
• • • •
0.20

• •
• •
• •
• •
• •
0.10

• •
5 10 15 20 5 10 15 20

Norsk Hydro (NYSE) Petroleum Geo-Services

0.10 0.20 0.30 0.40

• • • •
0.04 0.08 0.12 0.16

• • • •
• • • •
• • • •
• • • •
• • • •
• • • •
• •
• •
• •
• •
• •
• •
5 10 15 20 5 10 15 20

Figure 3. Estimated 1% Value-at-Risk based on half-year data (solid

lines) and all data (dotted lines).

[3℄ O. E. Barndor-Nielsen, Pro esses of Normal inverse Gaussian type, Finan e and Sto hasti s
2 (1998), 41-68.
[4℄ O. E. Barndor-Nielsen and O. Halgreen, Innite divisibility of the hyperboli and generalized
inverse Gaussian distributions, Z. Wahrs heinli hkeitstheorie Verw. Geb. 38 (1977), 309-312.
[5℄ P. Blsild and M. K. Srensen, Resear h Report No. 248. Department of Theoreti al Statisti s,
University of Aarhus (1992).
[6℄ E. Eberlein and U. Keller, Hyperboli distributions in nan e, Bernoulli 1(3) (1995), 281-299.
[7℄ E. Eberlein, U. Keller and K. Prause, New insights into smile, mispri ing and value at risk:
the hyperboli model, J. Business, 71(3) (1998), 371-405.
[8℄ Eron and Tibshriani, An introdu tion to the bootstrap. New York: Wiley, 1993.
[9℄ P. Embre hts, C. Kluppelberg and Th. Mikos h, Modelling Extremal Events. For Insuran e
and Finan e. Appli ations of Mathemati s, 33. Springer-Verlag, Berlin, 1997.
[10℄ P. E. Gill, W. Murray and M. H. Wright, Pra ti al optimization. A ademi Press, In ., London-
New York, 1981.
 QUANTIFICATION OF RISK IN NORWEGIAN STOCKS 13
[11℄ K. Prause, Modelling nan ial data using generalized hyperboli distributions, Preprint, Uni-
versitat Freiburg i. Br., (1997).
[12℄ K. Prause, How to use NIG laws to measure market risk, Preprint, Universitat Freiburg i. Br.,
(1999).
[13℄ W. H. Press, S. A. Teukolsky, W. T. Vetterling and B. P. Flannery, Numeri al Re ipes in C.
Cambridge: Cambridge University Press 1992.
[14℄ T. H. Rydberg, The normal inverse Gaussian Levy pro ess: Simulation and approximation.
Commun. Statist.{Sto hasti Models 13 (4) (1997), 887-910.
[15℄ B. W. Silverman, Density estimation for statisti s and data analysis. Monographs on Statisti s
and Applied Probability. Chapman & Hall, London-New York, 1986

(Erik Blviken)
Department of Mathemati s
University of Oslo
P.O. Box 1053, Blindern
N{0316 Oslo, Norway
and
Norwegian Computing Center
PO Box 114 Blindern
N-0314 Oslo, Norway

E-mail address : erikbmath.uio.no

(Fred Espen Benth)

Department of Mathemati s
University of Oslo
P.O. Box 1053, Blindern
N{0316 Oslo, Norway
and
Norwegian Computing Center
PO Box 114 Blindern
N-0314 Oslo, Norway

E-mail address : fredbmath.uio.no