Available online at www.sciencedirect.

com
Physica A 321 (2003) 286299
www.elsevier.com/locate/physa
From Brownian motion to operational risk:
Statistical physics and nancial markets
Johannes Voit
,1
Theoretische Physik 1, Universit at Bayreuth, D-95440 Bayreuth, Germany
Abstract
High-frequency returns of the DAX German blue chip stock index are used to test
geometric Brownian motion, the standard model for nancial time series. Even on a 15-s time
scale, the linear correlations of DAX returns have a zero-time delta function which carries 90%
of the weight, while the remaining 10% are positively correlated with a decay time of 53 s and
negatively correlated on a 9.4-min scale. The probability density of the returns possesses fat
tails with power laws whose exponents continuously increase with time scales. It is suggested
that hydrodynamic turbulence may provide a phenomenological framework for the description
of these data, and at the same time, open a way to use them for risk-management purposes, e.g.
option pricing and hedging. Option pricing also is the cornerstone of credit valuation, an area
of much practical importance not considered explicitly in most other physics-inspired papers on
nance. Finally, operational risk is introduced as a new risk category currently emphasized by
regulators, which will become important in many banks in the near future.
c 2002 Elsevier Science B.V. All rights reserved.
Keywords: Brownian motion; Turbulence; Financial markets; DAX stock index; Option pricing;
Credit default risk; Basel II capital accord; Operational risk
1. Introduction
Interactions between physics and nance have a long history. Apart from the in-
vestment ventures of physicists into stock markets, such as Isaac Newtons losing a
fortune in the South Sea bubble in summer 1720, the one-dimensional random walk

Tel.: +49-921-55-31-85; fax: +49-921-55-29-91.

E-mail address: Johannes.Voit@dsgv.de (J. Voit).
URL: http://www.phy.uni-bayreuth.de/

btp314
1
Present address: Deutscher Sparkassen- und Giroverband, Abteilung C IV Controlling, Simrockstrae 4,
D-53113 Bonn, Germany.
0378-4371/03/$ - see front matter c 2002 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0378-4371(02)01783-1
J. Voit / Physica A 321 (2003) 286299 287
was described successfully almost simultaneously in nance by Bachelier in 1900 [1,2]
and in physics by Einstein in 1905 [3]. Bacheliers work in nance was forgotten and
rediscovered in the 1950s, among others, by physicists [2,4]. Later, Mandelbrot ap-
plied L evy-stable stochastics processes and multifractal processes to the description of
nancial time series [5,6].
The application of statistical physics to the description of nancial markets [710]
and economics [1114] has become more systematic, broader and deeper during the
last couple of years. Topics discussed include:
statistical properties of nancial and economic time series, such as
probability densities and scaling of price changes,
linear and nonlinear correlations in and among time series,
universality in nancial time series, and their stochastic modeling;
applications to option pricing and hedging;
applications to portfolio theory;
simulation of microscopic market models;
theory and prediction of stock market crashes.
This new direction of research is driven by increasing evidence that an improved un-
derstanding of nancial markets can be gained by using parallels to phenomena in
nature such as normal and anomalous diusion, earthquakes, phase transitions, turbu-
lence, highly excited nuclei, etc., and the use of the specic models and techniques
that have been devised for their description [7].
In this paper, we will not review this work. Comprehensive reviews can be found
in a number of books which have appeared over the last years [710]. Instead, we will
sketch some of the important questions, both for fundamental and applied research,
which have driven past research or which are likely to become important in the fu-
ture. In some cases, we present new material from our own research. In other more
forward-looking cases, data are not yet available, and we limit ourselves to the outline
of directions of research.
2. Geometric Brownian motionthe standard model of asset returns
Fig. 1 displays the history, on a 15-s time scale, of the DAX German blue chip stock
index for the years 1999 and 2000. The DAX index is composed of the 30 biggest
companies in Germany in terms of market capitalization, and xed every 15 s.
Speculative investors will ask whether protable investments can be made both dur-
ing the upward and during the downward moves of the market, and if such investments
can be leveraged, i.e., the returns of an investment are bigger than those of the market,
what would be the success rates of such investments. These questions are not new,
and basically are those at the origin of the Ph.D. thesis of the French mathematician
Louis Bachelier [1]. For risk-management purposes, one would rather ask if there are
strategies to protect oneself from losses arising from adverse market moves once an
open position cannot be closed for strategic or commercial reasons (hedging).
288 J. Voit / Physica A 321 (2003) 286299
01.01.1999 01.07.1999 01.01.2000 01.07.2000 01.01.2001
time
4000
5000
6000
7000
8000
D
A
X

p
e
r
f
o
r
m
a
n
c
e

i
n
d
e
x
Fig. 1. Price chart of the DAX German blue chip index during 1999 and 2000. The performance index
includes dividend payments. Data are taken on a 15-s time scale. (From Ref. [21].)
In order to answer these questions, one has to solve a more basic problem: What is
the probability for an asset price S(t) at time t? Or better, because investors require a
percentage return over a certain time horizon t,
oS
t
(t) = ln
S(t + t)
S(t)
(1)
what is (oS
t
, t) [1,7]? Bachelier made two postulates: (i) A trade between two parties
is concluded only when its expectation value is the same for both parties (fair game
condition). Buy and sell orders having opposite signs, the expectation value of a trade
must be zero. Its time series then follows a martingale stochastic process. (ii) At any
traded price, the market does not believe in rising or falling prices, i.e., subsequent
price changes must be statistically independent. Somewhat roughly, both postulates
are summarized by the ecient market hypothesis: all available information is ac-
counted instantaneously in asset prices. By a series of arguments, Bachelier and later
researchers concluded that the statistically independent returns of an asset are drawn
from a Gaussian distribution
[oS
t
(t)] =
1

2o
2
t
exp

[oS
t
(t) jt]
2
2o
2
t

, (2)
i.e., prices follow geometric Brownian motion and are drawn from a lognormal dis-
tribution. Here, j is the drift rate of the returns, i.e., the growth rate of the asset value,
and o their standard deviation, or volatility.
J. Voit / Physica A 321 (2003) 286299 289
3. Tests of geometric Brownian motion
How well do real nancial markets follow geometric Brownian motion? There is
a large number of papers in the literature, both nancial [5,15] and physics oriented
[16,17,19,20], which provide evidence against Bacheliers model. Here, I present new
results from a study of high-frequency data taken at a 15-s interval of the German
DAX stock index during the years 1999 and 2000. The history is a combination of data
collected in a purpose-built database of German stock market data at the Department
of Physics, Bayreuth University, [21] and data provided by an economics database at
Karlsruhe University [22]. The data set investigated contained about 1.3 million entries,
the current data volume available is about 3 million entries.
Normalized returns are obtained from Eq. (1) by subtracting the expectation value
and dividing by the standard deviation of the return series
os
t
(t) =
oS
t
(t) oS
t
(t)

oS
2
t
(t) oS
t
(t)
2
. (3)
The history of normalized returns corresponding to Fig. 1 is shown in Fig. 2 [21]. It
is apparent that the high-frequency returns of the DAX index cannot follow geometric
Brownian motion: extreme signals occur much too frequently! Signals of the order
20, . . . , 40 o (where o is the standard deviation of the sample) are rather frequent, and
there are even events up to 160 o. For comparison, the probability of a 40o event
under hypothesis (2) is 1.5 10
348
and that of a 160o uctuation is 4.3 10
5560
.
Fig. 2. Returns os
15
(t) on a 15-s time scale of the DAX performance index, normalized to their standard
deviation. Notice the event at 160 o, and numerous events in the range 30 o, . . . , 60 o. (From Ref. [21].)
290 J. Voit / Physica A 321 (2003) 286299
0 5 10 15 20 25
[min]
-0.01
0
0.01
0.02
0.03
0.04
<

s
(
t
)

s
(
t
+

)
>
Fig. 3. Linear correlations function C
15
(t) for 15-s DAX returns (dots) with 3o-error bars. The solid line
is a t to Eq. (5) and demonstrates that the data are almost uncorrelated. (From Ref. [21].)
Linear correlations of 15-s returns os
15
,
C
15
(t) = os
15
(t)os
15
(t + t), (4)
are shown in Fig. 3 with 3o-error bars [21]. Correlations are positive with a short 53-s
correlation time and negative (overshooting) with a longer 9.4-min correlation time.
The remarkable feature of Fig. 3, however, is the small weight of these correlations!
The solid line represents a t of the data to
C
t
15

(t) = 0.89 o
t, 0
+ 0.12 e
t}53

0.01 e
t}9.4

, (5)
implying that the data are uncorrelated to almost 90%, even at a 15-s time scale.
Bacheliers postulate of statistically (linearly) uncorrelated price changes is satised
remarkably well.
While linear correlations are absent in the high-frequency DAX returns, nonlinear
correlations are signicant in nancial time series. Quite generally, one nds that both
correlations of the amplitude of the returns and of their square are important and decay
with power laws of time t
:
with : 0.3, . . . , 0.4 [19,23].
A double-logarithmic plot of the probability distribution of the absolute values |os
15
(t)|
of 15-s DAX returns is shown in Fig. 4 [21]. It is independent of whether only positive,
negative, or absolute returns are considered. The probability distribution is far from
Gaussian, characterized by fat tails and almost follow a power law. Using the Hill
J. Voit / Physica A 321 (2003) 286299 291
10
-2
10
-1
10
0
10
1
10
2
|s|
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
10
1
P
(

s
)
=2.33
Fig. 4. Probability density function of 15-s DAX returns (dots). The straight dotted line indicates a power
law with its exponent derived from extreme value theory (see text). The solid line is a t to a Student-t
distribution, again with its exponent determined independently by extreme value theory. (From Ref. [21].)
estimator from extreme value theory [1719], and writing the probability distribution
as
(os
t
) |os
t
|
1j
, (6)
we obtain an exponent j 2.33 at t=15

. This value is outside the range of exponents

characterizing stable L evy distributions [5,7], signicantly bigger than the benchmark
value j = 1.6 found by Mandelbrot in 1963, but smaller than the inverse cubic law
found by others [17,19].
The solid line in Fig. 4 is a t to a Student-t distribution, given the index j = 2.33
derived from extreme value theory [21]. The dotted line in Fig. 4 represents a pure
power law. Both the apparent curvature of the data away from the straight line, and
the convergence with a nite slope, of the Hill estimator to its innite uctuation limit
value suggest that the probability distribution of extreme returns is not a pure power law
but rather contains corrections varying more slowly than a power law. The existence of
such (e.g. logarithmic) corrections to power laws is well known in statistical physics
in the vicinity of critical points.
When the time scale of the returns is increased from 15 s to hours, days, and months,
the exponents j monotonically increase (Fig. 5), and distributions with exponents of the
order 10 or 15 are not signicantly dierent from exponential or Gaussian distributions
over the range of values considered. As previous work on US-markets [19], we thus nd
292 J. Voit / Physica A 321 (2003) 286299
10
0
10
1
10
2
10
3
10
4
10
5
10
6
[min]
0
2
4
6
8
10
12
14
I
n
d
e
x

high-frequency data
daily closing prices
Fig. 5. Dependence of index j of the power laws, Eq. (6), on the time scale t of the returns. (From
Ref. [21].)
indications of a gradual convergence to a Gaussian when the time scales are increased
and random numbers are aggregated. However, we do not nd a characteristic time
scale for the onset of convergence to a Gaussian, in our analysis [21]. Interestingly,
when 1-day returns are aggregated from our high-frequency data, the j values we derive
are bigger, i.e., distributions decay more rapidly than those derived from a second data
set of daily closes of 25 years of DAX history, or those found in other studies of the
German stock market [17]. The dierences may be due either to dierent historic times
investigated, or to the use of intraday quotes in our high-frequency sample as opposed
to closing quotes in the other samples.
The statistical approach to fully developed hydrodynamic turbulence may provide
a phenomenology for these data. Conceptually, there are two types of experiments in
turbulence which may be mapped onto nancial time series. In a turbulent jet or ow,
one may measure the velocity of ow at some position in space as a function of time.
This would be the equivalent of nancial time series. One may also perform a two-point
measurement where velocity time series are taken at two dierent positions, and then
converted into a time series of velocity dierences at a given spatial separation. This
separation then is varied in the course of the experiments. The equivalent in nance
is the time series of returns with the time scale mapped onto spatial separation in
turbulence [7,24].
For turbulence, a picture of energy cascades over spatial scales from a (large) driving
to a (small) dissipation scale has been proposed, and there have been speculations that
J. Voit / Physica A 321 (2003) 286299 293
information cascades from large to small nancial institutions might play a similar
role in nancial markets [7,24]. Leaving aside the question whether such cascades do
exist in nancial markets in the sense imagined, one can investigate the stochastic
process followed by the return history (i.e., the two-point measurement) as the time
scale is varied [7,20,25]. Checking the ChapmanKolmogorovSmoluchowski equation
demonstrates that the stochastic process across time scales is markovian. Consequently,
there is both a Langevin equation describing the time-scale evolution of the return
history and a FokkerPlanck equation describing the dependence of the probability
density of the returns on the time scales considered. When drift and diusion parameters
are determined from (foreign exchange) market data, one nds nontrivial contributions
of both additive and multiplicative noise to the diusion terms. When a t to the
probability density of market data at the longest time scales available is used as an input
for such a FokkerPlanck equation, its solutions correctly reproduce the probability
densities of the equivalent market data at shorter time scales [7,20]. The FokkerPlanck
equation therefore describes the changes in probability densities as the time scales are
contracting, i.e., in the direction opposite to what one would have for the aggregation
of random numbers where time scales expand. Statistical hypothesis tests also underpin
the explanatory power of such cascade models [26].
4. Option pricing
In a bullish stock market, prots can be made from speculative investments by buying
stock low and selling it high (a nontrivial task). Making prots in a bear market, or
leveraging an investment in a bull or bear market requires, however, the use of nancial
derivatives, such as forwards, futures, and options.
A forward contract (or simply forward) is a contract between two parties to sell
resp. buy an asset at some time 1 in the future (maturity of contract) for a price X
xed today. The contract is binding to both parties which makes pricing and hedging
independent of the properties of the stochastic process followed by the asset price S(t)
[7,27]. For our purposes, a futures contract is the same as a forward.
An call resp. put option is a contract between two parties giving the option holder
the right (but not the obligation) to buy resp. sell an asset at maturity 1 in the fu-
ture for the strike price X xed today. The writer of the option has the obligation to
honor the holders request. This asymmetry of rights and obligations, akin to insurance
contracts, makes option pricing a nontrivial problem, and dependent on the stochastic
process followed by the underlying asset [7,27]. Questions asked by the writing in-
stitutions would include: What is the appropriate risk premium to be charged to the
counterparty for accepting the liability in the contract? What is the appropriate hedge
of the short option position? The investor rather would ask: What is the appropri-
ate price for acquiring the right to exercise without having an obligation? What is
the appropriate fee for the limitation of losses without a corresponding limitation on
the prots?
The pricing and hedging of simple options was solved by Black, Merton and Scholes
assuming that the price process of the underlying asset was correctly described by
294 J. Voit / Physica A 321 (2003) 286299
geometric Brownian motion [7,2729]. The basic observation is that the price changes
of an option and its underlying are correlated. Let C(t) be the price of a (European)
call option. The correlation of the option price with that of the underlying is =9C}9S.
Therefore, by holding a long position in a suitable quantity (t) of the underlying asset
[worth S(t)] in a portfolio, together with the short option position [worth C(t)], it
should be possible to compensate the price movements of the option by those of
the underlying. By applying stochastic calculus to geometric Brownian motion, Black,
Merton, and Scholes showed that, indeed, all risk can be eliminated from the portfolio
if the price process is geometric Brownian motion and that in this case, there is no
need to charge a risk premium.
The dierential equation for the option price is
9C
9t
+ rS
9C
9S
+
1
2
o
2
S
2
9
2
C
9S
2
= rC. (7)
By substitutions, this equation can be reduced to a special nal value problem (deter-
mined by the options payo at maturity, C = max(S X, 0)) of the one-dimensional
diusion equation [7,2729]. The solution is
C(S, t) = SN(d
1
) Xe
r(1t)
N(d
2
). (8)
N(d) is the cumulative normal distribution
N(d) =
1

d
dx e
x
2
}2
, (9)
and its two arguments in (8) are given by
d
1
=
ln S}X +

r + o
2
}2

(1 t)
o

1 t
, (10)
d
2
=
ln S}X +

r o
2
}2

(1 t)
o

1 t
. (11)
r is the risk-free interest rate. The appearence of r instead of the drift rate j of the
historical time series of the underlying share price in Eq. (8) is a necessary consequence
of the construction of a risk-free portfolio.
This equation, however, has important limitations: (i) The derivation is based on
mathematical theorems satised by geometric Brownian motion but not by the more
general stochastic processes suggested by the empirical data on nancial markets, e.g.
those shown in Section 3. (ii) As the share price uctuates in time, so will do both
the option price C and its derivative with respect to share price. As a consequence,
a continuous adjustment of the hedge is necessary. This is possible, even in theory,
only in the absence of transaction costs. Whenever deviations occur, i.e., returns are
non-Gaussian or correlated, when rehedging is only done at discrete time steps, trans-
actions costs are nite, etc., there will be deviations from the BlackScholes option
prices and the -hedging strategy. The formation of a risk-free portfolio in general will
no longer be possible. The bank writing the option then should charge an appropriate
risk premium covering its residual risk.
J. Voit / Physica A 321 (2003) 286299 295
The option price, of course, varies as a function of time, and much of this variation
is determined by the price uctuations of the underlying. However, even when the price
of the underlying S(t)= const., C(t) becomes a monotonically decreasing function of
time (to maturity). A contracting time scale is one of the important factors determin-
ing the option price and hedging strategy. This would suggest that, in principle, the
FokkerPlanck equations derived from using the analogies between nancial markets
and hydrodynamic turbulence discussed in Section 3 could be useful to incorporate the
deviations from geometric Brownian motion observed in real markets into theories of
option pricing and hedging. One practical way to do so could be the integral (minimal
variance hedging) approach formulated by Bouchaud and Sornette [8,30].
5. Credit default risk and option pricing
There is another, surprising, application of option pricing with an importance at least
equal to derivative markets: credit markets. When a bank gives a credit, it faces two
essential risks: (i) Market interest rates may change while those charged to the debtor
are xed. As a consequence, the value of a credit changes with time (interest rate
risk). (ii) The debtor may be unable to either pay the interests on the credit at the
dates xed in the contract, or repay the credit at maturity (credit default risk). Interest
rate risk may be quantied essentially in the same way as the risk from stock market
uctuations, although the underlying stochastic processes are rather dierent.
More interesting is credit default risk. The basic problem is to quantitatively de-
termine the value of a credit when the debtor can default. The basic solutions go
backagainto Black and Scholes [28] and Merton [31]. Assume that there is a
company A which takes a credit. Its ability to repay the credit will depend on its value
at the time of maturity (in principle also on its value at all times where interests are
due). However, the value of company A is dicult to quantify: it comprises the value
of common stock it may have emitted, the value of machines and factories it possesses,
its human capital, its brand names, etc.
In order to make progress, assume that company A has issued stock and introduce
a company B whose sole purpose is to hold the stock of A. While the rm value of
A is dicult to measure, the rm value of B is simply the number of shares of A
it holds multiplied by the share price. Under the standard assumption of textbooks of
nance, the value of B therefore would follow geometric Brownian motion
dJ
B
= jJ
B
dt + oJ
B
d:. (12)
Eq. (12) is the transcription into a stochastic process of the statements made in
Section 3. Notice, with reference to Section 3, that this is an assumption not satised by
actual share prices, i.e., rm values.
In order to keep things simple, we simplify to the extreme and assume that tak-
ing a credit is essentially the same as issuing a bond. Moreover, we assume that the
bond/credit is a zero-coupon bond, i.e., there are no interest payments during the life-
time of the bond/credit. All interests are discounted into the price of the bond which is
lower than its nominal, to be repaid at maturity. At time t =0, company B thus issues
296 J. Voit / Physica A 321 (2003) 286299
a zero-coupon bond with nominal X, priced X where is the interests/spread. The
bond matures (the credit must be paid back) at t = 1. If the rm value J
B
(1) X,
the bond/credit is repaid in full. However, if the rm value J
B
(1) X, company B
defaults: it cannot pay back the entire bond X but only the fraction corresponding to
its value J
B
. The debtor or bond emitter (holder of the stock of company A) therefore
has acquired the right to sell company B to the bond holder at the price X even though
it may be worth less. Of course, this right is exercised only when J
B
(1) X, i.e.,
default has occurred. Taking a credit, resp. a short position in the bond, therefore is
equivalent to a long position in a (european style) put option on the company value.
When Eq. (12) is satised, the put option may be priced by the BlackScholes for-
mula. Stock prices, however, do not generally satisfy (12), and all the problems (and
solution paths) outlined above also apply to credit default risk valuation. Bonds/credits
with regular interest payments correspond to nested series of options, i.e., an option
on an option on . . . , etc.
While the consequences of the deviations of stock price histories from geometric
Brownian motion have been explored both for option pricing and hedging, and for
more general risk-management purposes for a long time, less work has been done on
similar aspects of credit default risk.
6. Basel IIthe new capital accord
To gauge the importance of future research on non-Gaussian rm value processes,
asset correlation, and default correlation in credit default risk, it may be useful to take
a look at the regulatory aspects of risk management.
Under current regulation (in Germany, most likely, however, in many other coun-
tries too), banking supervisors will not ask for specic capital provisions to cover
market risk, provided the bank has developed an internal model and this model has
been approved by regulators. Market risk includes stock price, currency, interest rate
uctuations, etc. An internal model is a systematic theory of the risk a bank faces
in a specic market, and its implementation in the bankss risk-management
infrastructure.
Credit default risk must be covered by a capital cushion, currently at 8% of
the risk-weighted assets. The main purpose of the new international capital accord
(Basel II) currently under negotiation at the Bank of International Settlements in Basel
[32] is to make this capital buer risk sensitive, i.e., dependent on the actual default
risk faced by a bank. To this end, internal (bank-specic) estimates of parameters,
such as the probability of default of a debtor (or a portfolio) or the loss given default,
may be combined with eective functions given by the supervisors, to estimate risk
and the capital charge. Even under Basel II, internal credit risk models will not be
acceptable.
There are many controversial issues in Basel II. The adequacy of the supervisory
risk-weight functions is disputed. The granularity adjustments (dependence of risk on
the number and diversity of credits) are under attack. Maturity corrections (dependence
of risk on a 1-year time scale, on the lifetime of a credit) are debated. All these
J. Voit / Physica A 321 (2003) 286299 297
controversies can be traced back to our insucient understanding of the dependence
of the value of a credit portfolio on rm values, their dependence on stock prices with
non-Gaussian statistics and asset correlations, and on default correlation.
It is likely that, with progress of research in these areas, banks will start to develop
internal credit risk models in the future, and that such models will gain supervisory
recognition once experience suggests that they can be relied on.
Finally, Basel II draws attention to a third category of risk besides market and credit
risk: operational risk.
7. Operational riskthe ultimate frontier?
Operational risk is dened as the risk of losses resulting from inadequate or failed
internal processes, people and systems or from external events and has been high-
lighted in the second consultative paper for the new Basel II capital accord [33].
Banks will be required to hold a capital cushion as a provision against operational
losses in the future. Examples include rogue traders, limit violations, insucient con-
trolling, fraud, IT failures and attacks, system inavailability, catastrophies such as re,
earthquakes, oods, etc. While the perception of operational risk is new in banking, it
is rather well known in industry where often hazardous processes are involved in the
production or transport of goods, e.g. in the chemical industry.
The principal challenges faced when attempting to describe operational risk are its
latent character, the absence of data, and the rarity of high-impact events. While for
market risk, plenty of data are publicly available, and for credit risk, sucient data
are available in banks internally, there are very few data available on operational risk.
There are one or two commercial databases which systematically gather descriptions
of those operational loss cases made public, e.g. in the press (e.g. [34]). Worse even,
for a given bank, stationary may be an impossibility when risk management improves,
e.g. as a consequence of losses suered.
Operational risk comprises two important aspects: (i) the frequency with which op-
erational losses occur, and (ii) the amount of loss suered in the case of an event.
Of course, both quantities will be stochastic. One therefore is interested in determin-
ing their probability density functions. Many operational risks can be insured. Some
inspiration can thus be gained from the standard model of actuarial science [35]. It pos-
tulates that the time interval between two insurance claims is random and distributed
according to a Poisson distribution and that the size of insurance claims is random and
drawn from a lognormal distribution!
Data collection therefore is an important focus of operational risk controlling. One
typically would build up data bases of operational risk losses across a bank or, better
even, across an entire banking federation. Such a programme, e.g. is being implemented
by Deutscher Sparkassen- und Giroverband for all 550 savings banks in Germany.
When operational, these data bases will yield the actual probability distributions of
the time intervals between losses, and of the amplitude of the losses suered by the
banks. The loss distribution functions then are generated by Monte Carlo simulation
and analyzed by standard statistical methods. The goal is to derive the established risk
298 J. Voit / Physica A 321 (2003) 286299
measures such as value at risk or expected shortfall, on a specied time horizon, e.g.
1 year.
A good operational risk-controlling program will, however, not rely on data alone.
Apparently extreme views would suggest not to rely primarily on loss data at all. One
problem is that data necessarily describe the past whereas risk controlling would prefer
to have a more dynamic picture including the consequences of management action on
future risks. Most serious, however, is the problem that there is risk even without data:
a bank may face signicant operational risks but may not have suered large losses in
the past. A data-based operational risk measure would grossly underestimate the risk
situation of the bank. Worse even, risk measures such as value at risk are strongly
aected by extreme losses which, hopefully, occur seldom enough to prevent good
data quality in that range. A qualitative self-assessment, i.e., expert interviews, is a
way out of this problem. When optimized in view of psychometric evaluation, such
questionnaires may provide more realistic risk estimates than data-based approaches.
Of course, methods such as fuzzy logic and Bayesian networks also allow to integrate
loss data with expert-based risk estimates for consolidated risk measures.
It is quite clear that physics-inspired computer simulations may be helpful in under-
standing the structure of operational risk time series and the connection of such time
series with the control structure of a bank. An approach has been developed where a
bank is modeled as a network, either regular or random [36]. Risk drivers (random
numbers) act on nodes which may be understood as elements of the internal organiza-
tion, processes, products and services, etc. Links between thoses nodes are interpreted
as paths of risk propagation. Rules determine how risk generates losses. Again, the
focus is on the statistical properties of operational loss time series, and on the prob-
ability distributions of the waiting times between two events, and of the losses gives
the events. Results will be published elsewhere [36]. Alternative approaches would
consider the error propagation in statistical networks [37].
Acknowledgements
Many of the results presented in Section 3 are based on a collaboration with
S. Dresel and H. B uttner. I also wish to acknowledge discussions with J.-P. Bouchaud,
W. Breymann, S. Ghashghaie, C.-K. Hu, L. Laloux, T. Lux, J. Peinke, M. Potters, and
H.E. Stanley. I was a Heisenberg fellow of Deutsche Forschungsgemeinschaft, and
additional nancial support was provided by Universit atsverein Bayreuth. The high-
frequency data used were partly provided by Universit at Karlsruhe [22].
References
[1] L. Bachelier, Th eorie de la Sp eculation, Ann. Sci. Ecole Norm. Sup., S er. 3 17 (1900) 21 (reprinted,
Jacques Gabay (Ed.), Paris, 1995. An English translation is available in Ref. [2]).
[2] P.H. Cootner (Ed.), The Random Character of Stock Market Prices, MIT Press, Cambridge MA, 1964.
[3] A. Einstein, Ann. Phys. (Leipzig) 17 (1905) 549.
[4] M.F.M. Osborne, Oper. Res. 7 (1959) 145 (reprinted in [2]).
J. Voit / Physica A 321 (2003) 286299 299
[5] B.B. Mandelbrot, J. Business 36 (1963) 394.
[6] B.B. Mandelbrot, J.W. van Ness, SIAM Rev. 10 (1968) 422.
[7] J. Voit, The Statistical Mechanics of Financial Markets, Springer, Berlin, Heidelberg, New York, 2001.
[8] J.-P. Bouchaud, M. Potters, Th eorie des risques nanciers, Al ea-Saclay, Paris, 1997. (An english
translation Theory of Financial Risk has appeared at Cambridge University Press, Cambridge, 2000).
[9] R.N. Mantegna, H.E. Stanley, An Introduction to Econophysics, Cambridge University Press, Cambridge,
2000.
[10] M. Levy, H. Levy, S. Solomon, Microscopic Simulation of Financial Markets, Academic Press, San
Diego, 2001.
[11] L.A.N. Amaral, S.V. Buldyrev, S. Havlin, M.A. Salinger, H.E. Stanley, Phys. Rev. Lett. 80 (1998)
1385.
[12] Y. Lee, L.A.N. Amaral, D. Canning, M. Meyer, H.E. Stanley, Phys. Rev. Lett. 81 (1998) 3275.
[13] J. Voit, Adv. Complex Systems 4 (2001) 149.
[14] R. Axtell, Science 293 (2001) 1818.
[15] E.F. Fama, J. Business 38 (1965) 34.
[16] R. Mantegna, H.E. Stanley, Nature 376 (1995) 46.
[17] T. Lux, Appl. Finance. Econom. 6 (1996) 463.
[18] B.M. Hill, Ann. Stat. 3 (1975) 1163.
[19] P. Gopikrishnan, V. Plerou, L.A.N. Amaral, M. Meyer, H.E. Stanley, Phys. Rev. E 60 (1999) 5305.
[20] Ch. Renner, J. Peinke, R. Friedrich, Phys. Rev. Lett. 84 (2000) 5224;
Ch. Renner, J. Peinke, R. Friedrich, Physica A 298 (2001) 499.
[21] S. Dresel, Die Modellierung von Aktienm arkten durch stochastische Prozesse, Diploma Thesis,
Universit at Bayreuth, 2001, unpublished.
[22] This database is operated by Institut f ur Entscheidungstheorie und Unternehmensforschung, Universit at
Karlsruhe, http://www-etu.wiwi.uni-karlsruhe.de/
[23] R. Cont, cond-mat/9705075.
[24] S. Ghashghaie, W. Breymann, J. Peinke, P. Talkner, Y. Dodge, Nature 381 (1996) 767.
[25] M. Ragwitz, H. Kantz, Phys. Rev. Lett. 87 (2001) 254501-1.
[26] T. Lux, Quantitative Finance 1 (2001) 632.
[27] J.C. Hull, Options, Futures, Other Derivatives, Prentice-Hall, Upper Saddle River, 1997.
[28] F. Black, M. Scholes, J. Polit. Econom. 81 (1973) 637.
[29] R.C. Merton, Bell J. Econom. Manag. Sci. 4 (1973) 141.
[30] J.-P. Bouchaud, D. Sornette, J. Phys. I (Paris) 4 (1994) 863.
[31] R.C. Merton, J. Finance 29 (1974) 449.
[32] The New Capital Accord, Consultative Document, The Basel Committee for Banking Supervision, Bank
of International Settlements, Basel, 2001, http://www.bis.org
[33] Working Paper on the Regulatory Treatment of Operational Risk, The Basel Committee for Banking
Supervision, Bank of International Settlements, Basel, 2001, http://www.bis.org
[34] FIRST data base operated by Zurich IC
2
, http://www.ic2.zurich.com
[35] S.A. Klugman, H.H. Panjer, G.E. Willmot: Loss ModelsFrom Data to Decisions, Wiley Series in
Probability and Statistics, New York, 1998.
[36] J. Voit, to be published.
[37] Hawoong Jeong, talk at this conference.