29.01.2015 Views

Fat Tails and Traditional Capital Market Theory - Gilles Daniel

Fat Tails and Traditional Capital Market Theory - Gilles Daniel

Fat Tails and Traditional Capital Market Theory - Gilles Daniel

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Fat</strong> <strong>Tails</strong> <strong>and</strong><br />

<strong>Traditional</strong> <strong>Capital</strong> <strong>Market</strong> <strong>Theory</strong><br />

Günter Bamberg<br />

Gregor Dorfleitner<br />

Heft 177/2001<br />

(3., abermals überarbeitete Auflage August 2001)<br />

Institut für Statistik<br />

und Mathematische Wirtschaftstheorie<br />

der Universität Augsburg<br />

Prof. Dr. Günter Bamberg<br />

Dr. Gregor Dorfleitner<br />

Universitätsstr. 16, D-86159 Augsburg<br />

Telefon: 0821/598-4151<br />

Telefax: 0821/598-4227<br />

E-mail: guenter.bamberg@wiso.uni-augsburg.de


<strong>Fat</strong> <strong>Tails</strong> <strong>and</strong> <strong>Traditional</strong> <strong>Capital</strong> <strong>Market</strong><br />

<strong>Theory</strong><br />

Günter Bamberg, Gregor Dorfleitner<br />

Overview<br />

Many empirical studies of stock price data conclude that log returns are fat-tailed.<br />

Statisticians mainly interested in time series analysis or curve fitting are not aware<br />

or do not bother about the consequences of fat-tailed log returns: expected stock<br />

prices fail to exist, expected returns (in the sense of percentage price changes) do<br />

not exist as well.<br />

If expected returns do not exist, then important theoretical underpinnings of capital<br />

market theory like mean-variance portfolio selection, CAPM, security market line,<br />

APT are no longer valid.<br />

These serious consequences are known to be true in the framework of stable (nonnormal)<br />

distributions. However, they are to be faced for arbitrary fat-tailed nonstable<br />

distributions.<br />

Acknowledgements. The authors would like to thank S.T. Rachev <strong>and</strong> two anonymous referees<br />

for valuable suggestions <strong>and</strong> improvements.<br />

1


A. Introduction <strong>and</strong> clarification of important notions<br />

This paper investigates the consequences of fat-tailed log returns for the percentage returns,<br />

on which traditional capital market theory is based. Before we come to our main<br />

statement in Section B <strong>and</strong> the conclusions in Section C, the different notions of returns<br />

<strong>and</strong> fat tails shall be clarified in this section.<br />

I. Percentage returns <strong>and</strong> log returns<br />

Let P t be the price of a certain security (adjusted for dividends, stock splits etc.) or the<br />

value of a stock index. The log return with respect to period t is defined as<br />

( )<br />

Pt<br />

R t = ln . (1)<br />

P t−1<br />

Hence the end-of-period price P t is<br />

P t = P t−1 e R t<br />

. (2)<br />

from which one can see that R t is the continuously compounded rate of return.<br />

equation<br />

The<br />

P t = P 0 e R 1+···+R t<br />

= P 0 e R 0t<br />

, (3)<br />

where the log return R 0t corresponds to the time interval [0,t], is valid from the definition<br />

of R t . Hence we have the additivity property<br />

R 1 + ···+ R t = R 0t . (4)<br />

Due to this property log returns are favorable if developments along the time axis are the<br />

topic of interest. The log return Rt<br />

P of a portfolio is related to the log returns R (i)<br />

t of the<br />

single stocks by the nonlinear formula<br />

R P t = ln<br />

( n∑<br />

a i e R(i) t<br />

i=1<br />

)<br />

, (5)<br />

where a 1 ,...,a n are the fractions of the invested capital. Because of this nonlinearity (5)<br />

traditional capital market theory prefers to work with discrete returns (or percentage price<br />

2


changes, or shortly: returns)<br />

DR t = P t − P t−1<br />

P t−1<br />

= P t<br />

P t−1<br />

− 1 = e R t<br />

− 1 . (6)<br />

With this notion of returns, we have a linear relation between the returns of the single<br />

stocks <strong>and</strong> the portfolio return, i.e.<br />

DR P t =<br />

n<br />

∑<br />

i=1<br />

a i DR (i)<br />

t . (7)<br />

Markowitz portfolio models (CAPM etc.) make full use of this simple linear structure.<br />

Obviously, the support of R t is the entire real line IR, whereas DR t is subject to the restriction<br />

−1 ≤ DR t < ∞ . (8)<br />

Therefore, the textbook approach to portfolio analysis of assuming discrete returns as<br />

normally distributed can only be considered as a rough approximation since it does not<br />

only contradict the property (8), but also ignores that the normal distribution is not closed<br />

under multiplications. The latter implies that the two-period return<br />

(1 + DR 1 ) · (1 + DR 2 ) − 1<br />

cannot be normally distributed simultaneously with DR 1 <strong>and</strong> DR 2 . Furthermore, the conventional<br />

techniques for annualizing parameters become very clumsy. 1<br />

Additionally, there are numerous papers (compare, for instance, the great amount of references<br />

given by Rachev/Mittnik (2000)) concluding that there is too much empirical<br />

evidence against normally distributed discrete returns. But what are the alternative distributions<br />

Of course, it is also heroic to consider the discrete return as a stable (non-normal) r<strong>and</strong>om<br />

variable though some eminent researchers (for instance Fama (1965, 1971), M<strong>and</strong>elbrot/Taylor<br />

(1967), Samuelson (1967)) worked with this hypothesis.<br />

The mentioned distributions, which are defined on the whole real line, are surely better<br />

suited for modeling the log (<strong>and</strong> not the discrete) returns. Moreover, due to property (4)<br />

3


log returns can more easily be h<strong>and</strong>led in statistical models. For these reasons, it has<br />

become very common to model the log returns with<br />

• normal distributions (i.e. the classical Black/Scholes world),<br />

• stable non-normal distributions (see e.g. Rachev/Mittnik (2000)),<br />

• (G)ARCH processes (see e.g. Taylor (1986)),<br />

• student distributions (see e.g Praetz (1972) or Blattberg/Gonades (1974)),<br />

to mention only a few.<br />

From (6) it is clear that DR is fully determined by R. However, the properties of the discrete<br />

return distribution are completely different from those of the log return distribution.<br />

This has severe consequences, as will be be pointed out in Section B. But firstly the term<br />

“fat tail” has to be defined.<br />

II. <strong>Fat</strong> <strong>Tails</strong><br />

Besides the considerable linguistic variety (fat tails, heavy tails, thick tails, long tails)<br />

there is also a variety of attempts to define fat tailedness exactly. 2<br />

Some definitions focus on a single moment, for instance<br />

R is fat-tailed ⇔ Var(R)=∞<br />

or<br />

R is fat-tailed ⇔ R is leptokurtic (i.e. kurtosis> 3) .<br />

Bryson (1982) states that these definitions are too crude <strong>and</strong> have to be replaced by approaches<br />

which take the tail behavior more explicitly into account, for instance the criterion<br />

of increasing conditional mean exceedance or the limiting distribution of the maximum<br />

of a r<strong>and</strong>om sample. The following five definitions seem to be rather important in<br />

the fields of insurance <strong>and</strong> finance. The resulting classes of fat-tailed distributions can be<br />

ordered with respect to inclusion as illustrated by Figure 1. Note that the family of stable<br />

(non-normal) distributions (set A in Figure 1) is a small subset of all those sets.<br />

4


Figure 1:<br />

The different classes of fat-tailed distributions<br />

F<br />

E<br />

D<br />

C<br />

B<br />

A<br />

A: stable (non-normal) distributions<br />

B: Pareto tails with α > 0<br />

C: regular variation with tail index α > 0<br />

D: subexponential distributions<br />

E: nonexistence of all exponential moments<br />

F: nonexistence of the exponential<br />

moment of order 1<br />

Corresponding to class B of Figure 1, a r<strong>and</strong>om variable R is said to have an (upper 3 )<br />

Pareto tail iff<br />

P(R > x) ∼ c<br />

x α (x → ∞) for some α > 0 . (9)<br />

Corresponding to class C of Figure 1, a r<strong>and</strong>om variable R is of regular variation iff<br />

P(R > x) ∼ L(x)<br />

x α (x → ∞) for some α > 0 , (10)<br />

where L(x) is a slowly varying function. 4 This definition of fat tailedness is rather<br />

widespread; compare, for instance, Shiryaev (1999), p. 334.<br />

The next section shows that both fat tail definitions prevent the existence of expected<br />

stock prices <strong>and</strong> expected (discrete) returns. Actually, the nonexistence of E ( e R) holds<br />

for a huge class of log return distributions.<br />

A third notion of fat tailedness is provided by the subexponential distributions. 5 A recent<br />

survey on this class is Goldie/Klüppelberg (1998). The survey makes it perfectly clear<br />

that, firstly, regular variation distributions with any tail index are subexponential, <strong>and</strong><br />

secondly, that subexponential distributions have no finite exponential moments of any<br />

order s, i.e.<br />

E ( e sR) = ∞ for all s > 0. (11)<br />

Property (11) can also be used as the fourth fat tail definition, which is wider<br />

5


than the requirement of a subexponential distribution. The monographs Rolski/Schmidili/Schmidt/Teugels<br />

(2000), p. 49, <strong>and</strong> Embrechts/Klüppelberg/Mikosch<br />

(1997), p. 405, define fat tailedness this way.<br />

In section B it is shown that some exponential distributions (though neither subexponential<br />

nor fat-tailed according to (11)) lead to a nonexisting exponential moment.<br />

Therefore, we need another relaxation of the term fat tailedness. Obviously, this definition<br />

is<br />

R is fat-tailed :⇔ E ( e R) = ∞ . (12)<br />

The definition corresponds to the biggest set F in Figure 1. Whereas definition (11) dem<strong>and</strong>s<br />

the nonexistence of the moment generating function at all points s > 0definition<br />

(12) only calls for the nonexistence of this function at the point s = 1.<br />

B. <strong>Fat</strong> tails <strong>and</strong> traditional capital market theory<br />

I. Nonexistence of Expected Prices <strong>and</strong> Returns<br />

The scope of this section is to show how a fat-tailed log return distribution of any kind<br />

leads to nonexisting expected prices <strong>and</strong> discrete returns. According to (2) <strong>and</strong> (6) both<br />

expected returns <strong>and</strong> expected stock prices are affine functions of the exponential moment<br />

E ( e R) of order 1 <strong>and</strong> thus their existence directly depends on a finite expectation of e R .<br />

Now we will carry out the proof for the case of a log return distribution with a Pareto<br />

tail (class B in Figure 1). We have to borrow only a little bit from probability theory (for<br />

instance Rényi (1966), p. 179), which tells us that if a r<strong>and</strong>om variable X ≥ 0 has finite<br />

expectation E(X), then the upper tail behavior must be of the following type:<br />

In the sequel, we use the negation of (13), i.e.<br />

lim x · P(X > x)=0 . (13)<br />

x→∞<br />

x · P(X > x) ↛ 0 for x → ∞ ⇒ E(X)=∞ . (14)<br />

With<br />

X = e R<br />

6


<strong>and</strong> R fat-tailed in the sense of formula (9), the left-h<strong>and</strong> side of (13) yields<br />

lim xP( e R > x ) = lim xP(R > lnx)=lim e z P(R > z)=lim e z c<br />

x→∞ x→∞ z→∞ z→∞ z α = ∞ . (15)<br />

The last equation results from the fact that the exponential function grows faster than any<br />

power function.<br />

Clearly, (13) is violated <strong>and</strong> hence with (14) we have<br />

E(X)=E ( e R) = ∞ . (16)<br />

So far we demonstrated the nonexistence of E ( e R) only for fat-tailed log returns in the<br />

sense (9) of a Pareto tail. However, the conclusion is valid for all types of fat tailedness<br />

illustrated by Figure 1. First, notice that neither the stability property nor the restriction<br />

0 < α < 2 has been used. Therefore, the nonexistence is also true for α ≥ 2 (despite the<br />

fact that the corresponding distributions belong to the domain of attraction of the normal<br />

law). Furthermore, the proof remains valid if the constant c is replaced by a slowly varying<br />

function L(z). One has to take into account the property (compare Ibragimov/Linnik<br />

(1971), p. 397):<br />

lim<br />

z→∞ zε · L(z)=∞ for all ε > 0 .<br />

Hence, (16) is also true for distributions belonging to set C of Figure 1. Moreover, (16) is<br />

valid for arbitrary subexponential distributions (set D in Figure 1). The first who discovered<br />

this fact was Chistyakov (1964). But his work has no relation whatsoever to finance.<br />

He was interested in technical applications (queuing theory, reliability etc.). Finally, for<br />

fat-tailed distributions according to (11) <strong>and</strong> (12) the validity of (16) is given by definition.<br />

The first who pointed to the consequence (of fat-tailed log returns) that expected returns<br />

do not exist was Agnew (1971). Agnew formulated his findings only in the framework<br />

of stable non-normal distributions (set A in Figure 1). Maybe this was the reason why<br />

his short note remained unnoticed by most researchers <strong>and</strong> textbook writers. To the<br />

best knowledge of the authors there exists only one additional reference pointing to (16),<br />

namely Smith (1976). His remark is hidden in a footnote. Like Agnew he confined himself<br />

to stable non-normal distributions. He was worried about the strange consequences<br />

for option pricing. But he did not mention (or was not aware of) the rather general validity<br />

of (16) <strong>and</strong> the far-reaching consequences for traditional capital market theory.<br />

7


Figure 2 shows some examples of log return distributions which are incompatible with finite<br />

expected prices. Some of the distributions (e.g. χ 2 , gamma, Weibull) are only defined<br />

on the positive axis. Of course, these distributions are not supposed to model the log return<br />

distribution directly. Rather, it is meant that the (maybe piecewise defined) log return<br />

distribution has an upper tail which is equivalent 6 to the tail of the respective distribution.<br />

Figure 2: Examples for fat-tailed (in the sense of nonexisting prices, grey area) <strong>and</strong><br />

light-tailed (white area) log return distributions<br />

• χ 2 <strong>and</strong> Cauchy distr.<br />

• lognormal <strong>and</strong> loggamma distr.<br />

• student <strong>and</strong> F distr.<br />

• empirical fat-tailed distr. with α ≥ 2<br />

• unconditional distributions of (G)ARCH models<br />

• exponential distr. with intensity parameter ≤ 1<br />

• gamma <strong>and</strong> Weibull distr. with appropriate<br />

values of the shape parameter<br />

• normal distr.<br />

• triangular distr.<br />

• truncated distr. of<br />

any kind<br />

• exp., gamma<br />

Weibull distr.<br />

with certain<br />

parameter<br />

values<br />

For the empirically found fat-tailed distributions mentioned in Figure 2 see e.g. Guillaume<br />

et al. (1997) where foreign exchange data are studied <strong>and</strong> tail indices of about 2.7 to3.5<br />

are estimated. A reference for the unconditional distributions of certain (G)ARCH models<br />

is Embrechts/Klüppelberg/Mikosch (1997), p. 461 ff, who proved that the unconditional<br />

distribution of the ARCH(1)-process<br />

R t = Z t<br />

√β + λRt−1 2 t ∈{1,2,...} ,<br />

where λ varies between 0 <strong>and</strong> about 3.56, has a Pareto tail. For instance a value for λ of<br />

0.1 (resp. 3.0) yields a tail index of α = 26.48 (resp. α = 0.15).<br />

II. Distribution families on the borderline between fat <strong>and</strong> light tails<br />

In Figure 2 the families of gamma <strong>and</strong> Weibull distributions <strong>and</strong> their cutting set, the<br />

familiy of exponential distributions, were mentioned. Indeed these families are on the<br />

borderline between fat-tailed <strong>and</strong> light-tailed distributions. We now want to exemplify<br />

8


this by the exponential distributions, which were already suggested by Agnew (1971) as a<br />

model for log returns. Moreover, the following proves that there are fat-tailed distributions<br />

which are not subexponential.<br />

From (15) we see, that if the tail P(R > z) multiplied by e z does not converge to zero, the<br />

exponential moment fails to exist. We now apply this criterion directly on exponential<br />

distributions.<br />

The two-sided exponential distribution has the density<br />

λ<br />

2 e−λ|z|<br />

<strong>and</strong> the upper tail<br />

P(R > z)= 1 2 e−λz .<br />

We get<br />

e z · P(R > z)= 1 2 ez · e −λz = 1 2 ez(1−λ) . (17)<br />

Obviously, (17) tends to ∞ if λ < 1 <strong>and</strong> to 1 2<br />

if λ = 1 such that<br />

E ( e R) = ∞ if λ ≤ 1 .<br />

For λ > 1, (17) tends to zero. The theorem from Section B.I gives no answer.<br />

But ( straightforward calculations ) show that in this case the exponential moment exists<br />

<strong>and</strong> is equal to . Thus the class of exponential distributions lies on the borderline<br />

λ2<br />

λ 2 −1<br />

between fat-tailed <strong>and</strong> thin-tailed distributions. Nevertheless, with respect to the existence<br />

of the exponential moment the class has to be subdivided into the two subclasses<br />

corresponding to λ ≤ 1 (in our sense fat-tailed) <strong>and</strong> λ > 1 (existence of the expected stock<br />

prices <strong>and</strong> returns).<br />

More generally, the class of Weibull distributions, of which the exponential distributions<br />

are a subclass, is also to be divided into fat-tailed <strong>and</strong> non-fat-tailed distributions. The<br />

same is true for the class of gamma distributions, the class of hyperbolic <strong>and</strong> the class<br />

of logistic distributions. However, the two latter classes do not contain the exponential<br />

distributions family.<br />

Figure (3) illustrates the situation.<br />

9


Figure 3:<br />

area)<br />

Distribution families on the borderline between fat (grey area) <strong>and</strong> light (white<br />

hyperbolic distr.<br />

exp. distr.<br />

Weibull distr.<br />

gamma distr.<br />

logistic distr.<br />

III. Markowitz compatibility of log return distributions<br />

Many empirical studies try to explain the expected (discrete) return through specified<br />

factors like the expected excess return of the market, the firm size, the industrial sector or<br />

certain ratios of the balance sheet.<br />

These studies are meaningful if the log return distribution does not belong to any of the<br />

fat-tailed distributions of Figure 1, which lead to nonexisting expected discrete returns.<br />

For other issues of traditional capital market theory (derivation of the efficient frontier,<br />

the CAPM, the security market line etc.) one needs the existence of the (discrete) return<br />

variance <strong>and</strong> the relevant covariances. These notions are meaningful if <strong>and</strong> only if E ( DR 2)<br />

is finite, or equivalently<br />

E ( e 2R) < ∞ . (18)<br />

For lack of an established technical term we call log returns satisfying (18) Markowitz<br />

compatible, <strong>and</strong> otherwise Markowitz incompatible. Since<br />

E ( e R) = ∞<br />

implies<br />

E ( e 2R) = ∞ ,<br />

the set of all Markowitz incompatible distributions includes all the sets in Figure 1 as<br />

10


subsets. Thus R is Markowitz incompatible if its distribution belongs to any of the sets in<br />

Figure 1.<br />

In terms of the density function f of R Markowitz compatibility is equivalent to<br />

∞<br />

lim<br />

z→∞ z<br />

e 2x f (x)dx = 0 (19)<br />

<strong>and</strong> to<br />

∞<br />

lim f<br />

z→∞ z<br />

( ) lnx<br />

dx = 0 . (20)<br />

2<br />

If we apply (20) to the two-sided exponential distribution (already discussed in Section<br />

B.II) we get<br />

( ) lnx<br />

f<br />

2<br />

= λ lnx<br />

e−λ 2 = λ 2 2 x− λ 2<br />

<strong>and</strong><br />

λ<br />

lim<br />

z→∞ 2<br />

∞<br />

z<br />

x − λ 2 dx = 0 ⇔ λ > 2 .<br />

Summarizing the findings with respect to exponential distributions, we have<br />

• 0 < λ ≤ 1: fat-tailed in the sense E ( e R) = ∞<br />

• 0 < λ ≤ 2: fat-tailed in the sense E ( e 2R) = ∞<br />

• λ > 2: light-tailed in the sense of Markowitz compatibility.<br />

Figure (4) illustrates the situation. The light grey area corresponds to the set F of Figure 1.<br />

The dark grey area corresponds to those Markowitz incompatible log return distributions<br />

for which E ( e R) is finite but E ( e 2R) is infinite. Of course, the white area (Markowitz<br />

compatibility) includes normal distributions <strong>and</strong> all distributions with finite support.<br />

11


Figure 4:<br />

The area of Markowitz incompatibility (of log returns)<br />

hyperbolic distr.<br />

Weibull distr.<br />

border between Markowitz incompatible<br />

<strong>and</strong> Markowitz compatible distributions<br />

gamma distr.<br />

logistic distr.<br />

C. Conclusions<br />

Curve fitting does not require restrictive assumptions with respect to the underlying distribution.<br />

Things are different if one intends to model the log return process R t (in discrete<br />

or continuous time). Then several assumptions are plausible or desirable, for instance:<br />

Assumption 1: The log returns R 1 ,R 2 ,... are independent (this corresponds to Fama’s<br />

weak efficiency hypothesis).<br />

Assumption 2: The log returns R 1 ,R 2 ,... (related to time intervals of equal length) have<br />

the same probability distribution (thus ruling out calendar anomalies).<br />

Assumption 3: The log returns R 1 ,R 2 ,... <strong>and</strong> the log return R 0t for the time span [0,t]<br />

differ only by location <strong>and</strong> scale parameters, i.e. the log return R 0t for the time span<br />

[0,t] has the same probability distribution as b t R 1 + a t , where b t > 0 <strong>and</strong> a t are suitable<br />

constants (this is required to generalize the empirical findings to a shorter or longer time<br />

horizon).<br />

Assumption 4: Expected stock prices are finite, i.e.<br />

E(P t )=P t−1 · E ( e R t ) < ∞ . (21)<br />

It is well known that the only class of probability distributions satisfying assumptions 1<br />

to 3 is the class of stable distributions.<br />

Under assumptions 1 to 3, stockprices <strong>and</strong> discrete returns are logstable or lognormal<br />

(shifted by −1).<br />

12


If we complement assumptions 1 to 3 with assumption 4, we end up with a unique <strong>and</strong><br />

well-known alternative, namely with normally distributed log returns <strong>and</strong> lognormally<br />

distributed stock prices, i.e. with the classical Black/Scholes world. Despite this clear-cut<br />

characterization the uniquely determined alternative is open to criticism. Indeed, problems<br />

related to lognormal returns are addressed in the papers by Feldstein (1969), Tobin<br />

(1969), Levy (1974) <strong>and</strong> others. The class of lognormal distributions is not closed under<br />

the formation of linear blends. Even if all basic securities have lognormal discrete returns,<br />

then portfolio discrete returns will not be lognormal. In order to be consistent with<br />

expected utility, one has to take resort to the implausible quadratic utility functions.<br />

On the other h<strong>and</strong>, if assumptions 1 to 3 are called in question 7 one falls out of the frying<br />

pan into the fire. Dropping the i.i.d. assumptions 1 <strong>and</strong> 2 hampers statistical inference<br />

considerably. As already mentioned, dropping assumption 3 the shape of log return distributions<br />

corresponding to longer periods could be completely different from those corresponding<br />

to shorter periods. The conventional techniques for annualizing parameters<br />

(like volatility) become obsolete.<br />

It seems to the authors that no ideal model exists which at the same time<br />

• makes sense from the perspective of probability theory,<br />

• is founded on decision theoretic roots,<br />

• is h<strong>and</strong>y even for practitioners <strong>and</strong> students,<br />

• is not empirically rejectable.<br />

Especially the last point is of some importance in our context. If fat tails of the log return<br />

distribution were actually beyond doubt from an empirical point of view, then portfolio<br />

analysis <strong>and</strong> capital market theory in the traditional sense would have to be fully rejected,<br />

since even expected prices (<strong>and</strong> also µ <strong>and</strong> σ values) would not exist. In such a case it<br />

would not make much sense to try to estimate these parameters by statistical procedures.<br />

The fact that statistical estimation of discrete µ <strong>and</strong> σ values often works pretty fine, may<br />

be a hint that the log return distribution could be light-tailed.<br />

Perhaps truncated fat tail distributions, which still have a fat tail behavior up to a certain<br />

value but vanish beyond that point, are the key to a better underst<strong>and</strong>ing of the observed<br />

phenomenon that the central peak is higher <strong>and</strong> tails are “fatter” than those of the normal<br />

law. Rachev/Mittnik (2000), Section 2.5, give some hints in this direction. Truncated fat<br />

13


tail distributions belong to the domain of attraction of the normal law, which implies an<br />

approximately normal long-term distribution (of log returns). The latter fits the remarks<br />

of Poddig/Dichtl/Petersmeier (2000) that monthly log returns are closer to a normal distribution<br />

than those of weekly or daily data, which usually have “fat tails”. If this is true in<br />

the sense of Markowitz incompatibility then the consequence is that for these time intervals<br />

no traditional capital market theory can be applied, whereas for longer time intervals<br />

this still is possible.<br />

Endnotes<br />

1 See Dorfleitner (2001) in this regard.<br />

2 Compare, for instance, Adler/Feldman/Taqqu (1998).<br />

3 Since we are primarily interested in the finiteness of expected stock prices or expected (discrete)<br />

returns, only the upper tail matters as will be completely clear in the next section. For<br />

other purposes (e.g. calculation of VaR) the lower tail is more important. For many distributions<br />

(normal, student, Cauchy etc.) the upper <strong>and</strong> the lower tail have the same properties.<br />

4 For the definition of a slowly varying function see Ibragimov/Linnik (1971), p. 394ff. An<br />

equivalent definition of (10) is given through<br />

1 − F(tx)<br />

lim<br />

t→∞ 1 − F(t) = x−α (x > 0) ,<br />

where F denotes the distribution function. The parameter α > 0 is called the tail index, which<br />

can be interpreted as an inverse measure of fat tailedness. Note that the power moment E ( |R| β)<br />

exists only for β < α.<br />

5 Subexponential distributions have tails which decrease more slowly than any exponential tail.<br />

For some equivalent definitions of this class we refer the reader to Goldie/Klüppelberg (1998).<br />

6 Two distribution functions F <strong>and</strong> G are called tail-equivalent if<br />

1 − F(x)<br />

lim<br />

x→∞ 1 − G(x) = c<br />

for some constant c > 0.<br />

7 See for instance Mittnik/Paolella (2000), p. 313.<br />

14


Literature<br />

Adler, R.J.; Feldman, R.E.; Taqqu, M.S. (Eds.) (1998): A Practical Guide to Heavy <strong>Tails</strong>:<br />

Statistical Techniques <strong>and</strong> Applications, Boston/Basel/Berlin.<br />

Agnew, R.A. (1971): Counter-Examples to an Assertion Concerning the Normal Distribution<br />

<strong>and</strong> a New Stochastic Price Fluctuation Model, Review of Economic Studies<br />

38, 381–383.<br />

Blattberg, R.C.; Gonedes, N.J. (1974): A Comparison of Stable <strong>and</strong> Student Distributions<br />

as Statistical Models for Stock Prices, Journal of Business 47, 244–280.<br />

Bryson, M.C. (1982): Heavy-Tailed Distributions, in: Johnson, N.L.; Kotz, S.; Read,<br />

C.B. (Eds.): Encyclopedia of Statistical Sciences, Vol. 3, New York et al., 598–601.<br />

Chistyakov, V.P. (1964): A Theorem on Sums of Independent Positive R<strong>and</strong>om Variables<br />

<strong>and</strong> Its Application to Branching R<strong>and</strong>om Processes, <strong>Theory</strong> of Probability <strong>and</strong> Its<br />

Applications 9, 640–649.<br />

Dorfleitner, G. (2001): Stetige versus diskrete Renditen: Finanzmathematische<br />

Überlegungen zur richtigen Verwendung beider Begriffe in Theorie und Praxis (3.,<br />

wesentlich überarbeitete Auflage), Arbeitspapiere des Instituts für Statistik und Mathematische<br />

Wirtschaftstheorie der Universität Augsburg, Heft 174/1999.<br />

Embrechts, P.; Klüppelberg, C.; Mikosch, T. (1997): Modelling Extremal Events for<br />

Insurance <strong>and</strong> Finance, Berlin/Heidelberg/New York.<br />

Fama, E.F. (1965): Portfolio Analysis in a Stable Paretian <strong>Market</strong>, Management Science<br />

11, 404–419.<br />

Fama, E.F. (1971): Risk, Return <strong>and</strong> Equilibrium, Journal of Political Economy 79, 30–<br />

55.<br />

Feldstein, M.S. (1969): Mean Variance Analysis in the <strong>Theory</strong> of Liquidity Preference<br />

<strong>and</strong> Portfolio Selection, Review of Economic Studies 36, 5–12.<br />

Goldie, C.M.; Klüppelberg, C. (1998): Subexponential Distributions, in: Adler, R.J.;<br />

Feldman, R.E.; Taqqu, M.S. (Eds.): A Practical Guide to Heavy <strong>Tails</strong>: Statistical<br />

Techniques <strong>and</strong> Applications, Boston/Basel/Berlin, 435–459.<br />

Guillaume, D.M.; Dacorogna, M.M.; Davé, R.R.; Müller, U.A.; Olsen, R.B.; Pictet, O.V.<br />

(1997): From the Bird’s Eye to the Microscope: A Survey of New Stylized Facts of<br />

the Intra-daily Foreign Exchange <strong>Market</strong>s, Finance <strong>and</strong> Stochastics 1, 95–129.<br />

Ibragimov, I.A.; Linnik, Y.V. (1971): Independent <strong>and</strong> Stationary Sequences of R<strong>and</strong>om<br />

Variables, Groningen.<br />

Levy, H. (1974): The Rationale of the Mean-St<strong>and</strong>ard Deviation Analysis: Comment,<br />

15


American Economic Review 64, 434–441.<br />

M<strong>and</strong>elbrot, B.; Taylor, H.M. (1967): On the Distribution of Stock Price Differences,<br />

Operations Research 15, 1057–1062.<br />

Mittnik, S.; Paolella, M.S. (2000): Conditional Density <strong>and</strong> Value-at-Risk Prediction of<br />

Asian Currency Exchange Rates, Journal of Forcasting 19, 313–333.<br />

Podding, T.; Dichtl, H.; Petersmeier, K. (2000): Statistik, Ökonometrie, Optimierung,<br />

Bad Soden/Ts.<br />

Praetz, P.D. (1972): The Distribution of Share Price Changes, Journal of Business 45,<br />

49–55.<br />

Rachev, S.T.; Mittnik, S. (2000): Stable Paretian Models in Finance, Chichester et al.<br />

Rényi, A. (1966): Wahrscheinlichkeitsrechnung, Berlin.<br />

Resnick, S. (1998): A Probability Path, Basel-Boston.<br />

Rolski, T.; Schmidili, H.; Schmidt, V.; Teugels, J. (2000): Stochastic Processes for Insurance<br />

<strong>and</strong> Finance, Chichester et al.<br />

Samuelson, P.A. (1967): Efficient Portfolio Selection for Pareto-Levy Investments, Journal<br />

of Financial <strong>and</strong> Quantitative Analysis 2, 107–122.<br />

Shiryaev, A.N. (1999): Essentials of Stochastic Finance. Facts, Models, <strong>Theory</strong>, Singapore<br />

et al.<br />

Smith, C.W. (1976): Option Pricing: A Review, Journal of Financial Economics 3, 3–51.<br />

Taylor, J.L. (1986): Modelling Financial Time Series, Chichester et al.<br />

Tobin, J. (1969): Comment on Borch <strong>and</strong> Feldstein, Review of Economic Studies 36,<br />

13–14.<br />

16

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!