Modeling Tail Dependence Using Copulas - Literature Review: Jan de Kort March 15, 2007
Modeling Tail Dependence Using Copulas - Literature Review: Jan de Kort March 15, 2007
Modeling Tail Dependence Using Copulas - Literature Review: Jan de Kort March 15, 2007
Jan de Kort
1
Contents
1 Bivariate Copulas 5
1.1 Sklars theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Frechet-Hoeffding bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Survival copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Multivariate Copulas 13
2.1 Sklars theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Frechet-Hoeffding bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3 Dependence 15
3.1 Linear correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Measures of concordance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.1 Kendalls tau . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.2 Spearmans rho . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.3 Ginis gamma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Measures of dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.1 Schweizer and Wolffs sigma . . . . . . . . . . . . . . . . . . . . . . . . 21
3.4 Tail dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.5 Multivariate dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2
6 Option Pricing with Copulas 32
6.1 Bivariate digital options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.2 Rainbow options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7 Project plan 34
B Implied quantities 41
B.1 Implied distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
B.2 Implied volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3
Introduction
Suppose we want to price a best-of option on corn and wheat, that is, a contract that
allows us to buy a certain amount of wheat or corn for a predetermined price (strike) on a
predetermined date (maturity). If the corn price is high, this is possibly due to unfavourable
weather conditions, also causing the wheat harvest to be bad. High corn prices thus are likely
to be observed together with high wheat prices. This phenomenon is called (upper) tail
dependence.
In commodities pricing it is common to model the dependence structure between assets using
a Gaussian copula. Copulas are a way of isolating dependence between random variables
(such as asset prices) from their marginal distributions. In section 4.2.1 it will be shown
that the Gaussian copula does not have tail dependence. This may cast some doubt on the
appropriateness of this model in case of the corn and wheat option, for the probability of
both crops having high prices will be underestimated by the Gaussian copula. This leads to
underpricing of the best-of contract.
Malevergne and Sornette [1] show that a Gaussian copula might indeed not always be a
feasible choice. They succeed in rejecting the hypothesis of the dependence between a number
of metals traded on the London Metal Exchange being described by a Gaussian copula.
Still, copulas provide a convenient way to model the dependence between assets, since the
marginal distributions of the underlyings can be dealt with separately so that their properties
(such as volatility smile) can be preserved.
The central question in this project is how to incorporate tail dependence in the pricing
of hybrid products. A first step would be to consider best-of contracts on two underlyings.
This literature review seeks to give an overview of the theory involved in answering the above
question. Sections 1 and 2 explain what copulas are and how they relate to multivariate
distribution functions. In section 3 it is described what kind of dependence is captured
by copulas. This, among other things, includes measures of concordance like Kendalls tau
and Spearmans rho. Next, section 4 summarizes the properties of a number of well-known
parametric families of copulas. Some calibration techniques are outlined in section 5. Section
6 finally describes how prices of hybrid contracts (e.g. digital options, best-of contracts) can
be expressed in terms of copulas.
Appendix A provides a brief introduction to derivatives pricing, in particular the role that
martingales play in this. Since the marginal distributions of the hybrid derivatives should
be chosen as to incorporate volatility smile, a short note on implied volatility / implied
distribution can be found in appendix B.
4
1 Bivariate Copulas
This section introduces copulas and describes how they relate to multivariate distributions
(i.e. Sklars theorem, section 1.1). Section 1.2 introduces the Frechet-Hoeffding upper and
lower bounds for copulas. It is also explained what it means for a multivariate distribution if
its copula is maximal or minimal. Finally, in section 1.3, survival copulas are defined which
will be useful in the discussion of tail dependence later on.
The extended real line R {, +} is denoted by R.
Definition 1.2 Suppose b1 = max S1 and b2 = max S2 exist. Then the margins F and G
of H are given by
F : S1 R, F (x) = H(x, b2 ),
G : S2 R, G(y) = H(b1 , y).
Definition 1.3 Suppose also a1 = min S1 and a2 = min S2 exist. H is called grounded if
H(a1 , y) = H(x, a2 ) = 0
and
H(x2 , y2 ) H(x2 , y1 ) H(x1 , y2 ) H(x1 , y1 ) (2)
for every [x1 , x2 ] [y1 , y2 ] S1 S2 . By setting x1 = a1 in (1) and y1 = a2 in (2) it can be
seen that
H( , y2 ) H( , y1 ),
H(x2 , ) H(x1 , ).
From lemma 1.4 it follows that (1) and (2) also hold in absolute value. Adding up these
inequalities and applying the triangle inequality (Figure 1) yields in particular
5
Figure 1: Schematic proof of lemma 1.5. Apply 2-increasingness to rectangle I and II and
combine the resulting inequalities via the triangle inequality. For the absolute value bars, use
that H is nondecreasing in both arguments (lemma 1.4).
Definition 1.7 A (two dimensional) copula is a subcopula whose domain is [0, 1]2 .
Remark 1.8 Note that reformulating lemma 1.5 in terms of subcopulas immediately leads to
the Lipschitz condition
6
Definition 1.10 The absolutely continuous component AC and the and the singular
component SC of the density are defined as
Z uZ v 2
C(s, t)
AC (u, v) = ds dt,
0 0 uv
SC (u, v) = C(u, v) AC (u, v).
The theorem under consideration in this section, due to Sklar in 1959, is the very reason
why copulas are popular for modeling purposes. It says that every joint distribution with
continuous margins can be uniquely written as a copula function of its marginal distributions.
This provides a way to seperate the study of joint distributions into the marginal distributions
and their joining copula.
Following Nelsen [2], we state Sklars theorem for subcopulas first, the proof of which is short.
The corresponding result for copulas follows from a straightforward, but elaborate, extension
that will be omitted.
Definition 1.13 If the derivative of the CDF of X exists, it is called the probability density
function (pdf ) of X.
Definition 1.14 Let X and Y be random variables. The joint distribution function of
X and Y is
H(x, y) := P[X x, Y y].
The margins of H are F (x) := limy H(x, y) and G(y) := limx H(x, y).
7
Lemma 1.16 Let H be a joint distribution function with margins F and G. Then there
exists a unique subcopula C 0 such that
and
H(x, y) = C 0 (F (x), G(y)) (3)
for all (x, y) R.
Proof For C 0 to be unique, every (u, v) RanF RanG should have only one possible image
C 0 (u, v) that is consistent with (3). Suppose to the contrary that C10 (u, v) 6= C20 (u, v) are both
2
consistent with (3), i.e. there exist (x1 , y1 ), (x2 , y2 ) R such that
2-increasingness:
Let u1 u2 be in Ran F and v1 v2 in Ran G. As CDFs are nondecreasing, there exist
unique x1 x2 , y1 y2 with F (x1 ) = u1 , F (x2 ) = u2 , G(y1 ) = v1 and G(y2 ) = v2 .
Remark 1.17 The converse of lemma 1.16 also holds: every H defined by (3) is a joint
distribution. This follows from the properties of a subcopula.
8
Theorem 1.18 (Sklars theorem) Let H be a joint distribution function with margins F
2
and G. Then there exists a unique 2-copula C such that for all (x, y) R
Proof Lemma 1.16 provides us with a unique subcopula C 0 satisfying (4). If F and G are
continuous, then RanF RanG = I 2 so C := C 0 is a copula. If not, it can be shown (see [2])
that C 0 can be extended to a copula C.
The converse is a restatement of remark 1.17 for copulas.
Now that the connection between random variables and copulas is established via Sklars
theorem, let us have a look at some implications.
Proof
Theorem 1.20 The continuous random variables X and Y are independent iff. their copula
is C (u, v) = uv.
C is called the product copula.
1.2 Fr
echet-Hoeffding bounds
In this section we will show the existence of a maximal and a minimal bivariate copula, usually
refered to as the Frechet-Hoeffing bounds. All other copulas take values in between these
bounds on each point of their domain, the unit square. The Frechet upper bound corresponds
to perfect positive dependence and the lower bound to perfect negative dependence.
9
1 1
0.75 1 0.75 1
0.5 0.5
0.8 0.8
0.25 0.25
0 0.6 0 0.6
0 0
0.2 0.4 0.2 0.4
0.4 0.4
0.6 0.2 0.6 0.2
0.8 0.8
10 10
Proof From lemma 1.4 we have C 0 (u, v) C 0 (u, 1) = u and C 0 (u, v) C 0 (1, v) = v, thus the
upper bound.
VH ([u, 1] [v, 1]) 0 gives C 0 (u, v) u + v 1 and VH ([0, u] [0, v]) 0 leads to C 0 (u, v) 0.
Combining these two gives the lower bound.
Plots of C + and C are provided in Figures 2 and 3. The remaining part of this section is
devoted to the question under what condition these bounds are attained.
2
Definition 1.22 A set S := S1 S2 R is called nondecreasing if for every (x1 , y1 ), (x2 , y2 )
S it holds that x1 < x2 y1 y2 . S is called nonincreasing if x1 > x2 y1 y2 .
An example of a nondecreasing set can be found in Figure 4.
Definition 1.23 The support of a distribution function H is the complement of the union
of all open subsets of R2 with H-measure zero.
Remark 1.24 Why not define the support of a distribution as the set where the joint density
function is non-zero?
1. The joint density does not necessarily exist.
2. The joint density can be non-zero in isolated points. These isolated points are not included in
definition 1.23.
Let X and Y be random variables with joint distribution H and continuous margins F : S1
2
R and G : S1 R. Fix (x, y) R . Suppose H is equal to the Frechet upper bound, then
either H(x, y) = F (x) or H(x, y) = G(y). On the other hand
It follows that either P[X x, Y > y] or P[X > x, Y y] is zero. As suggested by Figure 5
this can only be true if the support of H is a nondecreasing set.
This intuition is confirmed by the next theorem, a proof of which can be found in Nelsen [2].
10
Figure 4: Example of a nondecreasing set.
Theorem 1.25 Let X and Y be random variables with joint distribution function H.
H is equal to the upper Frechet-Hoeffding bound iff. the support of H is a nondecreasing
2
subset of R .
H is equal to the lower Frechet-Hoeffding bound iff. the support of H is a nonincreasing
2
subset of R .
Remark 1.26 If X and Y are continuous random variables, then the support of H cannot
have horizontal or vertical segments. Indeed, suppose the support of H would have a horizontal
line segment, then a relation of the form 0 < P[a X b] = P[Y = c] would hold, implying
that the CDF of Y had a jump at c.
Thus, in case of continuous X and Y , theorem 1.25 implies the support of H to be an al-
most surely increasing (decreasing) set iff. H is equal to the upper (lower) Frechet-Hoeffding
bound.
Remark 1.27 The support of H being an almost surely (in)(de)creasing set means that if
you observe X, there is only one Y that can be observed simultaneously, and vice versa.
Intuitively, this is exactly the notion of perfect dependence.
Every bivariate copula has a survival copula associated with it that gives the probability of
two random variables both to exceed a certain value.
11
Figure 5: In case of non-perfect positive dependence, the shaded area always contains points
with nonzero probability.
Indeed, C is a copula:
C(0, v) = 0 + v 1 + C(1, 1 v) = v 1 + 1 v = 0,
C(1, v) = 1 + v 1 + C(0, v) = v + 0 = v,
12
2 Multivariate Copulas
n
Definition 2.2 Let x, y R such that x y holds component-wise. Define the n-box [x, y]
by
[x, y] := [x1 , y1 ] [x2 , y2 ] . . . [xn , yn ].
The set of vertices ver([x, y]) of [x, y] consists of the 2n points w that have wi = xi or wi = yi
for 1 i n. The product
2 n
Y
sgn(w) := sgn(2wi xi yi )
i=1
13
2.1 Sklars theorem
2.2 Fr
echet-Hoeffding bounds
In the multidimensional case, the upper bound is still a copula, but the lower bound is not.
The following example, due
1 to Schweizer
1 and
Sklar [3], shows that C does not satisfy equation
(5). Consider the n-box 2 , 1 . . . 2 , 1 . For 2-increasingness, in particular, the H-volume
of this n-box has to be nonnegative. This is not the case for n > 2:
n o 1
max 1 + . . . + 1 n + 1, 0 n max + 1 + . . . + 1 n + 1, 0
| {z } 2
| {z }
=nn+1=1
= 12 +(n1)n+1= 12
n 1 1 1 1
+ max + + 1 + . . . + 1 n + 1, 0 + . . . . . . max + . . . + n + 1, 0
2 2 2 2 2
| {z } | {z }
=0 =0
n
= 1 .
2
On the other hand, for every u I n , n 3, there exists a copula C such that C(u) = C (u)
(see Nelsen [2]). This shows that a sharper lower bound does not exist.
14
3 Dependence
The dependence structure between random variables is completely described by their joint
distribution function. Benchmarks like linear correlation capture certain parts of this de-
pendence structure. Apart from linear correlation, there exist several other measures of
association. These, and their relation to copulas, are the subject of this section.
One could think of measures of association as one dimensional projections of the dependence
structure onto the real line. Consider for instance Scarsinis [4] definition:
Dependence is a matter of association between X and Y along any measurable function, i.e.
the more X and Y tend to cluster around the graph of a function, either y = f (x) or x = g(y),
the more they are dependent.
From this definition it is clear that there exists some freedom in how to define the extent
to which X and Y cluster around the graph of a function. In the following, some of the
standard interpretations of this freedom will be described.
Section 3.1 explains the concept of linear correlation. It measures how well two random
variables cluster around a linear function. A major shortcoming is that linear correlation is
not invariant under non-linear monotonic transformations of the random variables.
The concordance and dependence measures (e.g. Kendalls tau, Spearmans rho) introduced in
sections 3.2 and 3.3 reflect the degree to which random variables cluster around a monotone
function. This is a consequence of these measures being defined such as only to depend on the
copula see definition 3.5(6) and copulas are invariant under monotone transformations
of the random variables.
Finally, in section 3.4 dependence will be studied in case the involved random variables
simultaneously take extreme values.
From now on the random variables X and Y are assumed to be continuous.
Definition 3.1 For non-degenerate, square integrable random variables X and Y the linear
correlation coefficient is
Cov[X, Y ]
= 1
(Var[X]Var[Y ]) 2
Correlation can be interpreted as the degree to which a linear relation succeeds to describe
the dependency between random variables. If two random variables are linearly dependent,
then = 1 or = 1.
Example 3.2 Let X be a uniformly distributed random variable on the interval (0, 1) and
set Y = X n , n 1. X and Y thus are perfectly positive dependent.
The n-th moment of X is Z 1
n 1
E [X ] = xn dx = . (7)
0 1+n
15
The linear correlation between X and Y is
E [XY ] E [X] E [Y ]
= 1 1
(E [X 2 ] E [X]2 ) 2 (E [Y 2 ] E [Y ]2 ) 2
E X n+1 E [X] E [X n ]
= 1 1
(E [X 2 ] E [X]2 ) 2 (E [X 2n ] E [X n ]2 ) 2
(7)
3 + 6n
= .
2+n
(i). The linear correlation coefficient is not invariant under increasing, non-linear trans-
forms.
(ii). Random variables whose joint distribution has nondecreasing or nonincreasing support
can have correlation coefficient different from 1 or 1.
Definition 3.4
(i). Two observations (x1 , y1 ) and (x2 , y2 ) are concordant if x1 < x2 and y1 < y2 or if
x1 > x2 and y1 > y2 . An equivalent characterisation is (x1 x2 )(y1 y2 ) > 0. The
observations (x1 , y1 ) and (x2 , y2 ) are said to be discordant if (x1 x2 )(y1 y2 ) < 0.
(ii). If C1 and C2 are copulas, we say that C1 is less concordant than C2 (or C2 is more
concordant than C1 ) and write C1 C2 (C2 C1 ) if
In the remaining part of this section we will only consider bivariate copulas. Part (ii) of
definition 3.4 is then equivalent to C1 (u, v) C2 (u, v) for all u I 2 .
3. X,Y = Y,X ,
16
4. if X and Y are independent then X,Y = C = 0,
Lemma 3.6
(i). Measures of concordance are invariant under strictly monotone transformations of the
random variables.
(ii). A measure of concordance assumes its maximal (minimal) value if the support of the
joint distribution function of X and Y contains only concordant (discordant) pairs.
Note that these properties are exactly opposite to the conclusions in corollary 3.3 on the linear
correlation coefficient. The linear correlation coefficient thus is not a measure of concordance.
In the remaining part of this section, two concordance measures will be described: Kendalls
tau and Spearmans rho.
Let Q be the difference between the probability of concordance and discordance of two inde-
pendent random vectors (X1 , Y1 ) and (X2 , Y2 ):
In case (X1 , Y1 ) and (X2 , Y2 ) are iid. random vectors, the quantity Q is called Kendalls
tau .
17
Given a sample {(x1 , y1 ), (x2 , y2 ), . . . , (xn , yn )} of n observations from a random vector (X, Y ),
an unbiased estimator for is
cd
t := ,
c+d
where c is the number of concordant pairs and d the number of discordant pairs in the sample.
Nelsen [2] shows that if (X1 , Y1 ) and (X2 , Y2 ) are independent random vectors with (possibly
different) distributions H1 and H2 , but with common margins F , G and copulas C1 , C2
ZZ
Q=4 C2 (u, v) dC1 (u, v) 1. (9)
I2
It follows that the probability of concordance between two bivariate distributions (with com-
mon margins) minus the probability of discordance only depends on the copulas of each of
the bivariate distributions.
Note that if C1 = C2 := C, then, since we already assumed common margins, the distributions
H1 and H2 are equal which means that (X1 , Y1 ) and (X2 , Y2 ) are identically distributed. In
that case, (9) gives Kendalls tau for the iid. random vectors (X1 , Y1 ), (X2 , Y2 ) with copula C.
Furthermore it can be shown that
ZZ
C(u, v) C(u, v)
=14 du dv. (10)
I2 u v
In the particular case that C is absolutely continuous, the above relation can be deduced via
integration by parts.
As an example of the use of (10), consider
Lemma 3.7 C = C .
Proof
ZZ
C C
C = 14 du dv
I 2 u v
ZZ
C C
= 14 [1 ][1 ] du dv
I2 u v
ZZ
C C
= C 4 [1 ] du dv. (11)
I 2 u v
The second term of the integrand of (11) reduces to
ZZ Z 1 Z 1 Z 1
C 1
du dv = C(1, v) C(0, v) dv = C(1, v) dv = v dv = .
I 2 u 0 0 0 2
Similarly, ZZ
C 1
du dv = .
I2 v 2
Substituting in (11) yields the lemma.
Scarsini [4] proves axioms (1)(7) of definition 3.5 hold for Kendalls tau.
The next lemma does not hold for general concordance measures.
18
Lemma 3.8 Let H be a joint distribution with copula C.
C = C + iff. = 1,
C = C iff. = 1.
Proof We will prove the first statement, C = C + iff. = 1, via the following steps:
(i) = 1 H has nondecreasing support
(ii) H has nondecreasing support H = C +
(iii) H = C+ = 1
Step (ii) is immediate from theorem 1.25. Step (iii) follows from substitution of C + in formula
(9) and straightforward calculation. This step in fact also follows from axiom (6) in definition
3.5.
It remains to show that = 1 implies H having nondecreasing support. Suppose therefore
that
1 = = P[(X1 X2 )(Y1 Y2 ) > 0] P[(X1 X2 )(Y1 Y2 ) < 0].
Clearly,
P[(X1 X2 )(Y1 Y2 ) < 0] = 0. (12)
Now suppose that (x1 , y1 ) and (x2 , y2 ) are disconcordant and lie in the support of H. Inte-
grating (12) over I 2 yields
ZZ
0 = P[(X1 X2 )(Y1 Y2 ) < 0] du dv
2
ZZI
= P[(X2 x1 )(Y2 y1 ) < 0|(X1 , Y1 ) = (x1 , y1 )] dH(u, v)
2
ZZI n o
= P[X2 > x1 , Y2 < y1 |(X1 , Y1 ) = (x1 , y1 )] + P[X2 < x1 , Y2 > y1 |(X1 , Y1 ) = (x1 , y1 )] dH(u, v).
I2
It follows that there is no probability mass in the regions {(u, v) : u > x1 , v < y1 } and
{(u, v) : u < x1 , v > y1 }. In particular
P[Br (x1 , y1 )] = 0,
where r := 21 | min{x1 x2 , y1 y2 }| and Br (x, y) denotes an open 2-ball with radius r and
centre (x, y). Apparently (x1 , y1 ) is not in the complement of the union of open sets having
zero probability and therefore not in the support of H. This contradicts our assumption and
proves (i).
Let (X1 , Y1 ), (X2 , Y2 ) and (X3 , Y3 ) be iid. random vectors with common joint distribution H,
margins F , G and copula C. Spearmans rho is defined to be proportional to the probability
of concordance minus the probability of discordance of the pairs (X1 , Y1 ) and (X2 , Y3 ):
| {z } | {z }
Joint distr. H Independent
19
Note that X2 and Y3 , being independent, have copula C . By (9), three times the concordance
difference between C and C is
ZZ
S = 3 4 C(u, v)dC (u, v) 1
2
ZZ I
= 12 C(u, v) du dv 3. (13)
I2
Spearmans rho satisfies the axioms in definition 3.5 (see Nelsen [2]).
Let X F and Y G be random variables with copula C, then Spearmans rho is equiv-
alent to the linear correlation between F (X) and G(Y ). To see this, recall from prob-
ability theory that F (X) and G(Y ) are uniformly distributed on the interval (0, 1), so
E [F (X)] = E [G(Y )] = 1/2 and Var[F (X)] = Var[G(Y )] = 1/12. We thus have
(13)
S = 12 E [F (X), G(Y )] 3
Cherubini et al. [5] states that for Spearmans rho a statement similar to lemma 3.8 holds:
C = C iff. S = 1.
Whereas Spearmans rho measures the concordance difference between a copula C and inde-
pendence, Ginis gamma C measures the concordance difference between a copula C and
monotone dependence, i.e. the copulas C + and C . Using (9) this reads
ZZ ZZ
C = C(u, v)dC (u, v) + C(u, v)dC + (u, v)
I2 I2
Z 1 Z 1
= 4 C(u, 1 u)du u C(u, u) du .
0 0
Ginis gamma thus can be interpreted as the area between the secondary diagonal sections of
C and C , minus the area between the diagonal sections of C + and C.
20
3.3 Measures of dependence
2. 0 X,Y 1
3. X,Y = Y,X ,
6. if and are strictly monotone functions on Ran X and Ran Y respectively, then
X,Y = (X),(Y ) ,
(i). Concordance measures assume their maximal (minimal) values if the concerning random
variables are perfectly positive (negative) dependent. Dependence measures assume
their extreme values if and only if the random variables are perfectly dependent.
(ii). Concordance measures are zero in case of independence. Dependence measures are zero
if and only if the random variables under consideration are independent.
(iii). The stronger properties of dependence measures over concordance measures go at the
cost of a sign, or, in the words of Scarsini [4]:
[..] dependence is a matter of assocation with respect to a (strictly) monotone function
(indifferently increasing or decreasing). [..] Concordance, on the other hand, takes into
account the kind of monotonicity [..] the maximum is attained when a strictly increasing
relation exists [..] the minimum [..] when a relation exists that is strictly monotone
decreasing.
Schweizer and Wolffs sigma for two random variables with copula C is given by
ZZ
C = 12 |C(u, v) uv| dudv.
I2
Nelsen [2] shows this association measure to satisfy the properties of definition 3.9.
21
3.4 Tail dependence
Definition 3.10 Given two random variables X F and Y G with copula C, define the
coefficients of tail dependency
C(u, u)
L := lim P[F (X) < u|G(Y ) < u] = lim , (14)
u0 u u0
1 2u + C(u, u)
U := lim P[F (X) > u|G(Y ) > u] = lim . (15)
u1 u1 1u
C is said to have lower (upper) tail dependence iff. L 6= 0 (U 6= 0).
The interpretation of the coefficients of tail dependency is that it measures the probability of
two random variables both taking extreme values.
Lemma 3.11 Denote the lower (upper) coefficient of tail dependency of the survival copula
C by L (U ), then
L = U ,
U = L .
Proof
C(u, u) C(1 v, 1 v) 1 2v + C(v, v)
L = lim = lim = lim = U
u0 u v1 1v v1 1v
1 2u + C(u, u) 2v 1 + C(1 v, 1 v) C(v, v)
U = lim = lim = lim = L
u1 1u u0 v u0 v
C is differentiable in both (0, 0) and (1, 1), this is a sufficient condition for the limits (14)
and (15) to exist:
dC
L = (0)
du
d 2
= 2 u2 1 u=0 = 0 ,
= u
du u=0
dh u)
i dC
U = L = 2 u 1 + C(1 =2 (1)
du u=0 du
d 2
= 2 u = 2 [ 2 u2 1 ]u=1 = 2 2 .
du u=1
So the Gumbel copula has no lower tail dependency. It has upper tail dependency iff. 6= 1.
22
3.5 Multivariate dependence
Most of the concordance and dependence measures introduced in the previous sections have
one or more multivariate generalizations.
Joe [6] obtains the following generalized version of Kendalls tau. Let X = (X1 , . . . , Xm ) and
Y = (Y1 , . . . , Ym ) be iid. random vectors with copula C and define Dj := Xj Yj . Denote
by Bk,mk the set of m-tuples in Rm with k positive and m k negative components. A
generalized version of Kendalls tau is given by
m
X
C = wk P( (D1 , . . . , Dm ) Bk,mk )
k=b m+1
2
c
(i). C = 1 if C = C + ,
(ii). C = 0 if C = C ,
(i). wm = 1,
Pm m
(wk := wmk for k < b m+1
(ii). k=0 wk k =0 2 c).
The implication of (iii) is more involved (see [6], p. 18), though it is clear that at least
wm wm1 . . . wb m+1 c should hold.
2
For m = 3 the only weights satisfying (i)(iii) are w3 = 1 and w2 = 13 . The minimal value
of for m = 3 thus is 31 . For m = 4 there exists a one-dimensional family of generalizations
of Kendalls tau.
In terms of copulas, Joes generalization of Spearmans rho ([6], pp. 2224) for a m-multivariate
distribution function having copula C reads
Z Z
m
C = ... C(u) du1 . . . dum 2 (m + 1)1 + 2m .
Im
Properties (i) and (ii) are taken care of by the scaling and normalization constants and can be
checked by substituting C + and C . The increasingness of with respect to is immediate
from definition 3.4(ii).
There also exist multivariate measures of dependence. For instance, Nelsen [2] mentions the
following generalization of Schweizer and Wolffs sigma:
2m (m + 1)
Z Z
C = m ... |C(u1 , . . . , um ) u1 . . . um | du1 . . . dum ,
2 (m + 1) Im
where C is an m-copula.
23
4 Parametric families of copulas
This section gives an overview of some types of parametric families of copulas. We are
particularly interested in their coefficients of tail dependence.
The Frechet family (section 4.1) arises by taking affine linear combinations of the product
copula and the Frechet-Hoeffding upper and lower bounds. Tail dependence is determined by
the weights in the linear combination.
In section 4.2 copulas are introduced which stem from elliptical distributions. Because of
their symmetric nature, upper and lower tail dependence coefficients are equal.
Any function satisfying certain properties (described in section 4.3) generates an Archimedean
copula. These copulas can take a great variety of forms. Furthermore, they can have dis-
tinct upper and lower tail dependence coefficients. This makes them suitable candidates for
modeling asset prices, since in market data either upper or lower tail dependence tends to be
more profound.
Multivariate Archimedean copulas however are of limited use in practice as all bivariate
margins are equal. Therefore in section 4.4 an extension of the class of Archimedean copulas
will be discussed that allows for several distinct bivariate margins.
4.1 Fr
echet family
Every affine linear combination of copulas is a new copula. This fact can be used for instance
to construct the Frechet family of copulas
Indeed, the weight p (of C ) has negative sign and q (of C + ) has positive sign.
The Frechet family has upper and lower tail dependence coefficient q.
24
4.2 Elliptical distributions
Elliptical distributions are distributions whose density function (if it exists) equals
1
f (x) = || 2 g[(x )T 1 (x )], x Rn ,
f (tx)
lim = t .
x f (x)
In words: regularly varying functions have tails that behave like power functions.
Whether or not the generator being regularly varying is a necessary condition for tail depen-
dence is still an open problem, but Schmidt proves that to have tail dependency, at least one
bivariate margin must be O-regularly varying, that is it must satisfy
f (tx) f (tx)
0 < lim inf lim sup < ,
f (x) f (x)
x x
for every t 1.
where Z x Z y 2 2
1 2sts t
(x, y) = p e 2(12 ) ds dt
2 1 2
and denotes the standard normal CDF.
The Gaussian copula generates the joint standard normal distribution iff. u = (x) and
v = (y), that is iff. the margins are standard normal.
Gaussian copulas have no tail dependency unless = 1. This follows from Schmidts charac-
terisation of tail dependent elliptical distributions, since the density generator for the bivariate
Gaussian distribution ( 6= 1) is not O-regularly varying:
g(tx) 1
lim = lim exp{ x(t 1)} = 0, t 1.
u g(x) u 2
25
4.2.2 Bivariate Students t copula
Let t denote the central univariate Students t distribution function, with degrees of
freedom: Z x
(( + 1)/2) s2 +1
2
t (x) = 1+ ds,
(/2)
where is Euler function and t, , [0, 1], the bivariate distribution corresponding to t :
x y
s2 + t2 2st +2
Z Z
1
2
t, (x, y) = 1+ ds dt.
(1 2 )
p
2 1 2
T, (u, z) = t, (t1 1
(u), t (v)).
It follows that the Students t distribution has tail dependence for all > 0.
Every continuous, decreasing, convex function : [0, 1] [0, ) such that (1) = 0 is a
generator for an Archimedean copula. If furthermore (0) = +, then the generator is
called strict. Parametric generators give rise to families of Archimedean copulas.
Define the pseudo-inverse of as
1 (u),
0 u (0),
[1] =
0, (0) u .
The Gumbel copula from example 3.12 is Archimedean with generator (u) = ( log(u)) ,
[1, ). Some other examples are listed in table 1.
26
Table 1: One-parameter Archimedean copulas. The families marked with * include C , C
and C + .
The Frechet-Hoeffding lower bound C is Archimedean ((u) = 1 u), whereas the Frechet-
Hoeffding upper bound is not. To see this, note that [1] is strictly decreasing on [0, (0)].
Clearly, 2(u) > (u), so we have for the diagonal section of an Archimedean copula that
Nelsen [2] shows that if is a stict generator, then also (t ) (interior power family) and
[(t)] (exterior power family) are strict generators for (0, 1] and 1. If is twice
differentiable, then the interior power family is a strict generator for all > 0. Two-parameter
families of Archimedean copulas can now be constructed by taking
, = [(t )]
For = 1 this is (part of) the one-parameter Clayton family see table 1.
27
4.3.3 Multivariate Archimedean copulas
dk 1 (t)
(1)k 0 for all t [0, ), k = 1, . . . , n (18)
dtk
then
C A (u1 , . . . , un ) = 1 ((u1 ) + . . . + (un ))
is an n-copula.
For example, the generator (t) = t 1 ( > 0) of the bivariate Clayton family has inverse
1
1
(t) = (1 + t) which is readily seen to satisfy (18). Thus,
1
C (u1 , . . . , un ) = u
1 + . . . + u n n + 1
is a family of n-copulas.
It can be proven (see [10]) that Laplace transforms of distribution functions () satisfy (18)
and (0) = 1. The inverses of these transforms thus are a source of Archimedean n-copulas.
Archimedean n-copulas have practical restraints. To begin with, all k-margins are identical.
Also, since there are usually only two parameters, Archimedean n-copulas are not very flexible
to fit the n dimensional dependence structure. Furthermore, Archimedean copulas that have
generators with complete monotonic inverse, are always more concordant than the product
copula, i.e. they always model positive dependence.
There exist extensions of Archimedean copulas that have a number of mutually distinct
bivariate margins. This is discussed in the next section.
If we would choose the generator of the inner and the outer copula to be different, would
the composition then still be a copula? In other words, for which functions , is
28
a copula? If it is, then the (1,2) bivariate margin is different from the (1,3) margin, but the
(2,3) margin is equal to the (1,3) margin.
For n = 1, 2, . . . , , consider the following function classes:
n dk (t) o
Ln = : [0, ) [0, 1] (0) = 1, () = 0, (1)k 0 for all t [0, ), k = 1, . . . , n ,
dtk
n dk (t) o
Ln = : [0, ) [0, ) (0) = 0, () = , (1)k1 0 for all t [0, ), k = 1, . . . , n ,
dtk
Note that if 1 L1 , then is a strict generator for an Archimedean copula.
It turns out that if , L1 and 1 L , then (19) is a copula. For general n-copulas
similar conditions exist. In the n-dimensional case, n 1 of the 12 n(n 1) bivariate margins
are distinct.
29
5 Calibration of Copulas from Market Data
This section is concerned with the question which member of a parametric family of copulas
fits best to a given set of market data.
Consider a stochastic process {Yt , t = 1, 2, . . .} taking values in Rn . Our data consists in a
realisation {(x1t , . . . , xnt ) : t = 1, . . . , T } of the vectors {Yt , t = 1, . . . , T }.
The latter expression is often computationally more convenient. The vector that maximizes
l() is called the maximum likelihood estimator (MLE):
l()
=0
30
are possible candidates for MLE . But these solutions can also be local maxima, minima
or inflection points. On the other hand, maxima can occur at the boundary of (or if
|||| ), in discontinuity points and in points where the likelihood is not differentiable.
For joint distributions satisfying some regularity conditions, it can be shown (Shao [12]) that
if the sample size increases, the subsequent MLEs converge to a limit. This property is called
consistency.
The log-likelihood (21) consists of two positive parts. Joe and Xu [13] proposed the set
of parameters to be estimated in two steps: first the margins parameters and then the
copulas. By doing so, the computational cost of finding the optimal set of parameters reduces
significantly. This approach is called the Inference for the Margins (IFM) method.
T X
X n
1 = argmax log fi (xit ; 1 )
2 t=1 i=1
T
X
2 = argmax log c(F1 (x1t ), . . . , Fn (xnt ); 1 , 2 )
1 t=1
Set IFM := (1 , 2 ) to be the IFM estimator. In general, this estimator will be different from
the MLE, but it can be shown that the IFM estimator is consistent.
In the Canonical Maximum Likelihood (CML) method first the margins are estimated using
empirical distributions F1 , . . . , Fn . Then, the copula parameters are estimated using an ML
approach:
XT
CML := argmax log c(F1 (x1t ), . . . , Fn (xnt ); ).
t=1
31
6 Option Pricing with Copulas
In this section, existing pricing techniques for (bivariate) digital and rainbow options with
dependent underlyings will be restated in terms of copulas. Following Cherubini et al. [5], a
martingale approach will be used. An outline of martingale pricing is given in appendix A.4.
We will need a reformulation of Sklars theorem in terms of conditional copulas. Conditional
copulas are defined along the same lines as in section 1, details can be found in Patton [14].
Theorem 6.1 (Sklars theorem for conditional copulas) Let X|Ft F (|Ft ), Y |Ft
G(|Ft ) be random variables with joint conditional distribution H(X, Y |Ft ). Then there exists
a copula C such that
H(x, y|Ft ) = C(F (x|Ft ), G(y|Ft ) | Ft ).
Conversely, given two conditional distributions F (|Ft ) and G(|Ft ) the composition C(F (|Ft ), G(|Ft ))
is a conditional joint distribution function.
A digital option is a contract that pays the underlying asset or one unit cash if the price of
the underlying is above or below some strike level at maturity. Digital options paying a unit
cash are called cash-or-nothing (CoN) options, while those paying the asset are called
asset-or-nothing (AoN) options.
Consider two digital cash-or-nothing call options DC1 and DC2 with respective strikes K1 ,
K2 , written on the assets S1 , S2 having copula CLL . If at maturity Si > Ki , then DCi pays
one unit cash, i = 1, 2. The prices of these digital call options are
where Q is the risk-neutral measure associated with taking the discount bond B(t, T ), paying
one unit cash at maturity, as the numeraire.
Analogously, for i = 1, 2 , the prices of the digital cash-or-nothing put options DP1 and DP2
with strikes K1 , K2 , written on the assets S1 , S2 are given by
From (22) and (23) it follows that the value of a portfolio containing a digital call and a
digital put equals the price of the discount bond.
Letting CHH denote the survival copula to CLL , the price of a bivariate digital call option
paying one unit if both S1 > K1 and S2 > K2 , is
In the general bivariate case, one can no longer distinguish between put or call. Instead,
subscripts will be added to describe the payoff. For example, H(igher) L(ower) means that
32
one unit will be paid if the price of the first underlying is above a strike level and the second
is beneath a (possibly different) strike level. The price of this option in fact follows from the
sigma-additivity of Q:
DHL = DC1 DHH
DC1 DC2
= DC1 B(t, T )CHH ,
B B
DC1 DC1 DC2
= B CHH ,
B B B
DC1 DP2
= B CHL , ,
B B
where CHL (u, v) := u CHH (u, 1 v) satisfies the properties of a copula. In the last step the
relation DC1 + DP1 = B is used.
Using a similar argument one can show
DP1 DC2
DLH = B CLH , ,
B B
with CLH (u, v) := v CHH (1 u, v) the survival copula to CHL .
Finally, using the definition of survival copulas
DLL DC1 DC2 DC1 DC2
= 1 + CHH ,
B B B B B
DP1 DP2
= CLL , .
B B
Rainbow options, also known as best-of options, are multivariate contingent claims whose
underlying asset is the maximum or minimum in a set of assets.
Consider for example a put option on the maximum Z(T ) := max{S1 (T ), S2 (T )} of two
assets:
X (Z(T )) = max{0, K Z(T )}.
33
7 Project plan
The goal of the project is to examine whether in modeling the dependence structure of con-
tracts on tail dependent underlyings, the use of copulas different from the Gaussian copula
affects the price.
Although the final goal is to price contracts on any number of underlyings, a starting point
could be to fit a one-parameter bivariate copula to a given set of data using maximum like-
lihood. The margins will eventually have to be modeled parametrically so that volatility
smile can be incorporated, but if we choose a two-step approach in the maximum likelihood
estimation (IFM, section 5.2), the same estimation procedure can also be applied to empirical
margins (i.e. CML, section 5.3). To test the appropriateness of the resulting model, the
likelihood of observing the data in the new model can be compared to the case of a Gaus-
sian copula. Another obvious check would be to compare sample versions of Kendalls tau,
Spearmans rho and the TDCs to the corresponding quantities as implied by the fitted copula.
Furthermore, the copula estimate should not change too much in time and also sensitivity to
the parameters of the marginal distributions (SABR, displaced diffusion) has to be acceptable.
A next step could be to consider a mix of copulas, at least one of which has tail dependence.
Maximum likelihood estimation now is more involved, since there are several parameters to
fit. Hu (2004) suggests using the expectation maximization algorithm to carry out the second
step in the IFM method.
The calibrated mix of copulas could then be implemented in a pricing model to see if the new
copula indeed affects the price significantly.
Possible further steps include extension of the model to three or more dependent assets (since
many products have more than two underlyings) and performing a hedge test.
34
A Basics of derivatives pricing
defined on the probability space (, F, P). Also, let B(t) be the money account. For the time
being, assume the interest rate to be deterministic.
Let FtS denote the sigma algebra generated by S over the interval [0, t]:
Intuitively, an event belongs to the sigma algebra generated by S over [0, t] if, from the
trajectory of S(t) over [0, t], it is possible to decide whether the event has occured or not.
S -measurable random variable X .
A T-claim is any FT
Question: what should be the price (t; X ) of the T-claim X at time t?
To be able to assign a price to a derivative, the market is assumed to be arbitrage free, i.e.
it is not possible to make a risk-free profit. The next characterisation of risk-free markets will
be used extensively throughout this section.
Consider two assets driven by the same Wiener process:
dS1 = 1 dt + 1 dW,
dS2 = 2 dt + 2 dW.
35
The no arbitrage condition thus entails the market price of risk to be equal for all
assets in a market that are driven by the same Wiener process. This characterisation
will be used in section A.3 to derive the Black-Scholes fundamental PDE.
Note that for the above argument to be valid, the assets have to be tradable and the market
must be liquid, i.e. assets can be bought and sold quickly. Furthermore, it must to be possible
to sell a borrowed stock (short selling). It is also assumed that there are no transaction costs,
no taxes and no storage costs.
A proof can be found in Steele [15]. Particularly useful is the case when f (x, y) = x/y:
Corollary A.2 (Itos division rule) Assume the processes X(t) and Y (t) to follow the
dynamics
The price of a contingent claim can be recovered by solving the fundamental PDE associated
with the model.
36
As an example, consider the Black-Scholes model consisting of two assets with the following
dynamics:
dB(t) = rB(t)dt,
dS(t) = S(t)dt + S(t)dW (t). (25)
where F is a smooth function. Applying Itos formula to (26) and omitting arguments:
d = + dW,
Ft + SFS + 12 2 S 2 FSS
= ,
F
SFS
= .
F
No arbitrage implies the market price of risk to be the same for all assets driven by the same
Wiener process:
r r
= = ,
so
Ft + (r + )SFS + 21 2 S 2 FSS SFS
= =r+ = r + .
F F
This yields, after rearranging terms, the fundamental PDE for the Black-Scholes model:
Ft + rSFS + 12 2 S 2 FSS = rF ,
37
then there exists a measure Q, equivalent to P, such that
dQ RT 1 T 2
R
= e 0 (t) dW (t) 2 0 (t) dt ,
dP
dW (t) = (t)dt + dW Q (t).
P
Girsanovs theorem implies the existence of a new measure Q such that dt+dW P is a Wiener
process:
dS(t) = rS(t)dt + S(t)dW Q
Under the new measure, the instantaneous rate of return on the asset equals r:
Q dS(t)
E = rdt.
S(t)
Note that the risk-neutral measure Q only depends on the market price of risk ,
which is the same for all assets in the market. Thus, under Q, all assets in the market
have instantaneous rate of return equal to the instantaneous yield r of the risk free asset B.
From Itos division rule A.2 it follows that the process S(t)/B(t) has zero drift. If the volatility
of this process satifies the Novikov condition (27), then zero drift implies S(t)/B(t) to be a
martingale1 . Pricing formula (28) is an immediate consequence of this.
A measure like Q under which the prices of all assets in the market discounted by the risk-
neutral bond, are martingales, is called an equivalent martingale measure. Equivalent
means that P and Q agree on the same zero sets.
Theorem A.4 (First Fundamental Pricing Theorem) If a market model has a risk-
neutral probability measure, then it does not admit arbitrage.
Theorem A.5 (General pricing formula) The arbitrage free price process for the T-claim
X is given by " # " #
(t; X ) (T ; X ) X
= EQ Ft = EQ Ft (28)
S0 (t) S0 (T ) S0 (T )
1
In general, zero drift does not imply a stochastic process to be a martingale. The implication holds under
an extra condition, see [17] p. 79. For exponential martingales, this condition is equivalent to the (more
practical) Novikov condition.
38
where Q is the (not necessarily unique) martingale measure for the market S0 , S1 , . . . Sn with
S0 as the numeraire.
Suppose you have a Wiener process to which you add a drift term. How do you have to change
the measure to make the resulting Brownian motion a (driftless) Wiener process again under
the altered measure? This question was answered by Girsanovs theorem.
Now suppose you have a martingale that you multiply by a (positive) stochastic process. The
next lemma describes how you have to change the measure if you want the resulting process
to be a martingale again under a new measure.
Lemma A.6 (Change of num eraire) Assume that Q0 is a martingale measure for the
numeraire S0 (on FT ) and assume that S1 is a positive asset price process such that S1 (t)/S0 (t)
is a Q0 martingale. Define Q1 on FT by the likelihood process
S0 (0) S1 (t)
L10 (t) = , 0 t T. (29)
S1 (0) S0 (t)
Then Q1 is a martingale measure for the numeraire S1 .
Proofs of theorems A.4, A.5 and lemma A.6 can be found in Bjork [16].
A zero-coupon bond is an asset that pays one unit currency at maturity T . The price
p(t, T ) of such a bond at time t thus reflects the amount needed to ensure one unit currency
in the future.
Definition A.8 The risk-neutral martingale measure that arises from choosing the zero-
coupon bond with maturity T as the numeraire in lemma A.6 is called the T-forward mea-
sure QT .
The change of numeraire lemma A.6 provides us with a Radon-Nikomdym derivative
S0 (0) p(s, T )
LTQ = (30)
p(0, T ) S0 (s)
relating Q to QT . It follows that
thm. A.5
" #
(s) S0 (s) (s) S0 (s) Q (t)
= = E Fs
p(s, T ) p(s, T ) S0 (s) p(s, T ) S0 (t)
h i h i
S0 (0) Q (t) (30) Q LT (t) F
p(0,T ) E S0 (t) Fs E Q p(t,T ) s
= =
p(s,T ) S0 (0) LTQ
p(0,T ) S0 (s)
Bayes form.
" #
(t)
= ET Fs ,
p(t, T )
39
where ET denotes integration w.r.t. QT . In particular, as p(T, T ) = 1:
Lemma A.10 QT is equal to the risk-neutral measure associated with the fixed money account
iff. the interest rate r is deterministic.
Proof The two measures QT and Q being equal implies their Radon-Nikodym derivative to
be one. From equation (30) it can be seen that this is equivalent with the relation p(t, T ) =
p(0, T )S0 (t) to hold for all 0 t T . In an arbitrage free market with stochastic interest
rate such a relation cannot hold since it implies you can always exchange a position in the
bond for a position in the money account and vice versa, at no cost. If, on the other hand, the
interest rate is deterministic, then p(t, T )/S0 (t) 6= p(0, T ) for some t clearly leads to arbitrage
opportunities.
Lemma A.11 (GemanEl KarouiRochet) Given a financial market with stochastic short
rate r and a strictly positive asset price process S(t), consider a European call on S with ma-
turity T and strike K, i.e. a T -claim X = max{0, S(T ) K}. The option price is
Here QT denotes the T -forward measure and QS is the martingale measure for the numeraire
S(t)
process S(t). Under the assumption that the process p(t,T ) has deterministic volatility
S,T (t), equation (31) reduces to
where
q
d1 = d2 + 2S,T (T ), (33)
S(0) 1 2
log( Kp(0,T ) ) 2 S,T (T )
d2 = q , (34)
2S,T (T )
Z T
2
S,T (T ) = ||S,T (t)||2 dt. (35)
0
(36)
40
B Implied quantities
A market model and observed data (such as stock prices, option prices) together form an
overcomplete model. By removing the assumptions on some of the input parameters, the new
model can be used to retrieve the implied quantities, that is, implied by the market data
and by the assumptions in the model that were not removed.
This idea can for instance be used to check whether assumptions on input parameters are
reasonable.
How are the prices of the assets in the market distributed? The distribution cannot be
observed directly from market data, but it can be reconstruced from actual prices under
specific assumptions. These assumptions entail a certain market model to hold except
maybe for the distribution of the prices such that, together with the observed data, the
distribution is implied by the model. The implied distribution is sometimes referred to as the
implied measure.
As an example, consider the Black-Scholes model of Example A.3. The solution of the
SDE (25) is given by
1
S(t) = S(0) exp{( 2 )t + W (t)}.
2
Thus, under Black-Scholes, the price S(t) is lognormally distributed for all t. In practise
though, the distribution appears to be different. In particular, the tails of the distribution,
i.e. the probability of extreme values of the stock, are thicker than assumed in Black-Scholes.
Breeden and Litzenberger [18] proposed a way to reconstruct what would be the distribution,
if lognormality of the returns does not hold, but all the other assumptions of Black-Scholes
are satisfied. Denote CALL(S, t; T, K) the price of a call option with maturity T and strike
K on an asset S. Let Q be the EMM associated with the money account and let QS be the
martingale measure for the price process of the call option with numeraire S(t). From the
GemanEl Karoui-Rochet lemma (A.11), assuming furthermore constant volatility (2S,T =
2 ) and constant interest rate (QT = Q, lemma A.10), we have:
CALL(S, t; T, K)
er(T t) = Q(S(T ) K | Ft ). (37)
K
Similarly, for the price P U T (S, t; T, K) of a put option with underlying S, maturity T and
strike K it can be shown that
P U T (S, t; T, K)
er(T t) = Q(S(T ) K | Ft ). (38)
K
41
B.2 Implied volatility
One can also assume the Black-Scholes model to be correct, except for the constant volatility.
Assuming thus the pricing formula for a call option to be correct, one could for instance find
the volatility for which the Black-Scholes price coincides with the market price. Doing so
for different values of the strike yields the so called volatility smile, i.e. the effect that in
practise volatility is not constant, but relatively higher for extreme strike prices. Calculating
the implied volatility for different strikematurity pairs gives what is know as the volatility
surface.
Dupire [19] showed that under risk neutrality, there is a unique local volatility function
(t, T ) consistent with the implied distribution from the previous section. The implied volatil-
ity, however, is much easier to infer from the market and it also facilitates the handling of
time varying distributions.
42
References
[1] Y. Malevergne and D. Sornette. Testing the gaussian copula hypothesis for financial
assets dependences. Working paper, Institute of Geophysics and Planetary Physics,
2001.
[3] B. Schweizer and A. Sklar. Probabilistic Metric Spaces. Elsevier Science, 1983.
[5] U. Cherubini, E. Luciano, and W. Vecchiato. Copula Methods in Finance. Wiley, 2004.
[7] R. Schmidt. Tail dependence for elliptically contoured distributions. Math. Methods of
Operations Research, 55:301327, 2002.
[8] A.W. Marshall and I. Olkin. Families of multivariate distributions. Journal of the
American Statistical Association, 83:834841, 1988.
[10] W. Feller. An introduction to probability theory and its applications, vol. 2. Wiley, 1971.
[11] H. Joe. Multivariate models and dependence concepts. Chapman and Hall, 1997.
[13] H. Joe and J.J. Xu. The estimation method of inference functions for margins for multi-
variate models. Technical Report 166, Dept. of Statistics University of British Columbia,
1996.
[14] A.J. Patton. Modelling time-varying exchange rate dependence using the conditional
copula. Discussion paper, University of California, 2001.
[15] J.M. Steele. Stochastic Calculus and Financial Applications. Springer, 2000.
[16] T. Bjork. Arbitrage Theory in Continuous Time. Oxford University Press, 2004.
43