cutting edge.

OPTION PRICING

Information derivatives
Andrei Soklakov considers the problem of creating into account possible liquidity shortfalls – a serious concern in
setting complex hedging portfolios. In fact, the framework of
derivatives to provide tailored exposure to volatility risk. information derivatives forces us to clarify our liquidity
Information theory leads us to a whole class of such assumptions, and therefore acts as a safeguard against one of the
most common traps in volatility trading.
products. This class of ‘information derivatives’ includes We begin with a short introduction to the classical concept of
the standard volatility products (for example, variance information obtained from an observation of a random variable.
We then develop a simple framework describing the general
and gamma swaps) as special cases, which suggests evolution of a market variable from the point of view of an
individual trader. This allows us to calculate the amount of
the use of liquidity models as a basis for further information obtained by the trader from consecutive observations
refinement of standard volatility derivatives of the variable and introduce the corresponding concept of an
information derivative. As a particularly simple example, we
define an information swap as a forward contract on the amount
of information about the underlying variable revealed by
observations over a specified future period. We proceed by

It has long been recognised that the concept of

randomness is essential in describing the
evolution of asset prices. Traditionally, the amount of randomness
showing how standard volatility derivatives emerge within our
framework. This gives us the opportunity to illustrate the use of
our framework on particular known examples as well as highlight
in the returns of financial assets is measured by their volatility. the borderline where the extra flexibility of our framework can
Until recently, volatility was used merely as a descriptive concept be exploited. It turns out, for example, that variance, gamma
– an important theoretical characterisation of a tradable asset. The and conditional variance swaps can be viewed as information
situation has changed with the arrival of volatility derivatives such swaps that assume different liquidity models. The implied
as variance and gamma swaps that treat volatility as a tradable liquidity models are rather simplistic in all these cases, suggesting
asset in its own right (Jeffery (2005)). Recent blow-ups on the practical ways of tailoring a more manageable product. We
volatility markets have boosted an already growing demand for a conclude with a short summary of our framework and a further
larger set of financial instruments that provide a more tailored discussion of its applications.
exposure to the amount of randomness, or potential ‘surprise’, in
asset returns (Jung (2006b)). In this article, we introduce a broad General framework
class of such instruments that we call information derivatives. n Applied information theory. Let X = {x} be the set of possible
Just like volatility-based products, information derivatives outcomes for a random variable. We assume that all our knowledge
allow investors to hedge or get direct exposure to the amount of about the variable can be summarised in the form of probabilities
randomness in the underlying (market) variable. Thinking in {P(x)}x∈X for all possible outcomes of the variable. Upon
terms of information derivatives has the advantage of using a observation of a particular outcome, x, the amount of information
more flexible measure of randomness, which can be understood we receive, I(x), can therefore depend on x only via the
as the amount of information obtained by a trader from corresponding probability P(x). Mathematically:
continuous observations of the variable. This opens up possibilities
I ( x ) = f ( P ( x )) (1)
for using information derivatives wherever the classic notion of
volatility may not reflect the structure of the underlying random where f is some function that we require to be universal for all
processes (large deviations from lognormal behaviour, including random variables. If x and y are outcomes of two independent
in the credit, weather and emerging markets). Such flexibility random variables, the joint probability is P(x, y) = P(x)P(y). In
may be especially important in hedging periods of crisis when this case, we naturally require that the total information I(x, y)
standard volatility derivatives are believed to be particularly contained in the observation of both x and y should be the sum of
useful (Jung (2006b)) but may not be as manageable as we would I(x) and I(y):
like them to be.
f ( P ( x ) P ( y )) = f ( P ( x )) + f ( P ( y )) (2)
For a concrete example of where the framework of information
derivatives would be helpful, we should look no further than the Since this equation must hold for any pair of independent
relatively recent spikes in equity volatility that led to millions of random variables, we conclude that, up to an arbitrary
dollars in losses for some investors (Jung (2006a)). The situation multiplicative constant, f must coincide with the logarithmic
was exacerbated by the difficulties of hedging standard volatility function. The multiplicative constant (or, equivalently, the base of
instruments (variance swaps in particular), which have rather the logarithm) depends on the choice of information units and is
complicated replicating portfolios. We shall see that, in addition immaterial to our arguments. We follow the standard convention
to the view on volatility, information derivatives allow us to take and define1:

40 Nordic Risk Summer 2008

 context especially natural and straightforward. Incidentally, this
I ( x ) = − ln P ( x ) (3)
also turns out to be a useful exercise that forces us to describe the
The additivity property (2) is important not just because it dynamics of market variables in a more detailed way than is
results in a convenient mathematical framework: it provides a normally required for derivatives pricing.
crucial link between mathematics and practical intuition. In this Let t 0, t1, ... , tN be an increasing sequence of times and let
context, the concept of information takes a fundamental place: it x0, x1, ... , xN be the corresponding sequence of values for some
emerges as an essentially unique additive measure of surprise market variable, xi = x(ti). In the vast majority of financial
about random outcomes, which is universally applicable across a models, the relevant market variable is assumed to take values
very wide range of models. from the continuum of real numbers, xi ∈ R. As we have shown
Given a pair of random variables, let x and z be their in the previous section, such idealised models imply infinite
corresponding outcomes. In general, observation of z may information content for a typical scenario. We have also seen
provide us with some information about x. This manifests that this problem can be addressed by using mutual information,
itself as a change in the information content of x from the which requires two types of values for each market variable. In
original I(x) before the observation to I(x|z) = −ln P(x|z) after the following, we will refer to them as experienced and reference
the observation, where P(x|z) is the conditional probability of values. The following example definitions provide a good
observing x given z. The difference: practical illustration.
P ( x z) Let j be a trader who at the time ti makes a bid or offer by quoting
I ( z : x ) = I ( x ) − I ( x z ) = ln (4) a price of their choice. The trader may or may not be successful in
P(x) striking a deal for that price. The price is called ‘experienced’, and
gives us the amount of information ‘in z about x’. Because I(z : x) denoted zi(j), only if it refers to a specific successful transaction and
= I(x : z), that is, both z and x contain the same amount of only with respect to the particular trader who concluded the deal.
information about each other, the quantity (4) is also known as Different traders may follow different strategies, face different
mutual information. circumstances, or simply be more or less lucky. Other definitions of
In the case of continuous variables, the probability of observing experienced values can be introduced.
any particular outcome is infinitesimally small. According to (3), By reference values, we mean the values that are published with
the amount of information obtained from such an observation respect to a time period (for example, daily, monthly, etc.) by a
would be infinite. Indeed, a real number, for example, typically specialised agency as the overall (aggregate) experience of all
contains an infinite amount of information (infinitely many traders on the market. One can calculate, for example, the
decimal places) and the use of such numbers in real-life modelling arithmetic average:
(including finance) is merely a mathematical idealisation. 1 N
∑ zi(
j)
Information theory tells us that we need a more realistic model, a xi = (7)
N j =1
model that would not assume infinite accuracy.
One way of constructing such a model would be to refrain from where N is the number of participating traders, or consider other
using continuous variables. Fortunately, information theory ways of averaging across experiences of individual traders. An
presents us with a much better solution. We see from (4) that illustration of this approach for calculating reference values is given
mutual information can easily handle discrete as well as by the daily settlement price for a futures contract as quoted by the
continuous variables. Indeed, introducing probability densities Wall Street Journal. This is calculated as an average of all prices at
p(x), p(x|z) and an infinitesimal interval Δx, we obtain: which the contract is traded immediately before the end of daily
p ( x z ) ∆x p( x z) trading. It is important to note that the published quote for the
I ( z : x ) = ln = ln (5) daily settlement price appears in the press after the end of daily
p ( x ) ∆x p(x)
trading. Formally, we can publish xi only after the set {zi(j)}Nj = 1 is
We can also see that this flexibility of mutual information is not established. This retrospective publication is a fundamental
accidental and that mutual information is just the right tool for property of reference values of any market variable: the publication
real-life modelling. Using a physical analogy, the information we lag might be very small but it is always there.
get from raw measurement data z is always the information ‘in z
about x’, where x is the underlying physical quantity. In this setting, n Dynamics of market variables. We use the notation xi to
the effective accuracy of the measurement is limited by how well it denote the reference value for the variable at the time ti and use zi
reflects the underlying value of x, that is, by its correlation with x. for the corresponding experienced value (as seen by a particular
To make explicit reference to the background knowledge, such trader). As we have discussed, these values are correlated but not
as the results of previous observations, we can repeat the above necessarily identical. Without loss of generality, we can assume that
arguments for conditional probability densities, arriving at the the dependence between zi and xi is given by the conditional
conditional mutual information: probability density, p(zi|xi). Because p(zi|xi) describes the potential
p ( x z,b ) for disagreement between the experiences of the individual traders,
I ( z : x b ) = ln (6) we refer to p(zi|xi) as a model of market uncertainty. In addition, we
p( x b)
assume that, due to the retrospective nature of the publication
n Experienced versus reference values. Probabilistic modelling process, there is a small reporting lag so that the value of xi = x(ti)
is at the very heart of mathematical finance. This makes the becomes available only at the next time ti+1. Finally, we subscribe to
introduction of information-theoretic concepts in a financial the weak condition of market efficiency and restrict our attention
1
 ll logical components behind the detailed proof of (3) can be found in chapter 11 of Jaynes (2003),
A to (Markovian) market variables whose reference evolution at any
which is also available online. Although this book is not specifically designed as an introduction to
information theory, it reviews most relevant ideas with exceptional clarity, practical oversight and
given time can be described by a transition probability density
minimal mathematical overhead. p(xi|xi−1). We thereby arrive at the following extended model for the

risk.net/nordic 41
cutting edge. OPTION PRICING

dynamics of a Markovian market variable (as before, we formulate Because the process for xi is Markovian and because the values of
everything from the point of view of an individual trader): zk where k < i are irrelevant for prediction of xi (they become
n At any moment in time, ti, the market variable is described by obsolete as soon as xi−1 becomes known) we calculate:
two correlated values: the reference value xi and the experienced
value zi. n n p ( zi x i )
n Although both xi and zi refer to the same market variable and I t n = ∑ I ( zi : xi xi−1 ) = ∑ ln (13)
p ( zi xi−1 )
to the same time ti, only zi is known to the trader at ti. The i=1 i=1
_
reported reference value, xi, becomes known retrospectively at the One can also consider an average amount of information, I tn,
next time, ti+1, as an aggregate experience of all traders at ti. experienced by a typical trader. Separating the history hn into the
n The dependence between xi and zi is given by the conditional sequence of experienced and reference values, _ z1:n = z1, z2, ... , zn
probability density, p(zi|xi). The value zi is experienced by the trader and x0:n = x0, x1, ... , xn, one can define I tn as a conditional
given that the ‘fair’ market value for the variable in question is xi. expectation of Itn with respect to all possible z1:n given x0:n:
This probability density may change with time and be different for
different traders reflecting their individual circumstances.
I t n ( x 0:n ) = ∫ p ( z1:n x 0:n ) I t n ( z1:n , x 0:n ) dz1 Ldzn (14)
n The reference value of the market variable is assumed to follow where p(z1:m|x0:n) = p(z1|x1)p(z2|x2) ··· p(zn|xn). Using (13), this
a Markovian evolution. In other words, the probability density, definition can be rewritten as:
p(xi+1|xi), for the transition (xi, ti) → (xi+1, ti+1) depends only on
(xi, ti) and (xi+1, ti+1). I tn
Apart from the introduction of experienced values, this model n (15)
is identical to the standard one where we only ever use the =∑ ( ∫ p ( z x ) ln p ( z x ) dz − ∫ p ( z x ) ln p ( z
i i i i i i i i xi−1 ) dzi )
reference values. We see that the above model is completely i=1

specified by the reference dynamics p(xi+1|xi) and the model of _

market uncertainty p(zi|xi). In the following sections, we show Having defined Itn and I tn, we can
_ introduce the corresponding
how the definition of these two probability distributions translates derivatives contracts. Let Itn = I tn or Itn = Itn depending on
into contract specifications for information derivatives. whether we want our contract to be a derivative of the typical or
the actual experienced information. In either case, the procedure
n Realised information and its derivatives. We begin this for calculating Itn is specified by the dynamics p(xi+1|xi) and the
section by calculating the total amount of information experienced model of market uncertainty p(zi|xi) via (9), (13) and (15).
by the trader within the above framework. We then define the In the following, we will assume that all probabilities are
general concept of an information derivative and consider a calculated in the risk-neutral world. The fundamental theorem of
particularly simple example of an information swap. arbitrage-free pricing ensures safe use of the risk-neutral measure
As explained above, at the time ti the trader knows the past in our calculations. Indeed, under the penalty of arbitrage, any
reference value xi–1 and experiences zi, but the reference value xi transition xi−1 → xi that is possible in the real world must be
for ti remains unknown until the next time ti+1. Using the Bayes assigned a non-zero risk-neutral value p(xi|xi−1) (equivalence of
theorem, we obtain: risk-neutral and real-world measures).
Consider a European-style contract that upon maturity at tn
p ( zi xi ) p ( xi xi−1 ) pays the amount F(Itn). The present value of such a contract is
p ( xi zi , xi−1 ) = (8) given by:
p ( zi xi−1 )
t
where:  − ∫ n r ( τ )dτ 
p ( zi xi−1 ) = ∫ p ( zi xi′ )p ( xi′ xi−1 ) dxi′ (9)
V ( t 0 ) = E  e t0

F Itn ( )  (16)

The information in zi about xi given the value of xi–1 is: where r(t) is the risk-free rate and the risk-neutral expectation is
taken over all stochastic variables (including {zi}ni=1 where
p ( xi zi , xi−1 ) p ( zi x i ) appropriate). Generalisation to contracts with more complicated
I ( zi : xi xi−1 ) = ln = ln (10) exercise conditions (for example, American and Bermudan) as
p ( xi xi−1 ) p ( zi xi−1 )
well as path-dependent payouts (for example, barriers) can be
where we have used (8) for the last equality. This gives us the defined in the usual way.
amount of information received by the trader from their We introduce an information swap as a forward contract on
experience, zi, about the current ‘fair’ value, xi, while remembering Itn. In other words, the information swap is defined by the payout
the previous reference value xi–1. function:
By the time ti, the trader knows the history, hi−1, of all values up
to and including the time ti−1:
( ) (
Fswap I t n = A I t n − K ) (17)

hi−1 = ( x 0 ; x1 , z1 ; x 2 , z2 ;...; xi−1 , zi−1 ) where A is the nominal amount of the swap and K is the ‘fair
(11)
delivery value’. Setting the value of the swap to zero at time t 0,
The total amount of information, Itn, received by the trader about we obtain:
the consecutive values x1, x2, ... , xn in the process of experiencing
z1, z2, ... , zn is given by the sum: t
 − ∫ n r ( τ )dτ 
n
K = E  e t0 Itn  (18)
 
I t n = ∑ I ( zi : xi hi−1 ) (12)
i=1 For the sake of simplicity, we will also assume that the risk-free

42 Nordic Risk Summer 2008

rate r is deterministic. Using (10) and (13), together with the
∫ G ( x − α1 ,β1 ) ln G ( x − α 2 ,β 2 ) dx
2 2
linearity of the expectation, we therefore have:
(26)
−∫
tn
r ( τ )dτ −∫
tn
r ( τ )dτ
= − ln 2πβ 22 −
( α1 − α 2 )2 + β12
K=e t0
E  I t n  = e t0
E  I t n  (19) 2β 22
where the last expectation is the usual risk-neutral expectation (15) can now be calculated:
over the reference values, {xi}ni=1, of the underlying variable. It
now remains for us to specify the process for the reference values, n 
σ2
I t n = ∑  ln 1 + 2i δt i + σ i2
( δWi )2 − δti 
 (27)
p(xi|xi−1), and the model for the market uncertainty experienced
by the trader, p(zi|xi).
i=1 
 εi (
2 ε i2 + σ i2 δt i ) 

Decomposing the right-hand side of the above equation into

Gaussian market uncertainty and purely diffusive dynamics power series, keeping terms to the first order in δti, and taking the
Let zi be the experienced value of the market variable as seen by continuum limit (n → ∞, tn = T, δti → dt, (δWi)2 → dt), we obtain
the kth trader, zi = zi(k). (7) gives: a concise expression:

∑ zi(
j)
zi = Nxi − (20) ⌠ σ t2
T

j≠k IT = 
 2ε 2
dt (28)
⌡t t
where the sum is taken over the same range as in (7) except for 0

j = k. Assuming that the market is populated by a large number where the dependence of σt and et on time is now described using
N of independent statistically similar traders, and using the a continuous subscript t. In the case of constant et, the above
central limit theorem, we deduce from the previous expression expression is proportional to the realised variance. Realised
that p(zi|xi) is approximately Gaussian, that is: variance is currently used as the underlying of most pure volatility
derivatives: realised volatility, for example, is defined as the square
(
p ( zi xi ) ; G zi − xi , ε i2 ) (21) root of the realised variance. It follows that most pure volatility
_ be _viewed
derivatives can __ as derivatives on typical realised
where: information I T (or √ I T as in the case of a volatility swap). This
interpretation relies, of course, on the assumptions of purely
1  x2  diffusive dynamics (24) and flat market uncertainty (et = const).
(
G x, ε 2 = ) 2πε 2
exp  − 2 
 2ε 
(22) Relaxing either of these two assumptions leads us outside the
realm of pure volatility derivatives.
Within the model (21), the variance, e2i, can serve as a quantitative
measure of market uncertainty. This provides us with an n Examples: variance, gamma and conditional variance
opportunity to account for liquidity and other similar risks facing swaps. Substituting (28) into (19), we obtain:
the trader. In practice, one might replace ei with, for example, a
bid-offer spread or an alternative liquidity measure. It may also
T
∫t 0 r ( τ )dτ ⌠ T σ 2 
t
Ke ∝ E  dt  (29)
depend on xi. A monotonically decreasing function, ei(xi), for  
2
⌡t 0 ε t 
example, may help in modelling greater market uncertainty (big
ei), which frequently accompanies a sharp decline in the value of where we use the proportionality sign to indicate that constant
the underlying asset (small xi). factors, such as 1/T, which often appear in contract specifications,
In the purely diffusive model, the change of the reference values do not affect our arguments. In the case of constant market
δxi = xi − xi−1 is given by: uncertainty, et = const, this gives us, as expected, the fair delivery
value, Kvar, of a variance swap:
δxi = µ i δt i + σ i δWi (23)
T
∫t 0 r ( τ )dτ T
where μi and σi are (stochastic) functions of time as well
_ as_ other K var e ∝ E  ∫ σ t2 dt  (30)
(possibly stochastic) parameters, the value (δWi /√ δ ti) is  t 0 
distributed as a standard normal variable, and the time intervals The assumption of constant market uncertainty is a dangerous
are small enough so that μi and σi can be treated as constants one. When hedging a short volatility position in a variance swap
within any particular interval. The above model approximates with a portfolio of European-style vanilla options, we effectively
most of the pure diffusion processes considered in the financial make a bet that the underlying will stay in the area where liquidity
literature. For example, setting xi = lnSi, where Si is the price of variations can be safely ignored. Such an assumption will almost
an asset at the time ti, and taking the continuum limit, we recover certainly backfire if, for example, the market falls significantly,
the standard case of the locally lognormal price evolution. (23) moving into the area where vanilla options are illiquid. In this
can be rewritten as: case, we will find that most of our hedging portfolio is far out-of-
(
p ( xi xi−1 ) = G xi − xi−1 − µ i δt i , σ i2 δt i ) (24) the-money and the rest is too patchy to provide a reasonably
accurate hedge.
Using the Gaussian model for the market uncertainty (21), the We can improve on the assumption of constant market
integral in (9) can be calculated analytically to give: uncertainty by considering a better liquidity model. For
example, as we mentioned earlier, a monotonically decreasing
(
p ( zi xi−1 ) = G zi − xi−1 − µ i δt i , ε i2 + σ i2 δt i ) (25) function of the underlying asset value might be a better simple
model for et (see the paragraph following equations (21) and
Since for any constant α1, α2 , β1 and β2: (22)). Let xt = lnSt, where St is the underlying asset value at time

risk.net/nordic 43
cutting edge. OPTION PRICING

t. The simple choice e2t = 1/St leads us to the expression: Because market uncertainty is closely related to liquidity, this
T approach would also help with hedging during volatile periods
∫t 0 r ( τ )dτ T
Kγe ∝ E  ∫ St σ t2 dt  (31) (Jung (2006a)).
 t 0  For the sake of completeness, it is worth mentioning that, in
where Kγ is immediately recognised as the fair value of a gamma practice, it is often impossible to make direct references to
swap. Conditional variance swaps (Jung (2006b)) can be information such as that given, for example, by equation (28).
introduced in a completely analogous manner. For an up-var The standard solution to this is to use estimators. The integral on
swap, for example, we would have: the right-hand side of (30), for example, is often replaced with the
sum Sni=1[ln(Si/Si−1)]2 , where Si is the underlying stock price on
 const if St > S
*
day i. Together with the proportionality factor, 252/T, this leads
ε t2 =  (32) to the standard contract definition for the variance swap.
 ∞ otherwise
Similarly, the integral on the right-hand side of (31) does not
where S* is a pre-agreed level of the underlying spot market. We appear directly in the term sheet of a gamma swap. Instead, we
see that the seller of volatility in an up-var swap takes the would find an estimator, for example, [252/(TS 0)]Sni=1Si[ln(Si/
conservative view of assuming complete absence of liquidity Si−1)]2 . The choice of estimators becomes easier in the context of a
below the ‘crisis’ threshold (St < S*) and consequently refusing to particular product. If, however, the reader is interested in general
offer any payout in such ‘crisis’ periods. Above the threshold, estimators of information-theoretical quantities, Paninski (2003)
however, we are back in the carefree world of the variance swap. might be a good place to start. In general, once an expression for
the information measure is derived, for example, (28), and an
Concluding remarks appropriate estimator is proposed, basic contract definitions
Classical information theory introduces information as a simple become essentially straightforward and the relevant derivatives
intuitive measure of uncertainty in a random variable. This product is born.
measure is defined by its single fundamental property: it is Once the product is born, one must perform all standard risk
additive for independent random variables. We have shown how analysis in the usual way before trying to trade it. In the case of
to use this measure to quantify the uncertainty of market information swaps, this analysis can be done in the standard way
variables. In doing so we arrived at the notion of information using Carr & Madan’s (1998) replication approach. This approach
derivatives, which can be seen as a generalisation of many is general enough to cover large classes of smooth as well as
volatility-based derivatives. We have presented a simple discontinuous models of market uncertainty (see, for example,
mathematical framework for constructing information derivatives. Carr & Lewis (2004), Demeterfi et al, (1999) and Carr & Lee
The framework needs just two definitions: the type of dynamics (2003) and consider log St, St log St and Snt types of payout). n
of the underlying variable and a model of market uncertainty.
One can think of these two definitions as the stochastic process Andrei Soklakov has recently joined Goldman Sachs after working for
and the liquidity model for the underlying market variable. It is Lombard Risk, where this paper was completed. He would like to thank
important to stress that we take both of these definitions as given. Lombard Risk for supporting his research, with special thanks to Keith
After all, identifying the driving factors as well as liquidity Mitchell and Lee Wakeman for their interest in this work and many helpful
modelling for a market variable should be done regardless of discussions. Detailed feedback from two anonymous referees is greatly
whether or not we want to consider information derivatives for appreciated. The views expressed herein should not be considered as
that variable. Our framework tells us what information derivatives investment advice or promotion. They represent personal research of the
would look like for any given pair of such models. author and do not necessarily reflect the view of his employers, or their
Many volatility products can be viewed as information associates or affiliates. Email: [email protected]
derivatives. What would happen if we engage in trading volatility-
based products that have nothing to do with information
derivatives or, more realistically, imply poor liquidity models References
when viewed as information derivatives? Under realistic market
Carr P and R Lee, 2003 Jaynes E, 2003
conditions, such products would appear to be based on a measure Robust replication of volatility Probability theory: the logic of science
of randomness, which is significantly different from information. derivatives Cambridge University Press
The additivity property (2) for such a measure would be broken, Working paper, available at
allowing for non-linear interaction among independent factors. www.math.nyu.edu\research\carrp\papers Jeffery C, 2005
The evolution of variance
Although, in principle, there is nothing wrong with that, the Carr P and K Lewis, 2004 Risk November, pages 20–22
practical management of such products may become unnecessarily Corridor variance swaps
complicated. To illustrate this point, we considered the well- Risk February, pages 67–72 Jung J, 2006a
Vexed by variance
known examples of variance, gamma and conditional variance Carr P and D Madan, 1998 Risk August, pages 41–43
swaps for which the detailed comparative analysis is readily Towards a theory of volatility trading
available. We showed that easier-to-manage products, such as In Volatility: New Estimation Jung J, 2006b
Techniques for Pricing Derivatives, Volatility – viva la variance
gamma and conditional variance swaps, have more realistic edited by R Jarrow, Risk Publications, Risk October, pages 43–46
liquidity assumptions compared with their less manageable pages 417–427
sibling – the variance swap. Paninski L, 2003
Demeterfi K, E Derman, M Kamal Estimation of entropy and mutual
As we have shown, the difference between currently used and J Zou, 1999 information
volatility derivatives boils down to different models of market A guide to variance swaps Neural Computation 15,
uncertainty. Thinking in terms of market uncertainty would Risk June, pages 54–59 pages 1,191–1,253
therefore be the simplest way of fine-tuning the existing products.

44 Nordic Risk Summer 2008