440 Stochastic volatility modeling
11.4 The correlation swap
Consider a basket of stocks or indexes. A correlation swap of maturity
T
pays at
T
the average pairwise realized correlation of the basket components minus a xed
strike bρ
1
n(n −1)
X
i6=j
ρ
ij
− bρ (11.32)
where
bρ
is set so that the swap’s initial value vanishes. Correlation
ρ
ij
of
S
i
,
S
j
is
dened with the standard realized correlation estimator, using daily log-returns:
ρ
ij
=
Σr
i
k
r
j
k
q
Σ r
i
k
2
q
Σ r
j
k
2
(11.33)
where
r
i
k
= ln(S
i
k
/S
i
k−1
)
and the sums runs from
k = 1
to
k = N
, where
N
is the
number of returns used for estimating covariances and variances in (11.33).
n
is the number of securities: 2 or 3 when the components are indexes and up to
50 for a correlation swap on the constituents of the Euro Stoxx 50 index. We have
used in (11.32) the typical equal weighting.
Strike
bρ
is also called the implied correlation of the swap. Indeed, in a constant
volatility model, for daily returns, with all spot/spot correlations
ρ
SS
equal and a
large number of returns – thus a long maturity – bρ = ρ
SS
.
For shorter maturities, (11.33) is biased. Let us make the assumption of centered
log-returns:
r
i
k
= σ
i
√
∆tZ
i
k
, where
σ
i
is the (constant) volatility of asset
i
,
∆t
the
interval between two spot observations – here one day, and the
Z
i
k
are iid standard
normal random variables.
The bias and standard deviation of the correlation estimator are derived in
Appendix A, at lowest order in
1
N
:
E[ρ
ij
] = ρ
SS
1 −
1 −ρ
2
SS
2N
(11.34)
Stdev (ρ
ij
) =
1
√
N
1 −ρ
2
SS
(11.35)
where
N
is the number of returns in the historical sample. Typically the bias in
E[ρ
ij
] is about one point of correlation for ρ
SS
= 60% and T = 1 month.
Correlation swaps were introduced as a means of trading correlation and making
implied correlation an observable parameter.
As a measure of correlation, however, averaging pairwise correlations results in
a poorly dened estimator.
One would expect of an adequately dened estimator that its standard deviation
vanishes either in the limit of a large sample size (
N → ∞
) or as the number of