Professional Documents
Culture Documents
net/publication/46585947
CITATIONS READS
2 1,581
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Richard Warnung on 27 May 2014.
Abstract
In practice daily volatility of portfolio returns is transformed to longer
holding periods by multiplying by the square-root of time which assumes
that returns are not serially correlated. Under this assumption this pro-
cedure of scaling can also be applied to contributions to volatility of the
assets in the portfolio. Trading at exchanges located in distant time zones
can lead to significant serial cross-correlations of the returns of these assets
when using close prices as is usually done in practice. These serial correla-
tions cause the square-root-of-time rule to fail. Moreover volatility contri-
butions in this setting turn out to be misleading due to non-synchronous
correlations. We address this issue and provide alternative procedures for
scaling volatility and calculating risk contributions for arbitrary holding
periods.
1010 Vienna, Austria, and lecturer at Vienna University of Technology, Financial and Actu-
arial Mathematics, e-mail: richard.warnung@rcm.at resp. rwarnung@gmx.at. The contents of
this paper are the authors’ sole responsibility. They do not necessarily represent the views of
Raiffeisen Kapitalanlage-Gesellschaft m. b. H.
‡ Both authors thank the members of the risk management department of Raiffeisen
1
Building a portfolio using the weights λ the random portfolio return at time
t ∈ Z, denoted by Xt (λ), is given by
n
X
Xt (λ) := λT Xt = λi Xti , for t ∈ Z. (1.1)
i=1
For our analysis we will assume that (Xt )t∈Z and therefore, as we show in
Proposition 2.6, also (Xt (λ))t∈Z is a weakly stationary multivariate respectively
univariate time series. Thus mean and covariances exist and covariances depend
on the lag and not the absolute time points (for details see for example Box et al.
[2008], Brockwell and Davis [1986]).
In practice it is assumed that the component time series in (Xt )t∈Z do not
exhibit any serial correlation. Thus it is usually assumed that yesterday’s return
of asset i is not correlated with today’s return of asset j. But there are situations
where this assumption is clearly not fulfilled. Holding a globally diversified
portfolio of stock-index futures and analyzing the returns between close prices
one sometimes experiences significant correlations in lagged returns. This may
be due to non-overlapping trading times of various exchanges (e.g. New York
and Tokyo). For a detailed analysis of the correlation bias we refer to Kahya
[1997], Coleman [2007]. This problem can not be addressed by taking prices from
a point in time where all exchanges involved are open as such a moment might
simply not exist. This lead-lag effect is a prime example for the application of
multivariate time series analysis, see for example [McNeil et al., 2005, Chapter
4, Section 5]. In the following we give a concrete example and motivate our
focus on vector-autoregressive models and especially the VAR(1) case.
Example 1.1. Consider a portfolio of stock index futures traded in Japan, China,
Europe, South Africa and the United States. The following table shows the
names of stock indices in these regions, example weights and the closing times
of the exchanges in CET. For such a portfolio there is no point in time where
all exchanges are open. Using close prices at all exchanges one has to treat
the lagged correlations between the returns of different assets as it is illustrated
below.
The correlations of the returns1 of the indices underlying these futures con-
tracts are plotted in Figure 1 where the numbering corresponds to the numbering
in the table above. At lag zero we see that geographically close markets have
relatively strong correlation in returns. One recognizes an Asian (TOP,HSH), a
European (DJS,SWI) (also showing significant correlations with South Africa)
1 Estimated from a 250 days window ending in April 2010.
2
and an American block (RUS,NAS). In Figure 1 the correlations of the returns
(y-axis) to lagged returns (x-axis) are plotted. The auto-correlations on the
main diagonal (i.e. correlations of returns of one asset with lagged returns of
the same asset) are not significant which is consistent with a random walk. But
on the other hand this plot illustrates that, in our example, returns in Asian
markets have significant correlation with lagged returns of the European as well
as the US block. This illustrates the lead-lag effect. We will come back to this
portfolio in Section 3 and apply advanced procedures to analyze risk.
We continue with a rather heuristic analysis of the situation in the spirit of
Kahya [1997], Coleman [2007] and Bergomi [2010]. Let (B̃t )t∈R = (B̃t1 , . . . , B̃tn )t∈R
be an n−dimensional Brownian motion with the identity matrix as correlation
matrix, i.e.
Cov(B̃ti , B̃tj ) = 0, if i 6= j.
where we assume that assets are ordered with respect to their closing times,
i.e. x1 ≤ x2 ≤ · · · ≤ xn . If x1 = x2 · · · = xn then we can shift time by x1 by
stationarity of Brownian motion and get back the standard model with Σ as
the covariance matrix of the asset returns. Assuming without loss of generality
that xi > xj , i.e. that market i closes later than market j then we get for the
general case
[t − 1 + xj , t + xi ] = [t − 1 + xj , t − 1 + xi ) ∪ [t − 1 + xi , t + xj ) ∪ [t + xj , t + xi ]
3
for the Brownian motion. Thus the covariance matrix of Xt is given by
n
ΣX := (σi,j (1 − |xi − xj |))i,j=1 . (1.2)
Note that on the main diagonal we have σi,i = σi2 , i.e. variances are not biased
but covariances are.
Furthermore we consider the lag one covariance matrix of Xt . Let
Γ(1) = Cov(Xt+1 , Xt )
such that
i
Γ(1)i,j = Cov(Xt+1 , Xtj ), for i, j = 1, . . . , n.
With the assumed ordering of the markets Γ(1) is an upper triangular matrix
with zero main diagonal.
Finally note that the vector (Xt+1 , Xt ) is jointly Gaussian and the conditional
law of Xt+1 given Xt is Gaussian with mean Γ(1)Σ−1 X Xt and covariance matrix
ΣX − Γ(1)Σ−1 T
X Γ(1) .
with Φ = Γ(1)Σ−1R and a process (Zt )t∈Z which is a white noise process by
construction. This observation motivates our focus on vector-autoregressive
models in general and VAR(1)-models for the problem at hand.
The risk of the return Xt (λ) holding the portfolio for one day is often mea-
sured by volatility which is simply the square-root of variance. Stressing that
this volatility depends on the asset weights λ, we write
v
p
u
Xn
u
σ(λ) := Var(Xt (λ)) = tVar( λi Xti ), for t ∈ Z. (1.4)
i=1
4
7 NAS
6 RUS
5 JSE
values
0.2
4 SWI
0.4
y
0.6
0.8
3 DJS
1.0
2 HSH
1 TOP
0
TOP
HSH
DJS
SWI
JSE
RUS
NAS
1 2 3 4 5 6 7 8
x
7 NAS
6 RUS
5 JSE
values
−0.1
0.0
4 SWI
0.1
y
0.2
0.3
3 DJS
0.4
0.5
2 HSH
1 TOP
0
TOP
HSH
DJS
SWI
JSE
RUS
NAS
1 2 3 4 5 6 7 8
x
5
Regulatory rules require the calculation of the risk of a portfolio for a holding
period of d ≥ 1 days (e.g. 10 or 20 days) and it is common to quote volatility
as per annum (i.e. d = 250 or d = 252). For the clearness of presentation we
will assume that we model returns at t = 1, . . . , d when considering a holding
period of d days.
Definition 1.4 (Volatility for a holding period of d days). We denote the
volatility of the portfolio return given in (1.1) holding the assets for d days by
σ(λ, d) and thus
v
u
u Xd
σ(λ, d) := tVar( Xk (λ)). (1.5)
k=1
In practice it is assumed that the returns Xt (λ) are serially uncorrelated for
t ∈ Z and that variance is stationary, i.e. that Var(Xk ) = σ(λ) for k = 1, . . . , d.
These assumptions lead to the square-root-of-time rule.
Proposition 1.5 (Square-root-of-time rule). Under the above assumptions it
holds that
√
σ(λ, d) = σ(λ) d, for d ≥ 1, (1.6)
Cov(Xti , Xt (λ))
σi (λ) := λi p , for i = 1, . . . , n. (1.7)
Var(Xt (λ))
Remark 1.7. Using the same notation as in Remark 1.3 one can easily see that
(Σλ)i
σi (λ) = λi , for i = 1, . . . , n,
σ(λ)
6
Remark 1.9. Defining σ(λ) and σi (λ), i = 1, . . . , n as above it is easily seen that
we have full allocation for all holding periods d, i.e.
n
X
σ(λ, d) = σi (λ, d), for d ≥ 1.
i=1
7
Definition 2.1 (Auto-covariances of a univariate time series). Let (Xt (λ))t∈Z
denote a univariate weakly stationary stochastic process, then the auto-covariance
function and the auto-correlation function is denoted by
γ(k) := Cov(Xt (λ), Xs (λ)) and (2.1)
ρ(k) := Cor(Xt (λ), Xs (λ)) = γ(k)/γ(0), (2.2)
for t, s ∈ Z where |t − s| = k ≥ 0, respectively.
In the presence of auto-correlations in the portfolio returns the scaling by
the square-root of time is not accurate and the following proposition states the
correct scaling in this setting.
Proposition 2.2 (Volatility for a holding period of d days). Let (Xt (λ))t∈Z be
a univariate weakly stationary stochastic process with auto-covariance function
γ(·) as defined in Definition 2.1 then the volatility of the return when holding
the portfolio over d ≥ 1 days is given by
v
d
X
u
u d−1
X
σ(λ, d) = σ( Xk (λ)) = tdγ(0) + 2 (d − k)γ(k). (2.3)
k=1 k=1
The proof of Proposition 2.2 is given in the appendix and the correction to
scaling by the square-root of time is clearly seen in the following corollary.
Corollary 2.3 (Scaling rule for the univariate model). Let (Xt (λ))t∈Z as in
Proposition 2.2 and ρ(·) be its auto-correlation function then
σ(λ, d) = σ(λ)δ(d), (2.4)
where σ(λ) is the one-day volatility from (1.4) and the factor δ(d) is given by
v
u
u d−1
X
δ(d) = td + 2 (d − k)ρ(k). (2.5)
k=1
The short proof of this corollary can also be found in the appendix. Con-
sidering (2.5) we see that in our univariate model the scaling is given by the
square-root of time corrected by an expression taking into account all relevant
auto-correlations. If these are zero then (2.5) reduces to the well-known for-
mula (1.6). The next proposition gives a general estimate of the error when the
square-root-of-time rule is applied for scaling volatility when it is not appropri-
ate.
Corollary 2.4 (Error when using the square-root-of-time rule). Let (Xt (λ))t∈Z
as in Proposition 2.2 and δ(d) be the correct scaling factor, then the relative
error when applying the square-root-of-time rule can be estimated as
δ(d) √
√ ≤ d, for d ≥ 1. (2.6)
d
Proof. Considering (2.5) and noting that |ρ(k)| ≤ 1 for all k ∈ Z, we get
v
u
u d−1
X
δ(d) ≤ td + 2 (d − k) = d,
k=1
Pd−1
where we apply that k=1 (d − k) = 21 (d2 − d).
8
The above corollary states that the error that is made when applying the
square-root rule when the assumptions are not fulfilled grows with the square-
root of time. This conservative estimate does not assume anything non-trivial
about the auto-correlations. In Example 3.6 in Section 3 we will see that the
concrete picture is not always that bad.
for t ∈ Z.
Consider that the element at position (i, j) in the matrix Γ(k) is given by
i
Γ(k)i,j = Cov(Xt+k , Xtj ),
Γ(k) = Γ(−k)T ,
9
where Γ(k) denotes the covariance matrix of (Xt )t∈Z for k = 0, 1, . . .. The above
expression depends on the absolute value of the lag k only as
λT Γ(−k)λ = λT Γ(k)T λ = λT Γ(k)λ,
which concludes the proof of weak stationarity. To prove (2.8) consider that by
the bilinearity of covariance we get
i i
Cov(Xt+k , Xt (λ)) = Cov(Xt+k , λT Xt )
Xn
= i
λj Cov(Xt+k , Xtj ) = (Γ(k)λ)i ,
j=1
for i = 1, . . . n and k = 0, 1, . . ..
Using (2.8) we get useful expressions for the auto-covariance structure of the
portfolio returns as well as risk contributions in this setting.
Corollary 2.7 (Portfolio auto-covariance in the multivariate model). Let (Xt (λ))t∈Z
be the portfolio return where the asset returns have covariance structure as given
in (2.7) then the auto-covariance function (2.1) is given by
n
X
γ(k) = Cov(λT Xt+k , λT Xt ) = λT Γ(k)λ = λi γi (k) (2.9)
i=1
with γi (k) defined in (2.8) and δ(d) is the scaling factor for the respective port-
folio volatility.
10
The proofs of the two statements above can be found in the appendix.
Remark 2.10. Corollary 2.9 shows that in this modelling approach the relative
volatility contribution changes depending on the holding period and the whole
covariance structure of the returns involved. Thus assets whose relative risk
contribution increases with the holding period can be identified. This is a feature
that neither the model with uncorrelated asset returns nor the univariate model
has. Finally, it is easily seen that (2.11) reduces to
d √
δ(i, d) = √ = d, for i = 1, . . . , n,
d
in the case of no auto-correlations.
Considering the results in Proposition 2.2 and 2.8 in general we have to
estimate d − 1 auto-covariances respectively auto-covariance matrices to find a
correct scaling for d days. Considering volatility per annum this corresponds to
d − 1 = 249 or 251, which is clearly not appealing. In the following section we
apply these findings to classical auto-regressive time series models which reduces
the number of parameters tremendously.
11
process by VAR(p). In the one-dimensional case such a process is denoted by
ARMA(p, q) (auto-regressive moving average) and we write
p
X q
X
Xt − φk Xt−k = θk Zt−j + Zt , for t ∈ Z, (3.2)
k=1 j=1
Φ(z) := I − Φ1 z − · · · − Φp z p , (3.3)
are given by
After this preparation we can state the proposition that finally gives the link
needed. See Mauricio [1995] for a proof and an algorithm for a fast and exact
computation of the covariance matrices in the following proposition.
Proposition 3.4 (Yule-Walker equations). The auto-covariance matrices Γ(k)
for k = 0, 1, . . . of a causal VARMA(p, q) process are uniquely determined by the
following system of linear equations
p
X q
X
Γ(k) − Φj Γ(k − j) = Θj ΓXZ (j − k). (3.5)
j=1 j=k
12
3.1 Univariate models
First we consider univariate models and compare correct scaling to the square-
root-of-time rule.
Example 3.5 (ARMA(1, 0)-model). For the ARMA(1, 0)-model (or AR(1)) the
formula for ACF is explicitly found in many textbooks (e.g. [McNeil et al., 2005,
Chapter 4.2]) furthermore Proposition 3.4 gives
σ2
γ(0) = and
1 − φ21
γ(k) = φ1 γ(k − 1), for k = 1, 2 . . . (3.6)
The correct scaling factor in this setting is given by
s
φ1
δ(d) = d + 2 d(1 − φ1 ) + (φd1 − 1) ,
(φ1 − 1)2
as can be seen by plugging the ACF (3.6) into (2.5). Note that this kind
of scaling is already introduced in Alexander [2008]. Table 1 shows the error
as compared to the scaling by the square-root of time in the case of d = 10
for various values for φ1 . One sees that under- respectively overestimation of
volatility can be significant. Figure 3 illustrates this fact for various values of
φ1 .
Table 1: Errors in the case of the ARMA(1, 0)-model compared to the square-
root rule for varying φ1 for a holding period of d = 10 days.
φk−1
1 (φ1 + θ1 )(1 + φ1 θ1 )
γ(k) = , for k = 1, 2 . . .
1 + θ12 + 2φ1 θ1
which leads to the following scaling factor
s
(1 − φ1 θ1 )(φ1 − θ1 )
δ(d) = d + 2 (d(1 − φ1 ) + φd1 − 1).
(φ − 1)2 (1 + θ12 − 2φ1 θ1 )
Again considering the relative error for d = 10 in Table 2 we see that the under-
respectively overestimation of volatility in this model can be even more severe
13
Figure 3: Relative scaling error for various values for φ = φ1 and d = 1, . . . , 60
in the ARMA(1, 0)-model.
14
than in the AR(1)-model.
√ Considering the relative error as d grows the error is
bounded as δ(d) = O( d). To see this consider that in the ARMA(1, 1)-model
we have
s
(1−φθ)(φ−θ) d
δ(d) d + 2 (φ−1)2 (1+θ 2 −2φθ) [d(1 − φ) + φ − 1]
lim √ = lim
d→∞ d d→∞ d
s
(1 − φθ)(φ − θ)
= 1+2 .
(1 − φ)(1 + θ2 − 2φθ)
Table 2: Relative errors in the case of the ARMA(1, 1)-model compared to the
square-root rule for varying φ1 and θ1 for a holding period of d = 10 days.
As it will be shown in Example 3.10 the lead-lag effect requires in many cases
the usage of models of higher order than shown in the previous two examples. In
these cases handy formulas for the auto-covariance function can not be provided
any more, but the function ARMAacf in the software package R Development
Core Team [2010] or similar implementations may be used to find the ACF
after having estimated the ARMA-coefficients.
p̃ ≤ np and q̃ ≤ (n − 1)p + q.
The above theorem tells us that a a portfolio which is a simple linear transfor-
mation of a VARMA(p, q) process can not be guaranteed to have an ARMA(p, q)
representation of the same order. Furthermore it is important to note that
as a consequence the class of VAR(p)-models is not closed with respect to
linear transformation as, in general, the result of the transformation can be
15
some VARMA(p̃, q̃) process with q̃ > 0. Concerning multivariate models we
nevertheless focus on VAR(p)-models due to known identification problems of
VARMA(p, q)-models if q > 0 (see for example Lütkepohl [2004]). Furthermore
this model class arises naturally in our problem of global markets as we have
seen in Example 1.1.
As a preparation for our key result on portfolios constructed out of VAR(p)-
process we state the following proposition:
Proposition 3.8 (AR portfolios built from VAR processes). Let (Xt )t∈Z be a
VAR(p)-process in Rn for n > 0 of the form
p
X
Xt = Φk Xt−k + Zt
k=1
iff
λT Φk = φk λT for k = 1, . . . , p. (3.8)
Note that as in Corollary 2.7 (λT Zt )t∈Z is clearly a white noise process.
The full proof of the proposition can be found in the appendix. Condition (3.8)
means that only portfolios with λ being an eigenvector of all coefficient matrices
of the VAR(p)-process admit an AR(p) representation. The following corollary
concludes these considerations.
Corollary 3.9 (AR portfolios built from VAR processes). In the setting of
Proposition 3.8 the process (Xt (λ))t∈Z is an AR(p)-process for any portfolio
weighting λ ∈ Rn iff the coefficient matrices of the VAR-process are diagonal
and of the following form
Φk = φ k I n , for k = 1, . . . , p.
for some fixed value for a - all the same for each asset. Considering Example 1.1
it would be inconsistent to assume that dependence on lagged values is univari-
ate, which is implied by a diagonal structure of Φ1 . Furthermore the assumption
that this dependence is the same for all assets is clearly unrealistic. The con-
clusion is that the weighted VAR(1)-model is richer than an AR(1)-model.
16
After these general considerations we focus the VAR(1)-model of the form
Xt = Φ1 Xt−1 + Zt , for t ∈ Z,
since this model will be the natural choice to capture the characteristics of the
lead-lag effect (see Example 1.1). In this case we get the following expression
for the covariance matrices
and further calculations are relatively straight forward. The following concrete
example illustrates the trade-off when modelling such a portfolio whose time
series characteristics are essentially multivariate by univariate models.
Example 3.10. Consider a portfolio consisting of asset A traded in Tokyo and
asset B traded in New York. Although we could estimate the lag zero and lag
one covariance matrix on the data given but we rather use the heuristics de-
veloped in Example 1.1 for a numerical illustration. The time difference of the
15
observed closing prices is therefore given by (xB − xA ) = 24 day. For simplicity
we assume normal log returns. Furthermore assuming daily standard deviations
of σA = 0.012, σB = 0.018 (i.e. annualized volatilities of 20% and 30%) and con-
temporaneous correlation of ρ = 0.7 yields the complete dependence structure
of the returns by (1.2) and (1.3). Thus we consider
14.40 5.67 0 9.45
Γ(0) = × 10−5 and Γ(1) = × 10−5 .
5.67 32.40 0 0
We can model these two assets by a VAR(1)-model with the coefficient matrix
Φ1 given by
−1 −0.123 0.313
Φ1 = Γ(1)Γ(0) = .
0 0
Modelling a portfolio with the weighting λ = ( 12 , 21 )T the auto-covariance func-
tion of the resulting portfolio can be found by Corollary 2.7 and it is given by
γ(0) = 14.53 × 10−5 and for k ≥ 1
17
Using the results for scaling in univariate models and Equation (3.11) we
compare the resulting scaling constants in Table 3. We see that the AR(1)-model
performs much poorer in approximating the result of the multivariate model,
while the AR(2)-model can be seen as a valid approximation. This observation
is confirmed by considering the auto-covariance function of the three models in
Table 4.
Furthermore we can consider risk contributions in the VAR(1)-model: In
the multivariate model we get the scaling constants for the risk contributions of
each asset by Corollary 2.9 and (3.10), i.e.
d−1
!
1 X (Φk1 Γ(0)λ)i
δ(i, d) = d+2 (d − k) , (3.12)
δ(d) (Γ(0)λ)i
k=1
Applying (3.12) to our 2 asset example we see in Table 5 that the relative risk
contributions of asset A in the leading market decreases while the one of asset B
increases. This shows that the risk contribution of the lagging market increases
quicker with the holding period than the risk contribution of the leading one.
In fact the ratio of 1 : 2 without consideration of the lagged correlation (i.e.
d = 1) already changes significantly after considering one more day (d = 2) and
finally considering d = 250 the ratio of volatility contributions is given by 1 : 1.
Table 3: Volatility scaling factors δ(d) for VAR(1), AR(1) and AR(2) for various
holding periods d.
Table 4: ACF for various lags for VAR(1), AR(1) and AR(2).
After this in-depth analysis of the VAR(1)-model we conclude and come back
to our real world global portfolio of Example 1.1.
18
σ(d,A) σ(d,B)
d σ(d) σ(d)
1 34.52% 65.48%
2 43.68% 56.32%
5 47.10% 52.90%
10 48.18% 51.82%
30 48.88% 51.12%
90 49.05% 50.95%
250 49.18% 50.82%
Table 5: Relative risk contributions of assets A and B for various holding periods
d in the VAR(1)-model.
Example 3.11 (Example 1.1 continued). Considering Table 6 we see the con-
tributions to volatility p.a. on one hand by assuming zero serial correlations
between the assets in the portfolio and on the other hand by modelling the
asset returns in the portfolio as VAR(1)-process. Applying the square-root-of-
time rule to this portfolio we get a volatility p.a. of 19.82% while the volatility
p.a. increases by approximately 12% to 22.12% in the VAR(1)-model. Con-
cerning the analysis of sources of risk note that the risk contribution by the
geographically most distant market Japan of 1.52% looks quite small when ig-
noring serial cross-correlations. The risk contributions of the markets leading
the portfolio do not change dramatically. The increase of the portfolio volatility
can be attributed to the Asian assets whose risk contribution increases signifi-
cantly as serial cross correlations to the US and Europe are taken into account.
This example shows that not only the accuracy of total volatility of the porfolio
increases but also the attribution to single assets fits economic considerations
much better!
4 Conclusions
In this article we address the problem of calculating portfolio volatility for hold-
ing periods of more than one day. We show that ignoring serial correlations leads
on one hand to biased estimates of volatility and on the other hand to mislead-
ing risk contributions as Example 3.11 illustrates. Serial correlations naturally
19
occur when analyzing portfolios of geographically diversified assets traded in
distant time zones. In Example 1.1 we motivate the use of VAR(1)-models in
such situations. We propose formulas for calculating accurate volatility scaling
modelling the portfolio return as a univariate process as well as in a multi-variate
setting. Moreover in the multivariate setting we also provide explicit formulas
for genuine risk contributions that take the time series structure of the assets
involved into account. Modelling the asset returns as a vector autoregressive
process of order one we derive handy formulas and perform a complete analy-
sis of risk and risk contributions. Applying the findings of this article to the
calculation of the tracking error, i. e. the volatility of the additional return of
the portfolio above a given benchmark, can improve the analysis of relative risk
which is often an aim in asset management.
Besides the analysis of market risk our findings can be applied to portfolio
optimization as well as portfolio construction techniques such as risk-parity (also
known as equally-weighted risk contributions, see Maillard et al. [2008]) where
risk contributions by assets are the driving input. As another direction of further
research the findings of this article may also be applied to the VEC specification
of multivariate GARCH models, since they admit a VARMA representation
(see Lütkepohl [2006]).
Appendix: Proofs
Proof of Proposition 2.2 and Corollary 2.3. Considering that σ(λ, d) is the square-
Pd
root of Var( i=1 Xi (λ)) and writing down this variance in matrix form using
weak stationarity we get
Cov(X1 (λ), X1 (λ)) . . . Cov(X1 (λ), Xd (λ))
Xd Cov(X2 (λ), X1 (λ)) . . . Cov(X2 (λ), Xd (λ))
Xi (λ)) = 1T 1
Var( .. .. ..
i=1
. . .
Cov(Xd (λ), X1 (λ)) . . . Cov(Xd (λ), Xd (λ))
γ(0) γ(1) ... γ(d − 1)
γ(1) γ(0) ... γ(d − 2)
= 1T 1,
.. .. .. ..
. . . .
γ(d − 1) γ(d − 2) . . . γ(0)
where γ(·) denotes the auto-covariance function of (Xt (λ))t∈Z and 1 = (1, . . . , 1)T .
Summing up along the diagonals and using symmetry we get Equation (2.3).
For proving (2.5) note that the auto-covariances can be expressed in terms of
the auto-correlation and the variance in the following sense:
γ(k)
ρ(k) = ⇔ γ(k) = γ(0)ρ(k) = σ(λ)2 ρ(k),
γ(0)
for k = 0, 1, . . .
Proof of Proposition 2.8 and Corollary 2.9. Following the Euler allocation rule
the volatility contributions are given by
Pd Pd
Cov( k=1 Xki , k=1 Xk (λ))
σi (λ, d) = λi ,
σ(λ, d)
20
where we can calculate the denominator σ(λ, d) by (2.3) and (2.9). For the
numerator we get
where 1 = (1, . . . , 1)T and the γi (k) are given in (2.8) for i = 1, . . . , n and
k = 0, 1, . . . Now again use the symmetries and sum up along the diagonals to get
the result. To prove the corollary, recall that σ(λ, d) = σ(λ)δ(d). Plugging this
into the denominator of (2.10) and recalling that the one day risk contribution
is given by σi (λ) = λi γσ(λ)
i (0)
we get
d−1
!
γi (0) 2 X
σi (λ, d) = λi d+ (d − k)γi (k) /δ(d),
σ(λ) γi (0)
k=1
λT Xt = Xt (λ)
p
X p
X
⇔ λT Φk Xt−k + λT Zt = φk λT Xt−k + λT Zt
k=1 k=1
Xp
T
⇔λ (Φk − φk In )Xt−k = 0.
k=1
λT (Φk − φk In ) = 0 for k = 1, . . . , p.
To prove Corollary 3.9 consider that the above condition is true for arbitrary
λ ∈ Rn if and only if the rank of the matrix Φk − φk In is zero for k = 1, . . . , p
which concludes the proof.
References
Carol Alexander. Market Risk Analysis: Practical Financial Econometrics. Wi-
ley, second edition, June 2008. ISBN 9780470998014.
Lorenzo Bergomi. Correlations in Asynchronous Markets. SSRN eLibrary, 2010.
21
George E. P. Box, Gwilym M. Jenkins, and Gregory C. Reinsel. Time Series
Analysis: Forecasting and Control (Wiley Series in Probability and Statistics).
Wiley, fourth edition, June 2008. ISBN 0470272848. URL http://www.
worldcat.org/isbn/0470272848.
Peter J. Brockwell and Richard A. Davis. Time Series: Theory and Methods.
Springer-Verlag New York, Inc., New York, NY, USA, 1986. ISBN 0-387-
96406-1.
Peter J. Brockwell and Richard A. Davis. Introduction to Time Series and
Forecasting. Springer, March 2002. ISBN 0387953515. URL http://www.
worldcat.org/isbn/0387953515.
Thomas Coleman. Estimating the Correlation of Non-Contemporaneous Time-
Series. SSRN eLibrary, 2007.
Jon Danielsson and Jean-Pierre Zigrand. On time-scaling of risk and
the square-root-of-time rule. Journal of Banking & Finance, 30(10):
2701–2713, October 2006. URL http://ideas.repec.org/a/eee/jbfina/
v30y2006i10p2701-2713.html.
Francis X. Diebold, Andrew Hickman, Atsushi Inoue, and Til Schuermann.
Converting 1-day volatility to h-day volatility: Scaling by root-h is worse than
you think. Center for Financial Institutions Working Papers 97-34, Wharton
School Center for Financial Institutions, University of Pennsylvania, July
1997. URL http://ideas.repec.org/p/wop/pennin/97-34.html.
Feike C. Drost and Theodoor E. Nijman. Temporal Aggregation of GARCH
Processes. Econometrica, 61:909927, 1993.
Emel Kahya. Correlation of Returns in Non-contemporaneous Markets. Multi-
national Finance Journal, 1(2):123–135, 1997.
Helmut Lütkepohl. Introduction to Multiple Time Series Analysis. Springer,
second edition, 1993.
Helmut Lütkepohl. Forecasting with VARMA models. Open Access pub-
lications from European University Institute urn:hdl:1814/2805, European
University Institute, 2004. URL http://ideas.repec.org/p/ner/euiflo/
urnhdl1814-2805.html.
Helmut Lütkepohl. New Introduction to Multiple Time Series Analysis.
Springer, 1st ed. 2006. corr. 2nd printing edition, February 2006. ISBN
3540262393. URL http://www.worldcat.org/isbn/3540262393.
Sébastien Maillard, Thierry Roncalli, and Jérôme Teiletche. On the Properties
of Equally-Weighted Risk Contributions Portfolios. SSRN eLibrary, 2008.
José A. Mauricio. A corrected algorithm for computing the theoretical au-
tocovariance matrices of a vector arma model. Technical report, Instituto
Complutense de Análisis Económico, Universidad Complutense de Madrid,
March 1995.
Alexander McNeil, Rüdiger Frey, and Paul Embrechts. Quantative Risk Man-
agement: Concepts, Techniques, and Tools. Princeton University Press, 2005.
22
Terence C. Mills. The Econometric Modelling of Financial Time Series. Cam-
bridge University Press, 1993. ISBN 0521410487 0521422574. URL http:
//www.loc.gov/catdir/toc/cam021/92037624.html.
23