Dynamic Conditional Correlation:

A Simple Class of Multivariate

Generalized Autoregressive Conditional
Heteroskedasticity Models
Robert Engle
Department of Finance, New York University Leonard N. Stern School of Business, New York, NY 10012
and Department of Economics, University of California, San Diego ([email protected])

Time varying correlations are often estimated with multivariate generalized autoregressive conditional
heteroskedasticity (GARCH) models that are linear in squares and cross products of the data. A new
class of multivariate models called dynamic conditional correlation models is proposed. These have
the  exibility of univariate GARCH models coupled with parsimonious parametric models for the
correlations. They are not linear but can often be estimated very simply with univariate or two-step
methods based on the likelihood function. It is shown that they perform well in a variety of situations
and provide sensible empirical results.

KEY WORDS: ARCH; Correlation; GARCH; Multivariate GARCH.

1. INTRODUCTION need for bigger correlation matrices. In most cases, the number
Correlations are critical inputs for many of the common of parameters in large models is too big for easy optimization.
tasks of Ž nancial management. Hedges require estimates of In this article, dynamic conditional correlation (DCC) esti-
mators are proposed that have the  exibility of univariate
the correlation between the returns of the assets in the hedge.
GARCH but not the complexity of conventional multivariate
If the correlations and volatilities are changing, then the hedge
GARCH. These models, which parameterize the conditional
ratio should be adjusted to account for the most recent infor-
correlations directly, are naturally estimated in two steps—
mation. Similarly, structured products such as rainbow options a series of univariate GARCH estimates and the correla-
that are designed with more than one underlying asset have tion estimate. These methods have clear computational advan-
prices that are sensitive to the correlation between the under- tages over multivariate GARCH models in that the number of
lying returns. A forecast of future correlations and volatilities parameters to be estimated in the correlation process is inde-
is the basis of any pricing formula. pendent of the number of series to be correlated. Thus poten-
Asset allocation and risk assessment also rely on correla- tially very large correlation matrices can be estimated. In this
tions; however, in this case a large number of correlations is article, the accuracy of the correlations estimated by a variety
often required. Construction of an optimal portfolio with a set of methods is compared in bivariate settings where many
of constraints requires a forecast of the covariance matrix of methods are feasible. An analysis of the performance of the
the returns. Similarly, the calculation of the standard devia- DCC methods for large covariance matrices was considered
tion of today’s portfolio requires a covariance matrix of all by Engle and Sheppard (2001).
the assets in the portfolio. These functions entail estimation Section 2 gives a brief overview of various models
and forecasting of large covariance matrices, potentially with for estimating correlations. Section 3 introduces the new
thousands of assets. method and compares it with some of the cited approaches.
The quest for reliable estimates of correlations between Section 4 investigates some statistical properties of the
Ž nancial variables has been the motivation for countless aca- method. Section 5 describes a Monte Carlo experiment and
demic articles and practitioner conferences and much Wall results are presented in Section 6. Section 7 presents empiri-
Street research. Simple methods such as rolling historical cor- cal results for several pairs of daily time series, and Section 8
relations and exponential smoothing are widely used. More concludes.
complex methods, such as varieties of multivariate general-
2. CORRELATION MODELS
ized autoregressive conditional heteroskedasticity (GARCH)
or stochastic volatility, have been extensively investigated in The conditional correlation between two random variables
the econometric literature and are used by a few sophisticated r1 and r2 that each have mean zero is deŽ ned to be
practitioners. To see some interesting applications, exam- Etƒ1 4r11 t r21 t 5
121 t D q 0 (1)
ine the work of Bollerslev, Engle, and Wooldridge (1988),
Etƒ1 4r112 t 5Etƒ1 4r212 t 5
Bollerslev (1990), Kroner and Claessens (1991), Engle and
Mezrich (1996), Engle, Ng, and Rothschild (1990), Bollerslev,
© 2002 American Statistical Association
Chou, and Kroner (1992), Bollerslev, Engle, and Nelson Journal of Business & Economic Statistics
(1994), and Ding and Engle (2001). In very few of these arti- July 2002, Vol. 20, No. 3
cles are more than Ž ve assets considered, despite the apparent DOI 10.1198/073500102288618487

339
340 Journal of Business & Economic Statistics, July 2002

In this deŽ nition, the conditional correlation is based on infor- An alternative simple approach to estimating multivariate
mation known the previous period; multiperiod forecasts of models is the Orthogonal GARCH method or principle com-
the correlation can be deŽ ned in the same way. By the laws of ponent GARCH method. This was advocated by Alexander
probability, all correlations deŽ ned in this way must lie within (1998, 2001). The procedure is simply to construct uncondi-
the interval 6ƒ11 17. The conditional correlation satisŽ es this tionally uncorrelated linear combinations of the series r. Then
constraint for all possible realizations of the past information univariate GARCH models are estimated for some or all of
and for all linear combinations of the variables. these, and the full covariance matrix is constructed by assum-
To clarify the relation between conditional correlations and ing the conditional correlations are all zero. More precisely,
conditional variances, it is convenient to write the returns Ž nd A such that yt D Art 1 E4yt yt0 5 ² V is diagonal. Univari-
as the conditional standard deviation times the standardized ate GARCH models are estimated for the elements of y and
disturbance: combined into the diagonal matrix Vt . Making the additional
¢ q
strong assumption that Etƒ1 4yt yt0 5 D Vt , then
hi1 t D Etƒ1 ri12 t 1 ri1 t D hi1 t ˜i1 t 1 i D 11 23 (2)
0
˜ is a standardized disturbance that has mean zero and vari- Ht D A ƒ 1 V t A ƒ 1 0 (8)
ance one for each series. Substituting into (4) gives
In the bivariate case, the matrix A can be chosen to be trian-
Etƒ1 4˜11 t ˜21 t 5 gular and estimated by least squares where r1 is one compo-
121 t D q D Etƒ1 4˜11 t ˜21 t 50 (3)
Etƒ1 4˜211 t 5Etƒ1 4˜221 t 5 nent and the residuals from a regression of r1 on r2 are the
second. In this simple situation, a slightly better approach is
Thus, the conditional correlation is also the conditional covari- to run this regression as a GARCH regression, thereby obtain-
ance between the standardized disturbances. ing residuals that are orthogonal in a generalized least squares
Many estimators have been proposed for conditional cor- (GLS) metric.
relations. The ever-popular rolling correlation estimator is Multivariate GARCH models are natural generalizations of
deŽ ned for returns with a zero mean as this problem. Many speciŽ cations have been considered; how-
Ptƒ1 ever, most have been formulated so that the covariances and
sDtƒnƒ1 r11 s r21 s
O 121 t D q P ¢ Ptƒ1 ¢0 (4) variances are linear functions of the squares and cross prod-
tƒ1 2 2
r
sDtƒnƒ1 11 s r
sDtƒnƒ1 21 s ucts of the data. The most general expression of this type is
called the vec model and was described by Engle and Kroner
Substituting from (4) it is clear that this is an attractive estima- (1995). The vec model parameterizes the vector of all covari-
tor only in very special circumstances. In particular, it gives ances and variances expressed as vec4Ht 5. In the Ž rst-order
equal weight to all observations less than n periods in the case this is given by
past and zero weight on older observations. The estimator
will always lie in the 6ƒ11 17 interval, but it is unclear under 0
vec4Ht 5 D vec4ì5 C A vec4rtƒ1 rtƒ1 5 C B vec4H tƒ1 51 (9)
what assumptions it consistently estimates the conditional cor-
relations. A version of this estimator with a 100-day win-
dow, called MA100, will be compared with other correlation where A and B are n2 € n2 matrices with much structure
estimators. following from the symmetry of H . Without further restric-
The exponential smoother used by RiskMetrics uses declin- tions, this model will not guarantee positive deŽ niteness of the
ing weights based on a parameter ‹, which emphasizes current matrix H .
data but has no Ž xed termination point in the past where data Useful restrictions are derived from the BEKK representa-
becomes uninformative. tion, introduced by Engle and Kroner (1995), which, in the
Ptƒ1 tƒjƒ1 Ž rst-order case, can be written as
sD1 ‹ r11 s r21 s
O 121 t D q P ¢ P ¢0 (5) 0
tƒ1 tƒsƒ1 2
‹ r tƒ1 tƒsƒ1 2
‹ r Ht D ì C A4rtƒ1 rtƒ1 5A0 C BHtƒ1 B 0 0 (10)
sD1 11 s sD1 21 s

It also will surely lie in 6ƒ11 17; however, there is no guidance Various special cases have been discussed in the literature,
from the data for how to choose ‹. In a multivariate context, starting from models where the A and B matrices are simply
the same ‹ must be used for all assets to ensure a positive a scalar or diagonal rather than a whole matrix and continuing
deŽ nite correlation matrix. RiskMetrics uses the value of 094 to very complex, highly parameterized models that still ensure
for ‹ for all assets. In the comparison employed in this article, positive deŽ niteness. See, for example, the work of Engle and
this estimator is called EX .06. Kroner (1995), Bollerslev et al. (1994), Engle and Mezrich
DeŽ ning the conditional covariance matrix of returns as (1996), Kroner and Ng (1998), and Engle and Ding (2001).
In this study the scalar BEKK and the diagonal BEKK are
Etƒ1 4rt rt0 5 ² Ht 1 (6)
estimated.
these estimators can be expressed in matrix notation respec- As discussed by Engle and Mezrich (1996), these models
tively as can be estimated subject to the variance targeting constraint by
which the long run variance covariance matrix is the sample
1X n
covariance matrix. This constraint differs from the maximum
Ht D 4r r 0 5 and Ht D ‹4rtƒ1 rtƒ1
0
5 C 41 ƒ ‹5Htƒ1 0 (7)
n jD1 tƒj tƒj likelihood estimator (MLE) only in Ž nite samples but reduces
Engle: Dynamic Conditional Correlation 341

the number of parameters and often gives improved perfor- followed by the q’s will be integrated,
mance. In the general vec model of Equation (9), this can be
expressed as qi1 j1 t D 41 ƒ ‹54˜i1 tƒ1 ˜j1 tƒ1 5 C ‹4qi1 j1 tƒ1 51
(17)
1X i1 j1 t D
qi1 j1 t
0
vec4ì5 D 4I ƒ A ƒ B5vec4S51 where S D 4r r 0 50 (11) p
qii1 t q jj1 t
T t t t
A natural alternative is suggested by the GARCH(1, 1)
This expression simpliŽ es in the scalar and diagonal BEKK model:
cases. For example, for the scalar BEKK the intercept is
simply qi1 j1 t D N i1 j C 4˜i1 tƒ1 ˜j1 tƒ1 ƒ N i1 j 5 C ‚4qi1 j 1 tƒ1 ƒ N i1 j 5 (18)

ì D 41 ƒ  ƒ ‚5S0 (12) where N i1 j is the unconditional correlation between ˜i1 t and

˜j1 t . Rewriting gives

3. DCCs ³ ´
1ƒ ƒ‚ X̂
qi1 j1 t D N i1 j C  ‚sƒ1 ˜i1 tƒs ˜j1 tƒs 0 (19)
This article introduces a new class of multivariate GARCH 1ƒ‚ sD1
estimators that can best be viewed as a generalization of the
Bollerslev (1990) constant conditional correlation (CCC) esti- The average of qi1 j1 t will be N i1 j , and the average variance will
mator. In Bollerslev’s model, be 1.
nq o
qNi1 j û N i1 j 0 (20)
Ht D Dt RDt 1 where Dt D diag hi1 t 1 (13)
The correlation estimator
where R is a correlation matrix containing the conditional
correlations, as can directly be seen from rewriting this equa- qi1 j1 t
i1 j1 t D p (21)
tion as qi1 i1 t qj1 j1 t

Etƒ1 4˜t ˜0t 5 D Dtƒ1 Ht Dtƒ 1 D R since ˜t D Dtƒ1 rt 0 (14) will be positive deŽ nite as the covariance matrix, Qt with typ-
ical element qi1 j1 t , is a weighted average of a positive deŽ nite
The expressions for h are typically thought of as univari- and a positive semideŽ nite matrix. The unconditional expec-
ate GARCH models; however, these models could certainly tation of the numerator of (21) is N i1 j and each term in the
include functions of the other variables in the system as prede- denominator has expected value 1. This model is mean revert-
termined variables or exogenous variables. A simple estimate ing as long as  C ‚ < 1, and when the sum is equal to 1 it
of R is the unconditional correlation matrix of the standard- is just the model in (17). Matrix versions of these estimators
ized residuals. can be written as
This article proposes the DCC estimator. The dynamic cor-
relation model differs only in allowing R to be time varying: Qt D 41 ƒ ‹54˜tƒ1 ˜0tƒ1 5 C ‹Qtƒ1 (22)

H t D D t R t Dt 0 (15) and

Parameterizations of R have the same requirements as H , Qt D S41 ƒ  ƒ ‚5 C 4˜tƒ1 ˜0tƒ1 5 C ‚Qtƒ1 1 (23)
except that the conditional variances must be unity. The matrix
Rt remains the correlation matrix. where S is the unconditional correlation matrix of the epsilons.
Kroner and Ng (1998) proposed an alternative generaliza- Clearly more complex positive deŽ nite multivariate
tion that lacks the computational advantages discussed here. GARCH models could be used for the correlation parameter-
They proposed a covariance matrix that is a matrix weighted ization as long as the unconditional moments are set to the
average of the Bollerslev CCC model and a diagonal BEKK, sample correlation matrix. For example, the MARCH family
both of which are positive deŽ nite. of Ding and Engle (2001) can be expressed in Ž rst-order form
Probably the simplest speciŽ cation for the correlation as
matrix is the exponential smoother, which can be expressed as
Qt D S ž 4‰‰0 ƒ A ƒ B5 C A ž ˜tƒ1 ˜0tƒ1 C B ž Qtƒ1 1 (24)
Ptƒ1 s
sD1 ‹ ˜i1 tƒs ˜j1 tƒs
i1j1t D q P ¢ Ptƒ1 s 2 ¢ D 6Rt 7i1 j 1 (16) where ‰ is a vector of ones and ž is the Hadamard product
tƒ1 s 2
sD1 ‹ ˜ i1 tƒs sD1 ‹ ˜j1 tƒs of two identically sized matrices, which is computed simply
by element-by-element multiplication. They show that if A1 B,
a geometrically weighted average of standardized residuals. and 4‰‰0 ƒ A ƒ B5 are positive semideŽ nite, then Q will be
Clearly these equations will produce a correlation matrix at positive semideŽ nite. If any one of the matrices is positive
each point in time. A simple way to construct this correla- deŽ nite, then Q will also be. This family includes both earlier
tion is through exponential smoothing. In this case the process models as well as many generalizations.
342 Journal of Business & Economic Statistics, July 2002

4. ESTIMATION The volatility part of the likelihood is apparently the sum of

individual GARCH likelihoods
The DCC model can be formulated as the following statis-
n ³ ´
tical speciŽ cation: 1 XX ri12 t
LV 4ˆ5 D ƒ log42 5 C log4hi1 t 5 C 1 (30)
2 t iD1 hi1 t
rt —=tƒ1 N 401 Dt Rt Dt 51
Dt2 0
D diag8—i 9 C diag8Ši 9 ž rtƒ1 rtƒ1 C diag8‹i 9 ž Dtƒ1
2
1 which is jointly maximized by separately maximizing each
term.
˜t D Dtƒ 1 rt 1 (25) The second part of the likelihood is used to estimate the
Qt D S ž 4‰‰ 0
ƒ A ƒ B5 C A ž ˜tƒ1 ˜0tƒ1 C B ž Qtƒ1 1 correlation parameters. Because the squared residuals are not
dependent on these parameters, they do not enter the Ž rst-
Rt D diag8Qt 9ƒ1 Qt diag8Qt 9ƒ1 0 order conditions and can be ignored. The resulting estimator
The assumption of normality in the Ž rst equation gives rise to is called DCC LL MR if the mean reverting formula (18) is
a likelihood function. Without this assumption, the estimator used and DCC LL INT with the integrated model in (17).
will still have the Quasi-Maximum Likelihood (QML) inter- The two-step approach to maximizing the likelihood is to
Ž nd
pretation. The second equation simply expresses the assump-
tion that each asset follows a univariate GARCH process.
ˆO D arg max8LV 4ˆ59 (31)
Nothing would change if this were generalized.
The log likelihood for this estimator can be expressed as and then take this value as given in the second stage:
rt —=tƒ1 N 401 Ht 51 O ”590
max 8LC 4ˆ1 (32)
”
1X T
¢
LDƒ n log42 5 C log —Ht — C rt0 Htƒ 1 rt Under reasonable regularity conditions, consistency of the
2 tD1
Ž rst step will ensure consistency of the second step. The
1X T
maximum of the second step will be a function of the Ž rst-
Dƒ n log42 5 C log —Dt Rt Dt —
2 tD1 step parameter estimates, so if the Ž rst step is consistent, the
¢
C rt0 Dtƒ1 Rƒt 1 Dtƒ1 rt second step will be consistent as long as the function is con-
(26) tinuous in a neighborhood of the true parameters.
1X T
Newey and McFadden (1994), in Theorem 6.1, formulated
Dƒ n log42 5 C 2 log —Dt —
2 tD1 ¢ a two-step Generalized Method of Moments (GMM) prob-
C log —Rt — C ˜0t Rƒt 1 ˜t lem that can be applied to this model. Consider the moment
condition corresponding to the Ž rst step as ïˆ LV 4ˆ5 D 0.
1X T
Dƒ n log42 5 C 2 log —Dt — C rt0 Dtƒ1 Dtƒ1 rt The moment condition corresponding to the second step is
2 tD1 O ”5 D 0. Under standard regularity conditions, which
ï” LC 4ˆ1
¢ are given as conditions (i) to (v) in Theorem 3.4 of Newey
ƒ ˜0t ˜t C log —Rt — C ˜0t Rƒt 1 ˜t 1
and McFadden, the parameter estimates will be consistent,
which can simply be maximized over the parameters of the and asymptotically normal, with a covariance matrix of famil-
model. However, one of the objectives of this formulation is iar form. This matrix is the product of two inverted Hessians
to allow the model to be estimated more easily even when around an outer product of scores. In particular, the covariance
the covariance matrix is very large. In the next few para- matrix of the correlation parameters is
graphs several estimation methods are presented, giving sim-
ple consistent but inefŽ cient estimates of the parameters of V 4”5 D 6E4ï”” LC 57ƒ1
the model. SufŽ cient conditions are given for the consistency € E 8ï” LC ƒ E4 LC 56E4 LV 57ƒ1 ïˆ LV 9
and asymptotic normality of these estimators following Newey ¢
and McFadden (1994). Let the parameters in D be denoted € 8ï” LC ƒ E4 LC 56E4 LV 57ƒ1 ïˆ LV 90
ˆ and the additional parameters in R be denoted ”. The log- € 6E4ï”” LC 570 (33)
likelihood can be written as the sum of a volatility part and a
correlation part: Details of this proof can be found elsewhere (Engle and
Sheppard 2001).
L4ˆ1 ”5 D LV 4ˆ5 C LC 4ˆ1 ”50 (27) Alternative estimation approaches, which are again consis-
tent but inefŽ cient, can easily be devised. Rewrite (18) as
The volatility term is
1X ¢ ei1 j1 t D N i1 j 41 ƒ  ƒ ‚5 C 4 C ‚5ei1 j 1 tƒ1
LV 4ˆ5 D ƒ n log42 5 C log —Dt —2 C rt0 Dtƒ2 rt 1 (28)
2 t
ƒ ‚4ei1 j1 tƒ1 ƒ qi1 j1 tƒ1 5 C 4ei1 j1 t ƒ qi1 j1 t 51 (34)
and the correlation component is
where ei1 j1 t D ˜i1 t ˜j1 t . This equation is an ARMA(1, 1)
1X ¢ because the errors are a Martingale difference by construction.
LC 4ˆ1 ”5 D ƒ log —Rt — C ˜0t Rƒt 1 ˜t ƒ ˜0t ˜t 0 (29)
2 t The autoregressive coefŽ cient is slightly bigger if  is a small
Engle: Dynamic Conditional Correlation 343

positive number, which is the empirically relevant case. This 3. DCC IMA—DCC with integrated moving average esti-
equation can therefore be estimated with conventional time mation as in (35)
series software to recover consistent estimates of the parame- 4. DCC LL INT—DCC by log-likelihood for integrated
ters. The drawback to this method is that ARMA with nearly process
equal roots are numerically unstable and tricky to estimate. A 5. DCC LL MR—DCC by log-likelihood with mean revert-
further drawback is that in the multivariate setting, there are ing model as in (18)
many such cross products that can be used for this estimation. 6. MA100—moving average of 100 days
The problem is even easier if the model is (17) because then 7. EX .06—exponential smoothing with parameter D 006
the autoregressive root is assumed to be 1. The model is sim- 8. OGARCH—orthogonal GARCH or principle compo-
ply an integrated moving average ( IMA) with no intercept, nents GARCH as in (8).

ãei1 j1 t D ƒ‚4ei1 j1 tƒ1 ƒ qi1 j1 tƒ1 5 C 4ei1 j1 t ƒ qi1 j1 t 51 (35) Three performance measures are used. The Ž rst is simply
the comparison of the estimated correlations with the true cor-
which is simply an exponential smoother with parameter ‹ D relations by mean absolute error. This is deŽ ned as
‚. This estimator is called the DCC IMA.
1X
MAE D —O t ƒ t —1 (37)
5. COMPARISON OF ESTIMATORS T

In this section, several correlation estimators are com- and of course the smallest values are the best. A second mea-
pared in a setting where the true correlation structure is sure is a test for autocorrelation of the squared standardized
known. A bivariate GARCH model is simulated 200 times for residuals. For a multivariate problem, the standardized residu-
1,000 observations or approximately 4 years of daily data for als are deŽ ned as
each correlation process. Alternative correlation estimators are
t D Htƒ1=2 rt 1 (38)
compared in terms of simple goodness-of-Ž t statistics, multi-
variate GARCH diagnostic tests, and value-at-risk tests.
which in this bivariate case is implemented with a triangular
The data-generating process consists of two Gaussian
square root deŽ ned as
GARCH models; one is highly persistent and the other is not.
p
11 t D r11 t = H111 t 1
h11 t D 001 C 005r112 tƒ1 C 094h11 tƒ1 1
1 O t (39)
h21 t D 05 C 02r212 tƒ1 C 05h21 tƒ1 1 21 t D r21 t p ƒ r11 t p 0
³ ´ µ ³ ´¶ H221 t 41 ƒ O 2t 5 H111 t 41 ƒ O 2t 5
˜11 t 1 t
N 01 1 (36)
˜21 t t 1 The test is computed as an F test from the regression of 112 t
q q and 212 t on Ž ve lags of the squares and cross products of the
r11 t D h11 t ˜11 t 1 r21 t D h21 t ˜21 t 0 standardized residuals plus an intercept. The number of rejec-
tions using a 5% critical value is a measure of the perfor-
The correlations follow several processes that are labeled as
mance of the estimator because the more rejections, the more
follows:
evidence that the standardized residuals have remaining time
1. Constant t D 09 varying volatilities. This test obviously can be used for real
2. Sine t D 05 C 04 cos42 t=2005 data.
3. Fast sine t D 05 C 04 cos42 t=205 The third performance measure is an evaluation of the esti-
4. Step t D 09 ƒ 054t > 5005 mator for calculating value at risk. For a portfolio with w
5. Ramp t D mod 4t=2005 invested in the Ž rst asset and 41 ƒ w5 in the second, the value
These processes were chosen because they exhibit rapid at risk, assuming normality, is
changes, gradual changes, and periods of constancy. Some of q p
VaR t D 1065 4w 2 H111t C41 ƒw52 H221t C 2ü w41 ƒw5O t H111t H221 t 51 (40)
the processes appear mean reverting and others have abrupt
changes. Various other experiments are done with different
and a dichotomous variable called hit should be unpredictable
error distributions and different data-generating parameters but
based on the past where hit is deŽ ned as
the results are quite similar.
Eight different methods are used to estimate the
hitt D I4w ü r11 t C 41 ƒ w5ü r21 t < ƒVaRt 5 ƒ 0050 (41)
correlations—two multivariate GARCH models, orthogonal
GARCH, two integrated DCC models, and one mean revert- The dynamic quantile test introduced by Engle and Manganelli
ing DCC plus the exponential smoother from Riskmetrics and (2001) is an F test of the hypothesis that all coefŽ cients as well
the familiar 100-day moving average. The methods and their as the intercept are zero in a regression of this variable on its
descriptions are as follows: past, on current VaR, and any other variables. In this case, Ž ve
1. Scalar BEKK—scalar version of (10) with variance tar- lags and the current VaR are used. The number of rejections
geting as in (12) using a 5% critical value is a measure of model performance.
2. Diag BEKK—diagonal version of (10) with variance tar- The reported results are for an equal weighted portfolio with
geting as in (11) w D 05 and a hedge portfolio with weights 11 ƒ1.
344 Journal of Business & Economic Statistics, July 2002

1.0 1.0

0.9
0.8
0.8

0.6
0.7

0.6
0.4

0.5
0.2
0.4

0.0 0.3
200 400 600 800 1000 200 400 600 800 1000

RHO_SINE RHO_STEP

1.0 1.0

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
200 400 600 800 1000 200 400 600 800 1000

RHO_RAMP RHO_FASTSINE

Figure 1. Correlation Experiments.

6. RESULTS log-likelihood; overall, it is also the best method, followed by

the mean reverting DCC and the IMA DCC.
Table 1 presents the results for the mean absolute error
The value-at-risk test based on the long-short portfolio Ž nds
(MAE) for the eight estimators for six experiments with 200
that the diagonal BEKK is best for three of six, whereas the
replications. In four of the six cases the DCC mean reverting
DCC MR is best for two. Overall, the DCC MR is observed
model has the smallest MAE. When these errors are summed
to be the best.
over all cases, this model is the best. Very close second-
and third-place models are DCC integrated with log-likelihood From all these performance measures, the DCC methods are
estimation and scalar BEKK. either the best or very near the best method. Choosing among
In Table 2 the second standardized residual is tested for these models, the mean reverting model is the general winner,
remaining autocorrelation in its square. This is the more although the integrated versions are close behind and perform
revealing test because it depends on the correlations; the test best by some criteria. Generally the log-likelihood estimation
for the Ž rst residual does not. Because all models are mis- method is superior to the IMA estimator for the integrated
speciŽ ed, the rejection rates are typically well above 5%. For DCC models.
three of six cases, the DCC mean reverting model is the best. The conŽ dence with which these conclusions can be drawn
When summed over all cases it is a clear winner. can also be investigated. One simple approach is to repeat the
The test for autocorrelation in the Ž rst squared standardized experiment with different sets of random numbers. The entire
residual is presented in Table 3. These test statistics are typi- Monte Carlo was repeated two more times. The results are
cally close to 5%, re ecting the fact that many of these mod- very close with only one change in ranking that favors the
els are correctly speciŽ ed and the rejection rate should be the DCC LL MR over the DCC LL INT.
size. Overall the best model appears to be the diagonal BEKK
with OGARCH and DCC close behind.
The VaR-based dynamic quantile test is presented in Table 4
7. EMPIRICAL RESULTS
for a portfolio that is half invested in each asset and in Table 5
for a long-short portfolio with weights plus and minus one. Empirical examples of these correlation estimates are pre-
The number of 5% rejections for many of the models is close sented for several interesting series. First the correlation
to the 5% nominal level despite misspeciŽ cation of the struc- between the Dow Jones Industrial Average and the NASDAQ
ture. In Ž ve of six cases, the minimum is the integrated DCC composite is examined for 10 years of daily data ending
Engle: Dynamic Conditional Correlation 345

Table 1. Mean Absolute Error of Correlation Estimates

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT DCC IMA EX .06 MA 100 O-GARCH

FAST SINE .2292 .2307 .2260 .2555 .2581 .2737 .2599 .2474
SINE .1422 .1451 .1381 .1455 .1678 .1541 .3038 .2245
STEP .0859 .0931 .0709 .0686 .0672 .0810 .0652 .1566
RAMP .1610 .1631 .1546 .1596 .1768 .1601 .2828 .2277
CONST .0273 .0276 .0070 .0067 .0105 .0276 .0185 .0449
T(4) SINE .1595 .1668 .1478 .1583 .2199 .1599 .3016 .2423

Table 2. Fraction of 5% Tests Finding Autocorrelation in Squared Standardized Second Residual

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT DCC IMA EX .06 MA 100 O-GARCH

FAST SINE .3100 .0950 .1300 .3700 .3700 .7250 .9900 .1100
SINE .5930 .2677 .1400 .1850 .3350 .7600 1.0000 .2650
STEP .8995 .6700 .2778 .3250 .6650 .8550 .9950 .7600
RAMP .5300 .2600 .2400 .5450 .7500 .7300 1.0000 .2200
CONST .9800 .3600 .0788 .0900 .1250 .9700 .9950 .9350
T(4) SINE .2800 .1900 .2050 .2400 .1650 .3300 .8950 .1600

Table 3. Fraction of 5% Tests Finding Autocorrelation in Squared Standardized First Residual

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT DCC IMA EX .06 MA 100 O-GARCH

FAST SINE .2250 .0450 .0600 .0600 .0650 .0750 .6550 .0600
SINE .0804 .0657 .0400 .0300 .0600 .0400 .6250 .0400
STEP .0302 .0400 .0505 .0500 .0450 .0300 .6500 .0250
RAMP .0550 .0500 .0500 .0600 .0600 .0650 .7500 .0400
CONST .0200 .0250 .0242 .0250 .0250 .0400 .6350 .0150
T(4) SINE .0950 .0550 .0850 .0800 .0950 .0850 .4900 .1050

Table 4. Fraction of 5% Dynamic Quantile Tests Rejecting Value at Risk, Equal Weighted

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT DCC IMA EX .06 MA 100 O-GARCH

FAST SINE .0300 .0450 .0350 .0300 .0450 .2450 .4350 .1200
SINE .0452 .0556 .0250 .0350 .0350 .1600 .4100 .3200
STEP .1759 .1650 .0758 .0650 .0800 .2450 .3950 .6100
RAMP .0750 .0650 .0500 .0400 .0450 .2000 .5300 .2150
CONST .0600 .0800 .0667 .0550 .0550 .2600 .4800 .2650
T(4) SINE .1250 .1150 .1000 .0850 .1200 .1950 .3950 .2050

Table 5. Fraction of 5% Dynamic Quantile Tests Rejecting Value at Risk, Long-Short

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT DCC IMA EX .06 MA 100 O-GARCH

FAST SINE .1000 .0950 .0900 .2550 .2550 .5800 .4650 .0850
SINE .0553 .0303 .0450 .0900 .1850 .2150 .9450 .0650
STEP .1055 .0850 .0404 .0600 .1150 .1700 .4600 .1250
RAMP .0800 .0650 .0800 .1750 .2500 .3050 .9000 .1000
CONST .1850 .0900 .0424 .0550 .0550 .3850 .5500 .1050
T(4) SINE .1150 .0900 .1350 .1300 .2000 .2150 .8050 .1050
346 Journal of Business & Economic Statistics, July 2002

in March 2000. Then daily correlations between stocks and 1 .0

bonds, a central feature of asset allocation models, are
considered. Finally, the daily correlation between returns on 0 .9
several currencies around major historical events including the
launch of the Euro is examined. Each of these datasets has 0 .8

been used to estimate all the models described previously. The

0 .7
DCC parameter estimates for the integrated and mean revert-
ing models are exhibited with consistent standard errors from
0 .6
(33) in Appendix A. In that Appendix, the statistic referred to
as likelihood ratio is the difference between the log-likelihood 0 .5
of the second-stage estimates using the integrated model and
using the mean reverting model. Because these are not jointly 0 .4
maximized likelihoods, the distribution could be different from
its conventional chi-squared asymptotic limit. Furthermore, 0 .3
nonnormality of the returns would also affect this limiting 90 91 92 93 94 95 96 97 98 99
distribution.
D C C_ INT _ DJ_ NQ

Figure 3. Ten Years of Dow Jones–NASDAQ Correlations.

7.1 Dow Jones and NASDAQ
The dramatic rise in the NASDAQ over the last part of
the 1990s perplexed many portfolio managers and delighted can be seen that only the orthogonal GARCH correlations fail
the new internet start-ups and day traders. A plot of the to decline and that the BEKK correlations are most volatile.
GARCH volatilities of these series in Figure 2 reveals that the
NASDAQ has always been more volatile than the Dow but
that this gap widens at the end of the sample.
7.2 Stocks and Bonds
The correlation between the Dow and NASDAQ was esti-
mated with the DCC integrated method, using the volatili- The second empirical example is the correlation between
ties in Figure 2. The results, shown in Figure 3, are quite domestic stocks and bonds. Taking bond returns to be minus
interesting. the change in the 30-year benchmark yield to maturity, the
Whereas for most of the decade the correlations were correlation between bond yields and the Dow and NASDAQ
between .6 and .9, there were two notable drops. In 1993 the are shown in Figure 5 for the integrated DCC for the last 10
correlations averaged .5 and dropped below .4, and in March years. The correlations are generally positive in the range of
2000 they again dropped below .4. The episode in 2000 is .4 except for the summer of 1998, when they become highly
sometimes attributed to sector rotation between new economy negative, and the end of the sample, when they are about 0.
stocks and “brick and mortar” stocks. The drop at the end Although it is widely reported in the press that the NASDAQ
of the sample period is more pronounced for some estimators does not seem to be sensitive to interest rates, the data suggests
than for others. Looking at just the last year in Figure 4, it that this is true only for some limited periods, including the
Ž rst quarter of 2000, and that this is also true for the Dow.
Throughout the decade it appears that the Dow is slightly more
60 correlated with bond prices than is the NASDAQ.

50
7.3 Exchange Rates
40
Currency correlations show dramatic evidence of nonsta-
tionarity. That is, there are very pronounced apparent structural
30 changes in the correlation process. In Figure 6, the breakdown
of the correlations between the Deutschmark and the pound
20 and lira in August of 1992 is very apparent. For the pound
this was a return to a more normal correlation, while for the
10 lira it was a dramatic uncoupling.
Figure 7 shows currency correlations leading up to the
0 launch of the Euro in January 1999. The lira has lower cor-
90 91 92 93 94 95 96 97 98 99 relations with the Franc and Deutschmark from 93 to 96, but
then they gradually approach one. As the Euro is launched,
VO L_DJ_G ARCH VO L_NQ _G ARCH the estimated correlation moves essentially to 1. In the last
year it drops below .95 only once for the Franc and lira and
Figure 2. Ten Years of Volatilities. not at all for the other two pairs.
Engle: Dynamic Conditional Correlation 347

.9 .9

.8 .8

.7 .7

.6 .6

.5 .5

.4 .4

.3 .3
1999:07 1999:10 2000:01 1999:04 1999:07 1999:10 2000:01

D C C _IN T_D J_N Q D C C _M R _D J_N Q

.9 .8

.8
.7
.7

.6
.6

.5
.5

.4
.4
.3

.2 .3
1999:07 1999:10 2000:01 1999:04 1999:07 1999:10 2000:01

D IAG _BE K K_D J_N Q O G A R C H _D J_N Q

Figure 4. Correlations from March 1999 to March 2000.

.8

.6 1.0

0.9
.4
0.8
.2
0.7
.0
0.6
-.2
0.5
-.4
0.4

-.6
0.3
90 91 92 93 94 95 96 97 98 99 86 87 88 89 90 91 92 93 94 95

DC C_INT_DJ_BO ND DCC_INT _NQ _BO ND DCC_INT_RDEM_RGBP DCC_INT_RDEM_RITL

Figure 5. Ten Years of Bond Correlations. Figure 6. Ten Years of Currency Correlations.
348 Journal of Business & Economic Statistics, July 2002

1.0 If a 1% test is used re ecting the larger sample size, then
the number of rejections ranges from 7 to 21. Again the MA
0.8 100 is the worst but now the EX .06 is the winner. The DCC
LL MR, DCC LL INT, and diagonal BEKK are all tied for
0.6
second with 9 rejections each.
0.4
The implications of this comparison are mainly that a big-
ger and more systematic comparison is required. These results
0.2 suggest Ž rst that real data are more complicated than any of
these models. Second, it appears that the DCC models are
0.0 competitive with the other methods, some of which are difŽ -
cult to generalize to large systems.
-0.2

1994 1995 1996 1997 1998

8. CONCLUSIONS
DCC_INT_RDEM_RFRF DCC_INT_RFRF_RITL
DCC_INT_RDEM_RITL In this article a new family of multivariate GARCH mod-
els was proposed that can be simply estimated in two steps
Figure 7. Currency Correlations. from univariate GARCH estimates of each equation. A maxi-
mum likelihood estimator was proposed and several different
speciŽ cations were suggested. The goal for this proposal is to
Ž nd speciŽ cations that potentially can estimate large covari-
From the results in Appendix A, it is seen that this is the ance matrices. In this article, only bivariate systems were
only dataset for which the integrated DCC cannot be rejected estimated to establish the accuracy of this model for sim-
against the mean reverting DCC. The nonstationarity in these pler structures. However, the procedure was carefully deŽ ned
correlations presumably is responsible. It is somewhat surpris- and should also work for large systems. A desirable practical
ing that a similar result is not found for the prior currency feature of the DCC models is that multivariate and univari-
pairs. ate volatility forecasts are consistent with each other. When
new variables are added to the system, the volatility fore-
casts of the original assets will be unchanged and correlations
7.4 Testing the Empirical Models may even remain unchanged, depending on how the model is
revised.
For each of these datasets, the same set of tests that were
used in the Monte Carlo experiment can be constructed. In this The main Ž nding is that the bivariate version of this
case of course, the mean absolute errors cannot be observed, model provides a very good approximation to a variety of
but the tests for residual ARCH can be computed and the tests time-varying correlation processes. The comparison of DCC
for value at risk can be computed. In the latter case, the results with simple multivariate GARCH and several other estimators
are subject to various interpretations because the assumption shows that the DCC is often the most accurate. This is true
of normality is a potential source of rejection. In each case whether the criterion is mean absolute error, diagnostic tests,
the number of observations is larger than in the Monte Carlo or tests based on value at risk calculations.
experiment, ranging from 1,400 to 2,600. Empirical examples from typical Ž nancial applications are
The p-statistics for each of four tests are given in Appendix quite encouraging because they reveal important time-varying
B. The tests are the tests for residual autocorrelation in squares features that might otherwise be difŽ cult to quantify. Statisti-
and for accuracy of value at risk for two portfolios. The two cal tests on real data indicate that all these models are mis-
portfolios are an equally weighted portfolio and a long-short speciŽ ed but that the DCC models are competitive with the
portfolio. They presumably are sensitive to rather different multivariate GARCH speciŽ cations and are superior to moving
failures of correlation estimates. From the four tables, it is average methods.
immediately clear that most of the models are misspeciŽ ed for
most of the data sets. If a 5% test is done for all the datasets
on each of the criteria, then the expected number of rejec- ACKNOWLEDGMENTS
tions for each model would be just over 1 of 28 possibilities.
Across the models there are from 10 to 21 rejections at the This research was supported by NSF grant SBR-9730062
5% level! and NBER AP group. The author thanks Kevin Sheppard
Without exception, the worst performer on all of the tests for research assistance and Pat Burns and John Geweke for
and datasets is the moving average model with 100 lags. From insightful comments. Thanks also to seminar participants at
counting the total number of rejections, the best model is the New York University; University of California at San Diego;
diagonal BEKK with 10 rejections. The DCC LL MR, scalar Academica Sinica; Taiwan; C IRANO at Montreal; University
BEKK, O_GARCH, and EX .06 all have 12 rejections, and of Iowa; Journal of Applied Econometrics Lectures, Cam-
the DCC LL INT has 14. Probably, these differences are not bridge, England; CNRS Aussois; Brown University; Fields
large enough to be convincing. Institute University of Toronto; and Riskmetrics.
Engle: Dynamic Conditional Correlation 349

APPENDIX A

Mean Reverting Model Integrated Model

Asset 1 Asset 2 Asset 1 Asset 2

NQ DJ NQ DJ

Parameter T-stat Log-likelihood Parameter T-stat Log-likelihood

alphaDCC 0039029 60916839405 lambdaDCC 0030255569 4066248 18062079651

betaDCC 0941958 92072739572 1807905857 LR TEST 33057836423

Asset 1 Asset 2 Asset 1 Asset 2

RATE DJ RATE DJ

Parameter T-stat Log-likelihood Parameter T-stat Log-likelihood

alphaDCC 0037372 20745870787 lambdaDCC 002851073 30675969 13188063653

betaDCC 0950269 44042479805 13197082499 LR TEST 18037690833

Asset 1 Asset 2 Asset 1 Asset 2

NQ RATE NQ RATE

Parameter T-stat Log-likelihood Parameter T-stat Log-likelihood

alphaDCC 0029972 20652315309 lambdaDCC 0019359061 20127002 12578006669

betaDCC 0953244 46061344925 12587026244 LR TEST 18039149373

Asset 1 Asset 2 Asset 1 Asset 2

DM ITL DM ITL

Parameter T-stat Log-likelihood Parameter T-stat Log-likelihood

alphaDCC 00991 30953696951 lambdaDCC 0052484321 40243317 2097605062

betaDCC 0863885 21032994852 21041071874 LR TEST 1304250734

Asset 1 Asset 2 Asset 1 Asset 2

DM GBP DM GBP

Parameter T-stat Log-likelihood Parameter T-stat Log-likelihood

alphaDCC 003264 10315852908 lambdaDCC 0024731692 10932782 19480021203

betaDCC 0963504 37057905053 1950806083 LR TEST 56079255661

Asset 1 Asset 2 Asset 1 Asset 2

rdem90 rfrf90 rdem90 rfrf90

Parameter T-stat Log-likelihood Parameter T-stat Log-likelihood

alphaDCC 0059413 40154987386 lambdaDCC 0047704833 20880988 12416084873

betaDCC 0934458 59019216459 12426089065 LR TEST 20008382828

Asset 1 Asset 2 Asset 1 Asset 2

rdem90 ritl90 rdem90 ritl90

Parameter T-stat Log-likelihood Parameter T-stat Log-likelihood

alphaDCC 0056675 30091462338 lambdaDCC 0053523717 20971859 11442050983

betaDCC 0943001 5077614662 11443023811 LR TEST 10456541924

NOTE: Empirical results for bivariate DCC models. T -statistics are robust and consistent using (33). The estimates in the left column are DCC LL MR and the estimates in the right column are
DCC LL INT. The LR statistic is twice the difference between the log likelihoods of the second stage. The data are all logarithmic differences: NQDNasdaq composite, DJDDow Jones Industrial
Average, RATEDreturn on 30 year US Treasury, all daily from 3/23/90 to 3/22/00. Furthermore: DMDDeutschmarks per dollar, ITLDItalian Lira per dollar, GBPDBritish pounds per dollar, all from
1/1/85 to 2/13/95. Finally rdem90DDeutschmarks per dollar, rfrf90DFrench Francs per dollar, and ritl90DItalian Lira per dollar, all from 1/1/93 to 1/15/99.
350 Journal of Business & Economic Statistics, July 2002

APPENDIX B
P-Statistics From Tests of Empirical Models
ARCH in Squared RESID1

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT EX .06 MA100 O-GARCH

NASD&DJ .0047 .0281 .3541 .3498 .3752 .0000 .2748

DJ&RATE .0000 .0002 .0003 .0020 .0167 .0000 .0001
NQ&RATE .0000 .0044 .0100 .0224 .0053 .0000 .0090
DM&ITL .4071 .3593 .2397 .1204 .5503 .0000 .4534
DM&GBP .4437 .4303 .4601 .3872 .4141 .0000 .4213
FF&DM90 .2364 .2196 .1219 .1980 .3637 .0000 .0225
DM&IT90 .1188 .3579 .0075 .0001 .0119 .0000 .0010

ARCH in Squared RESID2

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT EX .06 MA100 O-GARCH

NASD&DJ .0723 .0656 .0315 .0276 .0604 .0000 .0201

DJ&RATE .7090 .7975 .8251 .6197 .8224 .0007 .1570
NQ&RATE .0052 .0093 .0075 .0053 .0023 .0000 .1249
DM&ITL .0001 .0000 .0000 .0000 .0000 .0000 .0000
DM&GBP .0000 .0000 .0000 .0000 .1366 .0000 .4650
FF&DM90 .0002 .0010 .0000 .0000 .0000 .0000 .0018
DM&IT90 .0964 .1033 .0769 .1871 .0431 .0000 .5384

Dynamic Quantile Test VaR1

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT EX .06 MA100 O-GARCH

NASD&DJ .0001 .0000 .0000 .0000 .0002 .0000 .0018

DJ&RATE .7245 .4493 .3353 .4521 .5977 .4643 .2085
NQ&RATE .5923 .5237 .4248 .3203 .2980 .4918 .8407
DM&ITL .1605 .2426 .1245 .0001 .3892 .0036 .0665
DM&GBP .4335 .4348 .4260 .3093 .1468 .0026 .1125
FF&DM90 .1972 .2269 .1377 .1375 .0652 .1972 .2704
DM&IT90 .1867 .0852 .5154 .7406 .1048 .4724 .0038

Dynamic Quantile Test VaR2

MODEL SCAL BEKK DIAG BEKK DCC LL MR DCC LL INT EX .06 MA100 O-GARCH

NASD&DJ .0765 .1262 .0457 .0193 .0448 .0000 .0005

DJ&RATE .0119 .6219 .6835 .4423 .0000 .1298 .3560
NQ&RATE .0432 .4324 .4009 .6229 .0004 .4967 .3610
DM&ITL .0000 .0000 .0000 .0000 .0209 .0081 .0000
DM&GBP .0006 .0043 .0002 .0000 .1385 .0000 .0003
FF&DM90 .4638 .6087 .7098 .0917 .4870 .1433 .5990
DM&IT90 .2130 .4589 .2651 .0371 .3248 .0000 .1454

NOTE: Data are the same as in the previous table and tests are based on the results in the previous table.

[Received January 2002. Revised January 2002.]

Ding, Z., and Engle, R. (2001), “Large Scale Conditional Covariance Matrix
Modeling, Estimation and Testing,” Academia Economic Papers, 29,
REFERENCES 157–184.
Engle, R., and Kroner, K. (1995), “Multivariate Simultaneous GARCH,”
Alexander, C. O. (1998), “Volatility and Correlation: Methods, Models and Econometric Theory, 11, 122–150.
Applications,” in Risk Manangement and Analysis: Measuring and Mod- Engle, R., and Manganelli, S. (1999), “CAViaR: Conditional Value at Risk by
elling Financial Risk, ed. C. O. Alexander, New York: John Wiley. Regression Quantiles,” NBER Working Paper 7341.
(2001), “Orthogonal GARCH,” Mastering Risk (Vol. 2), ed. C. O. Engle, R., and Mezrich, J. (1996), “GARCH for Groups,” Risk, 9, 36–40.
Alexander, New York: Prentice–Hall, pp. 21–38. Engle, R., and Sheppard, K. (2001), “Theoretical and Empirical Properties of
Bollerslev, T. (1990), “Modeling the Coherence in Short-Run Nominal Dynamic Conditional Correlation Multivariate GARCH,” NBER Working
Exchange Rates: A Multivariate Generalized ARCH Model,” Review of Paper 8554, National Bureau of Economic Research.
Economics and Statistics, 72, 498–505. Engle, R., Ng, V., and Rothschild, M. (1990), “Asset Pricing With a Fac-
Bollerslev, T., Chao, R. Y., and Kroner, K. F. (1992), “ARCH Modeling in tor ARCH Covariance Structure: Empirical Estimates for Treasury Bills,”
Journal of Econometrics, 45, 213–238.
Finance: A Review of the Theory and Empirical Evidence,” Journal of
Kroner, K. F., and Claessens, S. (1991), “Optimal Dynamic Hedging Portfo-
Econometrics, 52, 115–128. lios and the Currency Composition of External Debt,” Journal of Interna-
Bollerslev, T., Engle, R., and Wooldridge, J. M. (1988), “A Capital Asset Pric- tional Money and Finance, 10, 131–148.
ing Model With Time Varying Covariances,” Journal of Political Economy, Kroner, K. F., and Ng, V. K. (1998), “Modeling Asymmetric Comovements
96, 116–131. of Asset Returns,” Review of Financial Studies, pp. 817–844.
Bollerslev, T., Engle, R., and Nelson, D. (1994), “ARCH Models,” Handbook Newey, W., and McFadden, D. (1994), “Large Sample Estimation and Hypoth-
of Econometrics (Vol. 4), eds. R. Engle and D. McFadden, Amsterdam: esis Testing,” in Handbook of Econometrics (Vol. 4), eds. R. Engle and
North–Holland, pp. 2959–3038. D. McFadden, New York: Elsevier Science, pp. 2113–2245.