A Simple Method for

Predicting Covariance Matrices

of Financial Returns
Suggested Citation: Kasper Johansson, Mehmet G. Ogut, Markus Pelger,
Thomas Schmelzer and Stephen Boyd (2023), “A Simple Method for
Predicting Covariance Matrices
of Financial Returns”, : Vol. 12, No. 4, pp 324–407. DOI: 10.1561/0800000047.

Kasper Johansson
Stanford University
[email protected]
Mehmet G. Ogut
Stanford University
[email protected]
Markus Pelger
Stanford University
[email protected]
Thomas Schmelzer
Stanford University
Abu Dhabi Investment Authority
[email protected]
Stephen Boyd
Stanford University
[email protected]

This article may be used only for the purpose of research, teaching,
and/or private study. Commercial use or systematic downloading (by
robots or other automatic processes) is prohibited without explicit
Publisher approval.
Boston — Delft
 Contents

1 Introduction 325
1.1 Covariance prediction . . . . . . . . . . . . . . . . . . . . 325
1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . 326
1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

2 Some common covariance predictors 328

2.1 Rolling window . . . . . . . . . . . . . . . . . . . . . . . 328
2.2 EWMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
2.3 GARCH and MGARCH . . . . . . . . . . . . . . . . . . . 329
2.4 DCC GARCH . . . . . . . . . . . . . . . . . . . . . . . . 331
2.5 Iterated EWMA . . . . . . . . . . . . . . . . . . . . . . . 332

3 Combined multiple iterated EWMAs 334

3.1 Dynamically weighted prediction combiner . . . . . . . . . 334
3.2 Choosing the weights via convex optimization . . . . . . . 335

4 Evaluating covariance predictors 337

4.1 Mean squared error . . . . . . . . . . . . . . . . . . . . . 338
4.2 Log-likelihood . . . . . . . . . . . . . . . . . . . . . . . . 338
4.3 Log-likelihood regret . . . . . . . . . . . . . . . . . . . . . 339
4.4 Portfolio performance . . . . . . . . . . . . . . . . . . . . 340
5 Data sets and experimental setup 344
5.1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
5.2 Six covariance predictors . . . . . . . . . . . . . . . . . . 350

6 Results 352
6.1 CM-IEWMA component weights . . . . . . . . . . . . . . 352
6.2 Mean squared error . . . . . . . . . . . . . . . . . . . . . 355
6.3 Log-likelihood and log-likelihood regret . . . . . . . . . . . 355
6.4 Portfolio performance . . . . . . . . . . . . . . . . . . . . 363
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 372

7 Realized covariance 373

7.1 Combined multiple realized EWMAs . . . . . . . . . . . . 374
7.2 Data and experimental setup . . . . . . . . . . . . . . . . 375
7.3 Empirical results . . . . . . . . . . . . . . . . . . . . . . . 375

8 Large universes 381

8.1 Traditional factor model . . . . . . . . . . . . . . . . . . . 382
8.2 Fitting a factor model to a covariance matrix . . . . . . . 383
8.3 Data and experimental setup . . . . . . . . . . . . . . . . 386
8.4 Empirical results . . . . . . . . . . . . . . . . . . . . . . . 387

9 Smooth covariance predictions 391

9.1 Data and experimental setup . . . . . . . . . . . . . . . . 392
9.2 Empirical results . . . . . . . . . . . . . . . . . . . . . . . 392

10 Simulating returns 400

10.1 Data and experimental setup . . . . . . . . . . . . . . . . 400
10.2 Empirical results . . . . . . . . . . . . . . . . . . . . . . . 401

11 Conclusions 403

Acknowledgements 404

References 405
A Simple Method for
Predicting Covariance Matrices
of Financial Returns
Kasper Johansson1 , Mehmet G. Ogut2 , Markus Pelger3 ,
Thomas Schmelzer4 and Stephen Boyd5
1 Department of Electrical Engineering, Stanford University;
[email protected]
2 Department of Electrical Engineering, Stanford University;

[email protected]
3 Department of Management Science and Engineering, Stanford

University; [email protected]
4 Department of Electrical Engineering, Stanford University, and Abu

Dhabi Investment Authority; [email protected]

5 Department of Electrical Engineering, Stanford University;

[email protected]

Kasper Johansson, Mehmet G. Ogut, Markus Pelger, Thomas Schmelzer and

Stephen Boyd (2023), “A Simple Method for
Predicting Covariance Matrices
of Financial Returns”, : Vol. 12, No. 4, pp 324–407. DOI: 10.1561/0800000047.
©2024 A. Heezemans and M. Casey
324

ABSTRACT
We consider the well-studied problem of predicting the time-
varying covariance matrix of a vector of financial returns.
Popular methods range from simple predictors like rolling
window or exponentially weighted moving average (EWMA)
to more sophisticated predictors such as generalized autore-
gressive conditional heteroscedastic (GARCH) type methods.
Building on a specific covariance estimator suggested by En-
gle in 2002, we propose a relatively simple extension that
requires little or no tuning or fitting, is interpretable, and
produces results at least as good as MGARCH, a popular
extension of GARCH that handles multiple assets. To eval-
uate predictors we introduce a novel approach, evaluating
the regret of the log-likelihood over a time period such as
a quarter. This metric allows us to see not only how well a
covariance predictor does over all, but also how quickly it
reacts to changes in market conditions. Our simple predic-
tor outperforms MGARCH in terms of regret. We also test
covariance predictors on downstream applications such as
portfolio optimization methods that depend on the covari-
ance matrix. For these applications our simple covariance
predictor and MGARCH perform similarly.
 1
Introduction

1.1 Covariance prediction

We consider cross-sections, e.g., a vector time series of n financial returns,

denoted rt ∈ Rn , t = 1, 2, . . ., where (rt )i is the return of asset i from
t − 1 to t. We focus on the case where the mean E rt is small enough
that the second moment E rt rtT ∈ Rn×n is a good approximation of the
covariance matrix cov(rt ) = E rt rtT − (E rt )(E rt )T , where E denotes
expectation. This is the case for most daily, weekly, or monthly stock,
bond, and futures returns, factor returns, and index returns. We start
by focussing on the case where the number of assets n is modest, say,
on the order 10–100 or so; in chapter 8 we explain how to extend the
method to much larger universes using ideas such as factor models.
We model the returns rt as independent random variables with zero
mean and covariance Σt ∈ Sn++ (the set of symmetric positive definite
matrices). We focus on the problem of predicting or estimating Σt , based
on knowledge of r1 , . . . , rt−1 . The prediction is denoted as Σ̂t ∈ Sn++ .
The predicted volatilities of assets are given by

σ̂t = diag(Σ̂t )1/2 ∈ Rn ,

where diag with a matrix argument is the vector of diagonal entries of

325
326 Introduction

the matrix, and the squareroot of a vector above is elementwise. We

denote the predicted correlations as

R̂t = diag(σ̂t )−1 Σ̂t diag(σ̂t )−1 ,

where diag with a vector argument is the diagonal matrix with entries
from the vector argument.
Covariance estimation comes up in several areas of finance, including
Markowitz portfolio construction (Markowitz, 1952; Grinold and Kahn,
2000), risk management (McNeil et al., 2015), and asset pricing (Sharpe,
1964). Much attention has been devoted to this problem, and a Nobel
Memorial Prize in Economic Sciences was awarded for work directly
related to volatility estimation (Engle, 1982).
While it is well known that the tails of financial returns are poorly
modeled by a Gaussian distribution, our focus here is on the bulk of the
distribution, where the Gaussian assumption is reasonable. For future
use, we note that the log-likelihood of an observed return rt , under the
Gaussian distribution rt ∼ N (0, Σ̂t ), is
1 
lt (Σ̂t ) = −n log(2π) − log det Σ̂t − rtT Σ̂−1 rt . (1.1)
2 t

The Gaussian log-likelihood is closely related to a popular metric for

evaluating covariance predictors in econometrics, called the (Gaussian)
quasi-likelihood (QLIKE) (Patton, 2011; Patton and Sheppard, 2009;
Laurent et al., 2013). QLIKE is the negative log-likelihood, under the
Gaussian assumption, up to an additive constant and a positive scale
factor. Roughly speaking, we seek covariance predictors that achieve
large values of log-likelihood, or small values of QLIKE, on realized
returns. We will describe evaluation of covariance predictors in detail
in chapter 4.

1.2 Contributions

This monograph makes three contributions. First, we propose a new

method for predicting the time-varying covariance matrix of a vector of
financial returns, building on a specific covariance estimator suggested by
Engle in 2002. Our method is a relatively simple extension that requires
1.3. Outline 327

very little tuning and is readily interpretable. It relies on solving a small

convex optimization problem, which can be carried out very quickly
and reliably (Boyd and Vandenberghe, 2004). Our method performs as
well as much more complex methods, as measured by several metrics.
Our second contribution is to propose a new method for evaluating
a covariance predictor, by considering the regret of the log-likelihood
over some time period such as a quarter. This approach allows us to
evaluate how quickly a covariance estimator reacts to changes in market
conditions.
Our third contribution is an extensive empirical study of covari-
ance predictors. We compare our new method to other popular predic-
tors, including rolling window, exponentially weighted moving average
(EWMA), and generalized autoregressive conditional heteroscedastic
(GARCH) type methods. We find that our method performs slightly
better than other predictors. However, even the simplest predictors
perform well for practical problems like portfolio optimization.
Everything needed to reproduce our results, together with an open
source implementation of our proposed covariance predictor, is available
online at

https://github.com/cvxgrp/cov_pred_finance.

1.3 Outline

In chapter 2 we describe some common predictors, including the one that

our method builds on. We introduce our proposed covariance predictor
in chapter 3. In chapter 4 we discuss methods for validating covariance
predictors that measure both overall performance and reactivity to
market changes. We describe the data we use in our first empirical
studies in chapter 5, and give the results in chapter 6.
In the next chapters we discuss some extensions of and variations
on our method, including realized covariance prediction (chapter 7),
handling large universes via factor models (chapter 8), obtaining smooth
covariance estimates (chapter 9), and using our covariance model to
generate simulated returns (chapter 10).
 2
Some common covariance predictors

In this chapter we review some common covariance predictors, ranging

from simple to complex, with the goal of giving context and fixing our
notation. To simplify some formulas, we take rτ = 0 for τ ≤ 0.

2.1 Rolling window

The rolling window predictor with window length or memory M is the

average of the last M ≥ n outer products,
t−1
Σ̂t = αt t = 2, 3, . . . ,
X
rτ rτT ,
τ =t−M

where αt = 1/ min{t − 1, M } is the normalization constant. The rolling

window predictor can be evaluated via the recursion
αt+1
Σ̂t+1 = Σ̂t + αt+1 (rt rtT − rt−M rt−M
T
), t = 1, 2, . . . ,
αt

with initialization Σ̂1 = 0.

For t < n, the rolling window covariance estimate is not full rank.
To handle this, as well as to improve the quality of the prediction, we
can add regularization or shrinkage, for example by adding a positive

328
2.2. EWMA 329

multiple of diag(Σ̂t ) to our estimate (Ledoit and Wolf, 2004; Ledoit

and Wolf, 2003), or approximating the predicted covariance matrix by
a diagonal plus low rank matrix, as described in chapter 8.

2.2 EWMA

The exponentially weighted moving average (EWMA) estimator, with

forgetting factor β ∈ (0, 1), is
t−1
Σ̂t = αt t = 2, 3, . . . , (2.1)
X
β t−1−τ rτ rτT ,
τ =1

where !−1
t−1
1−β
αt = =
X
t−1−τ
β
τ =1
1 − β t−1
is the normalization constant. The forgetting factor β is usually ex-
pressed in terms of the half-life H = − log 2/ log β, for which β H = 1/2.
The half-life H is the number of periods when the exponential weight
has decreased by a factor of two. For example, for a half-life of one year,
the current observed return has twice the impact on our covariance
prediction as the return observed one year ago. The EWMA predictor
is widely used in practice; for example RiskMetrics suggests the for-
getting factor β = 0.94, which corresponds to a half-life of around 11
days (Menchero et al., 2011; Longerstaey and Spencer, 1996).
The EWMA covariance predictor can be computed recursively as
β − βt 1−β
Σ̂t+1 = Σ̂t + rt rT , t = 1, 2, . . . ,
1 − βt 1 − βt t
with initialization Σ̂1 = 0. Like the rolling window predictor, the EWMA
predictor is singular for t < n, which can be handled using the same
regularization methods described above.

2.3 GARCH and MGARCH

GARCH. The generalized autoregressive conditional heteroscedastic

(GARCH) predictor decomposes the return of a single asset as
rt = µ + ϵt ,
330 Some common covariance predictors

where µ is the mean return and ϵt is the innovation, and models the
innovation as
q p
ϵt = σt zt , =ω+ +
X X
σt2 aτ ϵ2t−τ 2
bτ σt−τ ,
τ =1 τ =1

where σt is the asset volatility, zt are independent N (0, 1), and q and p
(often both set to one in practice) determine the GARCH order (Boller-
slev, 1986). (Recall that we assume zero mean.) The model parameters
are ω, a1 , . . . , aq , and b1 , . . . , bp . Estimating the model parameters re-
quires solving a nonconvex optimization problem (Barratt and Boyd,
2022).
With p = 0 we recover the autoregressive conditional heteroscedastic
(ARCH) predictor, introduced in the seminal paper by Engle (1982).
This paper set the foundation for a wide variety of popular volatility
and correlation predictors and earned him the 2003 Nobel Memorial
Prize in Economic Sciences.

MGARCH. There are several ways of extending the GARCH predictor

to a multivariate or vector setting. The most popular is the dynamic
conditional correlation (DCC) predictor (Engle, 2002), which is a two-
step approach described below.
Many other MGARCH predictors have been proposed. The most
straightforward generalization from the univariate to multivariate pre-
dictors is the VEC predictor, where the covariance matrix is vectorized
and each element is modeled as a GARCH process with dependencies
on all other elements (Bollerslev et al., 1988). However, this extension
requires estimating n(n + 1)(n(n + 1) + 1)/2 ≈ n4 /2 parameters, which
can be impractical even for modest values of n.
Following the VEC extension of GARCH, multivariate GARCH
(MGARCH) predictors have been proposed in two lines of develop-
ment (Silvennoinen and Teräsvirta, 2009). The first line involves models
that impose restrictions on the parameters of the VEC predictor, in-
cluding DVEC (Bollerslev, 1986), BEKK (Engle and Kroner, 1995), FF-
MGARCH (Vrontos et al., 2003), O-GARCH (Alexander and Chibumba,
1997), and GO-GARCH (Weide, 2002), to name some. However, these
predictors have been shown to be hard to fit and can yield inconsistent
2.4. DCC GARCH 331

estimates (Brooks et al., 2003). (These inconsistencies may not have

much practical impact.) For detailed reviews of MGARCH predictors
we refer the reader to (Silvennoinen and Teräsvirta, 2009; Bauwens
et al., 2006)

2.4 DCC GARCH

The second line of extensions of GARCH to vector time series mod-

els conditional covariances through separate estimates of conditional
variances and correlations (Engle, 2002; Engle and Sheppard, 2001).
In (Bollerslev, 1990) Bollerslev introduced the constant conditional
correlation predictor (CCC) where the individual asset volatilities are
modeled as separate GARCH processes, while the correlation matrix
is assumed constant and equal to the unconditional correlation matrix.
This predictor was later extended to the dynamic conditional correlation
(DCC) predictor where the correlation matrix is allowed to change over
time (Engle, 2002). The DCC model has the form

Σt = Dt Rt Dt ,

where Dt is the diagonal matrix of standard deviations, i.e., (Dt )ii =

1/2
(Σt )ii , and Rt is the correlation matrix associated with Σt .
DCC GARCH models the diagonal elements of Dt as separate
univariate GARCH processes as described above. The correlation matrix
Rt is then modeled as a constrained multivariate GARCH (MGARCH)
process, e.g., as

Rt = diag(diag(Qt ))−1/2 Qt diag(diag(Qt ))−1/2 ,

Qt = Q̄(1 − a − b) + ar̃t r̃tT + bQt−1 ,

where Q̄ is the unconditional correlation matrix, a and b are the

MGARCH parameters, and r̃t are the volatility adjusted returns defined
as
r̃t = Dt−1 rt .
The parameters can be estimated in two steps via (quasi) maximum like-
lihood, but requires solving non-convex optimization problems (Engle,
2002). This predictor has become a popular choice amongst MGARCH
332 Some common covariance predictors

predictors due to its interpretability. Variants of the DCC predictor are

widely used in finance, where it is also often used in combination with
EWMA estimates. Conditional correlation predictors are easier to esti-
mate than other multivariate GARCH predictors, and their parameters
are more interpretable.

Iterated covariance estimation. DCC, which separately estimates the

volatilities and correlations, is closely related to the idea of iterated
covariance predictors (Barratt and Boyd, 2022). Iterated covariance
predictors estimate the covariance matrix in multiple iterations. In a
(1)
two-step iteration we first form a first covariance estimate Σ̂t of the
returns rt , at each time t, and form the whitened returns
(1) −1/2
 
r̃t = Σ̂t rt .
(2)
In the second iteration we form the covariance estimate Σ̂t of the
whitened returns r̃t . The final covariance estimate (of the returns rt ) is
then formed as
(1) 1/2 (2) (1) 1/2
   
Σ̂t = Σ̂t Σ̂t Σ̂t .
This procedure can be iterated further, and has been shown empirically
to improve the quality of the covariance estimate; see (Barratt and Boyd,
2022) for details. In DCC, Σ̂(1) is diagonal and models the volatilities;
Σ̂(2) is a correlation matrix.

2.5 Iterated EWMA

Iterated EWMA (IEWMA) was proposed by (Engle, 2002) and is

analogous to DCC GARCH but with EWMA estimates of the volatilities
and correlations instead of GARCH. Engle proposed IEWMA as an
efficient alternative to the DCC GARCH predictor, although he did not
refer to it as IEWMA; we use this term to emphasize its connection to
iterated whitening, as proposed in (Barratt and Boyd, 2022). Specifically,
IEWMA can be viewed as an iterated whitener, where we first use a
diagonal whitener (which estimates the volatilities) and then a full
matrix whitener (which estimates the correlations). This is analogous
(1)
to the two-step iterated covariance predictor where Σ̂t is the diagonal
2.5. Iterated EWMA 333

(2)
matrix of squared volatility estimates and Σ̂t estimates the correlation
matrix of the volatility adjusted returns.
First we form an estimate of the volatilities σ̂t = diag(Σ̂t )1/2 using
EWMA predictors for each asset. We denote the half-life of these
volatility estimates as H vol . We then form the marginally standardized
returns as
r̃t = D̂t−1 rt , (2.2)
where D̂t = diag(σ̂t ). These vectors should have entries with standard
deviation near one. It is common practice to winsorize the standardized
returns; a good rule of thumb is to clip r̃t at ±4.2, which corresponds
to clipping rt at ±4.2σ̂t .
Then we form a EWMA estimate of the covariance of r̃t , which
we denote as R̃t , using half-life H cor for this EWMA estimate. (We
use the superscript ‘cor’ since the diagonal entries of R̃t should be
near one, so R̃t is close to a correlation matrix.) From R̃t we form its
associated correlation matrix R̂t , i.e., we scale R̃t on the left and right
−1/2
by a diagonal matrix with entries (R̃t )ii . Since the diagonal entries
of R̃t should be near one, R̃t and R̂t are not too different.
Our IEWMA covariance predictor is

Σ̂t = D̂t R̂t D̂t , t = 2, 3, . . . .

This is the covariance predictor proposed in (Engle, 2002); replacing R̂t

with R̃t we obtain the iterated whitener proposed by Barratt and Boyd
in (Barratt and Boyd, 2022). As mentioned above, they are typically
quite close.
It is common to choose the volatility half-life H vol to be smaller
than the correlation half-life H cor . The intuition here is that we can
average over fewer past samples when we predict the n volatilities σ̂t ,
but need more past samples to reliably estimate the n(n − 1)/2 off-
diagonal entries of R̂t . Empirical studies on real return data confirm
that choosing a faster volatility half-life than correlation half-life yields
better estimates.
 3
Combined multiple iterated EWMAs

In this chapter we introduce a novel covariance predictor, which we

call combined multiple iterated EWMAs, for which we use the acronym
CM-IEWMA. The CM-IEWMA predictor is constructed from a modest
number of IEWMA predictors, with different pairs of half-lives, which
are combined using dynamically varying weights that are based on
recent performance.
The CM-IEWMA predictor is motivated by the idea that different
pairs of half-lives may work better for different market conditions. For
example, short half-lives perform better in volatile markets, while long
half-lives perform better for calm markets where conditions are changing
slowly.

3.1 Dynamically weighted prediction combiner

We first describe the idea in a general setting. We start with K different

(k)
covariance predictors, denoted Σ̂t , k = 1, . . . , K. These could be any
of the predictors described above, or predictors of the same type with
different parameter values, e.g., half-lives (for EWMA) or pairs of half-
lives (for IEWMA). In some contexts these different predictors are
referred to as a set of K experts (Hastie et al., 2009; Jordan and Jacobs,

334
3.2. Choosing the weights via convex optimization 335

1994).
We denote the Cholesky factorizations of the associated precision
(k) (k)
matrices (Σ̂t )−1 as L̂t , i.e.,
(k) −1
 
(k) (k)
Σ̂t = L̂t (L̂t )T , k = 1, . . . , K,
(k)
where L̂t are lower triangular with positive diagonal entries. We will
combine these Cholesky factors with nonnegative weights π1 , . . . , πK
that sum to one, to obtain
K
(k)
L̂t = (3.1)
X
πk L̂t .
k=1

From this we recover the weighted combined predictor

 −1
Σ̂t = L̂t L̂Tt . (3.2)

We will see below why we combine the Cholesky factors of the precision
matrices, and not the covariance or precision matrices themselves.

3.2 Choosing the weights via convex optimization

The log-likelihood (1.1) can be expressed in terms of the Cholesky factor

of the precision matrix L̂t as
n
lt (Σ̂t ) = −(n/2) log(2π) + log L̂t,ii − (1/2)∥L̂Tt rt ∥22 ,
X

i=1

where ∥ · ∥2 denotes the Euclidean norm. This is a concave function of

the weights π ∈ RK+ (Boyd and Vandenberghe, 2004).
We choose the weights at time t as the solution of the convex
optimization problem
 
maximize i=1 log L̂t−j,ii (1/2)∥L̂Tt−j rt−j ∥22
PN Pn
j=1 −
(j) (3.3)
subject to L̂τ = τ = t − 1, . . . , t − N
PK
j=1 πj L̂τ ,
π ≥ 0, 1T π = 1,
with variables π1 , . . . , πK , where N is the look-back, 1 denotes the
vector with entries one, and ≥ between vectors means entrywise. In
336 Combined multiple iterated EWMAs

words: we choose the (mixture) weights in each period so as to maximize

the average log-likelihood of the combined prediction over the trailing
N periods. The problem (3.3) is convex, and can be solved very quickly
and reliably by many methods (Boyd and Vandenberghe, 2004). The
covariance predictor is then recovered using (3.1) and (3.2).
The look-back N is a parameter that can be adjusted to give good
performance. Numerical experiments suggest that the predictor is not
very sensitive to the choice of N , and that a choice N = 10 seems to
work well for asset universes up to a few hundred assets.
We mention several extensions of the weight problem (3.3). First,
we can add one prediction which is diagonal, using any estimates of the
volatilities (including constant). This gives us shrinkage, automatically
chosen. We can also add a constraint or objective term that encourages
the weights to vary smoothly over time, as discussed more in chapter 9.
The CM-IEWMA predictor is a special case of the dynamically
weighted prediction combiner described above, where the K predictions
are each IEWMA, with different pairs of half-lives H vol and H cor .
 4
Evaluating covariance predictors

There are several ways of evaluating a covariance predictor, often di-

vided into two categories, direct and indirect (Patton and Sheppard,
2009), (Andersen et al., 2006, §7). Direct methods use a proxy for the
true covariance matrix to evaluate the predictor, while indirect meth-
ods use the covariance predictor on tasks of interest, such as portfolio
construction or portfolio tracking.
Popular direct methods are the Mincer-Zarnowitz (MZ) regression
and its variants, based on statistical tests of the regression coefficients
of a predicted variable on an observed variable (or in the case of vari-
ance and covariance, a proxy for the observed variable) (Mincer and
Zarnowitz, 1969; Theil, 1961). Direct methods also include the compari-
son between different predictors in terms of some loss function. Common
loss functions are the mean squared error (MSE) and quasi-likelihood
(QLIKE) (Patton, 2011; Patton and Sheppard, 2009). To select good
models, the model confidence set (MCS) is usually used (Hansen et al.,
2011), or the Ledoit–Wolf test (Ledoit and Wolf, 2008) to compare
Sharpe ratios.
Indirect methods use applications to rank covariance predictors, and
include the minimum variance and mean-variance portfolios, as well as

337
338 Evaluating covariance predictors

portfolio tracking tasks.

The difference in performance between various predictors can also
be evaluated using statistical tests. For a more detailed discussion of
both direct and indirect methods, we refer the reader to (Patton and
Sheppard, 2009).
In this chapter we discuss several evaluation metrics for covariance
predictors. The first three metrics are direct, and include the mean
squared error and two metrics based on a statistical measure, the log-
likelihood under a Gaussian distribution. The remaining metrics judge
a covariance predictor by the performance of a portfolio using a method
that depends on a covariance matrix. We are mainly interested in
illustrating how simple methods can perform just as well as or better
than more complex ones, rather than finding optimal predictors in
a statistical sense. Therefore we look at the absolute performance of
covariance predictors on these metrics.

4.1 Mean squared error

The mean squared error (MSE) is a common metric for evaluating a

covariance predictor Σ̂t , defined as
T
1X
∥rt rtT − Σ̂t ∥2F ,
T t=1
i.e., the average squared Frobenius norm of the difference between the
realized (rank one) covariance matrix rt rtT and the covariance predictor
Σ̂t . Lower values of MSE are better. One variation on the MSE error
assumes that Σ̂t is constant over some number of time periods and
replaces the rank one realized covariance rt rtT with an average of the
rank one terms over the periods, i.e., the realized empirical covariance.

4.2 Log-likelihood

A natural way of judging a covariance predictor is via its average

log-likelihood on realized returns,
T 
1 X 
− n log(2π) − log det Σ̂t − rtT Σ̂−1 rt ,
2T t=1 t
4.3. Log-likelihood regret 339

with larger values being better. This metric can be used to compare
different predictors.
To understand the performance of a covariance predictor over time
and changing market conditions, we can examine the average log-
likelihood over periods such as quarters, and look at the distribution of
quarterly average log-likelihood values. We are particularly interested
in poor, i.e., low values.

4.3 Log-likelihood regret

Recall that the best constant predictor, in terms of the log-likelihood,

is the empirical sample covariance
T
1X
Σemp = rt rT ,
T t=1 t
with value
1
− n(log(2π) + 1) − log det Σemp .


2
For any other constant Σ ∈ Sn++ , the log-likelihood is lower than the
log-likelihood of Σemp . We define the average log-likelihood regret as
the average log-likelihood of the (constant) empirical covariance, minus
the average log-likelihood of the covariance predictor. The regret is
a measure of how much the covariance predictor Σ̂t , t = 1, . . . , T ,
underperforms the best possible constant covariance predictor (i.e., the
sample covariance matrix). The term regret comes from the field of
online optimization; see, e.g., (Zinkevich, 2003; Mokhtari et al., 2016;
Hazan et al., 2007; Hazan, 2016).
We want our covariance predictor to have small regret. The regret
is typically positive, but it can be negative, i.e., our time-varying co-
variance can have higher log-likelihood than the best constant one. The
regret is not any more useful than the log-likelihood when comparing
predictors over one time interval, since it simply adds a constant and
switches the sign. But it is interesting when we compute the regret over
multiple periods, like months or quarters. The regret over multiple quar-
ters removes the effect of the log-likelihood of the empirical covariance
varying due to changing market conditions, and allows us to assess how
well the covariance predictor adapts.
340 Evaluating covariance predictors

4.4 Portfolio performance

We can also judge the performance of a covariance predictor by the in-

vestment performance of portfolio construction methods that depend on
the estimated covariance matrix. As with log-likelihood or log-likelihood
regret, we can examine the portfolio performance in periods such as
quarters, to see how evenly the performance is spread over time.
One obvious metric of interest is how close the ex-ante and re-
alized portfolio volatilities are. The metrics described above, MSE,
log-likelihood, and log-likelihood regret, are agnostic to the portfolio;
with specific real portfolios we can see how well our covariance predictors
predict portfolio volatility.
We will assess a covariance predictor using five simple portfolio
construction methods. The first is an equally weighted (or 1/n) portfolio,
which does not by itself depend on the covariance, but does when we
adjust it with cash to achieve a given ex-ante risk. The second, third,
and fourth portfolios depend only on the covariance matrix. They are
minimum variance, risk parity, and maximum diversification portfolios.
For an in depth discussion of these portfolios, see (Braga, 2015). The
last portfolio we consider is a mean-variance portfolio, using a very
simple mean estimator.
For each portfolio we look at four metrics: realized return, volatility,
Sharpe ratio, and maximum drawdown. The returns, volatilities, and
Sharpe ratios are reported in annualized values. The Sharpe ratio is
defined as the ratio of the excess return (over the risk-free rate), divided
by the volatility of the excess return,
p
τ =1 (rt − rtrf )
1 PT
T
 1/2 ,
1 PT p 1 PT p
2
T τ =1 rt − T τ =1 rt

where rtp and rtrf are the portfolio and risk-free returns at time t. The
maximum drawdown is defined as
Vtp2
max − 1,
1≤t1 <t2 ≤T Vtp1
where
Vtp = V0 (1 + r1p )(1 + r2p ) · · · (1 + rtp )
4.4. Portfolio performance 341

is the portfolio value at time t (with returns re-invested), starting with

value V0 > 0.
In addition to portfolio performance, we can also examine how well
the covariance prediction predicts the portfolio volatility. We compare
the realized or ex-post portfolio volatility
T
!1/2
1X
(rT wt )2 ,
T t=1 t

to the predicted or ex-ante portfolio volatility

T
!1/2
1X
wT Σ̂t wt ,
T t=1 t

where wt ∈ Rn are the portfolio weights. This directly measures the

ability of the estimated covariance matrix to predict portfolio risk.

Equal weight portfolio. We take the equal weight or 1/n portfolio

with w = (1/n)1. This portfolio does not depend on the covariance Σ̂t ,
but when we mix it with cash, as described below, it will.

Minimum variance portfolio. The (constrained) minimum variance

portfolio is the solution of the convex optimization problem

minimize wT Σ̂t w
subject to wT 1 = 1, ∥w∥1 ≤ Lmax , wmin ≤ w ≤ wmax
with variable w ∈ Rn , where Lmax ≥ 1 is a leverage limit, and wmin
and wmax are lower and upper bounds on the weights, respectively.

Risk-parity portfolio. The portfolio return volatility σ(w) = (wT Σ̂t w)1/2
can be broken down into a sum of volatilities (risks) associated with
each asset as
∂ log σ(w) ∂σ(w) wi wi (Σ̂t w)i
= = , i = 1, . . . , n.
∂wi σ(w) ∂wi wT Σ̂t w
The risk parity portfolio is the one for which these volatility attributions
are equal (Qian, 2011). This portfolio can be found by solving the convex
342 Evaluating covariance predictors

optimization problem (Boyd and Vandenberghe, 2023),

n
minimize (1/2)xT Σ̂t x − (1/n) log xi ,
X

i=1

with variable x, and then taking w = x⋆ /(1T x⋆ ).

Maximum diversification portfolio. The diversification ratio of a long-

only portfolio (i.e., one with w ≥ 0) is defined as
σ̂tT w
D(w) = .
(wT Σ̂t w)1/2
The diversification ratio tells us how much higher the portfolio volatility
would be if all assets were perfectly correlated. The maximum diversifi-
cation portfolio is the portfolio w that maximizes D(w), possibly subject
to constraints (Choueifaty and Coignard, 2008). Like the risk-parity
portfolio, the maximum diversification portfolio can be found via convex
optimization. We let x⋆ denote the solution of the convex optimization
problem (Boyd and Vandenberghe, 2023)
minimize xT Σ̂t x
subject to σ̂tT x = 1, x ≥ 0,
with variable x. The maximum diversification portfolio is w = x⋆ /1T x⋆ .

Volatility control with cash. We mix each of the four portfolios

described above with cash to achieve a target value of ex-ante volatility
σ tar . To do this we start with the portfolio weight vector wt , and
compute its ex-ante volatility σt = (wtT Σ̂t wt )1/2 . Then we add a cash
component so that the overall ex-ante volatility equals our target, i.e.,
we use the (n + 1) weights (with the last component denoting cash)
" #
θwt σ tar
, θ= .
(1 − θ) σt
This portfolio will have ex-ante volatility σ tar . Note that the cash weight
can be either positive (when it dilutes the portfolio volatility) or negative
(when it leverages the portfolio volatility to the desired level). The target
volatility σ tar should be chosen so as to avoid portfolios that are either
too diluted or too leveraged.
4.4. Portfolio performance 343

Mean variance portfolio. The last portfolio we consider is a basic

mean-variance portfolio, defined as the solution of the convex optimiza-
tion problem

maximize r̂tT w
1/2
subject to ∥Σ̂t w∥2 ≤ σ tar
1T w + c = 1, ∥w∥1 ≤ Lmax ,
wmin ≤ w ≤ wmax , cmin ≤ c ≤ cmax
with variable w, where r̂t is the predicted mean return vector at time t.
The vector w gives the weights of the non-cash assets and c denotes the
cash weight. The non-cash and cash weights are limited by wmin , wmax
and cmin , cmax , respectively. This portfolio does not need cash dilution,
since it includes cash in its construction. (If σ tar is chosen appropriately,
it will have ex-ante risk σ tar .) The mean-variance portfolio depends
not only on a covariance estimate, but also a return estimate. For this
we use one of the simplest possible return estimates, a EWMA of the
realized returns.
 5
Data sets and experimental setup

We illustrate our method on three different data sets: a set of 49

industry portfolios, a set of 25 stocks, and a set of 5 factor returns, each
augmented with cash (with the historical risk-free interest rate). For
each data set we show results for six covariance predictors. Everything
needed to reproduce the results is available online at

https://github.com/cvxgrp/cov_pred_finance.

5.1 Data sets

Industry portfolios. The first data set consists of the daily returns of
a universe of n = 49 daily traded industry portfolios, shown in table 5.1,
along with cash. The data set spans July 1st 1969 to December 30th,
2022, for a total of 13496 (trading) days. The data was obtained from
the Kenneth French Data Library (French, 2023).

Stocks. The second data set consists of the daily returns of n = 25

stocks and cash. The stocks were chosen to be the 25 largest stocks in the
S&P 500 at the beginning of 2010, listed in table 5.2. This data set spans
January 4th 2010 to December 30th, 2022, for a total of 3272 (trading)

344
5.1. Data sets 345

Table 5.1: Industry portfolios.

Agriculture Food products

Candy & soda Beer & liquor
Tobacco products Recreation
Entertainment Printing and publishing
Consumer goods Apparel
Healthcare Medical equipment
Pharmaceutical products Chemicals
Rubber and plastic products Textiles
Construction materials Construction
Steel works etc. Fabricated products
Machinery Electrical equipment
Automobiles and trucks Aircraft
Shipbuilding, railroad equipment Defense
Precious metals Non-metallic and industrial metal mining
Coal Petroleum and natural gas
Utilities Communication
Personal services Business services
Computers Computer software
Electronic equipment Measuring and control equipment
Business supplies Shipping containers
Transportation Wholesale
Retail Restaurants, hotels, motels
Banking Insurance
Real estate Trading
Other
346 Data sets and experimental setup

Table 5.2: List of companies and their tickers.

Ticker Company Name

XOM Exxon Mobil
WMT Walmart
AAPL Apple Inc.
PG Procter & Gamble
JNJ Johnson & Johnson
CHL China Mobile
IBM IBM
SBC AT&T
GE General Electric
CHV Chevron
PFE Pfizer
NOB Noble
NCB NCR
KO Coca-Cola
ORCL Oracle Corporation
HWP Hewlett-Packard
INTC Intel Corporation
MRK Merck & Co.
PEP PepsiCo
BEL Becton, Dickinson and Company
ABT Abbott Laboratories
SLB Schlumberger
P Pandora Media
PA Pan American Silver
MCD McDonald’s

days. The stock data was attained through the Wharton Research Data
Services (WRDS) portal (Wharton Research Data Services 2023).

Factor returns. The third data set consists of daily returns of the
five Fama-French factors taken from the Kenneth French Data Li-
brary (French, 2023), shown in table 5.3. The data set spans July 1st
1963 to December 30th, 2022, for a total of 14979 (trading) days.
5.1. Data sets 347

Table 5.3: The five Fama-French factors.

Factor Description
MKT-Rf market excess return over risk-free rate
SMB small stocks minus big stocks
HML high book-to-market stocks minus low book-to-
market stocks
RMW stocks with high operating profitability minus
stocks with low operating profitability
CMA stocks with conservative investment policies mi-
nus stocks with aggressive investment policies
348 Data sets and experimental setup

Cumulative returns. In figure 5.1 we show the cumulative returns of

the five factors, and the cumulative returns of five assets chosen from
each of the industry and stock data sets.
5.1. Data sets 349

(a) Industry data set.

(b) Stock data set.

(c) Factor data set.

Figure 5.1: Cumulative returns of five assets from each data set.
350 Data sets and experimental setup

5.2 Six covariance predictors

For each data set we evaluate six covariance predictors, described below.

• Rolling window estimates with 500-, 250-, and, 125-day windows

for the industry, stock, and factor data sets, respectively, denoted
RW in plots and tables.

• EWMA predictors with 250-, 125-, and, 63-day half-lives, for

the industry, stock, and factor data sets, respectively, denoted
EWMA.

• IEWMA predictors with half-lives (in days) H vol /H cor of 125/250,

63/125, and 21/63 for the three data sets, respectively, denoted
IEWMA.

• DCC GARCH predictor, denoted MGARCH, with parameters

re-estimated annually using the rmgarch package in R (Ghalanos,
2019).

• CM-IEWMA predictor with K = 5 IEWMA predictors and a

lookback of N = 10 days, with half-lives shown in table 5.4. For
each of the fastest IEWMA predictors we regularize the covariance
estimate by increasing the diagonal entries by 5%.

• Prescient predictor, i.e., the empirical covariance for the quarter

the day is in. This predictor maximizes log-likelihood for each
quarter, and achieves zero regret. It is of course not implementable,
and meant only to show a bound on performance with which to
compare our implementable predictors.

All the parameters above (e.g., half-lives) are chosen as reasonable

values that give good overall performance for each predictor. The results
are not sensitive to these choices.
For our experiments we use the first two years (500 data points) of
each data set to fit the MGARCH predictor and initialize the other
predictors. (After this initial MGARCH fit, we re-estimate its parameters
annually.) Hence, the evaluation period for our experiments below
ranges from June 24th 1971 to December 30th, 2022, for the industry
5.2. Six covariance predictors 351

Table 5.4: Half-lives for CM-IEWMA predictors, given as H vol /H cor , in days.

Data set Half-lives

Industries 21/63 63/125 125/250 250/500 500/1000
Stocks 10/21 21/63 63/125 125/250 250/500
Factors 5/10 10/21 21/63 63/125 125/250

portfolios, from December 28th, 2011, to December 30th, 2022, for the
stock portfolios, and from June 28th 1965 to December 30, 2022, for
the factor portfolios.
 6
Results

6.1 CM-IEWMA component weights

Figure 6.1 shows the weights for each of the five components of the
CM-IEWMA predictors, averaged yearly, for the three data sets.
We can see how the predictor adapts the weights depending on
market conditions. Substantial weight is put on the slower (longer half-
life) IEWMAs most years. During and following volatile periods like
the 2000 dot.com bubble or 2008 market crash, we see a big increase
in weight on the faster IEWMAs. We can illustrate these changes in
weights in response to market conditions via the effective half-life of
the CM-IEWMA, defined as the weighted average of the five (longer)
half-lives, shown in figure 6.2, averaged yearly.

352
6.1. CM-IEWMA component weights 353

(a) Industry data set.

(b) Stock data set.

(c) Factor data set.

Figure 6.1: Weights of the various IEWMA components in the CM-IEWMA

predictors on three data sets. The IEWMA components are represented as H vol /H cor
for the volatility and correlation half-lives, respectively.
354 Results

(a) Industry data set.

(b) Stock data set.

(c) Factor data set.

Figure 6.2: Effective half-lives of the CM-IEWMA predictor on three data sets.
6.2. Mean squared error 355

6.2 Mean squared error

Table 6.1 shows the average, standard deviation, and maximum of the
MSE computed over distinct quarters for the six covariance predictors
on the three data sets (with lower being better for all three metrics).
CM-IEWMA and MGARCH do better than the other predictors on
all metrics over all data sets, with MGARCH doing slightly better on
the industry data and CM-IEWMA slightly better on the stock data.
Interestingly, on the factor data set, the CM-IEWMA predictor does
better than the prescient predictor.

6.3 Log-likelihood and log-likelihood regret

Figure 6.3 shows the average quarterly log-likelihood for the different
covariance predictors over the evaluation period. Not surprisingly, the
prescient predictor does substantially better than the others. The differ-
ent predictors follow similar trends, with even the prescient predictor
experiencing a drop in log-likelihood during market turbulence. Close in-
spection shows that the CM-IEWMA and MGARCH predictors almost
always have the highest log-likelihood in each quarter.
Figure 6.4 shows the average quarterly log-likelihood regret for the
different covariance predictors over the evaluation period. Clearly, CM-
IEWMA and MGARCH perform best in volatile markets. Figure 6.5
illustrates the difference between CM-IEWMA and MGARCH. As seen,
CM-IEWMA consistently has lower regret on the industry and stock
data sets, while they perform similar on the factor data. More precisely,
CM-IEWMA has lower regret than MGARCH in 87% of the quarters
for the industry data, 71% for the stock data, and 51% for the factor
data.
Table 6.2 illustrates the differences in regret further, by showing the
average, standard deviation, and the maximum of the average quarterly
regret. As we can see, the average quarterly regret is lower for CM-
IEWMA than for the other predictors. The regret is also more stable for
CM-IEWMA, as the standard deviation is lower. Finally, the maximum
average quarterly regret is also lower for CM-IEWMA than for the other
predictors. These results are most prominent on the industry and stock
356 Results

Table 6.1: Metrics on the MSE, computed over distinct quarters, for six covariance
predictors on three data sets.

Predictor Average/10−4 Std. Dev./10−3 Max/10−2

RW 7.6 4.0 3.9
EWMA 7.5 4.0 3.9
IEWMA 7.4 3.9 3.9
MGARCH 6.8 3.6 3.8
CM-IEWMA 6.9 3.6 3.8
Prescient 6.6 3.5 3.7

Industry data set.

Predictor Average/10−7 Std. Dev./10−6 Max/10−5

RW 3.4 1.9 2.4
EWMA 3.4 1.9 2.4
IEWMA 3.3 1.8 2.4
MGARCH 3.2 1.8 2.4
CM-IEWMA 3.2 1.8 2.4
Prescient 3.1 1.8 2.3

Stock data set.

Predictor Average/10−4 Std. Dev./10−3 Max/10−2

RW 3.4 1.6 1.1
EWMA 3.3 1.6 1.1
IEWMA 3.2 1.6 1.1
MGARCH 3.0 1.4 1.0
CM-IEWMA 2.9 1.4 0.9
Prescient 3.0 1.5 1.0

Factor data set.

6.3. Log-likelihood and log-likelihood regret 357

(a) Industry data set.

(b) Stock data set.

(c) Factor data set.

Figure 6.3: The log-likelihood, averaged quarterly, for six covariance predictors and
three data sets.
358 Results

(a) Industry data set.

(b) Stock data set.

(c) Factor data set.

Figure 6.4: The regret, averaged quarterly, for five covariance predictors over the
evaluation periods for three data sets.
6.3. Log-likelihood and log-likelihood regret 359

(a) Industry data set.

(b) Stock data set.

(c) Factor data set.

Figure 6.5: The regret for MGARCH and CM-IEWMA, averaged quarterly over
the evaluation periods for three data sets.
360 Results

Table 6.2: Metrics on the average quarterly regret for six covariance predictors on
three data sets.

Predictor Average Std. dev. Max

RW 20.4 6.9 72.8
EWMA 19.4 6.2 70.1
IEWMA 18.2 3.6 41.4
MGARCH 17.9 3.0 32.8
CM-IEWMA 16.9 2.4 28.4
PRESCIENT 0.0 0.0 0.0

Industry data set.

Predictor Average Std. dev. Max

RW 7.0 4.8 37.0
EWMA 6.2 3.8 30.2
IEWMA 5.8 1.6 13.6
MGARCH 5.6 1.0 7.8
CM-IEWMA 5.3 1.0 7.6
PRESCIENT 0.0 0.0 0.0

Stock data set.

Predictor Average Std. dev. Max

RW 0.6 0.9 12.2
EWMA 0.6 0.7 9.5
IEWMA 0.4 0.3 4.1
MGARCH 0.4 0.3 3.1
CM-IEWMA 0.4 0.3 2.9
PRESCIENT 0.0 0.0 0.0

Factor data set.

6.3. Log-likelihood and log-likelihood regret 361

data, while MGARCH does similar on the factor data.

Figure 6.6 gives a final illustration of these results, by showing the
cumulative distribution functions of the average quarterly regret for
the different covariance predictors. Clearly, CM-IEWMA has the lowest
regret on the industry and stock data set, and MGARCH does similar
on the factor data.
362 Results

(a) Industry data set.

(b) Stock data set.

(c) Factor data set.

Figure 6.6: Cumulative distribution functions of average quarterly regret for five
covariance predictors on three data sets.
6.4. Portfolio performance 363

6.4 Portfolio performance

In this section we evaluate the covariance predictors on the portfolios

described in §4.4. In the minimum variance and mean-variance portfolios,
we use Lmax = 1.6 (which corresponds to 130:30 long:short), wmin =
−0.1 and wmax = 0.15 for the industry and stock return portfolios, and
wmin = −0.3 and wmax = 0.4 for the factor return portfolio. We use
target (annualized) volatilities of 5%, 10%, and 2% for the industry,
stock, and factor return portfolios, respectively.
For the mean-variance portfolio, our estimated returns are EWMAs
of the trailing realized returns. For the industry and stock data we use
250-day half-life EWMAs, winsorized at the 40th and 60th percentiles
(cross-sectionally), and for the factor data a 63-day half-life EWMA
(not winsorized).

Equal weight portfolio. Table 6.3 shows the metrics for the equal
weight portfolio. All predictors track the volatility targets well. MGARCH
attains the highest Sharpe ratios, although the results are very close.
The drawdowns are also very similar for all predictors, but MGARCH
and CM-IEWMA seem slightly better than the rest.

Minimum variance portfolio. Table 6.4 shows the metrics for the
minimum variance portfolio. For the factor data set, MGARCH does
best. On the industry and stock data sets, the three EWMA-based
predictors track the volatility target fairly well, while RW and MGARCH
underestimate volatility. CM-IEWMA and MGARCH both attain a
high Sharpe ratio. However, we note that the high Sharpe ratio for
MGARCH, as compared to the other predictors, is a consequence of the
high volatility. Finally, CM-IEWMA seems to consistently attain a lower
drawdown than the other predictors, although the other EWMA-based
approaches also do well.
To illustrate how the minimum variance trading strategy has evolved
over time, we show the yearly annualized Sharpe ratios for the CM-
IEWMA predictor in figure 6.7. We can see that the Sharpe ratio
achieved by the minimum variance portfolio decreases over time for the
364 Results

Table 6.3: Metrics for the equal weight portfolio performance for six covariance
predictors over the evaluation periods on three data sets.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 2.2 5.4 0.4 16
EWMA 2.2 5.1 0.4 15
IEWMA 2.2 5.1 0.4 15
MGARCH 2.4 5.1 0.5 14
CM-IEWMA 2.3 5.0 0.5 13
PRESCIENT 4.3 4.9 0.9 8

Industry data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 6.8 10.6 0.6 23
EWMA 6.4 10.0 0.6 21
IEWMA 6.7 10.1 0.7 20
MGARCH 7.2 9.4 0.8 15
CM-IEWMA 6.8 9.6 0.7 17
PRESCIENT 12.8 9.9 1.3 10

Stock data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 2.9 2.1 1.4 15
EWMA 2.9 2.0 1.4 15
IEWMA 3.0 2.0 1.5 14
MGARCH 3.2 2.0 1.6 12
CM-IEWMA 2.9 2.1 1.4 15
PRESCIENT 3.3 2.0 1.7 12

Factor data set.

6.4. Portfolio performance 365

Table 6.4: Metrics for the minimum variance portfolio performance for six covariance
predictors over the evaluation periods on three data sets.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 3.1 5.8 0.5 23
EWMA 3.1 5.4 0.6 19
IEWMA 3.3 5.5 0.6 19
MGARCH 4.3 6.1 0.7 20
CM-IEWMA 3.5 5.3 0.7 20
PRESCIENT 3.8 5.0 0.8 13

Industry data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 9.7 12.0 0.8 23
EWMA 8.9 11.1 0.8 20
IEWMA 9.7 11.3 0.9 19
MGARCH 11.3 12.3 0.9 18
CM-IEWMA 9.1 11.0 0.8 15
PRESCIENT 15.6 10.0 1.6 10

Stock data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 1.3 2.2 0.6 20
EWMA 1.4 2.1 0.7 18
IEWMA 1.2 2.1 0.6 17
MGARCH 1.8 2.1 0.9 15
CM-IEWMA 1.2 2.1 0.5 21
PRESCIENT 1.0 2.0 0.5 22

Factor data set.

366 Results

industry and stock data sets, with a small upward trend for the factor
data set.

Risk parity portfolio. The results for the risk-parity portfolio are shown
in table 6.5. Overall the results are similar for the various predictors.
There is very little that separates the predictors on the industry data
set. On the stock data, CM-IEWMA and MGARCH attain the highest
Sharpe ratios and lowest drawdowns. On the factor data set, MGARCH
has the best overall performance.

Maximum diversification portfolio. The maximum diversification

portfolio results are illustrated in table 6.6. On the industry and stock
data sets, CM-IEWMA and MGARCH do best in terms of Sharpe ratio,
drawdown, and tracking the volatility target. On the factor data set,
MGARCH does best overall.

Mean variance portfolio. The results for the mean-variance port-

folio are given in table 6.7. On the industry data set all predictors
underestimate volatility. The results are similar across predictors, with
CM-IEWMA and MGARCH performing slightly better than the rest
in terms of Sharpe ratio and drawdown. On the stock data set, CM-
IEWMA seems to do best overall. On the factor data set, the results
are almost identical between predictors.
Since we use simple EWMA return predictors, we can expect the
mean variance portfolio performance to vary over time. Intuitively it
should be better on historical data than more recent data. To illustrate
this, figure 6.8 shows the yearly annualized Sharpe ratios for the three
portfolios. There is a clear downward trend in the Sharpe ratios for the
industry and factor data sets, illustrating the difficulty of predicting
returns in recent years. This can be compared to the minimum variance
portfolios (figure 6.7) that have a more stable performance over time,
and notably do not depend on a mean estimate.
6.4. Portfolio performance 367

(a) Industry data.

(b) Stock data set.

(c) Factor data set.

Figure 6.7: Yearly annualized Sharpe ratios together with the linear trend for
minimum variance portfolios on three data sets.
368 Results

Table 6.5: Metrics for the risk parity portfolio performance for six covariance
predictors over the evaluation periods on three data sets.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 2.4 5.4 0.5 16
EWMA 2.4 5.1 0.5 15
IEWMA 2.5 5.1 0.5 14
MGARCH 2.7 5.1 0.5 14
CM-IEWMA 2.5 5.0 0.5 13
PRESCIENT 4.7 4.9 1.0 8

Industry data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 7.4 10.8 0.7 22
EWMA 6.8 10.1 0.7 21
IEWMA 7.2 10.2 0.7 20
MGARCH 7.9 9.7 0.8 15
CM-IEWMA 7.4 9.7 0.8 16
PRESCIENT 14.3 9.9 1.5 9

Stock data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 1.6 2.1 0.7 19
EWMA 1.7 2.1 0.8 18
IEWMA 1.6 2.1 0.8 18
MGARCH 2.0 2.1 1.0 16
CM-IEWMA 1.5 2.1 0.7 17
PRESCIENT 1.4 2.0 0.7 17

Factor data set.

6.4. Portfolio performance 369

Table 6.6: Metrics for the maximum diversification portfolio performance for six
covariance predictors over the evaluation periods on three data sets.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 2.1 5.5 0.4 16
EWMA 2.1 5.1 0.4 16
IEWMA 2.2 5.2 0.4 14
MGARCH 2.5 5.1 0.5 12
CM-IEWMA 2.3 5.0 0.5 12
PRESCIENT 3.8 5.0 0.8 10

Industry data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 8.4 11.2 0.8 22
EWMA 7.9 10.4 0.8 21
IEWMA 8.2 10.4 0.8 20
MGARCH 10.0 9.8 1.0 15
CM-IEWMA 8.8 10.0 0.9 16
PRESCIENT 13.5 9.9 1.4 11

Stock data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 1.4 2.2 0.7 19
EWMA 1.5 2.1 0.7 19
IEWMA 1.4 2.1 0.7 19
MGARCH 2.0 2.1 1.0 16
CM-IEWMA 1.4 2.1 0.7 18
PRESCIENT 1.3 2.0 0.7 18

Factor data set.

370 Results

Table 6.7: Metrics for the mean variance portfolio performance for six covariance
predictors over the evaluation periods on three data sets.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 5.6 6.2 0.9 16
EWMA 5.6 5.8 1.0 15
IEWMA 5.9 5.7 1.0 14
MGARCH 6.7 6.4 1.0 14
CM-IEWMA 6.1 5.6 1.1 13
PRESCIENT 4.6 5.0 0.9 10

Industry data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 6.1 11.9 0.5 26
EWMA 5.9 11.0 0.5 20
IEWMA 7.9 11.1 0.7 15
MGARCH 8.3 11.9 0.7 18
CM-IEWMA 7.3 10.9 0.7 13
PRESCIENT 14.3 9.9 1.4 9

Stock data set.

Predictor Return/% Risk/% Sharpe Drawdown/%

RW 7.5 2.2 3.3 4
EWMA 7.2 2.1 3.4 4
IEWMA 7.1 2.1 3.3 4
MGARCH 7.3 2.2 3.3 3
CM-IEWMA 6.9 2.2 3.2 4
PRESCIENT 6.5 1.9 3.3 4

Factor data set.

6.4. Portfolio performance 371

(a) Industry data.

(b) Stock data set.

(c) Factor data set.

Figure 6.8: Yearly annualized Sharpe ratios together with the linear trend for mean
variance portfolios on three data sets.
372 Results

6.5 Summary

In terms of log-likelihood and regret, CM-IEWMA performs best, fol-

lowed by MGARCH, which performs better than the simpler covariance
predictors. In downstream portfolio optimization experiments, CM-
IEWMA and MGARCH again perform better than the other predictors,
although in many cases not by much. In these experiments there is more
variation in the results, partly explained by the difference between our
prediction (of a covariance matrix) and our metrics (such as return, risk,
drawdown). Even the simplest covariance predictors do a reasonable
job of predicting the portfolio risk.
 7
Realized covariance

We have so far focused on predicting the covariance matrix of asset

returns using historical return data, i.e., we predict Σ̂t from r1 , . . . , rt−1 .
In this chapter we consider the use of additional data, specifically,
intraperiod returns. As an example, suppose the period is (trading)
days. The methods described in previous chapters predict the covariance
of the daily return from previous daily returns. In so-called realized
covariance, we predict the daily return covariance using intraday returns.
Instead of single period returns rt , we have multiple returns associated
with period t. It is not surprising that using multiple realized returns for
each period, instead of just one, can improve our covariance estimates.
Recent literature has shown that realized volatility and correlation
measurements (based on high-frequency intraperiod data) can improve
performance over traditional predictors that rely on a single realization
per period. Hansen et al. (2012) extend the univariate GARCH model
to the joint modeling of returns and realized measures of volatility,
and show empirically that this improves performance over the stan-
dard GARCH model. In (Bauwens et al., 2012) a multivariate realized
GARCH model is proposed. More recently, (Bollerslev et al., 2020)
propose a realized semicovariance GARCH model to allow for nuanced

373
374 Realized covariance

responses to positive and negative return shocks.

In this chapter we show that the dynamically weighted prediction
combiner of §3.1 readily handles multiple realized returns per period.
For simplicity we will assume each period has the same number of
intraperiod returns, equally spaced in time. We redefine the return
vector to be a return matrix rt ∈ Rn×m with columns that are the
m intraperiod return vectors, for times t = 1, . . . , T . The realized
covariance at time t is defined as

Ct = rt rtT ,

the same formula for the realized return when rt is a single (vector)
return. The realized covariance matrix Ct has rank m when the m return
vectors are linearly independent and m ≤ n; this can be compared to
the realized covariance when we do not have intraperiod returns, which
is rank one.

7.1 Combined multiple realized EWMAs

The dynamically weighted prediction combiner of §3.1 readily handles

multiple realized covariance predictors.

Realized EWMA. We define the realized EWMA (REWMA) predictor

as
t−1
Σ̂t = αt t = 2, 3, . . . ,
X
β t−1−τ Cτ ,
τ =1
where Cτ is the realized covariance at time τ , β ∈ (0, 1) is the forgetting
factor, and αt is the normalizing constant; see §2.2 for details. This
is the same formula as the usual EWMA covariance, with one return
per period, given in (2.1), with rt extended to be a matrix of multiple
returns.

Combined multiple realized EWMAs. The combined multiple realized

EWMA (CM-REWMA) predictor starts with a set of K REWMA
(k)
predictors Σ̂t with half-lives H (k) , k = 1, . . . , K, and combines them
using the dynamically weighted prediction combiner of §3.1.
7.2. Data and experimental setup 375

7.2 Data and experimental setup

Data set. We consider a universe of n = 39 assets with five-minute

intraday returns corresponding to m = 77. The assets were taken as
a subset of those used by Pelger (2020), and are available at (Pelger,
2023). The data set spans January 2nd 2004 to December 30th 2016,
for a total of 252021 data points over 3273 trading days. We list the
assets in table 7.1.

Four covariance predictors. We evaluate four covariance predictors,

described below.

• The CM-IEWMA predictor used for the stock data from §5.2. This
predictor only uses daily returns, and is not a realized covariance
predictor.

• An REWMA predictor with a half-life of H = 10 days, denoted

REWMA-10.

• A CM-REMWA predictor with five components with half-lives of

1, 5, 10, 21, and 63 days, respectively.

• Prescient predictor, i.e., the empirical covariance for the quarter

the day is in. As with the CM-IEWMA predictor, this predictor
uses daily return data. It is of course not implementable, and meant
only to show a bound on performance with which to compare our
implementable predictors.

7.3 Empirical results

CM-REWMA component weights. Figure 7.1 shows the component

weights for the CM-REWMA predictor, averaged annually. The weights
are fairly stable over time but a weight shift toward faster changing
EWMAs is seen in 2008, during the financial crisis.

MSE. The average, standard deviation, and maximum MSEs, com-

puted over distinct quarters for the four covariance predictors, are given
in table 7.2. The REWMA and CM-REWMA do slightly better than
376 Realized covariance

Ticker Company Name

JPM JPMorgan Chase
GS Goldman Sachs
KO The Coca-Cola Company
IBM International Business Machines Corporation
CAT Caterpillar Inc.
CVX Chevron Corporation
XOM Exxon Mobil Corporation
GE General Electric Company
MRK Merck & Co., Inc.
VZ Verizon Communications Inc.
PFE Pfizer Inc.
WMT Walmart Inc.
C Citigroup Inc.
HD The Home Depot, Inc.
BA The Boeing Company
MMM 3M Company
MCD McDonald’s Corporation
NKE NIKE, Inc.
JNJ Johnson & Johnson
INTC Intel Corporation
MSFT Microsoft Corporation
AAPL Apple Inc.
AMZN Amazon.com Inc.
CSCO Cisco Systems, Inc.
PG Procter & Gamble Co.
ABT Abbott Laboratories
VLO Valero Energy Corporation
HON Honeywell International Inc.
LMT Lockheed Martin Corporation
TXN Texas Instruments Inc.
COST Costco Wholesale Corporation
PEP PepsiCo, Inc.
UNP Union Pacific Corporation
WFC Wells Fargo & Co.
CVS CVS Health Corporation
ORCL Oracle Corporation
XRX Xerox Corporation
TMO Thermo Fisher Scientific Inc.
NSC Norfolk Southern Corporation

Table 7.1: Assets used in realized covariance study.

7.3. Empirical results 377

Figure 7.1: CM-REWMA component weights, averaged annually.

Predictor Average/10−4 Std. Dev./10−3 Max/10−3

CM-IEWMA 3.1 1.2 7.6
REWMA 3.0 1.1 7.3
CM-REWMA 3.0 1.1 7.2
PRESCIENT 3.0 1.1 7.1

Table 7.2: Metrics on the MSE, computed over distinct quarters.

the CM-IEWMA predictor, but overall there is not a big difference

between the predictors.

Regret. Figure 7.2 shows the average regret over distinct quarters
for the CM-IEWMA, REWMA, and CM-REWMA predictors. The
CM-REWMA predictor has the lowest regret in almost all quarters.
It has lower regret than the REWMA predictor in 41 out of the 50
quarters, and lower regret than the CM-IEWMA predictor in 39 out of
the 50 quarters.
Finally, figure 7.3 shows the cumulative distribution functions of
the average quarterly regret for the different covariance predictors. CM-
REWMA has lower regret than both the CM-IEWMA and REWMA
predictors, while REWMA has lower regret than CM-IEWMA.
378 Realized covariance

Figure 7.2: Average regret over distinct quarters for three covariance predictors.

Figure 7.3: Cumulative distribution functions of the average quarterly regret for
three covariance predictors.
7.3. Empirical results 379

Portfolio performance. Table 7.3 shows the portfolio metrics for five
different portfolio construction methods. CM-REWMA does better
than, or as well as, REWMA on almost all metrics, and better than
CM-IEWMA for all portfolios. However, the difference on portfolio tasks
is not large.

Summary. The results above show that using realized covariance,

i.e., intraperiod returns instead of just one return per period, gives
covariance estimates that are a bit better than those obtained using
only one return per period.
380 Realized covariance

Predictor Return/% Risk/% Sharpe Drawdown/%

Equal weight
CM-IEWMA 3.2 9.9 0.3 18
REWMA 3.7 10.3 0.4 16
CM-REWMA 4.4 10.6 0.4 16
PRESCIENT 6.7 9.9 0.7 13
Minimum variance
CM-IEWMA 10.7 11.0 1.0 25
REWMA 10.7 10.5 1.0 21
CM-REWMA 12.0 10.7 1.1 21
PRESCIENT 11.7 10.0 1.2 12
Risk parity
CM-IEWMA 4.1 10.0 0.4 18
REWMA 4.7 10.3 0.5 17
CM-REWMA 5.5 10.6 0.5 17
PRESCIENT 8.0 9.9 0.8 12
Maximum diversification
CM-IEWMA 3.6 10.2 0.4 25
REWMA 4.3 10.5 0.4 21
CM-REWMA 5.1 10.8 0.5 19
PRESCIENT 7.8 9.9 0.8 16
Mean variance
CM-IEWMA 8.6 10.5 0.8 22
REWMA 8.5 10.3 0.8 16
CM-REWMA 9.3 10.5 0.9 19
PRESCIENT 10.9 9.8 1.1 21

Table 7.3: Metrics for five different portfolio construction methods, using four
covariance predictors.
 8
Large universes

In a practical setting we often encounter a larger number of assets than

considered in the previous chapters, which has led to extensive research
in high-dimensional covariance estimation. One challenge in large di-
mensions is ensuring positive definiteness of the covariance matrix, in
particular with model-based approaches such as MGARCH (Bauwens
et al., 2006). Several techniques have been proposed for estimating
MGARCH models in large dimensions; see, e.g., (Engle et al., 2019;
De Nard et al., 2021; De Nard et al., 2022). Others have focused on
estimating realized covariance matrices in high dimensions; see, e.g.,
(Oh and Patton, 2016; Vassallo et al., 2021; Hautsch et al., 2015; deBrito
et al., 2018; Fan et al., 2016; Ait-Sahalia and Xiu, 2017). For a detailed
review of recent developments in high-dimensional covariance estimation,
we recommend (Bauwens and Otranto, 2023, §6).
The methods described in previous chapters can be adapted to
handle large universes of assets, say n larger than 100 or so. In this
chapter we describe two closely related methods for improving the
performance with large n. Both methods end up modeling Σ̂t as a low
rank plus diagonal matrix, in so-called factor form. Before describing
these methods, we mention that evaluating log-likelihood regret is

381
382 Large universes

complicated with large n. For the empirical covariance to be nonsingular

(which is needed to evaluate the regret), we need at least n periods; for
daily returns with n = 1000, this amounts to four years. Even if we
have n periods of data, we would only be able to evaluate the regret
a few times. For example, with n = 1000 (four years) we need at least
40 years of data to compute the average regret over 10 distinct periods.
The log-likelihood, however, can still be evaluated over fewer than n
periods.

8.1 Traditional factor model

In practice most return covariance matrices for large universes are

constructed from factors, with the model
rt = Ft ft + zt , t = 1, 2, . . . ,
where Ft ∈ Rn×k is the factor exposure matrix, ft ∈ Rk is the factor
return vector, zt ∈ Rn is the idiosyncratic return, and k is the num-
ber of factors, typically much smaller than n. The factor returns are
constructed or found by several methods, such as principal component
analysis (PCA), or by hand; see, e.g., (Bai and Ng, 2008; Bai, 2003;
Lettau and Pelger, 2020a; Lettau and Pelger, 2020b; Pelger and Xiong,
2022b; Pelger and Xiong, 2022a; Fama and French, 1993; Fama and
French, 1992). Thus we assume that the factor returns are known. Given
the factor returns, the rows of the factor exposure matrix are typically
found by least squares regression over a rolling or exponentially weighted
window (Cochrane, 2009). The idiosyncratic returns zt are then found as
the residuals in this least squares fit. The factor returns ft are modeled
as N (0, Σft ), and the idiosyncratic returns zt are modeled as N (0, Et ),
where Et is diagonal. It is also assumed that the factor returns and
idiosyncratic returns are independent across time and of each other.
We end up with a covariance matrix in factor form, i.e., rank k plus
diagonal,
Σt = Ft Σft FtT + Et . (8.1)
We can easily use the methods described above with a factor model.
Simply predict the factor return covariance Σ̂ft (using the factor returns
ft ) and the idiosyncratic variances Êt (using the entries of zt ), using
8.2. Fitting a factor model to a covariance matrix 383

the methods described in this monograph, and then form the covariance
estimate
Σ̂t = Ft Σ̂ft FtT + Êt .
The factor model (8.1) can be written in a simpler form as

Σt = F̃t F̃tT + Et , (8.2)

with F̃t = Ft (Σft )1/2 . This form does not include a factor covariance Σf ,
or equivalently, assumes Σft = I, i.e., the factors are independent with
standard deviation one. (The associated factors are called whitened
factors.) We will use the factor model form (8.2) in the sequel.
The factor model (8.2) has parameters F̃t and Et , which all together
include nk + n scalar parameters. (Some of these are redundant; for
example we can insist without loss of generality that F is lower trian-
gular.) The factor model contains substantially fewer scalar parameters
than a generic n × n covariance matrix, which contains n(n + 1)/2 scalar
parameters.
The smaller number of parameters is not the only reason for using a
factor model. Another is that it often gives better covariance estimates.
We can think of the low rank plus diagonal structure as regularization,
which can improve out-of-sample performance. In addition, the low
rank plus diagonal structure can be exploited in portfolio construction,
bringing the computational complexity down from O(n3 ) to O(nk 2 )
operations (Boyd and Vandenberghe, 2004). This makes portfolio opti-
mization with n = 1000 assets and k = 50 factors extremely fast, and
makes possible optimization of portfolios with much larger values of n.

8.2 Fitting a factor model to a covariance matrix

In this section we consider the problem of fitting a given covariance

matrix Σ by one in factor form, Σ̂ = F F T + E, where F ∈ Rn×k . This
corresponds to the model r = F f + z, with (factor return) f ∼ N (0, I),
and (idiosyncratic return) r ∼ (0, E), with E diagonal. We let θ = (F, E)
denote the parameters of our factor form model.
We seek F ∈ Rn×k and diagonal E ∈ Rn×n (with positive diagonal
entries) that minimize the Kullback-Leibler (KL) divergence between
384 Large universes

N (0, Σ) and N (0, Σ̂),

1 det Σ̂
!
K(Σ, Σ̂) = log − n + Tr Σ̂−1 Σ . (8.3)
2 det Σ

The KL divergence can be expressed in terms of the average log-

likelihood of N (0, Σ̂) under N (0, Σ) as

E ℓ (r) = −K(Σ, Σ̂) − (1/2)(n log 2π + n + log det Σ), (8.4)

r∼N (0,Σ) Σ̂

where ℓΣ̂ (r) is the log-likelihood of r under N (0, Σ̂). Hence minimizing
the KL-divergence (8.3) is equivalent to maximizing the expected log-
likelihood (8.4) of r under the model N (0, Σ̂).

Solution via EM. We can use the expectation-maximization (EM) algo-

rithm to approximately minimize K(Σ, F F T + E) over F and E (Demp-
ster et al., 1977; Rubin and Thayer, 1982). Usually EM is used to fit a
factor model to data, i.e., samples; here we use it to fit a given Gaussian
distribution N (0, Σ). The method described below was suggested and
derived by Emmanuel Candès. We are not aware of its appearance in
prior literature. A forthcoming paper on this method will include more
detail and applications.
The EM algorithm is an iterative method for maximizing (8.4).
Each iteration consists of two steps: the expectation or E-step, and
the maximization or M-step. We use the conventional symbols used to
describe EM, and use subscript j = 1, 2, . . . to denote iteration number.
(A good method for initializing the EM algorithm is provided below.)

E-step. In the E-step, we find the expected log-likelihood under the

current estimate of the parameters θj = (Fj , Ej ), over the true distribu-
tion of r:
Q(θ || θj ) = E E ℓθ (r, f ) (8.5)
r∼N (0,Σ) pθj (f |r)

where pθj (f | r) is the density of the conditional distribution of f under

the parameter estimates at iteration j, and ℓθ (r, f ) is the log likelihood
of the joint distribution with variable θ = (F, E).
8.2. Fitting a factor model to a covariance matrix 385

With our factor model the complete log-likelihood of (r, f ) is

1 
ℓθ (r, f ) = − (r − F f )T E −1 (r − F f ) + f T f + log det E
2
1 1
+ + − k/2.
(2π) n/2 (2π)k/2
The conditional distribution of f | r under θj is (Bishop, 2006)

f | r ∼ N (Bj r, Gj ),

where
Bj = Gj FjT Ej−1 , G−1 T −1
j = Fj Ej Fj + I. (8.6)
Hence, (8.5) becomes, up to an additive constant,
1 1
− Tr(E −1 (Crr − 2Crf F T + F Css F T )) − log det E, (8.7)
2 2
where
Crr = Σ, Crf = ΣBjT , Cf f = Bj ΣBjT + Gj . (8.8)

M-step. In the M-step (8.5) is maximized with respect to θ to obtain

the updated parameters:

θj+1 = argmax Q(θ || θj ).

θ

The maximizer of (8.7) is (Rubin and Thayer, 1982)

Fj+1 = Crf Cf−1

f,
Ej+1 = diag(diag(Crr − 2Crf Fj+1
T
+ Fj+1 Cf f Fj+1
T
)),

where the inner diag extracts the diagonal of its (matrix) argument, and
the outer diag creates a diagonal matrix from its (vector) argument.

EM iteration. The EM iteration has the form

Fj+1 = Crf Cf−1

f,
Ej+1 = diag(diag(Crr − 2Crf Fj+1
T
+ Fj+1 Cf f Fj+1
T
)),

where Crr , Crf , and Cf f come from (8.6) and (8.8).

386 Large universes

Initialization. To initialize the EM algorithm we use the following

method, based on low rank approximation via eigendecomposition. We
work with the correlation matrix of Σ, denoted
R = diag(σ)−1 Σ diag(σ)−1 ,
where σ = diag(Σ)1/2 (entrywise). First we express R in its eigende-
composition R = ni=1 λi qi qiT , with λ1 ≥ λ2 ≥ · · · ≥ λn . We then form
P

the rank k approximation

k
b=
X
R λi qi qiT .
i=1

We only need to compute the k dominant eigenvectors and eigenval-

ues, which can be done efficiently using for example the Lanczos algo-
rithm (Golub and Van Loan, 2013). Let
 
b = diag diag(R − R)
E b ,

which can be shown to have positive diagonal entries. Our low-rank

plus diagonal approximation of R is then R b + E.b It is also a correlation
matrix, i.e., has diagonal entries one. Our final factor approximation of
Σ is given by
b + E)
diag(σ)(R b diag(σ) = F F T + E,

with
F = diag(σ)[ λ1 q1 · · · λk qk ], E = diag(e ◦ σ 2 ),
p p

where ◦ denotes the elementwise (Hadamard) product, and σ 2 means

elementwise.
This initialization alone can serve as a basic method to fit a factor
model to a given covariance matrix. We will see below that in terms of
portfolio optimization, it serves just as well as a factor model fit using
the EM method.

8.3 Data and experimental setup

Data set. We gather the 500 largest NASDAQ stocks (by market
capitalization) at the beginning of 2000 from the WRDS portal (Wharton
8.4. Empirical results 387

Research Data Services 2023), compute the daily returns of these stocks
from January 3rd 2000 to December 30th 2022, and remove any stocks
with missing return values during this period. This gives us 238 stocks
over 5787 (trading) days. We acknowledge that we induce a survivor
bias, but the purpose of this empirical study is solely to demonstrate
the benefit of regularization in large universes, and not to backtest a
trading strategy.

Traditional factor model. We create a factor model using PCA as

follows. Every year, the k principal components of largest explanatory
power are computed, using the past two years of returns. These define
the columns of the factor exposure matrix Ft for the following year,
and the factor returns ft are the projections of the returns onto these
principal components. The idiosyncratic returns zt are the residuals. We
leverage the CM-IEWMA predictor to compute the factor covariance,
using three IEWMA components with half-lives (in days) H vol /H cor of
⌈k/2⌉/k, k/3k, and 3k/6k, where k denotes the number of factors. To
estimate the idiosyncratic variances a 21-day EWMA is used. We evalu-
ate the factor models on the average log-likelihood over the evaluation
period.

Fitting a factor model to the covariance matrix. We use a CM-

IEWMA covariance predictor with four IEWMA components with
half-lives 63/125, 125/250, 250/500, and 500/1000 days, respectively,
given as H vol /H cor . Given the CM-IEWMA estimate Σ̂t at time t, we
approximate it using a factor model as described in §8.2.
To evaluate the factor models, we look at the average log-likelihood
over the evaluation period and several performance metrics for a mini-
mum variance portfolio with Lmax = 1.6, wmin = −0.1, and wmax = 0.15,
diluted to a target risk of 10%.

8.4 Empirical results

Traditional factor model. Figure 8.1 shows the log-likelihood versus

the number of factors k for k between 2 and 75. A large increase
in log-likelihood is attained with around 20 factors, as compared to
388 Large universes

Figure 8.1: Log-likelihood versus the number of factors, using a conventional factor
model.

using the full covariance matrix. Thus using a traditional factor model
and applying our covariance estimation method to the factor returns
improves our overall covariance prediction.

Fitting a factor model to the covariance matrix. Figure 8.2 shows

the log-likelihood versus the number of factors (i.e., the rank of the
low-rank component) k for various k between 2 and 75, using the
eigendecomposition initialization and the EM algorithm. We see that
a rank of about r = 20 seems optimal for this data set, and achieves
a noticeably higher log-likelihood than using the full-rank covariance.
Moreover, the EM algorithm does better than just computing the
eigendecomposition.
Figure 8.3 shows the portfolio metrics for the minimum variance port-
folios. We can see that with roughly 10 factors or more, the performance
is essentially identical to that obtained using the full covariance matrix.
For these experiments we observed no notable difference between the
two factor model fitting methods, i.e., the simple eigendecomposition
based initialization and the more sophisticated EM method. While using
8.4. Empirical results 389

Figure 8.2: Log-likelihood versus the number of factors, obtained by fitting our
covariance estimate with a factor model.

the factor model does not improve portfolio performance, it greatly

speeds up the computation of the portfolio optimization problems.
390 Large universes

(a) Risk.

(b) Sharpe ratio.

(c) Drawdown.

Figure 8.3: Portfolio metrics for minimum variance portfolios constructed via factor
models with various number of factors.
 9
Smooth covariance predictions

We address here a secondary objective for a covariance prediction Σ̂t ,

which is that it vary smoothly across time. Perhaps the main reason
for desiring smoothness of the covariance estimate is that it can lead
to reduced trading in portfolio construction methods; it can also lead
to improved portfolio performance, even without taking into account
transaction costs.
To some extent smoothness happens naturally, since whatever
method is used to form Σ̂t from r1 , . . . , rt−1 is likely to yield a similar
prediction Σ̂t+1 from r1 , . . . , rt . It is also possible to further smooth the
predictions over time, perhaps trading off some performance, e.g., in
log-likelihood regret.
We have already mentioned that the weight optimization problem
(3.3) can be modified to encourage smoothness of the weights over
time. We can also directly smooth the prediction Σ̂t , to get a smooth
version Σ̂sm
t . A very simple approach is to let Σ̂t be a EWMA of Σ̂t ,
sm

with a half-life chosen as a trade-off between smoother predictions and

performance. This EWMA post-processing is equivalent to choosing
Σ̂sm
t to minimize
2 2
Σ̂sm
t − Σ̂t + λ Σ̂sm
t − Σ̂t−1
sm
,
F F

391
392 Smooth covariance predictions

where λ is a positive regularization parameter used to control the trade-

off between smoothness and performance, or equivalently, the half-life
of the EWMA post-processing. Here the first term is a loss, and the
second is a regularizer that encourages smoothly varying covariance
predictions.
We can create more sophisticated smoothing methods by changing
the loss or the regularizer in this optimization formulation of smoothing.
For example we can use the Kullback-Liebler (KL) divergence as a loss.
With regularizer λ∥Σ̂smt − Σ̂t−1 ∥F (no square in this case), we obtain a
sm

piecewise constant prediction, which roughly speaking only updates the

prediction when needed. This is a convex optimization problem which
can be solved quickly and reliably (Boyd and Vandenberghe, 2004).

9.1 Data and experimental setup

We consider again the Fama-French factor returns from §5.1, over the
same time horizon. We use the CM-IEWMA covariance predictor with
the same parameters as in §5.2.

Smoothly varying covariance. In the first experiment we smooth

the CM-IEWMA covariance estimates by applying a EWMA, which
2
corresponds to the λ Σ̂sm t − Σ̂t−1
sm regularizer. For different EWMA
F
half-lives we attain different levels of smoothness.

Piecewise constant covariance. In the second experiment we smooth

the CM-IEWMA covariance estimates by applying the λ Σ̂sm t − Σ̂t−1
sm
F
regularizer. For different values of λ we attain piecewise constant co-
variance predictors with different update frequencies.

9.2 Empirical results

Smoothly varying covariance. Figure 9.1 shows the regret versus

smoothness for various levels of smoothness. As seen, we can reduce
the smoothness by a factor of (roughly) four without losing much
performance in terms of regret. This can obviously be useful in practice
since a smoother covariance estimate would, for example, reduce trading.
9.2. Empirical results 393

Figure 9.1: Average regret versus smoothness when using EWMA smoothing of the
covariance predictor.

Table 9.1 shows the portfolio metrics for various values of λ for
the minimum variance portfolio with the same parameters as in §6.4;
here the turnover is defined as the average of 252 × ∥wt+1 − wt ∥1 /∥wt ∥1
over all times t in the evaluation period. Interestingly, the right amount
of smoothing not only reduces turnover, but also improves portfolio
performance in terms of Sharpe ratio and drawdown, while keeping the
desired volatility level. Too much smoothing, however, leads to reduced
portfolio performance. Figure 9.2 shows the yearly annualized Sharpe
ratios for λ = 10−4 , indicating a stable performance over time.
Figure 9.3 shows the portfolio weights for three different EWMA half-
lives. As seen, EWMA smoothing leads to smoothly varying portfolio
weights, while the weights vary significantly when no smoothing is
applied.

Piecewise constant covariance. Figure 9.4 shows the regret versus the
update frequency of the covariance estimate. There is a clear trade-off
394 Smooth covariance predictions

Table 9.1: Portfolio metrics for various EWMA half-lives used for smoothing the
covariance. Half-life 0 means no smoothing.

Half-life/days Return/% Risk/% Sharpe Drawdown/% Turnover/%

0 1.2 2.1 0.5 21 1855
10 1.4 2.1 0.7 16 310
100 1.8 2.1 0.9 15 56
250 2.1 2.1 1.0 13 30
5000 2.9 2.6 1.1 21 9

Figure 9.2: Yearly annualized Sharpe ratios for the minimum variance portfolio
when smoothing the CM-IEMA covariance predictor with a 250-day half-life EWMA.
9.2. Empirical results 395

(a) No smoothing.

(b) Half-life of 250 days.

(c) Half-life of 5000 days.

Figure 9.3: Portfolio weights for three different regularization parameters λ.

396 Smooth covariance predictions

Figure 9.4: Average regret versus time between covariance updates.

between the regret and update frequency. Roughly speaking, we could

update the covariance matrix weekly without losing much in terms of
regret.
As mentioned, a piecewise constant predictor can be desirable in
practice, since it encourages us not updating the portfolio weights, which
in turn reduces trading costs. Table 9.2 shows the portfolio metrics
for various values of λ for the minimum variance portfolio from §6.4.
As seen, smoothing can significantly reduce turnover, and interestingly
improve the Sharpe ratio and drawdown noticeably while maintaining

Table 9.2: Portfolio metrics for various regularization parameters λ. λ = 0 means

no smoothing.

λ Return/% Risk/% Sharpe Drawdown/% Turnover/%

0 1.2 2.1 0.5 21 1855
5 × 10−5 1.9 2.0 1.0 14 1190
10−4 2.4 1.9 1.3 9 112
10−3 2.6 2.1 1.2 17 7
7.5 × 10−3 3.0 4.8 0.6 31 0
9.2. Empirical results 397

Figure 9.5: Yearly annualized Sharpe ratios for the minimum variance portfolio
with a piecewise constant CM-IEMA covariance predictor using λ = 10−4 .

the correct risk level. Figure 9.5 shows the yearly annualized Sharpe
ratios for λ = 10−4 . The performance is relatively stable over time, with
a small downward trend.
To illustrate the impact of smoothing we show the portfolio weights
for three different values of λ in figure 9.6. Without smoothing the
portfolio weights are updated significantly every day. For λ = 10−4 the
weights are updated around once or twice a month. Finally, for λ = 10−5
the weights are updated on average every half a year, with only four
big weight updates over the whole trading period. Interestingly, the
weight updates for λ = 10−5 correspond precisely in time to the volatile
regime around 1980, the 2000 dot-com bubble, the 2008 financial crisis,
and the 2020 pandemic. In short, we can conclude from table 9.2 and
figure 9.6 that smoothing can lead to less trading and improve the
portfolio performance.
Finally, we note that there is some deviation between the regret
metric and portfolio performance. As seen from figure 9.4 regret increases
as we update the covariance matrix less than every other week. However,
as seen from table 9.2 and figure 9.6, portfolio performance can improve
398 Smooth covariance predictions

(a) λ = 0.

(b) λ = 10−3 .

(c) λ = 10−4 .

Figure 9.6: Portfolio weights for three different regularization parameters λ.

9.2. Empirical results 399

notably when updating the covariance matrix only every few months,
or even years.
 10
Simulating returns

Our model can be used to simulate future returns, when seeded by past
realized ones. To do this, we start with realized returns for periods
1, . . . , t − 1, and compute Σ̂t using our method. Then we generate
or sample rtsim from N (0, Σ̂t ). We then find Σ̂t+1 using the returns
r1 , . . . , rt−1 , rtsim . We generate rt+1
sim by sampling from N (0, Σ̂
t+1 ). This
continues.
This simple method generates realistic return data in the short term.
Of course, it does not include shocks or rapid changes in the return
statistics that we would see in real data, but the generative return
method has several practical applications. To mention just one, we can
simulate 100 (say) different realizations over the next quarter (say), and
use these to compute 100 performance metrics for our portfolio. This
gives us a distribution of the performance metric that we might see over
the next quarter.

10.1 Data and experimental setup

To illustrate the generative return method, we consider the five Fama-

French factor returns from §5.1. Using the same setup as in §5.2 we
compute CM-IEMWA covariance estimates, using data from January

400
10.2. Empirical results 401

1st 2011 to December 31 2013, i.e., over a three-year period. Returns

are then generated for 100 days, using the generative mode described
above.

10.2 Empirical results

We illustrate the results by looking at the SMB factor, i.e., we look

at the marginal distribution of this factor. Figure 10.1 shows the true
SMB factor returns and the simulated returns for two different random
number generator seeds. As seen, we attain realistic returns that could
be used to generate scenarios for downstream portfolio optimization
tasks, for example.
402 Simulating returns

(a) Observed returns.

(b) Obeserved returns (left) and

simulated returns (right).

(c) Observed returns (left) and

simulated returns (right).

Figure 10.1: Observed and simulated SMB factor returns for two different seeds. The
vertical line separates the in-sample (observed returns) and out-of-sample (simulated
returns) periods.
 11
Conclusions

We have introduced a simple method for predicting covariance matrices

of financial returns. Our method combines well known ideas such as
EWMA, first estimating volatilities and then correlations, and dynami-
cally combining multiple predictions. The method relies on solving a
small convex optimization problem (to find the weights used in the
combining), which is extremely fast and reliable. The proposed predictor
requires little or no tuning or fitting, is interpretable, and produces
results better than the popular EWMA estimate, and comparable to
MGARCH. Given its interpretability, light weight, and good practical
performance, we see it as a practical choice for many applications that
require predictions of the covariance of financial returns.

403
 Acknowledgements

Stephen Boyd acknowledges many helpful in-depth discussions with his

colleagues Trevor Hastie, Rob Tibshirani, Emmanuel Candès, Mykel
Kochenderfer, Kunal Menda, Misha Van Beek, Ron Kahn, and Gabriel
Maher. The authors thank Ron Kahn and Philipp Schiele for detailed
comments and suggestions on early drafts. The method of fitting a
factor model to a Gaussian described in §8.2 was suggested and derived
by Emmanuel Candès. The authors are indebted to two anonymous
reviewers for their detailed and helpful comments and suggestions.

404
 References

Ait-Sahalia, Y. and D. Xiu. (2017). “Using principal component analysis

to estimate a high dimensional factor model with high-frequency
data”. Journal of Econometrics. 201(2): 384–399.
Alexander, C. and A. Chibumba. (1997). “Multivariate orthogonal factor
GARCH”. University of Sussex, Mimeo.
Andersen, T., T. Bollerslev, P. Christoffersen, and F. Diebold. (2006).
“Volatility and Correlation Forecasting”. In: ed. by G. Elliott, C.
Granger, and A. Timmermann. Vol. 1. Handbook of Economic Fore-
casting. Elsevier. 777–878.
Bai, J. (2003). “Inferential theory for factor models of large dimensions”.
Econometrica. 71(1): 135–171.
Bai, J. and S. Ng. (2008). “Large dimensional factor analysis”. Founda-
tions and Trends® in Econometrics. 3(2): 89–163.
Barratt, S. and S. Boyd. (2022). “Covariance prediction via convex
optimization”. Optimization and Engineering.
Bauwens, L., S. Laurent, and J. Rombouts. (2006). “Multivariate
GARCH models: a survey”. Journal of Applied Econometrics. 21(1):
79–109.
Bauwens, L. and E. Otranto. (2023). “Modeling realized covariance
matrices: a class of Hadamard exponential models”. Journal of
Financial Econometrics. 21(4): 1376–1401.

405
406 References

Bauwens, L., G. Storti, and F. Violante. (2012). “Dynamic conditional

correlation models for realized covariance matrices”. CORE DP. 60:
104–108.
Bishop, C. (2006). Pattern recognition and machine learning. Vol. 4.
No. 4. Springer.
Bollerslev, T. (1986). “Generalized autoregressive conditional heteroskedas-
ticity”. Journal of Econometrics. 31(3): 307–327.
Bollerslev, T. (1990). “Modelling the Coherence in Short-Run Nominal
Exchange Rates: A Multivariate Generalized ARCH Model”. The
Review of Economics and Statistics. 72(3): 498–505.
Bollerslev, T., R. Engle, and J. Wooldridge. (1988). “A Capital Asset
Pricing Model with Time-Varying Covariances”. Journal of Political
Economy. 96(1): 116–131.
Bollerslev, T., A. Patton, and R. Quaedvlieg. (2020). “Multivariate lever-
age effects and realized semicovariance GARCH models”. Journal
of Econometrics. 217(2): 411–430.
Boyd, S. and L. Vandenberghe. (2004). Convex optimization. Cambridge
University Press.
Boyd, S. and L. Vandenberghe. (2023). “Convex Optimization Addi-
tional Exercises”. https://github.com/cvxgrp/cvxbook_additional_
exercises.
Braga, M. (2015). Risk-based approaches to asset allocation: concepts
and practical applications. Springer.
Brooks, C., S. Burke, and G. Persand. (2003). “Multivariate GARCH
models: software choice and estimation issues”. Journal of Applied
Econometrics. 18(6): 725–734.
Choueifaty, Y. and Y. Coignard. (2008). “Toward maximum diversifica-
tion”. The Journal of Portfolio Management. 35(1): 40–51.
Cochrane, J. (2009). Asset pricing: Revised edition. Princeton university
press.
De Nard, G., O. Engle R.and Ledoit, and M. Wolf. (2022). “Large dy-
namic covariance matrices: Enhancements based on intraday data”.
Journal of Banking & Finance. 138: 106426.
De Nard, G., O. Ledoit, and M. Wolf. (2021). “Factor models for
portfolio selection in large dimensions: The good, the better and the
ugly”. Journal of Financial Econometrics. 19(2): 236–257.
References 407

deBrito, D., M. Medeiros, and R. Ribeiro. (2018). “Forecasting Large

Realized Covariance Matrices: The Benefits of Factor Models and
Shrinkage”. Available at SSRN 3163668.
Dempster, A., N. Laird, and D. Rubin. (1977). “Maximum likelihood
from incomplete data via the EM algorithm”. Journal of the royal
statistical society: series B (methodological). 39(1): 1–22.
Engle, R. (1982). “Autoregressive Conditional Heteroscedasticity with
Estimates of the Variance of United Kingdom Inflation”. Economet-
rica. 50(4): 987–1007. (Accessed on 03/05/2023).
Engle, R. (2002). “Dynamic Conditional Correlation”. Journal of Busi-
ness & Economic Statistics. 20(3): 339–350.
Engle, R. and K. Kroner. (1995). “Multivariate Simultaneous General-
ized ARCH”. Econometric Theory. 11(1): 122–150.
Engle, R., O. Ledoit, and M. Wolf. (2019). “Large dynamic covariance
matrices”. Journal of Business & Economic Statistics. 37(2): 363–
375.
Engle, R. and K. Sheppard. (2001). “Theoretical and empirical proper-
ties of dynamic conditional correlation multivariate GARCH”.
Fama, E. and K. French. (1992). “The cross-section of expected stock
returns”. the Journal of Finance. 47(2): 427–465.
Fama, E. and K. French. (1993). “Common risk factors in the returns
on stocks and bonds”. Journal of financial economics. 33(1): 3–56.
Fan, J., A. Furger, and D. Xiu. (2016). “Incorporating global industrial
classification standard into portfolio allocation: A simple factor-
based large covariance matrix estimator with high-frequency data”.
Journal of Business & Economic Statistics. 34(4): 489–503.
French, K. (2023). “Kenneth French Data Library”. https://mba.tuck.
dartmouth.edu/pages/faculty/ken.french/data_ library.html#
Research.
Ghalanos, A. (2019). “rmgarch: Multivariate GARCH models”. R pack-
age version: 1–3.
Golub, G. and C. Van Loan. (2013). Matrix computations. JHU press.
Grinold, R. and R. Kahn. (2000). “Active portfolio management”.
Hansen, P., Z. Huang, and H. Shek. (2012). “Realized GARCH: a joint
model for returns and realized measures of volatility”. Journal of
Applied Econometrics. 27(6): 877–906.
408 References

Hansen, P., A. Lunde, and J. Nason. (2011). “The model confidence

set”. Econometrica. 79(2): 453–497.
Hastie, T., R. Tibshirani, and J. Friedman. (2009). The elements of
statistical learning: data mining, inference, and prediction. Vol. 2.
Springer.
Hautsch, N., L. Kyj, and P. Malec. (2015). “Do high-frequency data
improve high-dimensional portfolio allocations?” Journal of Applied
Econometrics. 30(2): 263–290.
Hazan, E. (2016). “Introduction to online convex optimization”. Foun-
dations and Trends® in Optimization. 2(3-4): 157–325.
Hazan, E., A. Agarwal, and S. Kale. (2007). “Logarithmic regret algo-
rithms for online convex optimization”. Machine Learning. 69(2-3):
169–192.
Jordan, M. and R. Jacobs. (1994). “Hierarchical mixtures of experts
and the EM algorithm”. Neural computation. 6(2): 181–214.
Laurent, S., J. Rombouts, and F. Violante. (2013). “On loss functions
and ranking forecasting performances of multivariate volatility mod-
els”. Journal of Econometrics. 173(1): 1–10.
Ledoit, O. and M. Wolf. (2003). “Improved estimation of the covariance
matrix of stock returns with an application to portfolio selection”.
Journal of empirical finance. 10(5): 603–621.
Ledoit, O. and M. Wolf. (2004). “Honey, I Shrunk the Sample Covariance
Matrix”. The Journal of Portfolio Management. 30(4): 110–119.
Ledoit, O. and M. Wolf. (2008). “Robust performance hypothesis testing
with the Sharpe ratio”. Journal of Empirical Finance. 15(5): 850–
859.
Lettau, M. and M. Pelger. (2020a). “Estimating latent asset-pricing
factors”. Journal of Econometrics. 218(1): 1–31.
Lettau, M. and M. Pelger. (2020b). “Factors that fit the time series
and cross-section of stock returns”. The Review of Financial Studies.
33(5): 2274–2325.
Longerstaey, J. and M. Spencer. (1996). Riskmetrics: Technical Docu-
ment. JP Morgan and Reuters.
Markowitz, H. (1952). “Portfolio Selection”. The Journal of Finance.
7(1): 77–91.
References 409

McNeil, A., R. Frey, and P. Embrechts. (2015). Quantitative risk man-

agement: concepts, techniques and tools-revised edition. Princeton
university press.
Menchero, J., D. Orr, and J. Wang. (2011). The Barra US equity model
(USE4), methodology notes. MSCI Barra.
Mincer, J. and V. Zarnowitz. (1969). “The evaluation of economic
forecasts”. In: Economic forecasts and expectations: Analysis of
forecasting behavior and performance. NBER. 3–46.
Mokhtari, A., S. Shahrampour, A. Jadbabaie, and A. Ribeiro. (2016).
“Online optimization in dynamic environments: Improved regret
rates for strongly convex problems”. In: 2016 IEEE 55th Conference
on Decision and Control (CDC). 7195–7201.
Oh, D. and A. Patton. (2016). “High-dimensional copula-based distribu-
tions with mixed frequency data”. Journal of Econometrics. 193(2):
349–366.
Patton, A. (2011). “Volatility forecast comparison using imperfect
volatility proxies”. Journal of Econometrics. 160(1): 246–256.
Patton, A. and K. Sheppard. (2009). “Evaluating volatility and cor-
relation forecasts”. In: Handbook of financial time series. Springer.
801–838.
Pelger, M. (2020). “Understanding systematic risk: A high-frequency
approach”. The Journal of Finance. 75(4): 2179–2220.
Pelger, M. (2023). Markus Pelger’s Data and Code. url: https : / /
mpelger.people.stanford.edu/data-and-code.
Pelger, M. and R. Xiong. (2022a). “Interpretable sparse proximate
factors for large dimensions”. Journal of Business & Economic
Statistics. 40(4): 1642–1664.
Pelger, M. and R. Xiong. (2022b). “State-Varying Factor Models of
Large Dimensions”. Journal of Business & Economic Statistics.
40(3): 1315–1333.
Qian, E. (2011). “Risk Parity and Diversification”. The Journal of
Investing. 20(1): 119–127. doi: 10.3905/joi.2011.20.1.119.
Rubin, D. and D. Thayer. (1982). “EM algorithms for ML factor analy-
sis”. Psychometrika. 47: 69–76.
Sharpe, W. (1964). “Capital asset prices: A theory of market equillibrium
under conditions of risk”. The Journal of Finance. 19(3): 425–442.
410 References

Silvennoinen, A. and T. Teräsvirta. (2009). “Multivariate GARCH

models”. In: Handbook of financial time series. Springer. 201–229.
Theil, H. (1961). Economic forecasts and policy. North-Holland Rotter-
dam.
Vassallo, D., G. Buccheri, and F. Corsi. (2021). “A DCC-type approach
for realized covariance modeling with score-driven dynamics”. Inter-
national Journal of Forecasting. 37(2): 569–586.
Vrontos, I., P. Dellaportas, and D. Politis. (2003). “A full-factor mul-
tivariate GARCH model”. The Econometrics Journal. 6(2): 312–
334.
Weide, R. van der. (2002). “GO-GARCH: A Multivariate General-
ized Orthogonal GARCH Model”. Journal of Applied Econometrics.
17(5): 549–564.
“Wharton Research Data Services”. (2023). https://wrds-web.wharton.
upenn.edu/wrds/.
Zinkevich, M. (2003). “Online Convex Programming and Generalized
Infinitesimal Gradient Ascent”. 2.