Estimation risk for the VaR of portfolios driven by semi-parametric multivariate models Christian Francq∗and Jean-Michel Zakoïan† July 17, 2018

Abstract Joint estimation of market and estimation risks in portfolios is investigated, when the individual returns follow a semi-parametric multivariate dynamic model and the asset composition is time-varying. Under ellipticity of the conditional distribution, asymptotic theory for the estimation of the conditional Value-at-Risk (VaR) is developed. An alternative method - the Filtered Historical Simulation - which does not rely on ellipticity, is also studied. Asymptotic confidence intervals for the conditional VaR, which allow for simultaneous quantification of the market and estimation risks, are derived. The particular case of minimum variance portfolios is analyzed in more detail. Potential usefulness, feasibility and drawbacks of the two approaches are illustrated via Monte-Carlo experiments and an empirical study based on stock returns.

JEL Classification: C13, C31 and C58. Keywords: Confidence Intervals for VaR, Dynamic Portfolio, Elliptical Distribution, Filtered Historical Simulation, Minimum Variance Portfolio, Model Risk, Multivariate GARCH.

Forthcoming in the Journal of Econometrics. ∗

CREST and University of Lille, BP 60149, 59653 Villeneuve d’Ascq cedex, France. E-Mail: christian.francq@univ-

lille3.fr † Corresponding author: Jean-Michel Zakoïan, University of Lille and CREST, 5 avenue Henry Le Chatelier, 91120 Palaiseau, France. E-mail: [email protected], Phone number: 33.1.70.26.68.46.

1

1

Introduction

A large strand of the recent literature on quantitative risk management has been concerned with risk aggregation (see for instance Embrechts and Puccetti (2010) and the references therein). For a vector of one-period profit-and-loss random variables y = (y1 , . . . , ym )′ , risk aggregation concerns the risk implied by an aggregate financial position defined as a real-valued function of y. For instance, under the terms of Basel II, banks often measure the risk of a vector y of financial positions by the Value-at-Risk (VaR) of a′ y = a1 y1 + · · · + am ym where the ai ’s define the composition of a portfolio. Exact calculation of the risk associated with an aggregate position can represent a difficult task, as it requires knowledge of the joint distribution of the components of y. It is even more difficult, in a dynamic framework, to evaluate the conditional risk of a portfolio of assets or returns. The current regulatory framework for banking supervision (Basel II and Basel III), allows large international banks to develop internal models for the calculation of risk capital. The so-called advanced approaches are based on conditional distributions, that is, conditional on the past, rather than marginal ones. The superiority of dynamic approaches over static methods based on marginal distributions has been demonstrated empirically, for instance in Kuester, Mittnik and Paolella (2006). The dynamics is not only present in the returns, yt instead of y, but also in the weights of the portfolio, at−1 instead of a. Such weights can be both time-varying and stochastic: the notation at−1 highlights the fact that investors may rebalance their portfolios at time t using, in particular, the information contained in the historical prices. To evaluate the conditional VaR of a portfolio, whose returns are defined by rt = a′t−1 yt , this paper focuses on multivariate semi-parametric approaches. Multivariate approaches are based on a time series model for the vector yt , instead of a univariate model for rt . As emphasized by Rombouts and Verbeek (2009), the advantage of multivariate approaches is to "take into account the dynamic interrelationships between the portfolio components, while the model underlying the VaR calculations is independent of the portfolio composition". Indeed, as shown by Santos, Nogales and Ruiz(2013), the multivariate approach is particularly relevant if the VaR has to be computed for a large number of portfolio compositions at−1 . Moreover, semi-parametric methods allow for more flexibility than fully parametric methods relying on a complete specification of the conditional distribution of yt . To our knowledge, the asymptotic properties of VaR estimators in a dynamic multivariate semiparametric framework are unknown. It seems however important to evaluate the accuracy of risk 2

estimators. Estimation risk refers to the uncertainty implied by statistical procedures in the implementation of risk measures. Uncertainty affects the estimation of risk measures, as well as the backtesting procedures used to assess the validity of risk measures. The new regulatory frameworks require that financial institutions take estimation risk into account (see e.g. Farkas, Fringuellotti and Tunaru (2016) and the references therein). The econometric literature devoted to the estimation risk in dynamic models is scant.

1

Christoffersen and Gonçalves (2005), and Spierdijk (2014)

used resampling techniques to account for parameter estimation uncertainty in univariate dynamic models. Escanciano and Olmo (2010, 2011) proposed corrections of the standard backtesting procedures in presence of estimation risk (and also of model risk). Gouriéroux and Zakoïan (2013) showed that estimation induces an asymptotic bias in the coverage probabilities and proposed a corrected VaR. Francq and Zakoïan (2015) introduced the notion of risk parameter and derived asymptotic confidence intervals for the conditional VaR of univariate returns. The first aim of this paper is to study the asymptotic properties of two multivariate semiparametric approaches for estimating the conditional VaR of a portfolio of risk factors (returns). One approach for estimating conditional VaR’s requires sphericity of the innovations distribution. An alternative approach, known as the Filtered Historical Simulation (FHS) method in the literature (see Barone-Adesi, Giannopoulos and Vosper (1999), Mancini and Trojani (2011) and the references therein), is assumption-free on the innovations distribution. The second aim is to provide methods based on the asymptotic theory or resampling schemes for constructing confidence intervals for the conditional VaR of portfolios. Such confidence intervals are in particular useful to visualize simultaneously the estimation and financial risks. As far as we know, our paper is the first one to study the asymptotic accuracy of conditional VaR estimators in a semi-parametric multivariate framework. The rest of this paper is organized as follows. Section 2 presents the general framework. Section 3 is devoted to the asymptotic properties of the estimators of the conditional VaR under the sphericity assumption. This assumption also allows us to extend the concept of risk parameter to multivariate semi-parametric models. Section 4 gives the asymptotic properties of the FHS method, which relaxes the sphericity assumption. A numerical illustration and an empirical study based on stock returns are proposed in Section 5. Section 6 concludes. Complementary results and proofs are collected in 1

For i.i.d. data, the literature is more voluminous, see Farkas, Fringuellotti and Tunaru (2016) for a recent

reference.

3

the Appendix.

2

Model and conditional VaR

Let pt = (p1t , . . . , pmt )′ denote the vector of prices of m assets at time t. Let yt = (y1t , . . . , ymt )′ be the corresponding vector of log-returns, with yit = log(pit /pi,t−1 ) for i = 1, . . . , m. Consider a portfolio of the m assets, whose return is given by rt =

m X

ai,t−1 yit = a′t−1 yt ,

(2.1)

i=1

where at−1 = (a1,t−1 , . . . , am,t−1 )′ is the vector of portfolio weights for the m assets. Such weights are assumed to be stochastic and measurable with respect to some information set It−1 containing the past prices (and possibly other variables). A portfolio is called crystallized when the number µi of units of each asset i is time independent. For such a portfolio we have ai,t−1 =

Pmµi pi,t−1 . j=1 µj pj,t−1

The conditional VaR of the portfolio’s return process (rt ) at risk level α ∈ (0, 1), denoted by (α)

VaRt−1 (rt ), is defined by

h i (α) Pt−1 rt < −VaRt−1 (rt ) = α,

(2.2) 2

where Pt−1 denotes the historical distribution conditional on It−1 . More generally, we denote by (α)

VaRt−1 (zt ) the conditional VaR of zt given It−1 , and by VaR(α) (z) the marginal VaR of a stationary process (zt ). Consider a general multivariate model for the vector of log-returns yt = mt (θ0 ) + ǫt ,

ǫt = Σt (θ0 )ηt ,

(2.3)

where (ηt ) is a sequence of independent and identically distributed (iid) Rm -valued variables with zero mean and identity covariance matrix; the m × m non-singular matrix Σt (θ0 ) and the m × 1 vector mt (θ0 ) are specified as functions depending on the infinite past of yt and parameterized by a d-dimensional parameter θ0 : mt (θ0 ) = m(yt−1 , yt−2 , . . . , θ0 ),

Σt (θ0 ) = Σ(yt−1 , yt−2 , . . . , θ0 ).

(2.4)

For the sake of generality, we do not consider a particular specification for the conditional mean mt and the conditional variance Ht (θ0 ) := Σt (θ0 )Σ′t (θ0 ), 2 3

3

but we assume

In this formula, we assumed for simplicity that the conditional cdf of rt is continuous and strictly increasing. The most widely used specifications of Multivariate GARCH (MGARCH) models are discussed in Bauwens,

4

A1:

(yt ) is a strictly stationary solution of Model (2.3)-(2.4), and ηt is independent from It−1 .

This assumption will be made explicit for particular classes of MGARCH models satisfying Model (2.3)-(2.4). Under (2.3)-(2.4), the portfolio’s return defined in (2.1) satisfies rt = a′t−1 mt (θ0 ) + a′t−1 Σt (θ0 )ηt , from which it follows that the portfolio’s conditional VaR at level α is given by

(2.5) 4


(α) (α) VaRt−1 (rt ) = −a′t−1 mt (θ0 ) + VaRt−1 a′t−1 Σt (θ0 )ηt .

(2.6)

The VaR formula can be simplified if the errors ηt have a spherical distribution, that is, P ηt and ηt have the same distribution for any orthogonal matrix P . Ellipticity of the conditional distribution of yt is equivalent to A2:

d

for any non-random vector λ ∈ Rm , λ′ ηt = kλkη1t , d

where k · k denotes the euclidian norm on Rm , ηit denotes the i-th component of ηt , and = stands for the equality in distribution.

5

Remark 2.1 (Restrictiveness of the sphericity assumption) Assumption A2 entails that though not independent (except in the Gaussian case), the components of ηt have the same symmetric distribution. This assumption is commonly used in finance and econometrics (see for instance Sentana (2003), Fiorentini and Sentana (2016)). The importance of the class of spherical and more generally elliptical - distributions to risk management is discussed in Bradley and Taqqu (2002). Examples of spherical distribution are the Gaussian N (0, Im ) distribution, and the standard multivariate Student distribution (see McNeil, Frey and Embrechts (2005) for details on spherical distributions). In fact, most parametric approaches for VaR estimation assume a spherical Gaussian Laurent and Rombouts (2006), Silvennoinen and Teräsvirta (2009), Francq and Zakoïan (2010, Chapter 11), Bauwens, Hafner and Laurent (2012), Tsay (2014, Chapter 7). Model (2.3)-(2.4) also includes multivariate extensions of the double-autoregressive models studied by Ling (2004). 4

The presence of the sign "−" in this formula comes from the fact that the VaR is defined in terms of returns

instead of loss variables. 5

Note that, in A2, the Euclidian norm cannot be replaced by any other norm N (·) under the assumption of unit d

covariance matrix for ηt . Indeed, if λ′ ηt = N (λ)η1t , we have Var(λ′ ηt ) = λ′ λ = N (λ)2 Var(η1t ) = N (λ)2 .

5

or Student error distribution, which is very restrictive in terms of kurtosis (for the Gaussian distribution) and more generally on the tails of the distribution. By contrast, Assumption A2 does not constrain (apart from symmetry) the size of the tails. It should also be noted that, while A2 entails that the components of ηt have the same symmetric distributions, this does not hold in general neither for the marginal nor for the conditional distribution of yt . In particular, this assumption is compatible with the usual leverage effect observed on financial returns. In other words, the elliptical model may incorporate asymmetric conditional second moments. Under the sphericity assumption A2 we have

(α) VaRt−1 (rt ) = −a′t−1 mt (θ0 ) + a′t−1 Σt (θ0 ) VaR(α) (η) ,

(2.7)

where VaR(α) (η) is the (marginal) VaR of η1t .

Remark 2.2 (Usefulness of sphericity for the VaR) By contrast with formula (2.6), the interest of (2.7) is to relate the VaR of any portfolio to the first two conditional moments of the portfolio’s return rt , and to a simple characteristic of the innovations distribution. Under the sphericity assumption, the VaR is indeed a function of three ingredients: the mean-volatility parameter, the quantile of the errors and the portfolio’s composition. In other words, (α)

VaRt−1 (rt ) = F (θ0 , VaR(α) (η) ; a′t−1 ),

(2.8)

where the first two components have to be estimated, while the third one is chosen by the risk manager. Such a decomposition does not hold in (2.6), which requires estimating a conditional quantile for any choice of the portfolio’s composition (see Section 4). As we will see in Section 3.2, for most time series models (2.8) can even be reduced to a formula of the form (α)

(α)

VaRt−1 (rt ) = F ∗ (θ0 ; a′t−1 ), (α)

with a new parameter θ0

(2.9)

of the same dimension as θ0 , henceforth called conditional VaR param-

eter. These simplifications, (2.8) and (2.9), of the general VaR formula (2.6) have obvious interest for risk management, in particular when several portfolios based on the same risk factors have to be managed simultaneously.

3

VaR estimation under conditional ellipticity

Formula (2.7) is well-known in the literature dealing with theoretical properties of VaR (see for instance McNeil et al., 2005), but its econometric implications have been surprisingly overlooked. 6

We now consider the statistical implementation of this formula. Under the sphericity assumption A2, a natural strategy for estimating the conditional VaR of a portfolio is to estimate θ0 by some consistent estimator θbn in a first step, to extract the residuals

and to estimate VaR(α) (η) in a second step. For the first step, we will consider a general estimator satisfying regularity conditions. For the second step, the sphericity assumption will allow us to interpret VaR(α) (η) as the (1 − 2α)-quantile ξ1−2α of the absolute residuals, and to estimate this quantile by an empirical quantile using all the components of the first-step residuals. Let Θ denote the parameter space, and assume θ0 ∈ Θ.

Let θbn denote an estimator of

parameter θ0 , obtained from observations y1 , . . . , yn and initial values ye0 , ye−1 , . . . .

The vec-

e −1 (θbn ){yt − m ft (θbn )}) = (b ft (θ) = η1t , . . . , ηbmt )′ , where m tor of residuals is defined by ηbt = Σ t m(yt−1 , . . . , y1 , ye0 , ye−1 , . . . , θ),

e t (θ) = Σ(yt−1 , . . . , y1 , ye0 , ye−1 , . . . , θ), for t ≥ 1 and θ ∈ Θ. Σ

For α ∈ (0, 1), let qα (S) denote the α-quantile of a finite set S ⊂ R. In view of (2.7), under the conditional ellipticity/sphericity assumption, an estimator of the conditional VaR at level α is ′ e t (θbn )kξn,1−2α , d (α) ft (θbn ) + ka′t−1 Σ VaR S,t−1 (rt ) = −at−1 m

(3.1)

where ξn,1−2α = q1−2α ({|b ηit |, 1 ≤ i ≤ m, 1 ≤ t ≤ n}). The latter estimator takes advantage of the fact that the components of ηt are identically distributed under A2.

3.1

Asymptotic joint distribution of θbn and a quantile of absolute returns

We start by introducing the assumptions that are employed to establish the asymptotic distribution of (θbn′ , ξn,1−2α ).

c

We now assume that the estimator θbn admits a Bahadur representation. Write a = b for a = b+c.

A3:

We have θbn → θ0 , a.s. Moreover, the following expansion holds  √  n θbn − θ0

oP (1)

=

n

1 X √ ∆t−1 V (ηt ), n t=1

(3.2)

where V (·) is a measurable function, V : Rm 7→ RK for some positive integer K, and ∆t−1 is a d × K matrix, measurable with respect to the sigma-field generated by {ηu , u < t}. The

variables ∆t and V (ηt ) belong to L2 with EV (ηt ) = 0, E∆t = Λ is full row rank.

7

var{V (ηt )} = Υ is nonsingular and

6

Assumption A3 holds for a variety of MGARCH models and estimators (see Appendix A for examples). The next assumption imposes smoothness of the functions m and Σ with respect to the parameter. A4: For any sequence (xi ), the functions θ 7→ m(x1 , x2 , . . . ; θ) and θ 7→ Σ(x1 , x2 , . . . ; θ) are continuously differentiable over Θ. The next theorem establishes the asymptotic normality of (θbn′ , ξn,1−2α ). Let    
∂ −1 ′ ′ ′ Ψ = E(∆t Υ∆t ), Ω = E vec Σt vec (Σt ) , Wα = Cov(V (ηt ), Nt ), ∂θ ′   P 1 − 1 + 2α , and, denoting by f the density of |η1t |, γα = var(Nt ), with Nt = m {|ηjt | 0, and any sequence (xi )i on Rm ,

m(x1 , x2 , . . . ; θ) = m(x1 , x2 , . . . ; θ ∗ ), KΣ(x1 , x2 , . . . ; θ) = Σ(x1 , x2 , . . . ; θ ∗ ), 6

and where θ ∗ = G(θ, K).

In the univariate setting, the asymptotic theory of estimation for GARCH parameters has been extensively

studied, in particular for the QMLE by Berkes, Horváth and Kokoszka (2003) and for the LAD (Least Absolute Deviation) estimator by Ling (2005). In the multivariate setting, the asymptotic properties of the QMLE or alternative estimators were established, for particular classes, by Comte and Lieberman (2003), Boswijk and van der Weide (2011), Francq and Zakoian (2012), Pedersen and Rahbek (2014), Francq, Horváth and Zakoian (2015), Francq and Zakoian (2016) among others.

8

In other words, a change of the scale in the components of η can be compensated by a change of the parameter. This assumption is obviously satisfied for all commonly used parametric forms of 7

Σt (θ).

Under sphericity and the stability by scale assumption A5, the conditional VaR can be

expressed in function of the expected returns vector and a reparameterized volatility matrix. Let α < 1/2, so that VaR(α) (η) > 0 under A2. It follows from A5 that a formula of the form (2.9) holds, namely (α)

(α)

(α)

VaRt−1 (rt ) = −a′t−1 mt (θ0 ) + ka′t−1 Σt (θ0 )k where (α)

θ0 (α)

The new parameter θ0

n o = G θ0 , VaR(α) (η) .

(3.4)

(3.5)

is referred to as the conditional VaR parameter, for a given risk level. It

does not depend on the portfolio composition. An estimator of the conditional VaR parameter can be defined as

  (α) (α) b d b θn = G θn , VaRn (η) (α)

with obvious notations. The asymptotic properties of θbn

are a direct consequence of Theorem 3.1.

Corollary 3.1 (CAN of the VaR-parameter estimator) Under the assumptions of Theorem i  h   √  (α) L (α) ˙ G˙ ′ where G˙ = ∂G(θ,ξ) . → N 0, Ξ∗ := GΞ 3.1, n θbn − θ0 ∂(θ′ ,ξ) (θ0 ,ξ1−2α )

Remark 3.1 (Usefulness of the conditional VaR parameter) Quantifying the estimation risk is in general a difficult task, due to the stochastic nature of the conditional risk. However, (α) when the VaR takes the form (2.9), the asymptotic distribution of θbn provides a quantification

of the estimation risk. It can be used to compare the relative asymptotic efficiencies of estimators. (i) Suppose, for instance, that estimators θbn , i = 1, . . . , m, satisfying (3.2) are available and let Ξ(i)

denote the corresponding asymptotic covariance matrices in (3.3). Then, we can say that, as far as the estimation of the conditional VaR at level α is concerned, the i-th estimator is asymptotically more efficient than the j-th iff

7

˙ (j) − Ξ(i) )G˙ ′ G(Ξ

is a positive semidefinite matrix.

For instance, in the case of the VAR(1) model yt = φyt−1 + ǫt with a BEKK-GARCH(1,1) model (3.7) for ǫt ,

and θ = (vec(φ)′ , vec(A)′ , vec(B)′ , vec(C)′ )′ , we find θ∗ = (vec(φ)′ , Kvec(A)′ , vec(B)′ , K 2 vec(C)′ )′ .

9

0.30 0.25 0.20

True and estimated VaR

700

720

740

760

780

800

Time

Figure 1: True 1%-VaR (full black line), estimated 1%-VaR (full blue line) and estimated 95%-confidence interval (dotted blue line), on a simulation of a fixed portfolio of a bivariate BEKK.

3.3

Asymptotic confidence intervals for the VaR’s of portfolios

b denote a consistent estimator of Ξ. Let α0 ∈ (0, 1). In view of (3.1), by the delta method, Let Ξ

an approximate (1 − α0 )% confidence interval (CI) for VaRt (α) has bounds given by o1/2 n b t−1 d (α) (rt ) ± √1 Φ−1 (1 − α0 /2) δ′ Ξδ , VaR t−1 S,t−1 n

(3.6)

where Φ−1 (u) denotes the u-quantile of the standard Gaussian distribution, u ∈ (0, 1), and " # e t (θbn ) ∂f m(θbn ) (at−1 ⊗ at−1 )′ ∂vecH ′ ′ ′ e t (θbn )k , δt−1 = at−1 + kat−1 Σ e t (θbn )k ∂θ ′ ∂θ ′ 2ka′t−1 Σ

ft (·) = Σ e t (·)Σ e ′ (·). See Beutner, Heinemann and Smeekes (2017) for an asymptotic justificawith H t tion of this type of CI incorporating parameter uncertainty. Drawing such CIs allows to take into

account the estimation risk inherent to the evaluation of the VaR of the portfolio. Note that the level α0 of risk estimation is independent from the market risk level α. An illustration is displayed in Figure 1, for the simulation of a bivariate BEKK model (see Appendix 3.4). The model parameters were estimated on 700 observations. The figure provides the true and estimated conditional 1%-VaRs, for t > 700, as well a CIs at 95% for the true conditional VaR, of a portfolio with fixed composition. This graph allows to visualize simultaneously the market risk (through the magnitude of the VaR) and the estimation risk (through the width of the CIs).

10

3.4

Estimating the asymptotic covariance matrix Ξ

In Theorem 3.1, most quantities involved in the asymptotic covariance matrix Ξ can be estimated by empirical means, replacing θ0 by the estimate θbn and the ηt ’s by the corresponding residuals. We focus on the estimation of Ω, which is the most delicate problem due to the presence of the

derivatives of Σt . If a recursive linear relationship between Σt and its past-values existed, then the derivatives could be computed recursively (as the derivatives of the σt or σt2 in standard univariate GARCH models). Unfortunately, the standard multivariate volatility models do not allow to derive such a recursive relationship. Let us distinguish two general class of models, depending on the type of stochastic recursive equation (SRE) involved in the dynamics. 3.4.1

Linear SRE on Ht

A typical example is the BEKK model of Engle and Kroner (1995). As in Pedersen and Rahbek (2014), we focus on the BEKK-GARCH(1,1) model, in which Σt (θ0 ) is the symmetric square root of Ht , given by 1/2

ǫt = Ht

ηt ,

Ht = C0 + A0 ǫt−1 ǫ′t−1 A′0 + B0 Ht−1 B0′

(3.7)

where A0 , B0 and C0 are real m × m matrices, with C0 positive definite, such that Ht is a positive

definite matrix. For some m × m matrices A, B and C > 0, let θ = (vec(A)′ , vec(B)′ , vec(C)′ )′ . The derivatives of vec(Ht ) can be computed as follows, omitting θ for ease of notation. From vec(Ht ) = vec(C) + (A ⊗ A)vec(ǫt ǫ′t ) + (B ⊗ B)vec(Ht−1 ), it follows that, for j = 1, . . . , 3d,

∂vec(C) ∂(A ⊗ A) + vec(ǫt ǫ′t ) ∂θj ∂θj ∂(B ⊗ B) ∂vec(Ht−1 ) + vec(Ht−1 ) + (B ⊗ B) . ∂θj ∂θj ′  ′ ∂M ′ . , . . . , For any m × n matrix M , let the dm × n matrix ∂M = ∂M ∂θ1 ∂θd ∂vec(Ht ) ∂θj

(vec′ (Ht ), {∂vec(Ht

where

=

Let Xt =

)}′ )′ .

We have, in block matrix notation,   B⊗B 0  Xt−1 + et , Xt =  ∂(B ⊗ B) Id ⊗ (B ⊗ B) 

et = 

vec(C) ∂vec(C)





+ 11

A⊗A ∂(A ⊗ A)



 vec(ǫt ǫ′t ).

(3.8)

Equation (3.8) allows to compute recursively the matrix Ht and its derivatives, provided that some initial values are chosen. 1/2

It remains to compute the derivatives of Σt = Ht

. Without generality loss, this matrix can

t be assumed to be symmetric and positive definite. We note that Σt ∂Σ ∂θi +

(Im ⊗ Σt + Σt ⊗ Im ) vec



∂Σt ∂θi



= vec



∂Ht ∂θi



∂Σt ∂θi Σt

=

∂Ht ∂θi .

Thus (3.9)

,

which allows to compute the derivative of Σt provided Im ⊗ Σt + Σt ⊗ Im is non-singular. In fact Im ⊗ Σt + Σt ⊗ Im = (Im ⊗ Σt )(Im2 + Σt ⊗ Σ−1 t ). The eigenvalues of Σ−1 and Σt being positive, the eigenvalues of the latter parenthesis are larger t than 1. The invertibility of Im ⊗ Σt + Σt ⊗ Im follows and we have     ∂Σt ∂Ht −1 vec = (Im ⊗ Σt + Σt ⊗ Im ) vec . ∂θi ∂θi 3.4.2

Linear SRE’s on the individual volatilities and the conditional correlation matrix 1/2

Consider parameterizations of the form Σt (θ) = Dt (θ)Rt (θ) where Dt (θ) is the diagonal matrix 1/2

of the individual volatilities (at θ0 ), and Rt (θ) denotes the symmetric positive definite square1/2

root of the conditional correlation matrix Rt (θ) (that is {Rt (θ)}2 = Rt (θ)). For all commonly used models, the derivatives of the individual volatilities (or their squares) can be straightforwardly computed, using the SRE on the vector of individual volatilities. The matrix

∂ ∂θi Dt (θ)

follows, for

1/2

any component θi of θ. Turning to the derivatives of Rt (θ), we note that, similar to (3.9), !   1/2 −1  ∂Rt ∂Rt 1/2 1/2 vec vec = Im ⊗ Rt + Rt ⊗ Im . ∂θi ∂θi Usual DCC models provide a SRE on the conditional correlation matrix Rt , from which the deriva1/2

tives of Rt

can be computed using the previous equality. Consider the cDCC model (see Appendix ∗−1/2

C). We have Rt = Qt

∗−1/2

Qt Qt

, and ∗1/2

∗1/2

−1 −1 Qt = (1 − α − β)S + αQt−1 Dt−1 ǫt−1 ǫ′t−1 Dt−1 Qt−1 + βQt−1 ,

where S is a correlation matrix. The diagonal terms of Qt are given by ! ǫ2i,t−1 qii,t = (1 − α − β) + α 2 + β qii,t−1 , σi,t−1 12

∗1/2

from which the derivatives of Q∗t can be recursively computed. The derivatives of Qt (3.9), which in the diagonal case reduces to

∗1/2

∂Qt ∂θi

=

∗ 1 ∗−1/2 ∂Qt 2 Qt ∂θi .

follow from

Now we turn to the non diagonal

terms. We have, for i 6= j, ǫi,t−1 √ ǫj,t−1 √ + βqij,t−1 , qij,t = (1 − α − β)Sij + α qii,t−1 qjj,t−1 σi,t−1 σj,t−1 from which the derivatives of qij,t follow recursively. The conclusion follows.

3.5

CI’s based on a conditional resampling scheme

For certain estimation methods/models the asymptotic distribution of the estimator θbn may not

be available. Even when it is, as shown in the previous section, the asymptotic variance Ξ may be

difficult to estimate. In particular, apart from the estimation of Ω, evaluating the density function f of the innovations distribution at the desired quantile may be delicate. An alternative, which we will now illustrate, is a bootstrap procedure. We will use the well-known result that, under the sphericity assumption, kηt k and ηt /kηt k are independent, the latter random variable being

uniformly distributed over the unit sphere S m−1 .

We consider the following resampling scheme, given observations y1 , . . . , yn and initial values: d (α) (rt ) =: VaR(r d t ). 1. Compute θbn = θbn (y1 , . . . , yn ), the residuals ηbt , and the estimator VaR S,t−1

2. Generate independent vectors s∗u , u = 1, . . . , n, that are uniformly distributed over S m−1 . e u∗ , that are uniformly distributed on (U e1 , . . . , U en ) where Independently, generate vectors U e u = Su−1/2 (ηbu − η b), Su is the sample covariance matrix of the residuals ηbu and η b is their U

e ∗ (θbn )η ∗ , where m e ∗ ks∗ and let y ∗ = m f∗u (θbn ) = f∗u (θbn ) + Σ sample mean. Compute ηu∗ = kU u u u u u ∗ , . . . , y∗ , y e ∗ (θbn ) = Σ(y ∗ , . . . , y ∗ , ye0 , ye−1 , . . . , θbn ). e−1 , . . . , θbn ) and Σ m(yu−1 u u−1 1 1 e0 , y

e ∗−1 (θb∗ ){y ∗ − m f∗u (θbn∗ )}) and 3. Compute θbn∗ = θbn (y1∗ , . . . , yn∗ ), the resampling residuals ηbu∗ = Σ u n u the estimator

′ b∗ d ∗ (rt ) = −a′ m e b∗ ∗ VaR t−1 f t (θn ) + kat−1 Σt (θn )kξn,1−2α ,

(3.10)

∗ |, 1 ≤ i ≤ m, 1 ≤ u ≤ n}), m ∗ ft (θbn∗ ) = m(yt−1 , . . . , y1 , ye0 , ye−1 , . . . , θbn∗ ) = q1−2α ({|b ηiu ξn,1−2α

e t (θbn ) = Σ(yt−1 , . . . , y1 , ye0 , ye−1 , . . . , θb∗ ). and Σ n

d ∗ (rt ), say. d ∗ (rt ), . . . VaR 4. Repeat B times Steps 1-3, resulting in VaR B 1 13

0.35 0.30 0.25 0.20

True and estimated VaR

True VaR Estimated VaR Asymptotic CI Bootstrap CI

700

720

740

760

780

800

Figure 2: Confidence Intervals based on asymptotic results vs bootstrap. Time

e t (·) and m ft (·) are built using the real data (not the Note that in (3.10), the conditional moments Σ

bootstraped ones), because we are estimating a conditional VaR (not a marginal one). Using the

pivot method (see e.g. Davison and Hinkley (1997)), we get a CI for the conditional VaR at the confidence level 1 − α0 as h

n o n oi ∗ d t ) − VaR d∗ d d d d VaR(r (r ) − VaR(r ) , VaR(r ) − VaR (r ) − VaR(r ) , t t t (1−α0 /2) t (α0 /2) t

with standard notation for the order statistic. An illustration is displayed in Figure 2, for the same setting as for Figure 1. As can be seen, the CI’s obtained by the bootstrap approach are similar to those obtained using the asymptotic results. For more complex models, or for estimators for which the asymptotic distribution is unknown or cannot be estimated, the latter CI’s would be impossible to derive, while the bootstrap approach described above could be implemented without further difficulties. The asymptotic validity of this procedure is however an open issue.

4

VaR estimation without the sphericity assumption

Rombouts and Verbeek (2009) proposed a semi-parametric method for evaluating the VaR of portfolios, which relies on: i) estimating θ0 , ii) using a Kernel estimator of the (multivariate) density of ηt , iii) evaluating by numerical integration the conditional VaR of a portfolio. While this approach seems attractive from a practical point of view, its asymptotic properties are unknown. Deriving

14

asymptotic confidence intervals for the VaR, which is the aim of this paper, would probably be extremely difficult with this method. In this section, we study an alternative semi-parametric method which is amenable to asymptotic properties. This approach, called FHS, relies on i) interpreting the conditional VaR at time t as the α-quantile of a linear combination (depending on t) of the components of the innovations; ii) replacing the innovations by the GARCH residuals and computing the empirical α-quantile of the estimated linear combination. The conditional VaR of the portfolio return is (α) (α)  VaRt−1 (rt ) = VaRt−1 bt (θ0 ) + c′t (θ0 )ηt

where bt (θ) = a′t−1 mt (θ) and c′t (θ) = a′t−1 Σt (θ). The conditional VaR at time t can thus be interpreted as the sum of the conditional mean and a quantile of a time-varying linear combination of the components of the iid noise. It can be estimated by d (α) VaR F HS,t−1 (rt ) = −qα

n bt (θbn ) + c′t (θbn )ηbs ,

1≤s≤n

o

(4.1)

.

Remark 4.1 (On the name FHS) In (4.1), all residuals are used to estimate the VaR. Alternabs ’s, tively, the conditional VaR could be estimated by randomly drawing N residuals among the η

for some specified number N (hence the term "simulation" in FHS).

Remark 4.2 (Higher horizons) The approach can be extended to higher horizons. For N in(N )

(1)

dependent draws of the ηbs ’s, N scenarios yt , . . . , yt

(i)

for yt are obtained. For each value yt , (i,1)

(i,N )

bs ’s, produces N scenarios yt+1 , . . . , yt+1 for yt+1 . Proanother set of N independent draws of the η (i ,...,i )

1 H for yt+H−1 , where ij ∈ {1, . . . , N }. ceeding recursively, at horizon H we get N H scenarios yt+H−1

Such scenarios allow to update the sequence of weights as , for s = t, . . . , t + H − 1. We deduce N H (i ,...,i )

1 H for rt+H−1 . The VaR of the portfolio at horizon H conditional on the available scenarios rt+H−1

information at time t − 1 can thus be estimated by d (H,α) (rt+H−1 ) = −qα VaR F HS,t−1

n (i1 ,...,iH+1 ) , rt+H−1

o ij ∈ {1, . . . , N } .

(4.2)

Let c : Θθ 7→ Rm and b : Θθ 7→ R denote continuously differentiable vector-valued functions. Let

ξα (θ) denote the theoretical α-quantile of b(θ) + c′ (θ)ηt (θ), where ηt (θ) = Σ−1 t (θ){yt − mt (θ)}. 15

Let ξn,α (θ) = qα ({b(θ) + c′ (θ)ηt (θ), 1 ≤ t ≤ n}). Let Aα = Cov(V (ηt ), 1{b(θ0 )+c′ (θ0 )ηt 0 or if a = 0 and bn < 0, and u = −1

otherwise. Thus ′ b ′ b |b ηkt | = ηkt − Mkt (θn − θ0 ) + oP (n−1/2 ) = |ηkt| − ukt Mkt (θn − θ0 ) + oP (n−1/2 ),

where ukt = ±1, the sign of ukt being equal to that of ηkt when ηkt 6= 0, and to the sign of ′ b −Mkt (θn − θ0 ) when ηkt = 0. Using the identity Z v  ρ1−2α (u − v) − ρ1−2α (u) = −v(1 − 2α − 1{u