Lu Cao. A Thesis in The Department of Mathematics and Statistics

Transcription

1 Multivariate Robust Vector-Valued Range Value-at-Risk Lu Cao A Thesis in The Department of Mathematics and Statistics Presented in Partial Fulllment of the Requirements for the Degree of Master of Science (Mathematics) at Concordia University Montreal, Quebec, Canada August 2017 c Lu Cao, 2017

2 CONCORDIA UNIVERSITY School of Graduate Studies This is to certify that the thesis prepared By: Lu Cao Entitled: Multivariate Robust Vector-Valued Range Value-at-Risk and submitted in partial fulllment of the requirements for the degree of Master of Science (Mathematics) complies with the regulations of the University and meets the accepted standards with respect to originality and quality. Signed by the nal Examining Committee: Dr. M. Mailhot Dr. F. Godin Dr. V. Bignozzi Thesis Supervisor Examiner Examiner Approved by 2017 Chair of Department or Graduate Program Director Dean of Faculty

3 iii Abstract Multivariate Robust Vector-Valued Range Value-at-Risk Lu Cao In a multivariate setting, the dependence between random variables has to be accounted for modeling purposes. Various of multivariate risk measures have been developed, including bivariate lower and upper orthant Value-at-Risk (VaR) and Tail Value-at-Risk (TVaR). The robustness of their estimators has to be discussed with the help of sensitivity functions, since risk measures are estimated from data. In this thesis, several univariate risk measures and their multivariate extensions are presented. In particular, we are interested in developing the bivariate version of a robust risk measure called Range Value-at-Risk (RVaR). Examples with dierent copulas, such as the Archimedean copula, are provided. Also, properties such as translation invariance, positive homogeneity and monotonicity are examined. Consistent empirical estimators are also presented along with the simulation. Moreover, the sensitivity functions of the bivariate VaR, TVaR and RVaR are obtained, which conrms the robustness of bivariate VaR and RVaR as expected.

4 iv Acknowledgments I would like to express my gratitude to all those who helped me during the writing of this thesis. First and foremost, I would like to thank Dr. Mélina Mailhot, my supervisor, for her constant encouragement and guidance. She has walked me through all the stages of the writing of this thesis. Without her consistent and illuminating instruction, this thesis could not have reached its present form. Sencond, special thanks should go to Prof. Valeria Bignozzi who put considerable time into the comments on this thesis. Also, I would like to express my gratitude to all the professors in the department. Last my thanks would go to my friends who have always been helping me out of diculties and supporting me.

5 CONTENTS List of Figures ix List of Tables ix 1. Introduction Univariate Risk Measures Preliminaries Denitions and Properties Robustness Numerical Examples Multivariate Risk Measures Copulas Bivariate VaR and TVaR Bivariate Lower Orthant RVaR Bivariate Upper Orthant RVaR Properties of Bivariate RVaR Empirical Estimators and Robustness of RVaR Empirical Estimator for Bivariate RVaR

6 Contents vi 4.2 Robustness of Multivariate Risk Measures Simulation Conclusion

7 LIST OF FIGURES 3.1 Gumbel and Frank Copulas with dependent parameters θ = 2 and θ = Clayton Copula with dependent parameters θ = Lower orthant VaR at level Upper orthant VaR at level Lower orthant VaR and TVaR at level Upper orthant VaR and TVaR at level Lower orthant VaR at level 0.95 and RVaR at level range [0.95, 0.99] for xed values of X Lower orthant VaR at level 0.95 and RVaR at level range [0.95, 0.99] for xed values of X Upper orthant VaR at level 0.99 and RVaR at level range [0.95, 0.99] for xed values of X Upper orthant VaR at level 0.99 and RVaR at level range [0.95, 0.99] for xed values of X Empirical estimator of the lower orthant VaR at level Empirical estimator of the upper orthant VaR at level Empirical estimator of the lower orthant RVaR at level range [0.95, 0.99] Empirical estimator of the upper orthant RVaR at level range [0.95, 0.99].. 62

8 List of Figures viii 4.5 Empirical estimator of the lower orthant RVaR at level range [0.95, 0.99] with dependent parameters θ = 1.4, θ = 1.5 and θ = Empirical estimator of the upper orthant RVaR at level range [0.95, 0.99] with dependent parameters θ = 1.4, θ = 1.5 and θ =

9 LIST OF TABLES 2.1 Probability distribution of the discrete variable X RV arα1,α 2 (X) and ρ(x)

10 1. INTRODUCTION Volatility and risks of nancial markets have recently increased signicantly with the globalization of economy and nancial innovation. For companies, risk management is crucial to their success. Entities are interested in risk measures in order to allocate capital and maintain solvency. Dierent univariate risk measures have been proposed in the literature. Consider a random loss variable X on a probability space (Ω, F, P) with its cumulative distribution function (cdf) F X. The term Value-at-Risk (VaR) found its way through the G-30 report published in July 1993 see [14] for details. It is the loss in market value that can only be exceeded with a probability of at most 1 α where α often takes value 0.95 or However, Artzner et al. (1999) show that VaR is not a coherent risk measure and it does not provide any information about the tail of the distribution, suggesting two specic risk measures called Tail Conditional Expectation (TCE) and Worst Conditional Expectation (WCE). TCE evaluates the average value of VaR over all condence levels greater than α while WCE is the expected loss under the condition that the set of worst events occurs. An alternative risk measure called Conditional Value-at-Risk (CVaR) is used to optimize portfolios by Rockafellar and Uryasev (2000). Another remedy for the deciencies of VaR is Expected Shortfall (ES) proposed by Acerbi and Tasche (2002). WCE is closely related to TCE, but in general does not coincide with it. For discrete random variables, the WCE could be greater than the TCE. WCE is only useful in a theoretical setting since it requires the knowledge of

11 1. Introduction 2 the whole underlying probability space while TCE is easy to compute but not coherent for discrete random variables. ES and CVaR are two dierent interpretations of the weighted average between the VaR and losses exceeding VaR. ES is more precise than CVaR and easier to compute, since it considers the eect of jump points. Specially, ES is continuous and monotonic with respect to the signicant level α. Although all these measures have been widely studied, univariate risk measures are not enough for current nancial markets since nancial risks are strongly interconnected and cannot only be managed individually or by aggregation. In reality, companies have to consider the dependence between risks so that they can get the accurate capital allocation, and systemic and global risk evaluation. Systemic risk refers to the risks imposed by interdependencies in a system. Univariate risk measures are unable to be used for heterogeneous classes of homogeneous risks. Therefore, multivariate risk measures have been developed in the last decade. An extension of the Worst Conditional Expectation (WCE) in Artzner et al. (1999) is called the Multivariate Worst Conditional Expectation (MWCE), which is proposed by Jouini et al. (2004). In the same framework, Bentahar (2006) introduces a quantile-based risk measure called vector-valued Tail Conditional Expectation (TCE). Furthermore, Tahar and Lépinette (2012, 2014) propose the Generalized Worst Conditional Expectation (GWCE). Embrechts and Puccetti (2006), Nappo and Spizzichino (2009) and Prékopa (2012) use the notion of quantile curves to dene a multivariate risk measure called upper and lower orthant VaR. Based on the same idea, Cossette et al. (2013) redene the upper and lower orthant VaR and propose the upper and lower orthant TVaR in Cossette et al. (2015). At the same time, Cousin and Di Bernardino (2013) develop a vectorized version of the upper and lower orthant VaR. Moreover, multivariate extensions of CTE and CoVaR are developed in Cousin and Di Bernardino (2014, 2015). CoVaR represents VaR for a nancial institution, conditional

12 1. Introduction 3 on the boundary of the α-level set, and measures a nancial institution's contribution to the system's risk. A drawback of multivariate VaR is that it represents the boundary of the α- level set and no information above is provided, similarly to the univariate VaR. Furthermore, relationships holding for univariate risk measures can be totally dierent in a multivariate setting. Thus, in this thesis, we will compare and summarize the relationships between these multivariate risk measures. Most risk measures are dened as functions of the loss distribution which should be estimated from data in applications. Cont et al. (2010) dene risk measurement procedure and analyze the robustness of dierent risk measures. They point out the conict between the subadditivity and robustness and propose a robust risk measure called weighted VaR (WVaR). Bignozzi and Tsanakas (2016) also suggest to use the truncated version of TVaR (or Range-Value-at-Risk (RVaR)) which is same as WVaR when the mean of the loss distribution is innite. We will develop the multivariate RVaR in this thesis, in order to provide a new robust multivariate risk measure.

13 2. UNIVARIATE RISK MEASURES 2.1 Preliminaries A risk measure ρ(x) for a univariate risk X corresponds to the required assets that have to be maintained such that the nancial position ρ(x) X is acceptable for regulators. Since there are several ways to dene risk measures, an appropriate choice becomes crucial for both regulators and entrepreneurs. Properties of coherent risk measures proposed by Artzner et al. (1999) can be signicant criteria. Denition For random variables X and Y, a risk measure ρ is a coherent risk measure if it satises the following four axioms, 1. (Translation invariance) For all c R, ρ(x + c) = ρ(x) + c. 2. (Positive homogeneity) If c 0, then ρ(cx) = cρ(x). 3. (Monotonicity) For X Y, then ρ(x) ρ(y ). 4. (Subadditivity) ρ(x + Y ) ρ(x) + ρ(y ). The interpretations of these axioms have been well documented in the literature (see, e.g., [2] for details). Translation invariance indicates that the addition of a certain amount of losses increases the risk measure by the same amount. Positive homogeneity indicates that

14 2. Univariate Risk Measures 5 the risk measure is proportional to the size of risk. For example, if we measure the losses in dierent currency units, the results will follow the same scale. Monotonicity indicates that the portfolio which always has higher losses should have a higher risk measure. Subadditivity indicates that the risk can be diversied by combining portfolios. Risk measures satisfying all these axioms are more reasonable and acceptable. Moreover, since risk measures are estimated from historical data in practice, the robustness of their estimators is a relevant question. Robust statistics can be dened as statistics that are not unduly aected by outliers. To clarify this denition, consider a sample generated from a log-normal distribution which is heavy-tailed. The occurrence of huge losses will signicantly shift up the sample average. Using sample average as the estimator of mean would lead to large mean squared error (MSE). Therefore, it is not a robust statistic. Now, consider a continuous random variable X with cdf F D where D is the convex set of cdfs. Notice that a risk measure is distribution-based if ρ(x 1 ) = ρ(x 2 ) when F X1 = F X2. Hence, we use ρ(f ) ρ(x) to represent the distribution-based risk measures. To quantify the sensitivity of a risk measure to the change in the distribution, the sensitivity function is used in this thesis. This method is used in Cont et al. (2010) and can be explained as the one-sided directional derivative of the eective risk measure at F in the direction δ z. Denition Consider ρ, a distribution-based risk measure of a continuous random variable X with distribution function F D. For ε [0, 1), set F ε = εδ z + (1 ε)f such that F ε D. δ z D is the probability measure which gives mass 1 to {z}. The distribution F ε is dierentiable at any x z and has a jump point at the point x = z. The sensitivity function is dened by ρ(f ε ) ρ(f ) S(z) = S(z; F ) lim, ε 0 + ε for any z R such that the limit exist.

15 2. Univariate Risk Measures 6 Furthermore, for a robust statistic, the value of sensitivity function will not go to innity when z becomes arbitrarily large. In other word, the bounded sensitivity function makes sure that the risk measure will not blow up when a small change happens. 2.2 Denitions and Properties Value-at-Risk (VaR) introduced in the G-30 report published in July 1993 provides the lower bound which covers the 100α% of the possible losses. In other words, it gives us the probable maximum loss under a given signicance level. Denition For a random variable X with cumulative distribution function (cdf) F X, the Value-at-Risk at signicance level α (0, 1) is given by V ar α (X) = inf {x R : F X (x) α}. Note that for a continuous random variable X with strictly increasing cdf, V ar α (X) = F 1 X (α), is also called the α-quantile, where F 1 X is the inverse function of cdf. It is well known that VaR is not a coherent risk measure since it is not subadditive. Moreover, VaR fails to give any information beyond the level α. Therefore, risk measures which quantify the magnitude of loss of the worst 100(1 α)% cases are developed, such as the Tail Conditional Expectation (TCE), the Worst Conditional Expectation (WCE), the Conditional Value-at-Risk (CVaR) and the Expected Shortfall (ES). Denition For a random variable X on a probability space (Ω, F, P) with cdf F X and V ar α (X) dened in Denition 2.2.1, the Tail Conditional Expectation at signicance level

16 2. Univariate Risk Measures 7 α [0, 1] is given by T CE α (X) = E [X X V ar α (X)]. If X is a continuous random variable, T CE α (X) could be wrote into the following form, T CE α (X) = T V ar α (X) = 1 1 α 1 α V ar u (X)du. TCE proposed by Artzner et al. (1999) measures the average loss given that the loss is no less than the 100α% of all possible cases. However, like VaR, it is not coherent since the subadditivity can only be satised when the random variable is continuous. To gure out this problem, the Worst Conditional Expectation is proposed in the same article. Denition For a random variable X on a probability space (Ω, F, P) with cdf F X, the Worst Conditional Expectation at signicance level α [0, 1] is dened by W CE α (X) = sup {E[X A] P (A) 1 α, A F}. WCE is the maximum expected loss of at least 100(1 α)% cases. Hence, it depends not only on the distribution of X but also on the structure of the underlying probability space. Thus, it seems hopeless to compute the value of it in practice, when the probability space is innite. Therefore, to nd a coherent risk measure which is computable, Rockafellar and Uryasev (2000) propose the CVaR. Denition For a random variable X on a probability space (Ω, F, P) with cdf F X, the Conditional Value-at-Risk at signicance level α [0, 1) is dened by } E[X a]+ CV ar α (X) = inf {a + : a R. 1 α

17 2. Univariate Risk Measures 8 Note the number a is selected to minimize the value of CVaR. CVaR is dened without VaR and it is subadditive, which makes this measure used to optimize portfolios. In addition, ES is proposed by Acerbi and Tasche (2002), making some modications on the denition of TCE such that ES is subadditive when the distribution is discrete. Denition For a random variable X on a probability space (Ω, F, P) with cdf F X and V ar α (X) dened as in Denition 2.2.1, the Expected Shortfall at signicance level α [0, 1) is dened by ES α (X) = 1 { E[X1{X V arα(x)}] + V ar α (X)[1 α P (X V ar α (X))] }. 1 α Acerbi and Tasche (2002) show that the ES is a coherent risk measure (see, e.g., [1] for details) and discuss the relationships between the TCE, WCE, CVaR and ES, considering that all of these risk measures are used to evaluate the same thing i.e. the expected losses of the worst 100(1 α)% cases. The dierences in denitions lead to the dierences in the numerical results (see, section 2.4 for details) and properties. Proposition For a discrete random variable X, W CE α (X), T CE α (X), CV ar α (X) and ES α (X) have the following relationships, T CE α (X) W CE α (X) ES α (X) CV ar α (X). Proposition For a continuous random variable X, W CE α (X), T CE α (X), CV ar α (X) and ES α (X) have the following relationships, T CE α (X) = W CE α (X) = ES α (X) = CV ar α (X). Note that Proposition and Proposition are proved by Acerbi and Tasche (2002) (see, e.g., [1] for details).

18 2. Univariate Risk Measures 9 In next section, we will discuss the robustness of the risk measures presented so far. 2.3 Robustness Proposition For a continuous random variable X with cdf F, the sensitivity function of V ar α (X) is given by S(z) = 1 α, f(v ar α(f )) α, f(v ar α(f )) z < V ar α(x) z > V ar α(x) 0, z = V ar α (X) which is a bounded function. Thus, VaR is a robust risk measure. Note, V ar α (F ) V ar α (X) since VaR is a distribution-based risk measure. The way to prove the Proposition is presented by Cont et al. (2010). The basic idea is to measure the eect of the small change at a point on the risk measures using the sensitivity function. A bounded sensitivity function can be obtained for the robust risk measure. Proof. Fix z R and set F ε = εδ z + (1 ε)f such that F 0 F, where F D and the direction of change δ z D. The distribution F ε is dierentiable at any x z with F ε(x) = (1 ε)f(x) > 0 and has a jump (of size ε [0, 1)) at the point x = z. Hence, F ( 1 α 1 ε), α < (1 ε)f (z) V ar α (F ε ) = Fε 1 (α) = F ( ) 1 α ε 1 ε, α (1 ε)f (z) + ε z, otherwise. Thus, the sensitivity function of V ar α (X) can be evaluated by

19 2. Univariate Risk Measures 10 [ ] V ar α (F ε ) V ar α (F ) d S(z) = lim = ε 0 + ε dε V ar α(f ε ) ε=0 1 α, z < V ar f(v ar α(f )) α(x) = α, z > V ar f(v ar α(f )) α(x) 0, z = V ar α (X). Note that the sensitivity function of V ar α (X) is bounded by two horizontal lines and has a jump point at z = V ar α (X). Consider a dataset with V ar α (X). Then, adding a value smaller than V ar α (X) would decrease V ar α (X) and vice versa. Proposition For a continuous random variable X, let ρ(x) = T CE α (X) = W CE α (X) = CV ar α (X) = ES α (X). Then, the sensitivity function of ρ(x) is given by V ar α (X) ρ(x), z < V ar α (X) S(z) = z αv ar α(x) ρ(x), z V ar 1 α α (X). Note, T CE α (X), W CE α (X), CV ar α (X) and ES α (X) have unbounded sensitivity functions, which means they are not robust. Proof. Let ρ(x) = T CE α (X) = W CE α (X) = ES α (X) = CV ar α (X), then the sensitivity function of ρ(x) is given by

20 2. Univariate Risk Measures 11 { 1 1 S(z) = lim ε α α = α = 1 1 α α 1 α lim ε 0 + } V ar u (F ε ) V ar u (F ) du ε { V aru (F ε ) V ar u (F ) ε [ ] d dε V ar u(f ε ) du ε=0 } du = α α 1 1 α [ d F ( )] 1 u ε du, F (z) < α dε 1 ε ε=0 { F (z) α [ d F ( 1 u dε 1 ε )]ε=0 du + 1 F (z) [ d F ( } 1 u ε dε 1 ε )]ε=0 du, F (z) α = V ar α (X) ρ(x), z < V ar α (X) z αv ar α(x) 1 α ρ(x), z V ar α (X). Note that the sensitivity function of ρ(x) is a linear function of z. When the jump happens on the right tail of the distribution, it goes to innity. Therefore, T CE α (X), W CE α (X), ES α (X) and CV ar α (X) are not robust. Since risk measures providing information on tails of distribution are not robust, Cont et al. (2010) present the Range Value-at-Risk. Denition For a continuous random variable X with cdf F X, the univariate Range

21 2. Univariate Risk Measures 12 Value-at-Risk at level range [α 1, α 2 ] [0, 1] is dened by RV ar α1,α 2 (X) = E [X V ar α1 (X) X V ar α2 (X)] = 1 α 2 α 1 α2 α 1 V ar u (X)du. Note that RV ar α1,α 2 (X) is not a coherent risk measure since it does not satisfy the subadditivity. For α 2 < 1, RV ar α1,α 2 (X) is always well bounded, RV ar α1,α 2 (X) = 1 α 2 α 1 1 α 2 α 1 α2 α 1 α2 α 1 V ar u (X)du V ar α2 (X)du = V ar α2 (X). Moreover, the robustness of RVaR can be proved in the same way as the one we used before. Proposition For a continuous random variable X with cdf F, RVaR is not sensitive to the small change at any point x. In other word, its sensitivity function is bounded and given by S(z) = (1 α 1 )V ar α1 (X) (1 α 2 )V ar α2 (X) α 2 α 1 RV ar α1,α 2 (X), z < V ar α1 (X) z α 1 V ar α1 (X) (1 α 2 )V ar α2 (X) α 2 α 1 RV ar α1,α 2 (X), V ar α1 (X) z V ar α2 (X) α 2 V ar α2 (X) α 1 V ar α1 (X) α 2 α 1 RV ar α1,α 2 (X), z > V ar α2 (X) which means RVaR is a robust risk measure. Proof. The sensitivity function of RV ar α1,α 2 (X) can be obtained as follows.

22 2. Univariate Risk Measures 13 { 1 α2 S(z) = lim ε 0 + α 2 α 1 = 1 α2 α 2 α 1 α 1 lim ε 0 + V ar u (F ε ) V ar u (F ) α 1 ε { V aru (F ε ) V ar u (F ) = 1 α2 [ ] d α 2 α 1 α 1 dε V ar u(f ε ) du ε=0 ε } du } du = = 1 α 2 α 1 α2 α 1 [ d F ( )] 1 u ε du, F (z) < α dε 1 ε ε=0 1 { 1 F (z) [ d F ( 1 u α 2 α 1 α 1 dε 1 ε )]ε=0 du + α 2 F (z) [ d F ( } 1 u ε dε 1 ε )]ε=0 du, α 1 F (z) α 2 1 α2 [ d F ( )] 1 u du, F (z) > α α 2 α 1 α 1 dε 1 ε ε=0 2 (1 α 1 )V ar α1 (X) (1 α 2 )V ar α2 (X) α 2 α 1 RV ar α1,α 2 (X), z < V ar α1 (X) z α 1 V ar α1 (X) (1 α 2 )V ar α2 (X) α 2 α 1 RV ar α1,α 2 (X), V ar α1 (X) z V ar α2 (X) α 2 V ar α2 (X) α 1 V ar α1 (X) α 2 α 1 RV ar α1,α 2 (X), z > V ar α2 (X). This result shows that the sensitivity function of RV ar α1,α 2 (X) is linear in z over the interval [V ar α1 (X), V ar α2 (X)] and constant over other intervals. It is bounded, which means RVaR is a robust risk measure.

23 2. Univariate Risk Measures Numerical Examples In this section, the examples for discrete and continuous random variables are presented, respectively. The results illustrate relationships between the risk measures we discussed in the previous sections. Example Let X, which is the loss of a policy, be a discrete random variable with the probability distribution as list in the following Table. Tab. 2.1: Probability distribution of the discrete variable X Ω ω 1 ω 2 ω 3 ω 4 ω 5 X P (X = x i ) 20% 30% 10% 30% 10% Let α = 80%, then V ar 0.8 (X) = 3 2 = 1.5 can make sure that 80% of losses can be covered. Then, T CE 0.8 (X) = E[X X V ar 0.8 ] = can be evaluated based on the value of V ar 0.8 (X) = = W CE 0.8 (X) can be calculated by maximizing the expectation of the events set which happens with probability larger than 0.2. Therefore, the events set A = {ω 3, ω 5 } will be selected after comparing the results of dierent combinations. Then, W CE 0.8 (X) = = 2.

24 CV ar α (X) = inf then {a + E[X a]+ 1 α 2. Univariate Risk Measures 15 } : a R, in this case we choose a = V ar 0.8 (X) = 3, 2 CV ar 0.8 (X) = (3 3 2 ) is the minima for any possible value of a. = 9 4 = 2.25 { ES α (X) = 1 E[X1{X V 1 α arα(x)}] + V ar α (X)[1 α P (X V ar α (X))] }, therefore { 1 3 ES 0.8 (X) = } ( ) = The results show that T CE α (X) W CE α (X) ES α (X) CV ar α (X), which coincides with the Proposition Example Consider a continuous random variable X with cdf F (x). According to the Denition 2.2.2, T CE α (X) = 1 1 α 1 α V ar u (X)du Furthermore, let A = {X V ar α (X)} such that W CE α (X) can be maximized. Then, Moreover, when a = V ar α (X), W CE α (X) = sup {E[X A] P (A) 1 α, A F} CV ar α (X) = a + = E[X X V ar α (X)] = T CE α (X). E[X a]+ 1 α = V ar α (X) + E[X V ar α(x)] + 1 α V ar = V ar α (X) + [x V ar α(x) α(x)]df (x) 1 α = 1 1 V ar u (X)du = T CE α (X). 1 α α

25 2. Univariate Risk Measures 16 Finally, ES α (X) = 1 { E[X1{X V arα(x)}] + V ar α (X) [1 α P (X V ar α (X))] } 1 α { } = 1 1 α = 1 1 α 1 α xdf (x) + V ar α (X) [1 α (1 α)] V ar α(x) V ar u (X)du = T CE α (X). In conclusion, for a continuous random variable, T CE α (X) = W CE α (X) = ES α (X) = CV ar α (X). Let Y N(µ, σ 2 ) such that X = e Y is a log-normal distributed random variable. Then the cumulative distribution function of X is given by ( ) ln x µ F (x) = Φ V ar α (X) = e σφ 1 (α)+µ, σ where Φ refers to the standard normal distribution. T CE α (X) = W CE α (X) = ES α (X) = CV ar α (X) = 1 1 α = 1 1 α = eµ+ 1 2 σ2 1 α 1 α 1 α V ar u (X)du e σφ 1 (u)+µ du [ Φ ( σ Φ 1 (α) )]. The result is obtained by using a change of variable of the form a = Φ 1 (u). In the next example, simulation results are obtained to show the robustness of risk measures. Example To estimate risk measures mentioned in the Example 2.4.2, we generate two data sets with sample size n = 100 from a log-normal distribution with the same mean E(X) = 100 and coecients of variation CV (X) = V ar(x)/e(x) taking values 1 and 2, respectively. Then the parameters (µ, σ) for each groups are (4.2586, ) and (3.8005, ).

26 2. Univariate Risk Measures 17 Let X (1), X (2),..., X ([nα]), X ([nα]+1),..., X (n) denote the data in the sample arranged in increasing order. Then, for n large enough, the empirical estimator of V ar α (X) will be the statistics X ([nα]+1). The empirical estimator of ρ(x) = T CE α (X) = W CE α (X) = ES α (X) = CV ar α (X) is given by ρ(x) = X ([nα]+1) X (n). n [nα] Moreover, RV ar α1,α 2 (X) can be calculated as RV ar α1,α 2 (X) = 1 α 2 α 1 V arα2 (X) V ar α1 (X) 1 t (ln t µ) 2 2πσ 2 te 2σ 2 dt = eµ+ 1 2 σ2 α 2 α 1 [ Φ ( σ Φ 1 (α 1 ) ) Φ ( σ Φ 1 (α 2 ) )], where Φ refers to the standard normal cumulative distribution function. The empirical estimator of RV ar α1,α 2 (X) is dened by RV ar α1,α 2 (X) = X ([nα 1 ]+1) X ([nα2 ]+1). [nα 2 ] [nα 1 ] + 1 The simulation results are presented in the Table 2.2, with α 2 = Tab. 2.2: RV ar α1,α 2 (X) and ρ(x) RV ar α1,α 2 (X) RV ar α1,α 2 (X) ρ(x) ρ(x) CV α 1 = α 1 = The dependence of these risk measures on dierent signicance levels is observed, risk measures are increasing with α 1. Furthermore, Table 2.2 shows that RV ar α1,α 2 (X) is more robust than T CE α (X), W CE α (X), ES α (X) and CV ar α (X) since large variance has a smaller impact on this statistic.

27 3. MULTIVARIATE RISK MEASURES We have reviewed several popular univariate risk measures and examined their properties and robustness. Before discussing properties and robustness of some established multivariate risk measures, we review copulas which are frequently used to model dependent random variables. 3.1 Copulas Denition Let X = {X 1, X 2,..., X d } be a random vector with cdf F. Set U i = F i (x i ) U[0, 1], i = 1,..., d. Then, the copula C : [0, 1] d [0, 1] of F is given by C(u 1,..., u d ) = P (U 1 u 1,..., U d u d ), u i [0, 1], i = 1,..., n. This denition is proposed by Nelsen (1999). Moreover, for a random vector (U 1,..., U d ) with cdf C, P (a 1 U 1 b 1,..., a d U d b d ) is non-negative. Furthermore, Sklar's theorem (see Sklar (1959)) states that any multivariate cumulative distibution function can be expressed in terms of its marginal distributions and a copula. Theorem (Sklar's Theorem) Let F be a d-dimensional distribution function with marginal distributions F 1,..., F d, then there exists a copula C such that F (x 1,..., x d ) = C(F 1 (x 1 ),..., F d (x d )).

28 3. Multivariate Risk Measures 19 Note, if F 1,..., F d are continuous, then C is unique. In this thesis we mainly discuss bivariate risk measures. Consider X 1 and X 2 with marginal cdf's F 1 and F 2, respectively. Then, we have F (x 1, x 2 ) = C(F 1 (x 1 ), F 2 (x 2 )). For what follows, the survival copula is dened by C(u 1, u 2 ) = P (U 1 > u 1, U 2 > u 2 ) = 1 P (U 1 u 1 ) P (U 2 u 2 ) + P (U 1 u 1, U 2 u 2 ) = 1 u 1 u 2 + C(u 1, u 2 ), (u 1, u 2 ) [0, 1] 2. Finally, we will evaluate the empirical estimators of bivariate risk measures in Chapter 4 with the empirical copula C n. For a random sample (X l1, X l2 ), l = 1,..., n, C n is dened by C n (u 1, u 2 ) = 1 n n 1 {Fn,1 (X l1 ) u 1,F n,2 (X l2 ) u 2 }, l=1 where F n,i represents the empirical cdf of X i = {X 1i,..., X ni }, i = 1, 2. Well known copulas and families of copulas are dened as follow. Fréchet Family First, we consider a pair of independent random variables. If X 1 is independent of X 2, then F (x 1, x 2 ) = F 1 (x 1 )F 2 (x 2 ), (x 1, x 2 ) R 2, or Π(u 1, u 2 ) = u 1 u 2, (u 1, u 2 ) [0, 1] 2.

29 3. Multivariate Risk Measures 20 Also, one has the upper and lower Fréchet-Hoeding bounds proposed in Fréchet (1951) and Hoeding (1940), respectively M(u 1, u 2 ) = min(u 1, u 2 ), W (u 1, u 2 ) = max(0, u 1 + u 2 1), for (u 1, u 2 ) [0, 1] 2. Theorem For an arbitrary bivariate copula C : [0, 1] 2 [0, 1] and any (u 1, u 2 ) [0, 1] 2, W (u 1, u 2 ) C(u 1, u 2 ) M(u 1, u 2 ). Note M(u 1, u 2 ) indicates that random variables are comonotonic (perfect positive dependence) whereas W (u 1, u 2 ) indicates that random variables are countermonotonic (perfect negative dependence). Archimedean Family The Archimedean Family introduced by Nelsen (2006) is of the form C(u 1, u 2 ; θ) = ψ 1 (ψ(u 1 ; θ) + ψ(u 2 ; θ); θ), (u 1, u 2 ) [0, 1] 2, θ Θ, where ψ is the generator function and satises following properties, (1) ψ(0) = and ψ(1) = 0, (2) ψ (t) < 0, (3) ψ (t) > 0. Moreover, the parameter θ dictates the dependence between the random variables. Below are presented some Archimedean copulas, most of which will be used throughout this thesis.

30 3. Multivariate Risk Measures 21 (1) Gumbel Copula For θ [1, ), the Gumbel copula is given by C(u 1, u 2 ; θ) = e {[ ln(u 1)] θ +[ ln(u 2 )] θ } 1 θ, with the generator ψ(t; θ) = ( ln (t)) θ and the inverse generator ψ 1 (t; θ) = e t 1 θ. Specially, for θ = 1, we have is the independent copula. (2) Frank Copula C(u 1, u 2 ) = e {[ ln(u 1)]+[ ln(u 2 )]} = u 1 u 2 = Π(u 1, u 2 ), For θ (, 0) (0, ), the Frank copula is dened by C(u 1, u 2 ; θ) = 1 θ ln [1 + (e θu 1 1)(e θu 2 1) ], e θ 1 with the generator ψ(t; θ) = ln( e θt 1 e θ 1 ) and the inverse generator ψ 1 (t; θ) = 1 θ ln [1 + e t (e θ 1)]. (2) Clayton Copula For θ [ 1, 0) (0, ), we have the Clayton copula dened by C(u 1, u 2 ; θ) = (u θ 1 + u θ 2 1) 1 θ, with the generator ψ(t; θ) = 1 θ (t θ 1) and the inverse generator ψ 1 (t; θ) = (1 + θt) 1 θ. Note, Π(u 1, u 2 ) is obtained if θ = 0. Moreover, W (u 1, u 2 ) and M(u 1, u 2 ) are attained by setting θ = 1 and θ, respectively. Figure 3.1 and Figure 3.2 show the dependence structure of variables with Gumbel, Frank and Clayton copulas, respectively. Frank copula is relatively symmetric, whereas Gumbel and Clayton copulas have stronger dependence in the right and left tails, respectively.

31 3. Multivariate Risk Measures 22 Fig. 3.1: Gumbel and Frank Copulas with dependent parameters θ = 2 and θ = 5

32 3. Multivariate Risk Measures 23 Fig. 3.2: Clayton Copula with dependent parameters θ = Bivariate VaR and TVaR First, let us introduce the bivariate orthant based VaR proposed by Embrechts and Puccetti (2006). Denition Let X = (X 1, X 2 ) be a random vector with joint cdf FX and joint suvivial function (sf) FX. At signicance level α [0, 1], the bivariate lower orthant VaR is dened by V ar α (X) = {x R 2 : FX(x) α}, (eq. 3.1) and bivariate upper orthant VaR is dened by V ar α (X) = {x R 2 : FX(x) 1 α}. (eq. 3.2)

33 3. Multivariate Risk Measures 24 Note, in a bivariate setting, the relationship between the cumulative distribution function and the survival function, namely F (x) = 1 F (x), does not hold. For this reason, bivariate VaR need to be dened as the lower and upper orthant VaR separately using either the cdf or sf. In addition, denotes the boundary of the set. To study the behavior of bivariate orthant based VaR and get the bivariate extension of TVaR, Cossette et al. (2013, 2015) rewrite Eq. 3.1 and Eq Denition Let X = (X 1, X 2 ) be a random vector with joint cdf FX and joint sf FX. At signicance level α [0, 1], the bivariate lower orthant VaR is dened by V ar α (X) = {(x 1, V ar α,x1 (X)), x 1 V ar α (X 1 )}, or V ar α (X) = {(V ar α,x2 (X), x 2 ), x 2 V ar α (X 2 )}, and the bivariate upper orthant VaR is dened by V ar α (X) = {(x 1, V ar α,x1 (X)), x 1 V ar α (X 1 )}, or For i, j = 1, 2 (i j), V ar α (X) = {(V ar α,x2 (X), x 2 ), x 2 V ar α (X 2 )}. V ar α,xi (X) = V ar α (X F Xi (x i j X ) i x i ), x i V ar α (X i ) and V ar α,xi (X) = V ar α F Xi (x i ) (X j X i x i ), x i V ar α (X i ). 1 F Xi (x i ) Example Consider the random vector (X 1, X 2 ) with joint cdf is dened with a Gumbel copula with dependent parameter θ = 1.5 and marginals X 1 Weibull (2, 50) and X 2

34 3. Multivariate Risk Measures 25 Weibull (2, 150). Then, we get V ar 0.95,x1 (X) on Figure 3.3 and V ar 0.99,x1 (X) on Figure 3.4. Let u x1 (respectively u x2 ) and l x1 (respectively l x2 ) denote the essential upper and lower support of X 1 (respectively X 2 ). Note, V ar 0.95,x1 (X) [V ar 0.95 (X 1 ), u x1 ] [V ar 0.95 (X 2 ), u x2 ], and V ar 0.99,x1 (X) [l x1, V ar 0.99 (X 1 )] [l x2, V ar 0.99 (X 2 )]. Also, Figure 3.3 and Figure 3.4 show the convexity of the curve which has been studied in Cossette et al. (2013). Fig. 3.3: Lower orthant VaR at level 0.95

35 3. Multivariate Risk Measures 26 Fig. 3.4: Upper orthant VaR at level 0.99

36 3. Multivariate Risk Measures 27 Proposition Let X = (X 1, X 2 ) be a continuous random vector. Let φ 1 and φ 2 be real functions dened on the supports of X 1 and X 2, respectively. 1. (Translation invariance) For all c = (c 1, c 2 ) R 2 and i, j = 1, 2, i j, then V ar α,xj +c j (X + c) = V ar α,xj (X) + c i, V ar α,xj +c j (X + c) = V ar α,xj (X) + c i. 2. (Positive homogeneity) For all c = (c 1, c 2 ) R 2 + and i, j = 1, 2, i j, then V ar α,cj x j (cx) = c i V ar α,xj (X), V ar α,cj x j (cx) = c i V ar α,xj (X). 3. (Negative transformations) For all c = (c 1, c 2 ) R 2 and i, j = 1, 2, i j, then V ar α,cj x j (cx) = c i V ar 1 α,xj (X), V ar α,cj x j (cx) = c i V ar 1 α,xj (X). In general, (1) For increasing functions φ 1 and φ 2, i, j = 1, 2, i j, V ar α,φj (x j )(φ(x)) = φ i (V ar α,xj (X)), V ar α,φj (x j )(φ(x)) = φ i (V ar α,xj (X)). (2) For decreasing functions φ 1 and φ 2, i, j = 1, 2, i j, V ar α,φj (x j )(φ(x)) = φ i (V ar 1 α,xj (X)), V ar α,φj (x j )(φ(x)) = φ i (V ar 1 α,xj (X)).

37 3. Multivariate Risk Measures (Monotonicity) Let X = (X 1, X 2 ) and X = (X 1, X 2) be two pairs of risks with joint cdf's FX and FX respectively. X is said to be more concordant than X, denoted X co X, if FX(x) FX (x) for all x R2. Then, V ar α (X ) V ar α (X), V ar α (X) V ar α (X ). Proof. The proof of Proposition is presented in Cossette et al. (2013). Denition Let X = (X 1, X 2 ) be a random vector with joint cdf FX and joint suvivial function (sf) FX. At signicance level α [0, 1], the bivariate lower orthant TVaR is given by T V ar α (X) = {( x i, T V ar α,xi (X) ), x i V ar α (X i ), i = 1, 2 }, and bivariate upper orthant TVaR is given by T V ar α (X) = { ( x i, T V ar α,xi (X) ), x i V ar α (X i ), i = 1, 2}. Note that for i, j = 1, 2 (i j), and T V ar α,xi (X) = E[X j X j > V ar α,xi (X), X i x i ] = 1 F Xi (x i ) α FXi (x i ) α V ar u,xi (X)du, T V ar α,xi (X) = E[X j X j > V ar α,xi (X), X i x i ] = 1 1 α 1 α V ar v,xi (X)dv. Example Consider the same random vector dened in Example with signicance level α = T V ar 0.99,x1 (X) (respectively T V ar 0.99,x1 (X)) is obtained based on the Denition For comparison, we also plot V ar 0.99,x1 (X) (respectively V ar 0.99,x1 (X)) on

38 3. Multivariate Risk Measures 29 Figure 3.5 and Figure 3.6. Note that T V ar 0.99,x1 (X) (respectively T V ar 0.99,x1 (X)) goes to T V ar 0.99 (X 2 ) when X 1 approaches to u x1 (respectively l x1 ). These results are obtained using the numerical integration tools in Matlab. Fig. 3.5: Lower orthant VaR and TVaR at level 0.99

39 3. Multivariate Risk Measures 30 Fig. 3.6: Upper orthant VaR and TVaR at level 0.99

40 3. Multivariate Risk Measures 31 Proposition Let X = (X 1, X 2 ) be a continuous random vector. 1. (Translation invariance) For all c = (c 1, c 2 ) R 2 and i, j = 1, 2, i j, then T V ar α,xj +c j (X + c) = T V ar α,xj (X) + c i, T V ar α,xj +c j (X + c) = T V ar α,xj (X) + c i. 2. (Positive homogeneity) For all c = (c 1, c 2 ) R 2 + and i, j = 1, 2, i j, then T V ar α,cj x j (cx) = c i T V ar α,xj (X), T V ar α,cj x j (cx) = c i T V ar α,xj (X). 3. (Monotonicity) Let X = (X 1, X 2 ) and X = (X 1, X 2) be two pairs of risks with joint cdf FX and FX respectively. If X co X, then T V ar α (X ) T V ar α (X), T V ar α (X) T V ar α (X ). Proof. The proof of Proposition can be found in Cossette et al. (2015). Jouini et al. (2004), Bentahar (2006) and Tahar and Lépinette (2012, 2014) extend the multivariate WCE and TCE. Furthermore, Cousin and Di Bernardino (2013, 2014, 2015) propose a series of multivariate risk measures developed from a dierent aspect. Denition Let X = (X 1, X 2 ) denote a bivariate random vector on the probability space (Ω, F, P). For α [0, 1], the vector-valued Worst Conditional Expectation at level α is dened by W CE α (X) = { x R 2 : E[x X A] 0, A F such that P (A) 1 α }. Denition Let X = (X 1, X 2 ) denote a bivariate random vector on the probability space (Ω, F, P). For α [0, 1], the vector-valued Tail Conditional Expectation at level α is

41 3. Multivariate Risk Measures 32 dened by T CE α (X) = {x R 2 : E[x X X A] 0, A Q α (X)}, where Q α (X) = {A R 2 : P (X A) 1 α}. Note that the vector-valued W CE α (X) and T CE α (X) are the natural extension of their real-valued versions. Therefore, they share similar properties. Denition Let X = (X 1, X 2 ) denote a bivariate random vector on the probability space (Ω, F, P). For α [0, 1], the Generalized Worst Conditional Expectation at level α is dened by GW CE α (X) = X W CE α ( X), where random variables X on ( Ω, F, P) having the same distribution as X. Proposition Let X = (X 1, X 2 ) denote a bivariate random vector on the probability space (Ω, F, P) having a continuous probability density. Then W CE α (X) = T CE α (X) = GW CE α (X). The proof of this Proposition is presented in [4]. Denition Let X = (X 1, X 2 ) denote a bivariate random vector on the probability space (Ω, F, P) with joint cdf FX and joint sf FX. The multivariate lower-orthant V ar at level α [0, 1] is dened by V ar α(x) = E[X FX(X) = α], and the multivariate upper-orthant V ar is dened by V ar α(x) = E[X FX(X) = 1 α].

42 3. Multivariate Risk Measures 33 Denition Let X = (X 1, X 2 ) denote a bivariate random vector on the probability space (Ω, F, P) with joint cdf FX and joint sf FX. The multivariate lower-orthant Conditional Tail Expectation at level α [0, 1] is dened by CT E α (X) = E[X FX(X) α], and the multivariate upper-orthant Conditional Tail Expectation is dened by CT E α (X) = E[X FX(X) 1 α]. Denition Let X = (X 1, X 2 ) denote a bivariate random vector on the probability space (Ω, F, P) with joint cdf FX and joint sf FX. multivariate lower-orthant CoVaR is dened by For a signicant level α [0, 1], the CoV ar α,ω (X) = V ar ω (X FX(X) = α), and the multivariate upper-orthant CoVaR is dened by CoV ar α,ω (X) = V ar ω (X FX(X) = 1 α), where ω = {ω 1, ω 2 } is the signicance level vector of marginal risk with ω i [0, 1]. Note that similar properties mentioned in the Proposition also hold for the multivariate upper and lower orthant VaR, CTE and CoVaR (see, e.g., [9], [10], [11] for details). 3.3 Bivariate Lower Orthant RVaR In the following part of this chapter, we will propose a new multivariate risk measure called bivariate lower and upper orthant RVaR based on the results in Cossette et al. (2013, 2015). Its properties, such as translation invariance, positive homogeneity and monotonicity, will be discussed.

43 3. Multivariate Risk Measures 34 Denition Consider a continuous random vector X = (X 1, X 2 ) on the probability space (Ω, F, P). Bivariate lower orthant RVaR at level range [α 1, α 2 ] [0, 1] is given by RV ar α1,α 2 (X) = ((x 1, RV ar α1,α 2,x 1 (X)), (RV ar α1,α 2,x 2 (X), x 2 )), where RV ar α1,α 2,x i (X) = E[X j V ar α1,x i (X) X j V ar α2 (X j ), X i x i ], for V ar α1,v ar α2 (X j )(X) x i V ar α2 (X i ), i, j = 1, 2(i j). This denition is based on the univariate RVaR and bivariate lower orthant VaR. As we can see, RV ar α1,α 2 (X) is the expectation of a random variable X given that it belongs to the interval [V ar α1 (X), V ar α2 (X)]. Hence, we start from bounding the X in the rectangle area [V ar α1 (X i ), V ar α2 (X i )] [V ar α1 (X j ), V ar α2 (X j )]. Considering the eect of the random variable X i on X j, V ar α1,x i (X) and V ar α2,x i (X) are applied. However, only V ar α1,x i (X) could lie in the above area, which has been shown in Example Therefore, we require V ar α1,x i (X) X j V ar α2 (X j ) to make sure that the lower bound at level α 1 can be achieved and the upper bound at level α 2 will not be exceeded. The next result shows that RV ar α1,α 2,x i (X) has a similar expression as T V ar α,xi (X) which can be expressed as the integration of V ar α,xi (X). Proposition For a continuous random vector X = (X 1, X 2 ) with joint cdf F (x 1, x 2 ) and marginal cdf's F X1 (x 1 ) and F X2 (x 2 ), RV ar α1,α 2,x i (X) can be restated as RV ar α1,α 2,x i (X) = 1 F (x i, V ar α2 (X j )) α 1 F (xi,v ar α2 (X j )) α 1 V ar u,xi (X)du,

44 3. Multivariate Risk Measures 35 for V ar α1,v ar α2 (X j )(X) x i V ar α2 (X i ), i, j = 1, 2(i j). Proof. RV ar α1,α 2,x i (X) = E[X j V ar α1,x i (X) X j V ar α2 (X j ), X i x i ] V arα2 (X j ) x j df Xi (x j ) = V ar α1,x (X) i F Xi (x i ) = F (x i, V ar α2 (X j )) α 1 F (x i,v ar α2 (X j )) F Xi (x i α 1 ) F Xi (x i ) V arα2 (X j ) V ar α1,x i (X) x j df Xi (x j ). Note, we have V ar α,xi (X) = V ar α (X F Xi (x i j X ) i x i ). Then, using u = F Xi (x j ) = P (X j x j X i x i ), F Xi (x i ) RV ar α1,α 2,x i (X) = F (x i, V ar α2 (X j )) α 1 1 = F (x i, V ar α2 (X j )) α 1 1 = F (x i, V ar α2 (X j )) α 1 1 = F (x i, V ar α2 (X j )) α 1 FXi (V ar α2 (X j )) F 1 X i (u)du α 1 /F Xi (x i ) FXi (V ar α2 (X j )F Xi (x i )) F 1 X i ( α 1 FXi (V ar α2 (X j )F Xi (x i )) α 1 F (xi,v ar α2 (X j )) α 1 u F Xi (x i ) )du V ar u,xi (X)du V ar u,xi (X)du. Example Consider the random vector X = (X 1, X 2 ) dened in Example Let the condence level range be α 1 = 0.95 and α 2 = Then, we get bivariate lower orthant RVaR in Figure 3.7 and Figure 3.8. For comparison, we plot V ar 0.95,xi (X) on the same graph.

45 3. Multivariate Risk Measures 36 Fig. 3.7: Lower orthant VaR at level 0.95 and RVaR at level range [0.95, 0.99] for xed values of X 1

46 3. Multivariate Risk Measures 37 Fig. 3.8: Lower orthant VaR at level 0.95 and RVaR at level range [0.95, 0.99] for xed values of X 2 The shape of the lower orthant RVaR curve is shown in Figure 3.7 and Figure 3.8. One can observe that RV ar α1,α 2,x i (X) converges to the univariate RVaR when x i (i = 1, 2) approaches innity. Also, when x i gets close to V ar α1 (X i ), RV ar α1,α 2,x i (X) approaches V ar α2 (X j ).

47 3. Multivariate Risk Measures Bivariate Upper Orthant RVaR Denition Consider a continuous random vector X = (X 1, X 2 ) on the probability space (Ω, F, P). Bivariate upper orthant RVaR at level range [α 1, α 2 ] [0, 1] is given by RV ar α1,α 2 (X) = ((x 1, RV ar α1,α 2,x 1 (X)), (RV ar α1,α 2,x 2 (X), x 2 )), where RV ar α1,α 2,x i (X) = E[X j V ar α1 (X j ) X j V ar α2,x i (X), X i x i ], for V ar α1 (X i ) x i V ar α2,v ar α1 (X j )(X), i, j = 1, 2(i j). Similarly as for the lower orthant RVaR we can dene the bivariate upper orthant RVaR. We consider the impact of the random variable X i on X j and require that V ar α1 (X j ) X j V ar α2,x i (X) to make sure that the upper level α 2 can be achieved and the lower level α 1 will not be exceeded. Note, here we only modify the upper bound of the interval since only the curve V ar α2,x i (X) could lie in the bounded area [V ar α1 (X i ), V ar α2 (X i )] [V ar α1 (X j ), V ar α2 (X j )]. The following result provides the expression of RV ar α1,α 2,x i (X) in the form of the integration of V ar α,xi (X). Proposition Consider X = (X 1, X 2 ) with joint sf F (x1, x 2 ) and marginal sf's F X2 (x 2 ) and F X2 (x 2 ), respectively. Then, RV ar α1,α 2,x i (X) can be expressed by for RV ar α1,α 2,x i (X) = 1 α2 α 2 (1 F V ar v,xi (X)dv, (x i, V ar α1 (X j ))) 1 F (x i,v ar α1 (X j )) V ar α1 (X i ) x i V ar α2,v ar α1 (X j )(X), i, j = 1, 2(i j).

48 3. Multivariate Risk Measures 39 Proof. We have that Note, one has RV ar α1,α 2,x i (X) = E[X j V ar α1 (X j ) X j V ar α2,x i (X), X i x i ] = = V arα2,x i (X) V ar α1 (X j ) x j df Xi (x j ) F (x i,v ar α1 (X j )) 1 F Xi (x i ) 1 α 2 1 F Xi (x i ) 1 F Xi (x i ) V arα2,xi (X) x j df Xi (x j ). F (x i, V ar α1 (X j )) 1 + α 2 V ar α1 (X j ) V ar α,xi (X) = V ar α F Xi (x i ) (X j X i x i ). Then, using v = F Xi (x j ) = P (X j x j X i x i ), 1 F Xi (x i ) α2 FXi 1 F Xi (x i ) RV ar α1,α 2,x i (X) = F (x i, V ar α1 (X j )) 1 + α 2 = 1 F (x i, V ar α1 (X j )) 1 + α 2 α2 1 = α 2 (1 F (x i, V ar α1 (X j ))) (x i ) 1 F Xi (x i ) 1 F Xi (V arα 1 (X j )) 1 F Xi (x i ) 1 F Xi (V ar α1 (X j )) α2 1 F (x i,v ar α1 (X j )) F 1 X i (v)dv F 1 X i ( v F X i (x i ) 1 F Xi (x i ) )dv V ar v,xi (X)dv. Example According Proposition 3.4.1, one gets the curve of RV ar α1,α 2,x 1 (X) and RV ar α1,α 2,x 2 (X) in Figure 3.9 and Figure 3.10, respectively. For comparison, we plot the curve of the upper orthant V ar in the same graph.

49 3. Multivariate Risk Measures 40 Fig. 3.9: Upper orthant VaR at level 0.99 and RVaR at level range [0.95, 0.99] for xed values of X 1

50 3. Multivariate Risk Measures 41 Fig. 3.10: Upper orthant VaR at level 0.99 and RVaR at level range [0.95, 0.99] for xed values of X 2 Figure 3.9 and Figure 3.10 show the shape of the upper orthant RVaR curve. RV ar α1,α 2,x i (X) converges to RV ar 0.95,0.99 (X j ) when x i (i = 1, 2) gets close to lower support of X i. Also, when x i approaches V ar α2 (X i ), RV ar α1,α 2,x i (X) approaches V ar α1 (X j ). As a result, the curves of bivariate RVaR are bounded by the curves of univariate VaR, which is similar to the univariate RVaR.

51 3. Multivariate Risk Measures Properties of Bivariate RVaR Proposition Let X = (X 1, X 2 ) be a continuous random vector. 1. (Translation invariance) For all c = (c 1, c 2 ) R 2 and i, j = 1, 2, i j, then RV ar α1,α 2,x j +c j (X + c) = RV ar α1,α 2,x j (X) + c i, RV ar α1,α 2,x j +c j (X + c) = RV ar α1,α 2,x j (X) + c i. 2. (Positive homogeneity) For all c = (c 1, c 2 ) R 2 + and i, j = 1, 2, i j, then RV ar α1,α 2,c j x j (cx) = c i RV ar α1,α 2,x j (X), RV ar α1,α 2,c j x j (cx) = c i RV ar α1,α 2,x j (X). 3. (Monotonicity) Let X = (X 1, X 2 ) and X = (X 1, X 2) be two pairs of risks with joint cdf's FX and FX respectively. If X co X, then RV ar α1,α 2 (X ) RV ar α1,α 2 (X), RV ar α1,α 2 (X) RV ar α1,α 2 (X ). Proof. Here we need to use the properties of bivariate VaR in Proposition to proof the above results. For Translation invariance, RV ar α1,α 2,x j +c j (X + c) = = = F (xj,v ar α2 (X i )) α 1 V ar u,xj +c j (X + c)du F (x j, V ar α2 (X i )) α 1 F (xj,v ar α2 (X i )) α 1 V ar u,xj (X) + c i du F (x j, V ar α2 (X i )) α 1 F (xj,v ar α2 (X i )) α 1 V ar u,xj (X)du + c i F (x j, V ar α2 (X i )) α 1 =RV ar α1,α 2,x j (X) + c i.

52 3. Multivariate Risk Measures 43 For Positive Homogeneity, RV ar α1,α 2,c j x j (cx) = = = c i F (xj,v ar α2 (X i )) α 1 V ar u,cj x j (cx)du F (x j, V ar α2 (X i )) α 1 F (xj,v ar α2 (X i )) α 1 c i V ar u,xj (X)du F (x j, V ar α2 (X i )) α 1 F (xj,v ar α2 (X i )) α 1 V ar u,xj (X)du F (x j, V ar α2 (X i )) α 1 =c i RV ar α1,α 2,x j (X). Using the same way, we could get similar results for upper orthant RVaR. Moreover, if X co X, then V ar α,xi (X ) V ar α,xi (X). According to the Denition 3.3.1, we have RV ar α1,α 2,x i (X) = E[X j V ar α1,x i (X) X j V ar α2 (X j ), X i x i ], RV ar α1,α 2,x (X) = E[X i j V ar α1,x (X ) X i j V ar α2 (X j), X i x i], and X, X have same marginal cdfs. Hence, the expectation over the interval [α 1, α 2 ] will following the same pattern, say RV ar α1,α 2 (X ) RV ar α1,α 2 (X). Also, since F x (X) = 1 F (X 1 ) F (X 2 ) + F x (X), F x (X ) = 1 F (X 1 ) F (X 2 ) + F x (X ), then F x (X) F x (X ). Thus, V ar α,xi (X) V ar α,xi (X ). Again, for the same reason, we can conclude that RV ar α1,α 2 (X) RV ar α1,α 2 (X ).

53 3. Multivariate Risk Measures 44 Consider the random vectors X M, X W and X Π which denote the monotonic, inverse monotonic and independent vector, respectively. They have following relationship X W co X Π co X M, which means according to the Proposition 3.5.1, we have RV ar α1,α 2 (X M ) RV ar α1,α 2 (X Π ) RV ar α1,α 2 (X W ), and RV ar α1,α 2 (X W ) RV ar α1,α 2 (X Π ) RV ar α1,α 2 (X M ). Example Consider a bivariate random vector (X 1, X 2 ) which is either comonotonic, counter-conomotonic or independent. We obtain the lower orthant RVaR based on the Proposition For i, j = 1, 2 (i j), α2 F Xi (x i ) 1 RV ar α1,α 2,x i (X Π ) = V ar u α 2 F Xi (x i ) α (X F 1 α Xi (x i j)du, ) 1 1 FXi (x i ) RV ar α1,α 2,x i (X M ) = V ar u (X j )du, F Xi (x i ) α 1 α 1 and 1 FXi (x i )+α 2 1 RV ar α1,α 2,x i (X W ) = V ar u FXi (x F Xi (x i ) + α 2 1 α i )+1(X j )du. 1 If the random vector above is dened with exponential marginal cdfs, i.e. X i Exp(λ i ), then we get the following results. α 1