On Asymptotic Diversification Effects for Heavy-Tailed Risks
Dissertation zur Erlangung des Doktorgrades der Fakult¨atf¨urMathematik und Physik der Albert-Ludwigs-Universit¨atFreiburg im Breisgau
vorgelegt von Georg Mainik
Februar 2010 Dekan: Prof. Dr. Kay K¨onigsmann Referenten: Prof. Dr. Ludger R¨uschendorf Prof. Dr. Paul Embrechts
Datum der Promotion: 29. April 2010 Contents
1 Introduction1 1.1 Motivation...... 1 1.2 Overview of related theory...... 2 1.3 Central results and structure of the thesis...... 3 1.4 Acknowledgements...... 7
2 Multivariate regular variation9 2.1 Basic notation and model assumptions...... 9 2.2 Canonical forms of exponent and spectral measures...... 14 2.3 Dependence functions...... 18 2.4 Copulas...... 20 2.5 Spectral densities of Gumbel copulas...... 24 2.6 Spectral densities of elliptical distributions...... 27
3 Extreme risk index 35 3.1 Basic approach...... 35 3.2 Representations in terms of spectral measures...... 39 3.3 Portfolio optimization and diversification effects...... 42 3.4 Minimization of risk measures...... 49
4 Estimation 55 4.1 Basic approach...... 55 4.2 Main results...... 59 4.3 Empirical processes with functional index...... 63 4.4 Proofs of the main results...... 76 4.5 Examples and comments...... 78
5 Stochastic order relations 87 5.1 Introduction...... 87 5.2 Ordering of extreme portfolio losses...... 90 5.3 Ordering of spectral measures...... 94
i ii CONTENTS
5.4 Convex and supermodular orders...... 106 5.5 Examples...... 111
6 Modelling and simulation 117 6.1 Objectives and design...... 117 6.2 Estimation of the tail index...... 118 6.3 Models...... 121 6.4 Simulation results...... 126 6.5 Conclusions...... 147
A Auxiliary results 149 A.1 Regular variation...... 149 A.2 Empirical processes...... 149
Bibliography 155
Index 163 Chapter 1
Introduction
1.1 Motivation
Management of financial risks is one of the central challenges in the area of finance and insurance. In particular, any market agent has a natural inter- est in protection against extreme losses arising from market crashes, natural catastrophes, or political turbulences. Furthermore, financial crashes have proved to be a serious danger not only to the prosperity of share holders and financial corporations, but also to the general economic and political sta- bility of entire societies. Thus, additionally to the microeconomic interest, management of extreme financial risks is an issue of macroeconomic, politi- cal, and social importance. This issue is even greater in a rapidly changing world with advancing globalization of financial markets and acceleration of communication networks. In consequence of these developments, permanent validation and adjustment of models used in the risk management is essential to the market agents and the regulating authorities as well. Portfolio diversification is one of the basic microeconomic approaches to the reduction of non-systemic risk. The first probabilistic concept of diversi- fication was given by Markowitz(1952). Quantifying risk and dependence by variance and correlation, respectively, Markowitz gave arguments explaining why diversification decreases the portfolio risk. Designed for the multivariate Gaussian model assumption, this approach shaped the intuition of portfolio diversification in economics (cf. Sharpe, 1964; Lintner, 1965) and has proved to be appropriate to many application areas. However, the mean-variance portfolio theory cannot be applied to all diversification problems. First, this approach assumes existence of second moments, which is questionable in some applications. There are even real- world risk data sets suggesting infinite first moments (cf. Moscadelli, 2004;
1 2 Chapter 1. Introduction
Neˇslehov´aet al., 2006). Another limitation is the two-sided view of risk, as it automatically follows from measuring risk by variance. This may be misleading for asymmetric distributions and extreme risks related to rare events. Moreover, measuring dependence by correlation is also problematic in many applications (cf. Embrechts et al., 2002). Indeed, since moments depend on the entire distribution of a random variable in its value domain, they may be inappropriate to the quantification of risk and dependence in the tail region. The limitations of the Markowitz approach are well known and addressed by recent developments in financial risk theory. In particular, the two-sided view of risk is commonly replaced by the notion of downside risk and re- lated risk measures such as the value-at-risk or the expected shortfall. The theory of coherent risk measures founded by Artzner et al.(1999) provides an axiomatic characterization of risk measures that favours portfolio diver- sification. This objective can be considered as an extension of the original motivation of Markowitz, who wanted to find a mathematical reasoning for portfolio diversification (cf. Markowitz, 1991). Talking about rare events and extreme losses, one has to mention the heavy-tailed distributions. The evidence of heavy tails in finance and in- surance is manifold and the application area is so vast that an outline of the relevant literature would go far beyond the scope of this introduction. To mention only the basic facts, it should be said that heavy-tailed models are generally accepted in the realm of insurance (cf. Hogg and Klugman, 1984; Embrechts et al., 1997) and vividly discussed in mathematical finance since its earliest days (cf. Mandelbrot, 1963; Fama, 1965). The equivocality concerning heavy-tailed modelling of financial returns is a result of the high complexity of financial markets. On the one hand, extremal behaviour of the market data seems to suggest heavy-tailed distributions (cf., among others, Longin, 1996; Beirlant et al., 2004b). On the other hand, reliable estima- tion of tail parameters in non-stationary time series with clustering effects is known to be difficult (cf. Malevergne et al., 2006). This implies that indica- tion for heavy tails in financial return data should be handled very carefully. Still, despite the ongoing scientific discussion, it is commonly agreed that heavy tails are worth considering.
1.2 Overview of related theory
Extremal behaviour of random variables is a subject of vital research since the very beginnings of probability theory. The core basis of extreme value theory is the Fisher–Tippett theorem, which characterizes the limit distribu- 1.3. Central results and structure of the thesis 3 tions for maxima of i.i.d. random variables (Fisher and Tippett, 1928; Gne- denko, 1943). The Fisher–Tippett theorem states that there are only three possible limit distribution types: the Fr´echet, the Gumbel, and the Weibull distributions. This result was applied to the estimation of high quantiles and probabilities of extremal events in many application areas, such as hydrol- ogy, network modelling, and risk management in finance and insurance. An elaborate overview of the latter application area is given by Embrechts et al. (1997). The probabilistic structure of multivariate extremes was characterized by de Haan and Resnick(1977). This seminal result provided a sound theoretical basis for various approaches to the modelling and the estimation of extremal dependence. In particular, modelling concepts based on multivariate extreme value theory and copulas have been developed for an adequate description and analysis of risks and risk portfolios. Comprehensive elaborations on this topic are given in McNeil et al.(2005) and Malevergne and Sornette(2006). Further developments in modelling extremal dependence are based on the notions of tail dependence and tail copulas, multivariate excess distributions, and the empirical distribution of excess directions. See, among others, Falk et al.(1994); Schmidt and Stadtm¨uller(2006); Kl¨uppelberg and Resnick (2008); Hauksson et al.(2000). The current state of extreme value theory in both univariate and multivariate settings with particular emphasis on statistical applications is presented in de Haan and Ferreira(2006), whereas heavy-tailed models are specially treated by Resnick(2007). There are also various applications of multivariate extreme value theory related to sums of random variables (cf. W¨uthrich, 2003; Barbe et al., 2006; B¨ocker and Kl¨uppelberg, 2008; Kortschak and Albrecher, 2009a). However, the focus of these results is put on loss aggregation rather than on portfolio diversification. Although tightly related, these problems are obviously not identical. In particular, characterization of risk aggregated in the sum of random variables does not provide a direct link to the location of the optimal portfolio. Furthermore, as highlighted in Remark 5.31, the influence of the tail index on diversification effects turns out to be different from the influence on the aggregated risk.
1.3 Central results and structure of the thesis
This thesis is dedicated to the diversification of extreme risks. In contrast to the Markowitz approach, the problem is reduced to the comparison and minimization of the extreme portfolio risks, while the consideration of average gains is omitted. Thus application of these results in practical portfolio 4 Chapter 1. Introduction optimization might need additional modelling of gains in the non-extreme region and a trade-off between the risks and the expected gains. This problem statement should not be misunderstood as a prudent view of risk. The high abstraction level of the present approach is motivated by the interest in a pure mathematical concept of diversification effects for extreme risks. To achieve this, all conditions that are not essential to the extremal be- haviour are dropped. The results are derived upon the assumption of mul- tivariate regular variation, which specifies only the tail behaviour and the asymptotic dependence structure in the tail region. This setting is rather different from the Markowitz optimization problem, which was primarily de- signed for the multivariate Gaussian case. In particular, the distributions considered here need not be symmetric or have finite variances, and even the case of infinite first moments is included. The general problem setting can be described as follows. Given a multi- variate regularly varying random vector X = (X(1),...,X(d)), one is inter- ested in the asymptotic comparison of portfolio losses
ξ>X := ξ(1)X(1) + ... + ξ(d)X(d) with portfolio weights ξ(i) ∈ R satisfying ξ(1) +...+ξ(d) = 1. This comparison is based upon the limit
P ξ>X > t γξ := lim . t→∞ P {kXk1 > t} The existence of this limit is obtained from the multivariate regular variation of X, which also implies the asymptotic quantile relation
← Fξ>X (1 − λ) 1/α lim = γξ . λ↓0 F ← (1 − λ) kXk1
Thus γξ determines the first-order term for the approximation of high port- folio loss quantiles. As recently shown by Degen et al.(2010), first-order approximation may need further improvement by second-order terms. Still, analysis of first-order properties remains important. The basic result underlying this thesis is the characterization of the limit ratio γξ as a functional of the portfolio vector ξ and the characteristics of the multivariate regular variation, given by the tail index α and the spectral measure Ψ. This characterization also includes short positions, i.e., negative portfolio weights. Due to its extraordinary role, the functional γξ is called extreme risk index of the portfolio ξ. The mapping ξ 7→ γξ characterizes the problem of portfolio optimization with respect to extreme risks. It turns 1.3. Central results and structure of the thesis 5 out that this function is convex if the loss expectations exist and concave otherwise under the additional assumption that the portfolio weights are non-negative and there are no extreme gains. Moreover, positive dependence of loss components decreases the total portfolio risk in case of infinite loss expectations. In the general case, i.e., with possible extreme gains and short positions, the properties of the function ξ 7→ γξ are less definite. The con- vexity for α ≥ 1 remains, but the qualitative behaviour for α < 1 depends on the spectral measure. Examples are presented where ξ 7→ γξ is convex, concave, or none of that. These results contradict the intuition of diversification within the Marko- witz theory and the theory of coherent risk measures. Similar effects are al- ready known from aggregation results in multivariate regularly varying mod- els (cf. Rootz´enand Kl¨uppelberg, 1999; Neˇslehov´aet al., 2006; Embrechts et al., 2009a,b). Thus the extreme risk index γξ can be considered as an extension of these findings to diversification problems. Moreover, γξ allows to compare extreme portfolio risks in the infinite mean case, where coherent risk measurement is not possible (cf. Delbaen, 2009). The next series of results is related to the estimation of the extreme risk index γξ from i.i.d. observations of the random loss vector. The estimation of the optimal portfolio ξopt minimizing γξ is also included. A semiparamet- ric estimator γbξ is proposed that exhibits the same convexity and concavity properties as the true extreme risk index γξ. Thus the qualitative results on the optimization of γξ remain valid for the estimate γbξ. Asymptotic normal- ity and strong consistency of the estimator γbξ are obtained uniformly over compact sets of portfolio vectors ξ. The uniform asymptotic normality means weak convergence of the properly normalized estimation error to a Gaussian limit process with index ξ. The proofs incorporate empirical process the- ory presented in van der Vaart and Wellner(1996). The uniform strong consistency of γbξ immediately yields the strong consistency of the estimated optimal portfolio ξbopt. In addition to the comparison of risks for different portfolio vectors, this thesis addresses the comparison of stochastic models. A new notion of stochastic ordering is introduced, appropriate to the ordering of random vectors with respect to the resulting extreme portfolio risks. This asymptotic portfolio loss order apl is characterized for multivariate regularly varying random vectors. The equivalent criterion obtained here incorporates order- ing of marginal distributions with respect to apl and ordering of spectral measures with respect to a specific integral order relation that is linked to the extreme risk index γξ. In addition to the criteria based upon spectral measures, sufficient criteria for apl are obtained in terms of some well-known 6 Chapter 1. Introduction multivariate stochastic order relations. The statistical results are accomplished by an extensive simulation study. Monte Carlo performance tests for the estimators γbξ and ξbopt are imple- mented in two exemplaric models. The models are chosen to illustrate how the estimates of the tail index α and the spectral measure Ψ influence the bias and the variability of the estimator γbξ. Furthermore, the optimization quality achieved by the estimator ξbopt is compared for different values of the tail index and the degree of dependence between the loss components. In particular, simulation results demonstrate that portfolio optimization may be problematic if α is close to 1 and the loss components are non-negative. The optimization results obtained in other cases are reliable. The thesis is organized as follows. Chapter2 provides the basic nota- tion of multivariate regular variation and multivariate extreme value theory, followed by an overview of alternative approaches to the modelling of depen- dence structures, auxiliary results, and examples. Chapter3 is dedicated to the characterization of diversification effects in multivariate regularly vary- ing models and the resulting optimization problem. The extreme risk index γξ is introduced, analysed and applied to the minimization of risk measures. Chapter4 comprises the statistical results, including consistency and asymp- totic normality of the estimated extreme risk index γbξ and consistency of the estimated optimal portfolio ξbopt. Stochastic ordering with respect to extreme portfolio losses is discussed in Chapter5. Finally, the simulation study is presented in Chapter6. Auxiliary results on regular variation and empirical processes are collocated in AppendixA. Some of the results presented in Chapter3 and a special case of the results obtained in Chapter4 are published in Mainik and R¨uschendorf(2010). 1.4. Acknowledgements 7 1.4 Acknowledgements
I would like to thank my scientific advisor Professor Dr. Ludger R¨uschendorf for drawing my attention to extremal dependence structures, for encouraging me to work independently, and for the support he gave me by his valuable comments and decisive questions. I thank my colleagues at the Department of Mathematical Stochastics for the friendly atmosphere, especially appreciating the time and the discussions I had with Wolfgang Kluge, Nataliya Koval, Joachim Schneegans, Eva-Maria Schopp, and Volker Pohl. Further thanks go to the department’s heart and soul, Mrs. Monika Hat- tenbach, for stylistic proof reading and serving as a wandering encyclopaedia of LATEX layout tricks. I thank my friends Katja Guschanski, Adrian Kantian, Florian Dennert, Hilmar B¨ohm,and Roman Iakoubov for the time they shared with me and for all the moments of humour and truth. Greatest thanks go to my family. I thank my parents, Johannes and Ludmila Mainik, and my brother Andreas for their unconditional support that gave me so much confidence. Enormous thanks are due to my wife, Annette, for walking this path along with me and being a constant source of power, will, and joy. 8 Chapter 1. Introduction Chapter 2
Multivariate regular variation
This chapter introduces the heavy-tailed multivariate regularly varying prob- ability distributions, which constitute the theoretical framework of the thesis. The definition and the basic properties of univariate and multivariate regular variation are given in Section 2.1, whereas further extensions, related notions, examples, and auxiliary results needed later are collocated in a sequence of sections addressing special topics. Thus, the so-called canonical standardizations of exponent and spectral measures are introduced in Section 2.2, while the notion of dependence func- tions is subject of Section 2.3, and the copula approach to the modelling of multivariate extremes is discussed in Section 2.4. Highlighting the intercon- nections between the different dependence notions, these sections put them into a common theoretical framework. The chapter is concluded by two sections that deal with specific models appearing in examples of Chapters5 and6. Section 2.5 presents an exem- plary computation of the canonical spectral densities associated with Gumbel copulas. Finally, Section 2.6 provides a general representation for the spectral densities of multivariate regularly varying elliptical distributions.
2.1 Basic notation and model assumptions
Let X be a random vector in Rd with components X(1),...,X(d) representing the gains and the losses of risky assets. Focusing on the risks, let X be a random loss vector, i.e., X(i) > 0 quantifies losses and X(i) < 0 quantifies gains generated by the i-th asset. According to the application area (finan- cial or actuarial), it is natural to distinguish between the general case with components X(i) taking both positive and negative values (loss-gain case) and the case when the value domains of all components X(i) are restricted
9 10 Chapter 2. Multivariate regular variation to R+ := [0, ∞)(pure loss case). As it turns out further, portfolio diversifi- d cation for loss vectors X ∈ R+ exhibits some remarkable properties that do not hold in the general case. Denoting the weight of the i-th asset in the portfolio by ξ(i), one obtains d d the portfolio vector ξ ∈ R+ or ξ ∈ R if negative portfolio weights (short positions) are permitted. The portfolio loss is given by the scalar product of the portfolio vector ξ and the loss vector X. In the following, vectors are regarded as columns and the portfolio loss is written as ξ>X: ξ>X := ξ(1)X(1) + ... + ξ(d)X(d). As a special case, this notation includes components X(i) representing relative losses of assets Z(i): (i) (i) (i) Z0 − ZT X = (i) , i = 1, . . . , d. Z0 In this setting the scalar product ξ>X equals the random loss generated by investing the value ξ(i) in the i-th asset for i = 1, . . . , d:
d (i) > X ξ (i) (i) ξ X = (i) Z0 − ZT . i=1 Z0
> Pd (i) Furthermore, the relative loss of the portfolio ξ is given by ξ X/ i=1 ξ . The probability distributions of the loss components X(i) are assumed to be heavy-tailed in the sense that the tail index of each X(i) is finite:
n (i) β o αi := sup β ∈ [0, ∞):E X < ∞ < ∞. (2.1) Sometimes it is more convenient to operate with the upper and the lower tail index of a random variable Y , obtained from the positive part Y+ := Y · 1 {Y > 0} and the negative part Y− := |Y | · 1 {Y < 0}, respectively. It is obvious that the total tail index of Y defined in (2.1) is equal to the minimum of the upper and the lower one. A random variable Y is called heavier-tailed than another random variable Z if the tail index of Y is lower than that of Z. It is well known that in the case of unequal component tail indices αi the contribution of lighter tails to the portfolio loss ξ>X is asymptotically negligible if all portfolio weights ξ(i) are positive and there is no mutual neutralization of the heavier tails due to linear dependence between gains and losses. Consequently, the study of asymptotic diversification effects can be reduced to the non-trivial case by assuming that the component tail indices αi are equal:
α1 = ... = αd =: α. (2.2) 2.1. Basic notation and model assumptions 11
The heavy-tail property (2.1) is strengthened by the assumption of (uni- variate) regular variation. Definition 2.1. A non-negative random variable Y is called regularly varying with tail index α ∈ [0, ∞) if the following condition is satisfied:
P{Y > tx} ∀x > 0 → x−α, t → ∞. (2.3) P{Y > t}
It is easy to see that regular variation (2.3) implies the heavy-tail prop- erty (2.1) and that the tail index α characterizing these two properties is necessarily the same. Analogously to the heavy-tail property, regular varia- tion can also be considered separately for upper and lower tails. The notion of regular variation for random variables and corresponding probability distributions is intimately related to the regular variation of func- tions. A function f defined on a neighbourhood of ∞, f :(c, ∞) → R, c ∈ R, is called regularly varying (at ∞) if there exists a function g : (0, ∞) → R such that f(tx) ∀x > 0 lim = g(x). (2.4) t→∞ f(t) It is well known that for measurable functions f any solution g of (2.4) is a power function, β g(x) = x , β ∈ R, (2.5) and that the regular variation index β in (2.5) is unique for each f. In the special case β = 0 the function f is called slowly varying, and any measurable regularly varying function f has necessarily the form
f(t) = l(t) · tβ (2.6) with a slowly varying function l. For more details on regular variation of functions see Bingham et al.(1987). Thus, in terms of the probability distribution function corresponding to a random variable Y ,
FY (t) := P {Y ≤ t} , regular variation of Y with tail index α ∈ [0, ∞) in the upper or in the lower tail means regular variation with index β = −α of the function t 7→ 1−FY (t) at t = ∞ or of the function t 7→ FY (t) at t = −∞, respectively. In order to obtain a non-trivial dependence structure in the tails, the univariate regular variation of the asset losses X(i) is strengthened by the assumption of multivariate regular variation. 12 Chapter 2. Multivariate regular variation
Definition 2.2. A random vector X taking values in Rd is called multivariate regularly varying with tail index α ∈ (0, ∞) if there exist a sequence an → ∞ and a (non-zero) Radon measure ν on the Borel σ-field B([−∞, ∞]d \{0}) such that ν([−∞, ∞]d \ Rd) = 0 and, as n → ∞,
−1 v nP an X → ν on B([−∞, ∞]d \{0}), (2.7)
v −1 where → denotes the vague convergence of Radon measures and P an X is the −1 probability distribution of an X. It should be noted that random vectors with non-negative components yield limit measures ν that are concentrated on [0, ∞]d \{0}. Therefore multivariate regular variation in this special case can also be defined by vague convergence on B([0, ∞]d \{0}). For a full account of technical details related to the notion of multivariate regular variation, vague convergence, and the Borel σ-fields on the punctured spaces [−∞, ∞]d \{0} and [0, ∞]d \{0} the reader is referred to Resnick(2007) or Lindskog(2004). It is well known that the limit measure ν obtained in (2.7) is unique except for a constant factor, has a singularity in the origin in the sense that ν((−ε, ε)d) = ∞ for any ε > 0, and exhibits the scaling property
ν(tA) = t−αν(A) (2.8) for all sets A ∈ B [−∞, ∞]d \{0} that are bounded away from 0. It is also well known that (2.7) implies that the random variable kXk with an arbitrary norm k·k on Rd is univariate regularly varying with tail index α. Moreover, the sequence an can always be chosen as
← an := FkXk(1 − 1/n), (2.9)
← where FkXk is the quantile function of kXk. The resulting limit measure ν is d normalized on the set Ak·k := {x ∈ R : kxk > 1} by ν Ak·k = 1. (2.10)
Thus, after normalizing ν by (2.10), the scaling relation (2.8) yields an equivalent rewriting of the multivariate regular variation condition (2.7) in terms of weak convergence:
−1 w L t X | kXk > t → ν|Ak·k on B Ak·k (2.11)
for t → ∞, where ν|Ak·k is the restriction of ν to the set Ak·k. 2.1. Basic notation and model assumptions 13
Additionally to (2.7) it is assumed that the limit measure ν is non-degen- erate in the following sense: