E4602: IEOR Quantitative Risk Management c 2012 by Martin Haugh
Spring 2012
Risk Measures, Risk Aggregation and Capital Allocation
We consider risk measures, risk aggregation and capital allocation in these lecture notes and build on our earlier introduction to Value-at-Risk (VaR) and Expected Shortfall (ES). We will follow Chapter 6 of Quantitative Risk Management by MFE closely. This chapter, however, contains considerably more material than we will cover and it should be consulted if further details are required.
1
Coherent Measures of Risk
In 1999 Artzner et al. proposed a list of properties that any good risk measure should have and this list gave rise to the concept of coherent and incoherent measures of risk. Since then a substantial body of research has developed on the theoretical properties of risk measures and we describe some of these results here. Let M denote the space of random variables representing portfolio losses over some fixed time interval, . We assume that M is a convex cone so that if L1 M and L2 M then L1 + L2 M and L1 M for every > 0. A risk measure is then a real-valued function, : M R, that satisfies certain desirable properties. (L) may be interpreted as the riskiness of a portfolio or the amount of capital that should be added to a portfolio with a loss given by L, so that the portfolio can then be deemed acceptable from a risk point of view. Note that under this latter interpretation, portfolios with (L) < 0 are already acceptable and do not require capital injections. In fact, if (L) < 0 then capital could even be withdrawn while the portfolio would still remain acceptable. The following properties of a risk measure merit special attention: Axiom 1 : (Translation Invariance) For all L M and every constant a R, we have (L + a) = (L) + a. This property is necessary if the risk-capital interpretation we stated above is to make sense. Axiom 2 : (Subadditivity) For all L1 , L2 M, we have (L1 + L2 ) (L1 ) + (L2 ). This axiom reflects the idea that pooling risks helps to diversify a portfolio. While this has been the most debated of the risk axioms, it allows for the decentralization of risk management. For example, if a risk manager has a total risk budget of B, he can divide B into B1 and B2 where B1 + B2 = B. He can then allocate risk budgets of B1 and B2 to different trading desks or operating units in the organization, safe in the knowledge that the firm-wide risk will not exceed B. Axiom 3 : (Positive Homogeneity) For all L M and every > 0 we have (L) = (L). This axiom is also somewhat controversial and has been criticized for not penalizing concentration of risk and any associated liquidity problems. In particular, if > 0 is very large, then some people claim that we should require (L) > (L). However, such a result would be inconsistent with the subadditivity axiom. This is easily seen if we write (nL) = (L + + L) n(L) (1) where n N and the inequality follows from subadditivity. The positive homogeneity assumption states that we must have equality in (1). This reflects he fact that there are no diversification benefits when we hold multiples of the same portfolio, L.
Risk Measures, Risk Aggregation and Capital Allocation Axiom 4 : (Monotonicity) For L1 , L2 M such that L1 L2 almost surely, we have (L1 ) (L2 ). It is clear that any risk measure should satisfy this axiom.
2
Definition 1 A risk measure, , acting on the convex cone M is called coherent if it satisfies the translation invariance, subadditivity, positive homogeneity and monotonicity axioms. Remark 1 The criticisms of the subadditivity and positive homogeneity axioms have led to the study of convex risk measures. A convex risk measure satisfies the same axioms as a coherent risk measure except that the subadditivity and positive homogeneity axioms are replaced by the convexity axiom: Axiom 5 : (Convexity) For L1 , L2 M and [0, 1], (L1 + (1 - )L2 ) (L1 ) + (1 - )(L2 ). It is possible within the convex class to find risk measures that satisfy (L) (L) for > 1.
Value-at-Risk
Value-at-risk is not a coherent risk measure because it fails to be subadditive. This is perhaps the principal1 criticism that is made of VaR when it is compared to other risk measures. We will see two examples below that demonstrate this. We first recall the definition of VaR. Definition 2 Let (0, 1) be some fixed confidence level. Then the VaR of the portfolio loss, L, at the confidence interval, , is given by VaR := q (L) = inf{x R : FL (x) }. where FL () is the CDF of the random variable, L. Example 1 Consider2 two assets, X and Y , that are usually normally distributed but are subject to occasional shocks. In particular, assume that X and Y are independent and identically distributed with { 0, with prob .991 X = + where N(0, 1) and = -10, with prob .009. Consider a portfolio consisting of X and Y . Then VaR.99 (X + Y ) = 9.8 > VaR.99 (X) + VaR.99 (Y ) = 3.1 + 3.1 = 6.2 thereby demonstrating the non-subadditivity of VaR. Exercise 1 Confirm that the VaR values of 3.1 and 9.8 in the previous example are correct. We now give a more meaningful and disturbing example of how VaR fails to be sub-additive. Example 2 (VaR for a Portfolio of Defaultable Bonds (E.G. 6.7 in MFE)) Consider a portfolio of n = 100 defaultable corporate bonds where the probability of a default over the next year is identical for all bonds and is equal to 2%. We assume that defaults of different bonds are independent from one another. The current price of each bond is 100 and if there is no default, a bond will pay 105 one year from now. If the bond defaults then there is no repayment. This means we can define Li , the loss on the ith bond, as Li := 105Yi - 5 where Yi = 1 if the bond defaults over the next year and Yi = 0 otherwise. By assumption we also see that P (Li = -5) = .98 and P (Li = 100) = .02. Consider now the following two portfolios:
1 Of course, the general criticism that summarizing an entire loss distribution with just a single number can be applied to all risk measures, coherent or not. Furthermore, there is the implicit assumption that we know the loss distribution when determining the value of a risk measure. This assumption is often unjustifiable: it can and indeed has often led to financial catastrophe! 2 This example is taken from "Subadditivity Re-Examined: the Case for Value-at-Risk" by Dan ielsson et al.
Risk Measures, Risk Aggregation and Capital Allocation A: A fully concentrated portfolio consisting of 100 units of bond 1. B: A completely diversified portfolio consisting of 1 unit of each of the 100 bonds. We can compute the 95% VaR for each portfolio as follows: Portfolio A: The loss on portfolio A is given by LA = 100L1 so that VaR.95 (LA ) = 100VaR.95 (L1 ). Note that P (L1 -5) = .98 > .95 and P (L1 l) = 0 < .95 for l < -.5. We therefore obtain VaR.95 (L1 ) = -5 and so VaR.95 (LA ) = -500. So the 95% VaR for portfolio A corresponds to a gain(!) of 500. Portfolio B: The loss on portfolio B is given by LB =
100 i=1 100 i=1
3
Li = 105
Yi - 500
100 100 and so VaR.95 (LB ) = 105 VaR.95 ( i=1 Yi ) - 500. Note that M := i=1 Yi Bin(100, .02) and by inspection we see that P (M 5) .984 > .95 and P (M 4) .949 < .95. Therefore VaR.95 (M ) = 5 and so VaR.95 (LB ) = 525 - 500 = 25. So according to VaR.95 , portfolio B is riskier than portfolio A. This is clearly nonsensical. Note that we have shown that ( 100 ) 100 VaR.95 Li 100 VaR.95 (L1 ) = VaR.95 (Li )
i=1 i=1
demonstrating again that VaR is not subadditive. Remark 2 Let be any coherent risk measure that depends only on the distribution of L. Then we obtain ( 100 ) 100 Li (Li ) = 100(L1 )
i=1 i=1
and so in the previous example, would correctly classify portfolio A as being riskier than portfolio B. We now describe a situation where VaR is always subadditive. Theorem 1 (Subadditivity of VaR for Elliptical Risk Factors (Theorem 6.8 in MFE)) Suppose that X En (, , ) and let M be the set of linearized portfolio losses of the form M := {L : L = 0 + Then for any two losses L1 , L2 M, and 0.5 < 1, VaR (L1 + L2 ) VaR (L1 ) + VaR (L2 ). Proof: Without loss of generality we may assume that 0 = 0. Recall also that if X En (, , ) then X = AY + where A Rnk , Rn and Y Sk () is a spherical random vector. Any element L M can therefore be represented as L = T X = T AY + T ||T A|| Y1 + T
n i=1
i Xi , i R}.
(2)
where (2) follows from part 3 of Theorem 2 in the Multivariate Distributions lecture notes. Now the translation invariance and positive of homogeneity VaR imply VaR (L) = ||T A|| VaR (Y1 ) + T .
Risk Measures, Risk Aggregation and Capital Allocation Suppose now that L1 := T X and L2 := T X. The triangle inequality implies 1 2 ||(1 + 2 )T A|| ||T A|| + ||T A|| 1 2 and since VaR 0 for .5 (why?), the result follows from (2). Remark 3 It is a widely held belief that if the individual loss distributions under consideration are continuous and symmetric then VaR is subadditive. This is not true and a counterexample may be found in Section 6.2 of MFE. The loss distributions in the counterexample are smooth and symmetric but the copula is highly asymmetric. VaR can also fail to be subadditive when the individual loss distributions have heavy tails.
4
Expected Shortfall
We now show that expected shortfall (ES) or CVaR is a coherent measure of risk. We first recall the definition of ES. Definition 3 For a portfolio loss, L, satisfying E[|L|] < the expected shortfall at confidence level (0, 1) is given by 1 1 ES := qu (FL ) du. 1- The relationship between ES and VaR is therefore given by 1 1 ES := VaRu (L) du 1- from which it is clear that ES (L) VaR (L). When the CDF, FL , is continuous then a more well known representation of ES (L) is given by ES := E [L; L q (L)] = E [L | L VaR ] . 1- (4)
(3)
The following result demonstrates that expected shortfall is a coherent risk measure. We again follow the proof in MFE. Theorem 2 Expected shortfall is a coherent risk measure. Proof: The translation invariance, positive homogeneity and monotonicity properties all follow from the representation of ES in (3) and the same properties for quantiles. We therefore only need to demonstrate subadditivity. Let L1 , . . . , Ln be a sequence of random variables and let L1,n Ln,n be the associated sequence of order statistics. Note that
m i=1
Li,n = sup{Li1 + + Lim : 1 i1 < < im n}
(5)
~ where m N satisfying 1 m n is arbitrary. Now let (L, L) be a pair of random variables with joint CDF, F , ~ ~ and let (L1 , L1 ), . . . , (Ln , Ln ) be an IID sequence of bivariate random vectors with this same CDF. Then
m ~ (L + L)i,n i=1
~ ~ = sup{(L + L)i1 + + (L + L)im : 1 i1 < < im n} ~ ~ sup{Li1 + + Lim : 1 i1 < < im n} + sup{Li1 + + Lim : 1 i1 < < im n} m m ~ = Li,n + Li,n . (6)
i=1 i=1
Risk Measures, Risk Aggregation and Capital Allocation m 1 Now set3 m = n(1 - ) and let n . It may be shown4 that m i=1 Li,n ES (L), m ~ m 1 1 ~ ~ ~ i=1 Li,n ES (L) and m i=1 (L + L)i,n ES (L + L). The subadditivity of ES then follows m immediately from (6). There are many other examples of risk measures that are coherent. They include, for example, risk measures based on generalized scenarios and spectral risk measures of which expected shortfall is an example.
5
2
Bounds for Aggregate Risk
Let L = (L1 , . . . , Ln ) denote a vector of random variables, each one representing a loss on a particular trading desk, portfolio or operating unit within a firm. Sometimes we wish to aggregate these losses into a single random variable, (L), say. Common examples of the aggregating function, (), include: n The total loss so that (L) = i=1 Li . The maximum loss where (L) = max{L1 , . . . , Ln }. n The excess-of-loss treaty so that (L) = i=1 (Li - ki )+ . n + The stop-loss treaty in which case (L) = ( i=1 Li - k) . We wish to understand the risk of the aggregate loss function, ((L)), but to do so we need to know the distribution of (L). In practice, however, we often know only the distributions of the Li 's and have little or no information about the dependency or copula of the Li 's. In this case we can try to compute lower and upper bounds on ((L)). In particular we can formulate the two problems min max := := inf{((L)) : Li Fi , i = 1, . . . , n} sup{((L)) : Li Fi , i = 1, . . . , n}
where Fi is the CDF of the loss, Li . Problems of this type are referred to as Frechet problems and solutions are available in some circumstances. Indeed, when we studied copulas we saw an example of such a problem when we addressed the question of attainable correlations given known marginal distributions. In a risk management n context, these problems have been studied in some detail when (L) = i=1 Li and () is the VaR function. Results related to this problem are generally of more theoretical than practical interest and so we will not discuss them any further. Results and references, however, can be found in Section 6.2 of MFE.
3
Capital Allocation
n Consider again a total loss given by L = i=1 Li and suppose we have determined the risk, (L), of this loss. The capital allocation problem seeks a decomposition, AC1 , . . . , ACn , such that (L) =
n i=1
ACi
(7)
and where ACi is interpreted as the risk capital that has been allocated to the ith loss, Li . This problem is important in the setting of performance evaluation where we want to compute a risk-adjusted return on capital (RAROC). This return might be estimated, for example, by Expected Profit / Risk Capital and in order to
4 See, 3 x
is defined to be the largest integer less than or equal to x, i.e. the floor of x. for example, Lemma 2.20 in MFE.
Risk Measures, Risk Aggregation and Capital Allocation
6
compute this we must determine the risk capital of each of the Li 's. Obviously, we would require the corresponding risk capitals to sum to the total risk capital so that (7) is satisfied. n More formally, let L() := i=1 i Li be the loss associated with the portfolio consisting of i units of the loss, Li , for i = 1, . . . , n. The loss on the actual portfolio under consideration is then given by L(1). Let () be a risk measure on a space M that contains L() for all , an open set containing 1. Then the associated risk measure function, r : R, is defined by r () = (L()). We have the following definition. Definition 4 Let r be a risk measure function on some set Rn \ 0 such that 1 . Then a mapping, f r : Rn , is called a per-unit capital allocation principle associated with r if, for all , we have
n i=1
i fi () = r ().
r
(8)
We then interpret fi as the amount of capital allocated to one unit of Li when the overall portfolio loss is r L(). The amount of capital allocated to a position of i Li is therefore i fi and so by (8), the total risk capital is fully allocated. Definition 5 (Euler Capital Allocation Principle) If r is a positive-homogeneous risk-measure function which is differentiable on the set , then the per-unit Euler capital allocation principle associated with r is the mapping r r f r : Rn : fi () = (). i The Euler allocation principle is seen to be a full allocation principle since a well-known property of any positive n r homogeneous and differentiable function, r() is that it satisfies r() = i=1 i i (). The Euler allocation principle therefore gives us different risk allocations for different positive homogeneous risk measures. It should also be mentioned that there are good economic reasons5 for employing the Euler principle when computing capital allocations. We will not discuss those reasons here, however. We end by briefly describing some examples below but Section 6.3 of MFE should be consulted for proofs and further details if necessary. Standard Deviation and the Covariance Principle Let rsd () = std(L()) be our risk measure function and write for the variance-covariance matrix of ( )1/2 L1 , . . . , Ln . Then rsd () = T and using the Euler allocation principle it follows that n ()i Cov(Li , L()) rsd j=1 Cov(Li , Lj )j rsd () = = = fi () = i rsd () rsd () Var(L()) and the actual capital allocation, ACi , for Li is obtained by setting = 1 in (9). This is then known as the covariance principle. Value-at-Risk and Value-at-Risk Contributions
If rV aR () = VaR (L()) is our risk measure function, then subject to technical conditions it can be shown6 that rV aR r fi V aR () = () = E [Li | L() = VaR (L())] , for i = 1, . . . , n. (10) i Expected Shortfall and Shortfall Contributions If rES () = E [L() | L() VaR L()] is our risk measure function, then subject again to technical conditions it can be shown that 1 rES r () = E [Li | L() VaR (L())] , for i = 1, . . . , n. fi ES () = i 1-
r
(9)
(11)
We therefore have the capital allocation ACi = E [Li | L VaR (L)] for the risk, Li , where L := L(1).
Section 6.3.3 of MFE for these reasons. Simulations with Exact Means and Covariances (2009) by A. Meucci for an application where the VaR of a portfolio is allocated to the individual securities in the portfolio using the Euler Capital Allocation Principal in (10).
6 See 5 See