Risk Minimizing Portfolio Optimization and Hedging
with Conditional Value-at-Risk
Jing Li, University of North Carolina at Charlotte, Department of Mathematics and Sta-tistics, Charlotte, NC 28223, USA. Email: [email protected].
Mingxin Xu1
, University of North Carolina at Charlotte, Department of Mathematics and Statistics, Charlotte, NC 28223, USA. Email: [email protected].
January 23, 2008
Abstract: We look at the problem of how to find a dynamic optimal portfolio so that the Conditional Value-at-Risk (CVaR) is minimized under the condition where the returns are bounded. CVaR is a risk measure based on the popular VaR that is coherent. In a complete market setting, we derive the exact optimal conditions. Then we provide applications in two classic complete market models: the Binomial model and the Black-Scholes model. In these cases, the procedures to find the optimal strategies are given with exact formulas. Numerical results show, as expected, dynamic portfolio provide much lower CVaR risk than static portfolios.
1. Introduction
Portfolio optimization has been a long-standing classic topic in financial theory since Markowitz’s ground-breaking work [12] in 1952. His choice of variance as the risk measure of a portfolio is no longer compatible with today’s risk management practice. One risk measure, first proposed by J.P.Morgan in RiskMetrics, that the regulators (BASEL II) make extensive
1The work of Mingxin Xu is supported by National Science Foundation under the grant SES-0518869 and
John H. Biggs Faculty Fellowship.
use of is Value-at-Risk (VaR). VaR is defined as “the loss level that will not be exceeded with a certain confidence level during a certain period of time”. As a quantile of the portfolio, VaR is a threshold and does not give information about the loss size when the loss exceeds VaR. In addition, VaR is in general neither sub-additive or super-additive which either encourage or discourage actions of diversification or merger, and portfolio optimization with respect to VaR leads to non-smooth results. Recent research in the area of mathematical finance, following the economic theory of utility functions, developed into an axiomatic approach for risk measure: Coherent Risk Measure was first proposed by Artzner, et al. [3] and [4]. The axioms of coherence parallel the axioms of rational investors. Conditional Value-at-Risk (CVaR), sometimes called Shortfall Value-at-Risk, defined as “the expected loss during a certain period of time, conditional that the loss is greater than a loss threshold corresponding to a certain confidence level”, studied by Rockafellar and Uryasev [13] and Acerbi and Tasche [1], turned out to be a distribution-based coherent risk measure. More importantly, CVaR is a vast improvement over VaR in producing smooth portfolio optimization results. The wide use of VaR and the advantage of CVaR has lead many financial institutions to consider using VaR for reporting purpose and CVaR for internal risk control.
Numerical implementation of an optimization problem with quantile-based objective func-tion does not have to be easy. The contribufunc-tion of Rockafellar and Uryasev [13] is that they found an equivalent formula for CVaR as a convex function which opened the door for convex programming methods. Using Monte-Carlo simulation for a one-time step model with mul-tiple assets, they formulated the portfolio optimization problem into a linear programming problem which can be efficiently implemented with standard programming software. The contribution of our paper is to work out the portfolio optimization problem with dynamic trading strategies (including multi-period models) in closed-form solutions for complete mar-ket models. In the classic setup of portfolio optimization, expected returns are maximized given limits on the risk of the portfolio. It can be made formal, but with some technical conditions, it is equivalent to minimize the risk of the portfolio given requirements for its
return. Our setup will focus on the later approach which takes focus on risk minimization as the objective. An attempt to employ the dynamic programming method for multi-period models was made by Ruszczy´nski and Shapiro [17]. They took the approach to modify the risk measure CVaR into a dynamic version ‘conditional risk mappings for CVaR’. In this paper, we will stick to the original CVaR and only let the portfolio composition adjusts over time because optimizing CVaR at each time period conditionally is very different than optimizing CVaR of the final wealth of a dynamic portfolio. In practice, portfolio rebalance occurs much more often than financial reports on return and risk. Our solution will an-chor on duality methods based on risk neutral measures similar to those employed in option pricing and utility maximization problems. This martingale approach is well-studied in re-cent mathematical finance research partly because it allows for solutions to a wider ranges of problems which does not possess Markovian property and thus does not meet dynamic programming principles.
Section 2 establishs the general risk minimization problem and provides the solution. Section 3 details the solutions to two popular complete market models: the Binomial model and the Black-Scholes model. Section 4 provide numerical examples and Section 5 concludes.
2. Dynamic Portfolio Optimization and Hedging with CVaR
All results in this section work for multi-dimensional market models. In an attempt to keep the notations simple, we will present them in a one-dimensional case where there is a single risky asset. Let us consider a self-financing portfolio consisting of investments in a money market account and a stock account. Assume the money market account earns a risk-free rate r and the stock price St is an adapted process on a filtered probability space (Ω, F , (Ft)t≥0, P ) that satisfies the usual conditions. It is standard for continuous time models, as in the Black-Scholes model, to write the dynamics of this self-financing portfolio as
where ξt is the number of shares of stocks held at time t. The riskiness of the return of this portfolio will be measured by CVaR. A good reference for CVaR is in the book written by F¨ollmer and Schied [9] where CVaR is given a third name Average VaR. We define VaR at level λ to be
V aRλ(Z) = inf{ m | P (Z + m < 0) ≤ λ }, λ ∈ (0, 1).
From a risk management perspective, it can be viewed as the smallest capital reserve to add to a financial position Z for it to be admissible: the probability for the combined position to default is less than λ. Mathematically, it is the negative value of a λ-quantile of Z. CVaR is defined as a simple average of VaR
(1) CV aRλ(Z) = 1 λ Z λ 0 V aRγ(Z)dγ.
With this definition, it is easy to see why CVaR is different than VaR that it is smooth with respect to the change of the confidence level λ. To relate to the definition of expectation on the tail as stated in the Introduction, we have to be a little careful when we write down the equivalent form of (1) for the case when the probability space could be discrete
CV aRλ(Z) = − 1 λ E[Z1{Z<V aRλ(Z)}] + V aRλ(Z)(1 − P (Z < V aRλ(Z)) .
The question is how we should trade the shares throughout a finite holding period [0, T ] so that we can achieve minimal risk at time T , conditioned that the return is within acceptable range? Mathematically, we want to find the optimal investment strategy with initial capital X0 = x:
ξt∗ = arg inf ξt
CV aRλ(XT), (2)
This is equivalent to the problem of minimizing CVaR on the return RT
ξt∗ = arg inf ξt
CV aRλ(RT),
subject to rd ≤ Rt≤ ru, for all t ∈ [0, T ],
whether it be percentage return RT = XTX−X0 0 or log return RT = lnXXT0 because we only need to identify the one-to-one correspondence between the quantiles of XT and RT. A more noted difference here to a classic setup is that we require the returns to be bounded by two constants rd and ru and we are not putting constraints on the expected returns. Even when the expected return is high enough, a particular realized return still have a chance to turn out to be very low. We are requiring the realized return to be above xd in all cases if we take xu = ∞. Practically, this corresponds to a fund with a guaranteed return.
The same question can be asked for a discrete time trading model where adjustments of the portfolio take place at 0 = t0 < t1 < ... < tN = T , and the self-financing portfolio is
Xtn+1 = ξtnStn+1+ (Xtn− ξtnStn)(1 + r(tn+1− tn)).
Since the discrete time trading model can be viewed as a special case of the continuous time trading model where the stock price and the investment strategy are held piece-wise constant over time intervals, we will derive all the theorems in the continuous time model, which can be readily applied to discrete time models.
When we have an existing portfolio Ht consisting of investments in derivatives, we can ask a second question as how to hedge our risk with a self-financing admissible portfolio. Let H be the random variable representing the final value of the existing portfolio. It is FT-measurable since H = HT. The optimal hedging problem is to solve
ξ∗t = arg inf ξt
CV aRλ(H + XT), (3)
Thus when we combine the original portfolio and the hedging portfolio, the risk is minimized. It is straight-forward to see that if XT∗ is the final wealth of the optimal portfolio for problem (2), by which we mean that
min ξt
CV aRλ(XT) = CV aRλ(XT∗),
then XT∗ − H is the optimal solution for problem (3) if it is the result of a self-financing strategy. In a complete market model, this is not an issue because all derivatives can be replicated. Therefore, we will focus on solving problem (2) in a complete market model in this paper.
As mentioned in Section 1, problem (2) is complicated by the fact that the objective function involves quantile function and the corresponding numerical methods will have to involve ordering the returns. Rockafellar and Uryasev [13] and [14] found it to be the Fenchel-Legendre dual of expected shortfall where standard convex analysis applies. For λ ∈ (0, 1), definition (1) is equivalent to CV aRλ(Z) = 1 λx∈Rinf(E[(x − Z) + ] − λx),
when Z is integrable. Now we reformulate problem (2) into a more tractable convex opti-mization problem ξ∗t = arg inf ξt 1 λx∈Rinf(E[(x − XT) + ] − λx), (4)
subject to xd≤ Xt≤ xu, for all t ∈ [0, T ].
If we let the stock price St be as general as a semimartingale, then the market is complete when the risk neutral measure is unique. These fundamental theorems are developed in Delbaen and Schachermayer [5], El Karoui and Quenez [6], Kramkov [11] and F¨ollmer and Kabanov [7]. Define the set of risk neutral measures as
Now we make the main assumption of the section.
Assumption 2.1. S is a semimartingale on the filtered probability space (Ω, F , Ft, P ). M has a unique element ˜P .
Under this assumption, the no-arbitrage price of any derivative security H is ˜E[H] where the expectation is taken under the risk neutral measure ˜P , and a corresponding hedging strategy can be found when the price is finite. Now we are ready to further simplify the main optimization problem (2).
2.1. Static Formulation of the Dynamic Problem. Let the initial wealth be X0 = xr. Under Assumption 2.1, the space of final outcomes of self-financing strategies are those FT -measurable random variables X such that ˜E[X] = xr. Therefore, the dynamic problem (4) is equivalent to the static problem
X∗ = arg inf X 1 λx∈Rinf(E[(x − X) +] − λx), (5) subject to E[X] = x˜ r, xd ≤ X ≤ xu,
where the constants satisfy xd< xr < xu. Let XT∗ = X
∗. Then X∗
T is the final value of the optimal portfolio for problem (2):
CV aRλ(XT∗) = inf ξt CV aRλ(XT) = inf X 1 λx∈Rinf(E[(x − X) +] − λx) = 1 λx∈Rinf(E[(x − X ∗ )+] − λx).
Martingale representation theorem applied to Xt∗ = ˜E[X∗|Ft] will produce the optimal hedging strategy ξt∗ for problem (2). Problem (5) is intrinsically much simpler than problem (2) because it looks for an optimal random variable X∗ with convex objective function. The above simplification steps we have taken is based on classic duality theory (or sometimes called martingale approach) in mathematical finance. By duality, we mean there are two important spaces: the primal space consists of dynamics of self-financing portfolios and the dual space consists of risk neutral measures. The optimization problem in the primal space
is translated into an optimization problem in the dual space where it is always simpler in a complete market when the dual space consists of a singleton.
2.2. Solution to the Static Formulation. Rewrite the above static problem (5) by in-terchanging the order of infimum:
inf ξt CV aRλ(XT) = 1 λx∈Rinf inf X E[(x − X) + ] − λx (6) subject to E[X] = x˜ r, xd≤ X ≤ xu.
we arrive to the final form of the main optimization problem (2) where we provide a direct solution in two steps:
Step 1: Minimization of Expected Shortfall
v(x) = inf
X E[(x − X) +]
subject to E[X] = x˜ r, xd ≤ X ≤ xu,
Step 2: Minimization of CVaR
inf ξt
CV aRλ(XT) = 1
λx∈Rinf(v(x) − λx).
Schied [18] solved a general law invariant risk minimization problem of the type (6). We solve the CVaR minimization with the above two-step approach where we do not require the probability space to be atomless so the tree models are included. We give explicit computation methods for the Black-Scholes and Binomial model in Section 3.
The solution for Expected Shortfall Minimization is studied in a semimartingale model in F¨ollmer and Leukert [8] and Xu [20]. Apply Proposition 4.1 in [8] to the above shortfall problem, we get the following result.
Theorem 2.2 (Solution to Expected Shortfall Minimization Problem). Define a set A = n
where the constant a is to be determined on a case by case basis. X∗ defined below is the optimal solution to the shortfall problem in Step 1 and the corresponding value function v(x) is given.
Case 1: x ≤ xd:
X∗ = any random variable X with values in [xd, xu] where ˜E[X] = xr, and v(x) = 0. Case 2: xd≤ x ≤ xr < xu:
X∗ = any random variable X with values in [x, xu] where ˜E[X] = xr, and v(x) = 0. Case 3: xd< xr ≤ x ≤ xu:
X∗ = xdIAx+ kxICx+ xIBx, where Ax, Bx, Cx are decided by level ax defined as
ax = sup a : ˜P (B) ≤ xr− xd x − xd ,
and kx is chosen so that the constraint
xr = ˜E[X∗] = xdP (A˜ x) + kxP (C˜ x) + x ˜P (Bx) is satisfied: kx = xr− xdP (A˜ x) − x ˜P (Bx) ˜ P (Cx) 1{ ˜P (C x)>0}. v(x) = (x − xd)P (Ax) + (x − kx)P (Cx). Case 4: x ≥ xu:
X∗ = xdIA+ kIC + xuIB, where A, B, C are decided by level a defined as
a = sup a : ˜P (B) ≤ xr− xd xu− xd ,
and k is chosen so that the constraint xr = xdP (A) + k ˜˜ P (C) + xuP (B) is satisfied:˜
k = xr− xd ˜ P (A) − xuP (B)˜ ˜ P (C) 1{ ˜P (Cx)>0}. v(x) = (x − xd)P (A) + (x − k)P (C) + (x − xu)P (B).
Remark 2.3. The global minimum for function v(x) is 0. For the first two cases where x ≤ xr, the minimal value of 0 can be easily achieved by an admissible X∗ ≥ x, including the special example of X∗ ≡ xr that naturally satisfies the constraint of ˜E[X∗] = xr. For the latter two cases where x > xr, the solution is provided by Neyman-Pearson lemma. A part of the X∗ should be as large as possible to minimize v(x) on the good set ‘B’, while the other part should be taken at the lower bound to offset this large number so that the risk-neutral expectation of X∗ is guaranteed to stay at xr. In Case 3, we can equivalently define ax as
ax= inf a : ˜P (A) ≤ x − xr x − xd ,
because for fixed x level, Ax is the largest set satisfying ˜P (Ax) ≤ x−xx−xrd, and Bx is the largest set satisfying ˜P (Bx) ≤ xx−xr−xd
d . When there is one or more point mass at ax, set Cx is not empty and kx has to be chosen to satisfy the constraint of ˜E[X∗] = xr. When there is no point mass at ax, set Cx is empty, and we have exact equality ˜P (Ax) = x−xx−xr
d and ˜P (Bx) =
xr−xd
x−xd. Note that the sets A, B, C and the number k in Case 3 are functions of x, while in Case 4 they are not.
Remark 2.4. Using Theorem 2.2 to solve Step 2, we need to find the global minimum among four cases:
1 λx≤xinfd (v(x) − λx) = 1 λx≤xinfd (0 − λx) = −xd, 1 λxd≤x≤xinf r (v(x) − λx) = 1 λxd≤x≤xinf r (0 − λx) = −xr ≤ −xd, 1 λxr≤x≤xinf u (v(x) − λx) = 1 λxr≤x≤xinf u (x − xd)P (Ax) + (x − kx)P (Cx) − λx, 1 λx≥xinfu (v(x) − λx) = 1 λx≥xinfu (x − xd)P (A) + (x − k)P (C) + (x − xu)P (B) − λx.
The solutions for the first two cases are simple where we observe the second case dominate the first case. We rewrite the thirds case as
1 λxr≤x≤xinf u (v(x) − λx) = 1 λxr≤x≤xinf u ((x − xd)P (Ax) + (x − kx)P (Cx) − λx) = −xr+ 1 λxr≤x≤xinf u ((x − xd)P (Ax) + (x − kx)P (Cx) − λx + λxr) = −xr+ 1 λxr≤x≤xinf u ((x − xd)(P (Ax) − λ ˜P (Ax)) + (x − kx)(P (Cx) − λ ˜P (Cx))) = −xr+ 1 λxr≤x≤xinf u h(x)
where h(x) = (x − xd)(P (Ax) − λ ˜P (Ax)) + (x − kx)(P (Cx) − λ ˜P (Cx)), and solve the problem
inf xr≤x≤xu
h(x)
in Lemma 2.5. In case four, A, B, C and k are irrelevant to x. The minimization is simpler because it is linear in x with positive slope so the minimum is obtained when x = xu:
1 λx>xinfu (x − xd)P (A) + (x − k)P (C) + (x − xu)P (B) − λx = 1 λ (xu− xd)P (A) + (xu− k)P (C) − λxu ≥ 1 λxr<x≤xinf u (x − xd)P (Ax) + (x − kx)P (Cx) − λx.
We have shown here that the minimum obtained in the fourth case will not provide the global minimum because it is dominated by the results from the third case. It is easy to see that case two is also dominated by case three because it coincides with the result in case three when x = xr. Therefore, after Lemma 2.5, we arrive naturally to the result of Step 2 in Theorem 2.7.
Lemma 2.5. Recall sets A, B, C are related to the number a in the same manner as in Theorem 2.2, namely A = nω ∈ Ω : d ˜dPP(ω) > ao, B = nω ∈ Ω : d ˜dPP(ω) < ao, and C =
n
ω ∈ Ω : d ˜dPP(ω) = ao. For any fixed x, define
ax = inf a : ˜P (A) ≤ x − xr x − xd , kx = xr− xdP (A˜ x) − x ˜P (Bx) ˜ P (Cx) 1{ ˜P (Cx)>0},
where the sets are associated to x as Ax = n
ω ∈ Ω : d ˜dPP(ω) > ax o
, etc. Denote the pa-rameters amax, kmax, Amax, Bmax, Cmax corresponding to x = xu as the max-system; those parameters amin, kmin, Amin, Bmin, Cmin corresponding to x = xr as the min-system, those parameters a∗, k∗, A∗, B∗, C∗ corresponding to x = x∗ as the star-system. The solution to the minimization problem
inf xr≤x≤xu h(x), where h(x) = (x − xd)(P (Ax) − λ ˜P (Ax)) + (x − kx)(P (Cx) − λ ˜P (Cx)), is: • If d ˜P dP ≤ 1
λ, P −a.s., then infxr≤x≤xuh(x) = h(xr) = 0.
• Otherwise, if amax≥ 1− ˜λ−P (AP (Amaxmax)), then infxr≤x≤xuh(x) = h(xu). If amax<
1− ˜P (Amax)
λ−P (Amax), then infxr≤x≤xuh(x) = h(x
∗), where a∗ = supna : a < 1− ˜P (A) λ− ˜P (A) o , A∗ =nω ∈ Ω : d ˜P dP(ω) > a ∗o, B∗ =nω ∈ Ω : d ˜P dP(ω) < a ∗o, C∗ =nω ∈ Ω : d ˜P dP(ω) = a ∗o, x∗ = xr−xd( ˜P (A∗)+ ˜P (C∗)) ˜
P (B∗) , are the parameters that form the star-system defined above.
Remark 2.6. In the case where the probability space is atomless and the Radon-Nikod´ym derivative d ˜P
dP(ω) has continuous distribution, we have C = ∅, B = A
c, so the constraint
xr = xdP (A) + x ˜˜ P (B) + k ˜P (C) becomes xr = xdP (A) + x ˜˜ P (Ac), i.e., ax can be chosen for the precise equality ˜P (Ax) = x−xx−xrd to hold and kx = 0. The solution to the minimization problem infxr≤x≤xuh(x) where h(x) = (x − xd)(P (Ax) − λ ˜P (Ax)) is:
• If d ˜P dP ≤
1
λ, P −a.s., then infxr≤x≤xuh(x) = h(xr) = 0. • Otherwise, if amax≥
1− ˜P (Amax)
λ−P (Amax), then infxr≤x≤xuh(x) = h(xu). If amax < 1− ˜λ−P (AP (Amaxmax)), then infxr≤x≤xuh(x) = h(x
∗), where x∗ = xr−xdP (A˜ ∗)
1− ˜P (A∗) and A
∗ = n
Proof for Remark 2.6. Let us first prove 2.6 in the continuous distribution case. Suppose d ˜P
dP ≤ 1
λ, P −a.s. Then for any x ∈ [xr, xu],
˜ P (Ax) = Z Ax d ˜P dP(ω)dP (ω) ≤ 1 λP (Ax).
Thus h(x) = (x − xd)(P (Ax) − λ ˜P (Ax)) ≥ 0. When x = xr, P (Amin) = ˜P (Amin) = 0 and h(xr) = 0. Therefore,
inf xr≤x≤xu
h(x) = h(xr) = 0.
Now suppose P (d ˜dPP > 1λ) > 0. In this special case when d ˜dPP has a continuous distribution, we have the equality ˜P (Ax) = x−xx−xr
d where Ax = n
ω : d ˜dPP(ω) > ax o
, and we observe the following:
• Ax increases as x increases; ax decreases as x and Ax increases. Define function f (x) = x−xr
x−xd. We see that f (x) is an increasing function since f
0(x) = xr−xd
(x−xd)2 > 0. Notice that the probability function ˜P (Ax) is an increasing function of Ax, so x %⇔ f (x) %⇔ ˜P (Ax) %⇔ Ax %. It’s obvious from the definition that the set Ax %⇔ ax &.
• ˜P (Amin) = 0 and ˜P (Amax) = xxuu−x−xrd.
x ranges over the interval [xr, xu], thus f (x) range over h
0,xu−xr
xu−xd i
. Amin corresponds to the case x = xr. By the definition of function f (x), we know that f (xr) = 0 =
˜
P (Amin). It is obvious that ˜P (Amax) = f (xu) = xxuu−x−xrd. • d ˜P (Ax) dx = xr−xd (x−xd)2; D −P (A x) = ax−1 d ˜P (Adxx), D+P (Ax) = ax+1 d ˜P (Adxx). Use the definition of f (x), d ˜P (A)dx = f0(x) = xr−xd
(x−xd)2. When the Radon-Nikod´ym derivative d ˜dPP(ω) has continuous distribution, ax is a decreasing function of x where at times it can jump downward. Therefore, the left- and right-limit ax−and ax+exist. In fact, when ax− > ax+, ˜P (ax+ < d ˜dPP(ω) < ax−) = 0. Since P and ˜P are equivalent, P (ax+ < d ˜dPP(ω) < ax−) = 0 and P (Ax) = P (dPd ˜P(ω) > ax−) = P (d ˜dPP(ω) > ax+). If we
denote the left- and right-derivatives as D−P (Ax) = lim %0 P (Ax+) − P (Ax) , D+P (Ax) = lim &0 P (Ax+) − P (Ax) , we have D−P (Ax) = lim %0 P (Ax+) − P (Ax) = lim%0 E[1Ax+] − E[1Ax] = lim %0 ˜ E[dP d ˜P1d ˜dPP>ax+] − ˜E[ dP d ˜P1d ˜dPP>ax−] = lim%0 ˜ E[dP d ˜P1ax−<d ˜dPP≤ax+] − Since ˜ E[dP d ˜P1ax<d ˜dPP≤ax+] − ≥ 1 ax+ ˜ E[1a x<d ˜dPP≤ax+] − = 1 ax+ ˜ P (Ax+) − ˜P (Ax) → 1 ax− d ˜P (Ax) dx , and ˜ E[dP d ˜P1ax−<d ˜dPP≤ax+] − < 1 ax− ˜ E[1a x−<d ˜dPP≤ax+] − = 1 ax− ˜ P (Ax+) − ˜P (Ax) → 1 ax− d ˜P (Ax) dx , as % 0. We conclude the left derivative is
D−P (Ax) = 1 ax−
d ˜P (Ax) dx . Similarly, the right derivative is
D+P (Ax) = 1 ax+
d ˜P (Ax) dx .
If ax is continuous at x, i.e., ax+ = ax− = ax, then the derivative exists dP (Ax) dx = 1 ax d ˜P (Ax) dx = xr− xd ax(x − xd)2 .
Now let’s turn to the first derivative and second derivative of the function we would like to minimize:
h(x) = (x − xd)(P (Ax) − λ ˜P (Ax)),
and show it to be a convex function. On x ∈ (xr, xu), when ax is continuous at x, we have
h0(x) = (P (Ax) − λ ˜P (Ax)) + (x − xd) dP (Ax) dx − λ d ˜P (Ax) dx ! = P (Ax) − λ x − xr x − xd + (x − xd) xr− xd ax(x − xd)2 − λ xr− xd (x − xd)2 = P (Ax) − λ x − xr x − xd + 1 ax − λ xr− xd x − xd = P (Ax) − λ + 1 ax (1 − ˜P (Ax)).
When ax is discontinuous at x, we can define the left- and right-derivatives
D−h(x) = lim %0 h(x + ) − h(x) , D+h(x) = lim &0 h(x + ) − h(x) .
Similarly to the above calculation, we get
D−h(x) = P (Ax) − λ + 1 ax− (1 − ˜P (Ax)), D+h(x) = P (Ax) − λ + 1 ax+ (1 − ˜P (Ax)). When ax is continuous at x, ˜P (Ax) = ˜P (d ˜dPP(ω) > ax) = 1 − ˜P (d ˜dPP(ω) ≤ ax) = 1 − ˜F (ax), where ˜F (·) is the cumulative distribution function of the Radon-Nikod´ym derivative d ˜dPP. Since ˜P (Ax) = x−xx−xr
d, ˜F (ax) =
xr−xd
x−xd. We have also started by assuming
d ˜P
distribution, therefore the derivative of ˜F (·) exists and is the probability density function ˜
f (·). When ax− = ax+, ˜P (Ax) is strictly increasing as x increases, thus ˜f (ax) > 0. By Implicit Differentiation Theorem, the derivative of ax exists and can be computed as
(ax)0 = −
xr− xd ˜
f (ax)(x − xd)2 < 0.
With Chain Rule, we know
1 ax 0 = −a 0 x a2 x > 0. Now we can compute the second derivative of h(x):
h00(x) = dP (Ax) dx + 1 ax 0 (1 − ˜P (Ax)) − 1 ax d ˜P (Ax) dx = 1 ax 0 (1 − ˜P (Ax)) > 0.
Here 1 − ˜P (Ax) is positive since the maximum value of ˜P (Ax) is ˜P (Amax) = xxuu−x−xr
d < 1 by definition. Clearly, the second derivative indicates that h0(x) is strictly decreasing at those points x ∈ (xr, xu) where ax is continuous. When ax is discontinuous, we have
D−h(x) = P (Ax) − λ + 1 ax− (1 − ˜P (Ax)) < D+h(x) = P (Ax) − λ + 1 ax+ (1 − ˜P (Ax)).
We recognize that this is a kink point for h(x). Finally, we conclude h(x) is convex on (xr, xu). It is easy to see that h(x) is continuous at both left and right end points. Therefore, it is con-vex on the closed interval [xr, xu]. If we can find x∗ ∈ [xr, xu], where 0 ∈ [D−h(x∗), D+h(x∗)], then it is the minimum. Otherwise if D+h(x
r) ≥ 0 the infimum is obtained at x = xr; if D−h(xu) ≤ 0 the infimum is obtained at x = xu. If the derivative of h(x) exists at x = x∗, then the above condition collapse to h0(x∗) = 0, or equivalently, a∗ = λ−P (A1− ˜P (A∗∗)). When the derivative does not exist, it is obvious that 0 ∈ [D−h(x∗), D+h(x∗)] is still equivalent to the existence of a∗ ∈ [ax∗−, ax∗+] where a∗ = 1− ˜P (A
∗)
λ−P (A∗), and the corresponding x
∗ can be computed
from the equation ˜P (A∗) = x∗−xr
x∗−x d.
Recall that P (d ˜P dP > 1 λ) > 0. Since ˜P (Amin) = ˜P ( d ˜P dP > amin) = 0, amin > 1 λ. So D−h(xr) = P (Amin) − λ +amin1 (1 − ˜P (Amin)) = −λ +amin1 < 0. On the other end, D+h(xu) = P (Amax) − λ +amax1 (1 − ˜P (Amax)). We have D+h(xu) > 0 if and only if amax<
1− ˜P (Amax)
λ−P (Amax). In this case, the minimum occurs at x∗ ∈ (xr, xu) where a∗ = 1− ˜P (A
∗)
λ−P (A∗). If amax ≥ λ−P (A1− ˜P (Amaxmax)), the minimum is achieved at the right end point with the value h(xu).
Proof for Lemma 2.5. As in the proof for Remark 2.6, let ˜F (·) be the cumulative distribution function of the Radon-Nikod´ym derivative d ˜dPP. Then for fixed x, we have ˜F (ax) = 1 − ˜P (Ax). In the proof for Remark 2.6, we have assumed that d ˜P
dP has a continuous distribution. This essentially dealt with case where ˜F (·) is continuous: it could either be strictly increasing or flat. Now to deal with the general case for the distribution of d ˜dPP, we only need to discuss the remaining case where ˜F (·) has a jump, i.e., there is a point mass at d ˜dPP = ax. See Fig. 1:
Figure 1. ˜F (a) is the cumulative distribution function of the Radon-Nikod´ym derivative d ˜dPP.
Recall the definitions ax = inf a : ˜P (A) ≤ x − xr x − xd , Ax = ( ω ∈ Ω : d ˜P dP(ω) > ax ) , Cx = ( ω ∈ Ω : d ˜P dP(ω) = ax ) , Bx = ( ω ∈ Ω : d ˜P dP(ω) < ax ) , kx = xr− xdP (A˜ x) − x ˜P (Bx) ˜ P (Cx) 1{ ˜P (Cx)>0}, h(x) = (x − xd)(P (Ax) − λ ˜P (Ax)) + (x − kx)(P (Cx) − λ ˜P (Cx)),
and we would like to find
inf xr≤x≤xu
h(x).
When dPd ˜P has a point mass at ax, i.e., ˜P (dPd ˜P = ax) = ˜P (Cx) > 0, the distribution function ˜F has a jump at ax: ˜F (ax) − ˜F (ax−) = ˜P (Cx). As in the proof for Remark 2.6, we first discuss the case d ˜dPP ≤ 1
λ, P −a.s. Similarly we can show ˜P (Ax) ≤ 1
λP (Ax) and ˜P (Cx) ≤ 1
λP (Cx). It is easy to check that x − kx ≥ 0 when xr ≤ x ≤ xu. We conclude,
inf xr≤x≤xu
h(x) = h(xr) = 0.
Now suppose P (d ˜dPP > λ1) > 0. As soon as the equality ˜P (A) = x−xr
x−xd, we have ˜P (B) =
xr−xd
x−xd and ˜P (C) = 0. Therefore, we need only discuss two cases:
• ˜P (Ax) = x−xx−xrd, ˜P (Bx) = xx−xr−xdd and ˜P (Cx) = 0, • ˜P (Ax) < x−xx−xr
d, ˜P (Bx) <
xr−xd
x−xd and ˜P (Cx) > 0.
In the case when ˜P (A) < x−xr
x−xd, ˜P (B) <
xr−xd
x−xd and ˜P (C) > 0, the sets A, B, C as well as the point ax do not change with a small change in x. However, in this case where x increases from xr−xdP (A)˜ ˜ P (B)+ ˜P (C) to xr−xd( ˜P (A)+ ˜P (C)) ˜ P (B) , kx decreases from xr−xdP (A)˜ ˜
h(x) is easily calculated as h0(x) = (P (A) − λ ˜P (A)) + (1 −dkx dx )(P (C) − λ ˜P (C)) = P (A) − λ + 1 ax (1 − ˜P (A))
Since ax is not changing with respect to x in this case, h0(x) is constant on
x ∈ xr− xdP (A)˜ ˜ P (B) + ˜P (C), xr− xd( ˜P (A) + ˜P (C)) ˜ P (B) ! .
This formula is exactly the same as the one in the continuous case. Therefore, the convexity of h(x) is preserved on the interval
x ∈ " xr− xdP (A)˜ ˜ P (B) + ˜P (C), xr− xd( ˜P (A) + ˜P (C)) ˜ P (B) ! , i.e., D−h( xr− xd ˜ P (A) ˜ P (B) + ˜P (C)) ≤ h 0 (x), for x ∈ xr− xd ˜ P (A) ˜ P (B) + ˜P (C), xr− xd( ˜P (A) + ˜P (C)) ˜ P (B) ! .
At the other end point x = xr−xd( ˜P (A)+ ˜P (C))
˜
P (B) , kx dropped to rd and the old sets A and C combine to produce the new set Ax = AS C and we encounter the other case where
˜
P (Ax) = x−xx−xrd, ˜P (Bx) = xx−xr−xdd and ˜P (Cx) = 0. The left derivative at this point is computed above D−h(xr− xd( ˜P (A) + ˜P (C)) ˜ P (B) ) = P (A) − λ + 1 ax (1 − ˜P (A)),
where A is the old set before x reach the right end point. The right derivative is computed in the proof for Remark 2.6:
D+h(xr− xd( ˜P (A) + ˜P (C)) ˜ P (B) ) = P (Ax) − λ + 1 ax+ (1 − ˜P (Ax)).
The difference P (Ax) − λ + 1 ax (1 − ˜P (Ax)) − P (A) − λ + 1 ax (1 − ˜P (A)) = P (C) − 1 ax ˜ P (C) = 0,
by the definition of set C. Since a1 x+ ≥
1
ax, we keep the convexity of h(x) at x =
xr−xd( ˜P (A)+ ˜P (C)) ˜ P (B) because D−h(xr− xd( ˜P (A) + ˜P (C)) ˜ P (B) ) ≤ D +h(xr− xd( ˜P (A) + ˜P (C)) ˜ P (B) ).
We also conclude after this discussion that when there is a point mass at d ˜dPP = ax, the convex function h(x) becomes linear. In any case, combining the results we have just shown and those in the proof of Remark 2.6, we know that h(x) is convex all the time on x ∈ [xr, xu]. As shown in Fig. 2, in a case like the Binomial model where there are only point masses, h(x) is a piecewise constant convex function; in a case like the Black-Scholes model where the distribution is continuous and spans the whole positive part of the real line, h(x) is a continuously differentiable convex function.
Figure 2. The left picture is how h(x) look like in the Binomial model; the right pictures is for the Black-Scholes model.
The discussion about minimizing h(x) in the proof of Remark 2.6 is also valid here and we conclude our proof by verifying the star-system when it happens at the right end point
0 ∈ (D−h(xr−xd( ˜P (A)+ ˜P (C))
˜
P (B) ), D
+h(xr−xd( ˜P (A)+ ˜P (C))
˜
P (B) )). This means axis the largest value where
P (A) − λ + 1 ax
(1 − ˜P (A)) < 0.
Therefore, we get the desired condition
a∗ = sup ( a : a < 1 − ˜P (A) λ − ˜P (A) ) .
Theorem 2.7 (Solution to CVaR Minimization Problem (6)). Define the sets A, B, C and the numbers ax, kx the same way as in Lemma 2.5 and Theorem 2.2. Denote the min-system, max-system, star-system as in Lemma 2.5. The solution to problem (6) is as follows:
• If d ˜P dP ≤
1
λ, P −a.s., then X ∗ = x
r is optimal, and the minimal risk is CV aRλ(X∗) = −xr.
• Otherwise, if amax ≥
1− ˜P (Amax)
λ−P (Amax), then the optimal risk is achieved by the max-system. X∗ = xdIAmax+ kmaxICmax + xuIBmax is optimal, and the associated minimal risk is
CV aRλ(X∗) = −xr+ 1
λ[(xu−xd)(P (Amax) − λ ˜P (Amax)) + (xu−kmax)(P (Cmax) − λ ˜P (Cmax))]. If amax <
1− ˜P (Amax)
λ−P (Amax), then the optimal risk is achieved by the star-system. Set a
∗ = sup n a : a < 1− ˜λ− ˜P (A)P (A) o , A∗ = n ω ∈ Ω : d ˜dPP(ω) > a∗ o , B∗ = n ω ∈ Ω : d ˜dPP(ω) < a∗ o , C∗ = nω ∈ Ω : d ˜dPP(ω) = a∗o, x∗ = xr−xd( ˜P (A∗)+ ˜P (C∗)) ˜ P (B∗) . Then X ∗ = x dIA∗ + k∗IC∗ + x∗IB∗ is optimal, and the associated minimal risk is
CV aRλ(X∗) = −xr+ 1 λ[(x
∗ − x
d)(P (A∗) − λ ˜P (A∗)) + (x∗− k∗)(P (C∗) − λ ˜P (C∗))].
Remark 2.8. For the continuous case as in Remark 2.6 where C = ∅, B = Ac, and the constraint xr = xdIA+ xIAc, we can simplify the result as follows:
• If d ˜P dP ≤
1
λ, P −a.s., then X ∗ = x
r is optimal, and the minimal risk is
CV aRλ(X∗) = −xr.
• Otherwise, if amax ≥
1− ˜P (Amax)
λ−P (Amax), then set x
∗ = x
u. The optimal portfolio is X∗ = xdIAmax + x
∗I Ac
max, and the optimal risk is
CV aRλ(X∗) = −xr+ 1 λ(xu− xr)(P (Amax) − λ ˜P (Amax)). If amax < 1− ˜P (Amax) λ−P (Amax), then x ∗ = xr−xdP (A˜ ∗) 1− ˜P (A∗) and A ∗ = nω ∈ Ω : d ˜P dP(ω) > a ∗o satisfies
a∗ = λ−P (A1− ˜P (A∗∗)). The optimal portfolio is X
∗ = x
dIA∗+ x∗IA∗c, and the optimal risk is
CV aRλ(X∗) = −xr+ (xr− xd)(1 − 1 λa∗).
Proof. Theorem 2.2 gives the solution to the shortfall problem in Step 1. Leveraging these results, we have discussed in Remark 2.4 that the solution for step 2 is achieved by finding the solution to the third case of
1 λxr<x≤xinf u (v(x) − λx) = −xr+ 1 λxr<x≤xinf u h(x),
where h(x) = (x − xd)(P (Ax) − λ ˜P (Ax)) + (x − kx)(P (Cx) − λ ˜P (Cx)). Now combine the solution to
inf xr<x≤xu
h(x),
found in Lemma 2.5, we quickly arrive to the conclusion.
3.1. Binomial Model. In the recombining tree shown below, we have the following dy-namics for the stock and the self-financing portfolio:
Sn+1(H) = uSn, with P (ωn = H) = p and ˜P (ωn = H) = ˜p Sn+1(T ) = dSn, with P (ωn = T ) = q and ˜P (ωn= T ) = ˜q
Xn+1= ξnSn+1+ (Xn− ξnSn)(1 + r),
where ˜p, ˜q are risk-neutral probability, p, q are physical probability, u, d are the step size for up move and down move respectively, and r is the risk-free interest earned for one time step.
Our main goal is first to find
inf ξn
CV aRλ(XN) s.t. ˜E[XN] = xr, xd≤ XN ≤ xu,
then to find the corresponding dynamic hedging ξn.
The optimal final portfolio XN∗ and the minimal risk infξnCV aRλ(XN) can be found easily according to Theorem 2.7. To find the corresponding optimal hedging ξn, we first calculate Xn= (1+r1 )(N −n)E[X˜ N∗|Fn] for n ∈ {1, 2, ..., N − 1}, then compute ξn= XSn+1n+1(H)−X(H)−Sn+1n+1(T )(T ) by martingale representation theorem.
Algorithm 3.1. CVaR Minimization for Binomial Model
(1) We first arrange the states ω ∈ Ω such that PP˜(ω) is descending: PP˜(ω1) ≥ ˜ P
P(ω2) ≥ .... (2) If PP˜(ω1) ≤ 1λ, then the optimal risk is infξnCV aRλ(XN) = −xr, and the optimal
hedge is ξn= 0 for all n. Stop.
(3) If there is some ω such that PP˜(ω) > λ1, then we need to find the max-system.
(4) If amax > λ−P (A1− ˜P (Amaxmax)), then the optimal portfolio is XN∗ = xdIAmax+kmaxICmax+xuIBmax, and the optimal risk is infξnCV aRλ(XN) = −xr+
1
λ[(xu−xd)(P (Amax)−λ ˜P (Amax))+ (xu − kmax)(P (Cmax) − λ ˜P (Cmax))]. Go to (6).
(5) If amax < λ−P (A1− ˜P (Amaxmax)), then find the star-system where A∗ is the smallest set satisfying a < 1− ˜λ−P (A)P (A). Compute x∗ from the equation ˜P (A∗) = x−xr
x−xd, and set the rest elements of the star-system. The optimal portfolio is XN∗ = xdIA∗ + k∗IC∗ + x∗IB∗, and the optimal risk is infξnCV aRλ(XN) = −xr +
1 λ[(x
∗ − x
d)(P (A∗) − λ ˜P (A∗)) + (x∗ − k∗)(P (C∗) − λ ˜P (C∗))].
(6) Calculate Xn = (1+r1 )(N −n)E[X˜ N∗|Fn] for all n. (7) Calculate ξn = XSn+1(H)−Xn+1(T )
n+1(H)−Sn+1(T ).
3.2. Black-Scholes Model. Let’s turn our attention to the Black-Scholes model, a com-plete market model with continuously distributed stock price. The dynamics of the stock price and the self-financing portfolio are as follows:
dSt= St(µdt + σdWt)
dXt= ξtdSt+ (Xt− ξtSt)rdt.
Our main goal is first to find
inf ξt
CV aRλ(XT) s.t. ˜E[XT] = xr, xd ≤ XT ≤ xu,
then to find the corresponding dynamic hedging ξt.
According to Theorem 2.2, and 2.7 of the general complete market, we need to check whether d ˜P dP(ω) T ≤ 1
λ for all ω ∈ Ω, and if not, then whether amax >
1− ˜P (Amax)
λ−P (Amax). The Radon-Nikod´ym derivative for geometric Brownian motion is Zt := d ˜dPP|t= e−θWt−
θ2 2 t, where θ = µ−r σ . Obviously, d ˜dPP(ω) T > 1
λ for some ω ∈ Ω since ess sup ZT = ∞. To check the second inequality, and possibly to find solution to the equation a = 1− ˜λ−P (A)P (A), we need to find the explicit relation among the three elements ˜P (A), P (A) and a. Notice that
A = ( d ˜P dP T > a ) = ( WT √ T < − θ√T 2 − ln a θ√T ) = ( ˜W T √ T < θ√T 2 − ln a θ√T ) ,
we have P (A) = N (−θ √ T 2 − ln a θ√T), ˜ P (A) = N (θ √ T 2 − ln a θ√T),
where N (·) is the cumulative distribution function for normal distribution. Link the above two equations with the definition ˜P (A) = x−xr
x−xd, we can solve for ˜P (Amax), amax, P (Amax) sequentially: ˜ P (Amax) = xu− xr xu− xd amax = θ √ T " θ√T 2 − N −1 ( ˜P (Amax)) # P (Amax) = N (− θ√T 2 − ln amax θ√T ), and solve the equation a = 1− ˜λ−P (A)P (A) through its explicit form:
a = 1 − N ( θ√T 2 − ln a θ√T) λ − N (−θ √ T 2 − ln a θ√T) .
A direct application of Theorem 2.7 gives the optimal final portfolio XT∗ and the minimal CV aR. To find the corresponding optimal hedging ξt∗, we apply the martingale representa-tion theorem.
Xt = er(T −t)E[X˜ T∗|Ft] = e−r(T −t){x∗P˜
t(Ac∗) + xdP˜t(A∗)}
= e−r(T −t){x∗+ (xd− x∗) ˜Pt(A∗)},
where ˜Pt(A∗) is the conditional probability under risk-neutral measure. Since A := {ω : ZT > a}, where Ztis the Radon-Nikod´ym derivative process, we have then ZT = Zte−θ(WT−Wt)−
θ2
a on A. So ˜ Pt(A∗) = ˜Pt(Zte−θ(WT−Wt)− θ2 2(T −t)> a∗) = ˜Pt ˜WT − ˜Wt √ T − t < − lnZa∗ t θ√T − t + θ 2 √ T − t ! = N (− ln a∗ Zt θ√T − t + θ 2 √ T − t).
And Zt can be represented by the stock price St:
Zt= e−θWt− 1 2θ 2t ⇒ Wt = − ln Zt+12θ2t θ St = S0eσWt+(µ− 1 2σ 2)t ⇒ Wt = lnSt S0 − (µ − 1 2σ 2)t σ So, Zt = e θ σ[ µ+r−σ2 2 t−ln St S0],
Give the stock price St at time t, the optimal strategy ξt∗ is given by:
ξt∗ = dXt dSt = [dXt dZt ][dZt dSt ] = [e−r(T −t)(xd− x∗)( 1 √ 2πe −1 2(− ln a∗ Zt θ√T −t+ θ√T −t 2 ) 2 )( 1 Zt θ√T − t)][Zt(− θ σSt )] = e−r(T −t)(xd− x∗)( 1 √ 2πe −1 2(− ln a∗ Zt θ√T −t+ θ√T −t 2 ) 2 )(− 1 σSt √ T − t) = e−r(T −t)(xd− x∗)( 1 √ 2πe −12{ 1 θ√T −t[− ln a ∗+θ σ( µ+r−σ2 2 t−ln St S0)+ θ2 2(T −t)]} 2 )(− 1 σSt √ T − t). Algorithm 3.2. CVaR Minimization for Black-Scholes Model
(1) Calculate ˜P (Amax), P (Amax), amax.
If amax > λ−P (A1− ˜P (Amaxmax)), then set x∗ = xu, a∗ = amax, and infξtCV aRλ(XT) = −xr +
1
λ(xu − xr)(P (Amax) − λ ˜P (Amax)). Otherwise, find the unique solution x
∗, a∗, A∗ to
the equation a = λ−P (A)1− ˜P (A), and infξtCV aRλ(XT) = −xr+ (xr− xd)(1 −
1 λa∗).
(2) Compute optimal hedge ξt∗ = e−r(T −t)(xd− x∗)( 1 √ 2πe −1 2{ 1 θ√T −t[− ln a ∗+θ σ( µ+r−σ2 2 t−ln St S0)+ θ2 2(T −t)]} 2 )(− 1 σSt √ T − t). 4. Results: Dynamic & Static Hedging
4.1. Binomial Model. Dynamic Hedging
Take a 2-step binomial model as an example, where the sampling times are t = t0, t1, t2. Also assume p = 78, q = 18, u = 2, d = 12, r = 14, S0 = 4, X0 = 1, xu = 2, xd = 1, λ = 0.1. We calculate the risk-neutral probability ˜p = 1+r−du−d = 12, ˜q = u−1−ru−d = 12, and xr = (1 + r)2 = 2516.
˜ P (T T ) = 1 4, P (T T ) = 1 64, ˜ P P(T T ) = 16 ˜ P (T H) = 1 4, P (T H) = 7 64, ˜ P P(T H) = 16 7 ˜ P (HT ) = 1 4, P (HT ) = 7 64, ˜ P P(HT ) = 16 7 ˜ P (HH) = 1 4, P (HH) = 49 64, ˜ P P(HH) = 16 49.
By Algorithm 3.1, we first put the states in descending order: {(TT), (TH,HT), (HH)}. Since ˜
P
P(T T ) > 1
λ, we jump to step 3 and calculate the max-system. Amax is the largest set such that ˜P (A) < xu−xr
xu−xd = 0.4375, so Amax= {(T T )}. Similarly, Bmaxis the largest set such that ˜
P (B) < xr−xd
xu−xd = 0.5625, so Bmax= {(HH)}. And consequently Cmax = {(HT, T H)}, amax=
˜ P P(C) = 16 7 , kmax = xr−xdP (A˜ max)−xuP (B˜ max) ˜ P (Cmax) = 13
8. With the above calculated max-system, we see that amax < λ−P (A1− ˜P (Amaxmax)), so we proceed with step 5. The only set A that’s smaller than Amax is A = ∅. Its corresponding a = 16, and we have the inequality a > λ−P (A)1− ˜P (A). Therefore, the star-system includes A∗ = ∅, a∗ = 16 in this simple 2-step recombining binomial tree. We calculate the corresponding x∗ = 74, k∗ = 47, and the optimal risk is infξtCV aRλ(Xt) = −xr+λ1[(x∗−xd)(P (A∗)−λ ˜P (A∗))+(x∗−k∗)(P (C∗)−λ ˜P (C∗))] = −1.6328. Now we calculate the portfolio value X1(H), X1(T ). We know that X2(T T ) = xd, X2(T H) = X2(HT ) = k∗,
and X2(HH) = x∗, so X1(H) = 1 1 + r(˜px ∗ + ˜qk∗) = 1.4, X1(T ) = 1 1 + r(˜pk ∗+ ˜qx d) = 1.1.
And the hedging
ξ1(H) = X2(HH) − X2(HT ) S2(HH) − S2(HT ) = 0 ξ1(T ) = X2(T H) − X2(T T ) S2(T H) − S2(T T ) = 0.25 ξ0 = X1(H) − X1(T ) S1(H) − S1(T ) = 0.05. Static Hedging
If we can only determine the hedging at the beginning, i.e., ξ0, then the portfolio values along the binomial tree are X2 = ξ0S2+ (1 + r)2(X0− ξ0S0). To constrain X2 on the interval [xd, xu], we require ξ0 ∈ [−0.0577, 0.0449], and the resulting CV aRλ(X2) with λ = 0.1, is tabulated as follows: P (ω) X2(ω) ξ0 = -0.0577 ξ0 ∈ (−0.0577, 0) ξ0 = 0 ξ0 ∈ (0, 0.0449) ξ0 = 0.0449 49 64 X2(HH) 1 % 1.5625 % 2 7 64 X2(HT ) 1.6923 & 1.5625 & 1.4615 7 64 X2(T H) 1.6923 & 1.5625 & 1.4615 1 64 X2(T T ) 1.8654 & 1.5625 & 1.3269 CV aR0.1(X2) -1 -1 -1.5625 % -1.4405
Obviously the static hedging is not as good as the dynamic hedging since the optimal risk is infxi0CV aRλ(X2) = −1.5625, achieved at ξ0 = 0.
4.2. Black-Scholes Model. Assume λ = 0.1, T = 0.5, r = 0.05, S0 = 10, X0 = 10, xd = 5, xu = 30. Let’s compare the optimal risk from dynamic hedge and static hedge for three examples:
Example 1 Example 2 Example 3 µ 0.1 0.3 0.4 σ 0.2 0.1 0.08 Static CV aR -0.2532 -1.2131 -4.5710 Dynamic CV aR -10.2531 -11.8619 -27.2300 5. Conclusion
We have so far found exact solutions for CVaR minimization in complete market models like the Binomial model and the Black-Scholes model. When we compare the minimal risk resulting from dynamic portfolio management to a static portfolio management, we get very different numbers in magnitude. This suggests that the procedure provided in this paper help to find a dynamic solution that is far better than a static solution. The procedure
also works for higher dimensional case as long as the market is complete. However, the incomplete market is still an open question.
References
[1] Acerbi, C., D. Tasche (2002): “On the coherence of expected shortfall,” Journal of Banking and Finance, 26, 1487–1503.
[2] Alexander, S., T. F. Coleman, Y. Li (2004): “Derivative portfolio hedging based on CVaR”, Risk Measures for the 21st Century, John Wiley & Sons, 339–363.
[3] Artzner, P., F. Delbaen, J.-M. Eber, D. Heath (1997): “Thinking coherently,” Risk, 10, 68–71. [4] Artzner, P., F. Delbaen, J.-M. Eber, D. Heath (1999): “Coherent measures of risk,”
Mathe-matical Finance, 9, 203–228.
[5] Delbaen, F., W. Schachermayer (1994): “A general version of the fundamental theorem of asset pricing,” Mathematische Annalen, 300, 463–520.
[6] El Karoui, N., M. C. Quenez (1995): “Dynamic programming and pricing of contingent claims in an incomplete market,” SIAM Journal on Control and Optimization, 33, 1, 29–66.
[7] F¨ollmer, H., Yu. M. Kabanov (1998): “Optional decomposition and lagrange multipliers,” Finance and Stochastics, 2, 69–81.
[8] F¨ollmer, H., P. Leukert (2000): “Efficient hedging: Cost versus shortfall risk,” Finance and Stochastics, 4, 117–146.
[9] F¨ollmer, H., A. Schied (2002): Stochastic finance - an introduction in discrete time, Walter de Gruyter, Studies in Mathematics, 27.
[10] Glasserman, P. (2004): Monte Carlo Methods in Financial Engineering, Springer.
[11] Kramkov (1996): “Optional decomposition of supermartingales and hedging contingent claims in incomplete security markets,” Probability Theory and Related Fields, 105, 459–479.
[12] Markowitz, H. (1952): “Portfolio Selection”, The Journal of Finance, 7,1, 77–91.
[13] Rockafellar, R. T., S. Uryasev (2000): “Optimization of Conditional Value-at-Risk,” The Journal of Risk, 2, 4, 21–51.
[14] Rockafellar, R. T., S. Uryasev (2002): “Conditional value-at-risk for general loss distributions,” Journal of Banking and Finance, 26, 1443–1471.
[15] Ruszczy´nski, A., G. CH. Pflug (2004): “A risk measure for income proceses”, Risk Measures for the 21st Century, John Wiley & Sons, 249–269.
[16] Ruszczy´nski, A., G. CH. Pflug (2005): “Measuring risk for income streams”, Computational Optimization and Applications, 32, 249–269.
[17] Ruszczy´nski, A., A. Shapiro (2006): “Conditional risk mapping,” Mathematics of Operations Re-search, 31, 3, 544–561.
[18] Schied, A. (2004): “On the Neyman-Pearson problem for law-invariant risk measures and robust utility functionals,” The Annals of Applied Probability, 14, 3, 1398–1423.
[19] Shreve, S. E. (2004): Stochastic Calculus for Finance II - Continuous-Time Models, Springer. [20] Xu, M. (2004): “Minimizing shortfall risk using duality approach - an application to partial hedging
