Moment Explosions and Long-Term Behavior of Affine Stochastic Volatility Models
aa r X i v : . [ q -f i n . P R ] O c t MOMENT EXPLOSIONS AND LONG-TERM BEHAVIOR OFAFFINE STOCHASTIC VOLATILITY MODELS
MARTIN KELLER-RESSEL
Abstract.
We consider a class of asset pricing models, where the risk-neutraljoint process of log-price and its stochastic variance is an affine process in thesense of Duffie, Filipovic, and Schachermayer [2003]. First we obtain condi-tions for the price process to be conservative and a martingale. Then wepresent some results on the long-term behavior of the model, including an ex-pression for the invariant distribution of the stochastic variance process. Westudy moment explosions of the price process, and provide explicit expressionsfor the time at which a moment of given order becomes infinite. We discussapplications of these results, in particular to the asymptotics of the impliedvolatility smile, and conclude with some calculations for the Heston model, amodel of Bates and the Barndorff-Nielsen-Shephard model. Introduction
Duffie, Pan, and Singleton [2000] introduced the notion of an affine jump-diffusion,which is a jump-diffusion process, whose drift vector, instantaneous covariance ma-trix and arrival rate of jumps all depend in an affine way on the state vector.Duffie, Pan, and Singleton remark that models built on affine processes provide abalanced tradeoff between analytical tractability and complexity, making them anattractive choice for applications in mathematical finance. In particular they men-tion applications to the pricing of options in stochastic volatility models and notethat the models of Heston [1993], Bates [1996, 2000] , and Bakshi, Cao, and Chen[1997] fall into the affine class. To these, we could add the more recent models ofBarndorff-Nielsen and Shephard [2001] and Carr and Wu [2004], which are also ofaffine type.Duffie, Filipovic, and Schachermayer [2003] subsequently extended the class of affinejump-diffusions, defining an affine process as a time-homogenous Markov process,whose characteristic function is the exponential of an affine function of the statevector. It turns out that this class coincides for a large part with the class of affinejump-diffusions, but also allows for infinite activity of jumps and for killing or ex-plosions of the process. Duffie, Filipovic, and Schachermayer aim to give a rigorousmathematical foundation to the theory of affine processes, covering many aspects,such as the characterization of an affine process in terms of the ‘admissible param-eters’ (comparable to the characteristic triplet of a L´evy process) and properties ofthe ordinary differential equations (‘generalized Riccati equations’) that are impliedby the process.
Key words and phrases. affine process, stochastic volatility, moment explosions, implied volatilitysmile.Supported by the Austrian Science Fund (FWF) through the START programm Y328.
In this article we study stochastic volatility models, comprised of a log-priceprocess ( X t ) t ≥ and a stochastic variance process ( V t ) t ≥ , such that the joint pro-cess ( X t , V t ) t ≥ is an affine process. We will show that many properties of such amodel, including its long-term behavior and moment explosions, can be analyzedby studying differential equations of the generalized Riccati type. Our results onthe long-term behavior are formulated as asymptotic results for the cumulant gen-erating function of the stock price, as time goes to infinity. Asymptotics of thistype have been used by Lewis [2000] to obtain large-time-to-maturity results forthe implied volatility smile of stochastic volatility models via a saddlepoint expan-sion. The issue of moment explosions in stochastic volatility models has recentlyreceived much attention, due to the articles of Andersen and Piterbarg [2007] andLions and Musiela [2007]. Moment explosions are intimately connected to large-strike asymptotics of the implied volatility smile via results of Lee [2004], that havelater been expanded by Benaim and Friz [2006].In the first part of the paper we introduce our main assumption, that the jointprocess ( X t , V t ) t ≥ is affine, and recapitulate the main results of Duffie et al. [2003].We derive necessary and sufficient conditions for conservativeness of the processand for the martingale property of the discounted price process S t = exp( X t ). Atthe end of Section 2 we add two assumptions, and give a precise definition of theclass of affine stochastic volatility models, which constitutes the main subject of thisarticle. In Section 3 we derive our central results on long-term properties of an affinestochastic volatility model, providing conditions for the existence of an invariantdistribution of the stochastic variance process, and characterizing this distributionin terms of its cumulant generating function. We also give results on the long-termproperties of the price process, showing that as time tends to infinity, the marginaldistributions of the price process approach those of an exponential-L´evy process.The characteristic exponent of this L´evy process can be derived directly from thespecification of the affine stochastic volatility model. Both results are obtained byapplying qualitative ODE theory to the generalized Riccati equations introducedin the first part.In Section 4 we study moment explosions of the price process, and show that anexplicit representation for the time of moment explosion can be given – not onlyfor the primary model, but also for the model in the stationary variance regime.In Section 5 we outline applications of our results to the asymptotics of impliedvolatilities and of implied forward volatilities. We briefly discuss the results of Lee[2004] and point out the connection between the stationary variance regime andthe pricing of forward-start options, when the time until the start of the contractis large. We conclude in Section 6 with explicit calculations for several models towhich our results apply, such as the Heston model, a Heston model with addedjumps, a model of Bates, and the Barndorff-Nielsen-Shephard model.2. Affine Stochastic Volatility Models
Definition and the generalized Riccati equations.
We consider an asset-pricing model of the following kind: The interest rate r is non-negative and constant,and the asset price ( S t ) t ≥ is given by S t = exp( rt + X t ) t ≥ , FFINE STOCHASTIC VOLATILITY MODELS 3 such that ( X t ) t ≥ is the discounted log-price process starting at X ∈ R a.s. Thediscounted price process is simply exp( X t ), such that we will assume in the remain-der that r = 0, and that ( S t ) t ≥ is already discounted. Denote by ( V t ) t ≥ anotherprocess, starting at V > X t ) t ≥ , but may also control the arrival rate of jumps. The followingassumptions are made on the joint process ( X t , V t ) t ≥ : A1: ( X t , V t ) t ≥ is a stochastically continuous, time-homogeneous Markov pro-cess. A2:
The cumulant generating function Φ t ( u, w ) of ( X t , V t ) is of a particularaffine form: We assume that there exist functions φ ( t, u, w ) and ψ ( t, u, w )such thatΦ t ( u, w ) := log E [ exp( uX t + wV t ) | X , V ] = φ ( t, u, w ) + V ψ ( t, u, w ) + X u for all ( t, u, w ) ∈ R > × C , where the expectation exists.By convention, the logarithm above denotes the principal branch of the complexlogarithm. Assumptions A1 and A2 make ( X t , V t ) t ≥ an affine process in the senseof Duffie et al. [2003]. The term X u in the cumulant generating function Φ t ( u, w )corresponds to a reasonable homogeneity assumption on the model: If the startingvalue X of the price process is shifted by x , also X t is simply shifted by x for any t ≥
0. Note that Assumption A2 also implies that the variance process ( V t ) t ≥ is aMarkov process in its own right. We do not yet make the assumption that ( S t ) t ≥ is conservative (i.e. without explosions or killing) or even a martingale. Instead itwill be our first goal in Section 2.2 to obtain necessary and sufficient conditions forthese properties.Applying the law of iterated expectations to Φ t ( u, w ) yields the following ‘flow-equations’ for φ and ψ : (see also Duffie et al. [2003, Eq. (3.8)–(3.9)]) φ ( t + s, u, w ) = φ ( t, u, w ) + φ ( s, u, ψ ( t, u, w )) ,ψ ( t + s, u, w ) = ψ ( s, u, ψ ( t, u, w )) , (2.1)for all t, s ≥
0. The following result will be crucial:
Theorem 2.1.
Suppose that | φ ( τ, u, η ) | < ∞ and | ψ ( τ, u, η ) | < ∞ for some ( τ, u, η ) ∈ R > × C . Then, for all t ∈ [0 , τ ] and w ∈ C with Re w ≤ Re η | φ ( t, u, w ) | < ∞ , | φ ( t, u, w ) | < ∞ , and the derivatives (2.2) F ( u, w ) := ∂∂t φ ( t, u, w ) (cid:12)(cid:12)(cid:12)(cid:12) t =0+ , R ( u, w ) := ∂∂t ψ ( t, u, w ) (cid:12)(cid:12)(cid:12)(cid:12) t =0+ exist. Moreover, for t ∈ [0 , τ ) , φ and ψ satisfy the generalized Riccati equations ∂ t φ ( t, u, w ) = F ( u, ψ ( t, u, w )) , φ (0 , u, w ) = 0(2.3a) ∂ t ψ ( t, u, w ) = R ( u, ψ ( t, u, w )) , ψ (0 , u, w ) = w . (2.3b)The above theorem is ‘essentially’ proven in Duffie et al. [2003], but under slightlydifferent conditions . Note that the differential equations (2.3) follow immediatelyfrom the flow equations (2.1) by taking the derivative with respect to s , and evalu-ating at s = 0. They are called generalized Riccati Equations since they degenerate Duffie et al. assume differentiability of φ and ψ with respect to t (’regularity’) a priori, while inour case we can deduce it directly from Assumption A2. A proof is given in the appendix. MARTIN KELLER-RESSEL into (classical) Riccati equations with quadratic functions F and R , if ( X t , V t ) t ≥ is a pure diffusion process.Note that the first Riccati equation is just an integral in disguise, and φ may bewritten explicitly as(2.4) φ ( t, u, w ) = Z t F ( u, ψ ( s, u, w )) ds . Also the solution ψ of the second Riccati equation can be represented at leastimplicitly in the following way: Suppose that ψ ( t, u, w ) is a non-stationary localsolution on [0 , δ ) of (2.3b). Then R ( u, ψ ( t, u, w )) = 0 for all t ∈ [0 , δ ), and ψ ( t, u, w )is a strictly monotone function of t ; dividing both sides of (2.3b) by R ( u, ψ ( t, u, w )),integrating from 0 to t < δ , and substituting η = ψ ( s, u, w ) yields(2.5) Z ψ ( t,u,w ) w dηR ( u, η ) ds = t . Another important result that can be found in Duffie et al. [2003] states that F and R must be of L´evy-Khintchine form, i.e. F ( u, w ) = ( u, w ) · a · (cid:18) uw (cid:19) + b · (cid:18) uw (cid:19) − c (2.6a) + Z D \{ } (cid:18) e xu + yw − − ω F ( x, y ) · (cid:18) uw (cid:19)(cid:19) m ( dx, dy ) ,R ( u, w ) = ( u, w ) · α · (cid:18) uw (cid:19) + β · (cid:18) uw (cid:19) − γ (2.6b) + Z D \{ } (cid:18) e xu + yw − − ω R ( x, y ) · (cid:18) uw (cid:19)(cid:19) µ ( dx, dy )where D = R × R > , and ω F , ω R are suitable truncation functions, which we fixby defining ω F ( x, y ) = (cid:18) x x (cid:19) and ω R ( x, y ) = (cid:18) x x y y (cid:19) . Moreover the parameters ( a, α, b, β, c, γ, m, µ ) satisfy the following admissibilityconditions: • a, α are positive semi-definite 2 × a = a = a = 0. • b ∈ D and β ∈ R . • c, γ ∈ R > • m and µ are L´evy measures on D , and R D \{ } (cid:0) ( x + y ) ∧ (cid:1) m ( dx, dy ) < ∞ .The affine form of the cumulant generating function, the generalized Riccatiequations and finally the L´evy-Khintchine decomposition (2.6) lead to the follow-ing interpretation of F and R : F characterizes the state-independent dynamic ofthe process ( X t , V t ) while R characterizes its state-dependent dynamic. Both F and R decompose into a diffusion part, a drift part, a jump part and an instan-taneous killing rate. Hence a + αV t can be regarded as instantaneous covariancematrix of ( X t , V t ) t ≥ , b + V t β as the instantaneous drift, m ( dx, dy ) + V t µ ( dx, dy )as instantaneous arrival rate of jumps with jump heights in ( dx × dy ), and finally c + γV t as the instantaneous killing rate. FFINE STOCHASTIC VOLATILITY MODELS 5
The following Lemma establishes some important properties of F and R as func-tions of real-valued arguments. A proof is given in the appendix. Lemma 2.2. (a) F and R are proper closed convex functions on R .(b) F and R are analytic in the interior of their effective domain.(c) Let U be a one-dimensional affine subspace of R . Then F | U is either a strictlyconvex or an affine function. The same holds for R | U .(d) If ( u, w ) ∈ dom F , then also ( u, η ) ∈ dom F for all η ≤ w . The same holds for R .Remark . As usual in convex analysis, we regard F and R as functions definedon all of R , that may attain values in R ∪ { + ∞} . The set { ( u, w ) : F ( u, w ) < ∞} is called effective domain of F , and denoted by dom F .We define a function χ ( u ), that will appear in several conditions throughout thisarticle. Corollary 3.5 gives an interpretation of χ as a rate of convergence for theasymptotic behavior of the cumulant generating function of ( X t ) t ≥ . Definition 2.4.
For each u ∈ R where R ( u, < ∞ , define χ ( u ) as χ ( u ) := ∂R∂w ( u, w ) (cid:12)(cid:12)(cid:12)(cid:12) w =0 .χ ( u ) is well-defined at least as a limit as w ↑
0, possibly taking the value + ∞ ;it can be written explicitly as χ ( u ) = α u + β + Z D \{ } y (cid:18) e xu −
11 + y (cid:19) µ ( dx, dy ) . Note that also χ ( u ) is a convex function.2.2. Explosions and the martingale property.
We are interested in conditionsunder which S t = exp( X t ) is conservative and a martingale. If such conditions aresatisfied, ( S t ) t ≥ may serve as the price process under the risk-neutral measure inan arbitrage-free asset pricing model. The following theorem gives sufficient andnecessary conditions: Theorem 2.5.
Suppose ( X t , V t ) satisfies Assumptions A1 and A2. Then the fol-lowing holds:(a) ( S t ) t ≥ is conservative if and only if F (0 ,
0) = R (0 ,
0) = 0 and (2.7) Z − dηR (0 , η ) = −∞ ; (b) ( S t ) t ≥ is a martingale if and only if it is conservative, F (1 ,
0) = R (1 ,
0) = 0 and (2.8) Z − dηR (1 , η ) = −∞ . Remark . The notation R − denotes an integral over an arbitrarily small leftneighborhood of 0.By (2.6) the condition F (0 ,
0) = R (0 ,
0) = 0 is equivalent to c = γ = 0, i.e.obviously the killing rate has to be zero for the process to be conservative. Aswill be seen in the proof, the integral conditions (2.7) and (2.8) are related to a MARTIN KELLER-RESSEL uniqueness condition for non-Lipschitz ODEs, which has been discovered by Osgood[1898].The following Corollary gives easy-to-check sufficient conditions:
Corollary 2.7.
Suppose ( X t , V t ) satisfies Assumptions A1 and A2.(a) If F (0 ,
0) = R (0 ,
0) = 0 and χ (0) < ∞ then ( S t ) t ≥ is conservative.(b) If ( S t ) t ≥ is conservative, F (1 ,
0) = R (1 ,
0) = 0 and χ (1) < ∞ , then ( S t ) t ≥ isa martingale.Proof. For a proof of 2.5a we refer to [Filipovi´c, 2001, Th. 4.11]. Statement 2.5bcan be shown in a similar way:Since ( X t , V t ) is Markovian, we have for all 0 ≤ s ≤ t , that E [ S t |F s ] = S s exp ( φ ( t − s, ,
0) + V s ψ ( t − s, , . We have assumed that V > S t ) t ≥ is a martingale if and only if( X t ) t ≥ is conservative and ψ ( t, ,
0) = φ ( t, , ≡ t ∈ R > .We show Corollary 2.7 and the first implication of 2.5b: Suppose that ( S t ) t ≥ isconservative and that F (1 ,
0) = R (1 ,
0) = 0. By Theorem 2.1 ψ ( t, , w ) solves thedifferential equation ∂∂t ψ ( t, , w ) = R (1 , ψ ( t, , w )) , ψ (0 , , w ) = w (2.9)for all w ≤
0. Since R (1 ,
0) = 0 it is clear that e ψ ( t, , ≡ w = 0. To deduce that e ψ ( t, ,
0) = ψ ( t, ,
0) however, we need toknow whether the solution is unique. Since R (1 , w ) is continuously differentiablefor w <
0, it satisfies a Lipschitz condition on ( −∞ , χ (1) < ∞ , the Lipschitzcondition can be extended to ( −∞ , ψ ( t, , ≡ χ (1) < ∞ , we substitute Lipschitz’ condition byOsgood’s condition (2.8): Suppose that (2.8) holds, and there exists a non-zerosolution e ψ such that e ψ ( t , , < t >
0. Then for all t < t such that ψ remains non-zero on [ t, t ] we have (similarly to (2.5)) that(2.10) Z e ψ ( t, , e ψ ( t , , dηR (1 , η ) = t − t . Assume that t ≥ t such that e ψ ( t , , w ) = 0. Letting t ↓ t , the left side of (2.10) tends to −∞ , whereas the right side remains bounded,leading to a contradiction. We conclude that ψ ( t, , ≡ F (1 ,
0) = 0 yields that also φ ( t, , ≡ t ∈ R > and we have shown that ( S t ) t ≥ is a martingale.For the other direction of 2.5b note that ( S t ) t ≥ being a martingale impliesthat φ = ψ ≡ F (1 ,
0) = R (1 ,
0) = 0. It remains to show (2.8). Assume that (2.8) does not hold. Then,for each t >
0, (2.10) with t = 0 implicitly defines a solution e ψ ( t, ,
0) of thegeneralized Riccati equation (2.9), satisfying e ψ ( t, , < t >
0. By unique-ness of the solution ψ ( t, , w ) for w < e ψ ( t + s, ,
0) = ψ ( t, , e ψ ( s, , t, s small enough. Letting s ↓ ψ ( t, ,
0) = e ψ ( t, , <
0, which is a contradiction to ψ ≡ (cid:3) See Osgood [1898]
FFINE STOCHASTIC VOLATILITY MODELS 7
We add now two assumptions to A1 and A2 and complete our definition of anaffine stochastic volatility model:
A3:
The discounted price process S t = e X t is a martingale. A4: R ( u, = 0 for some u ∈ R .Assumption A4 excludes models where the distribution of ( X t ) t ≥ does not de-pend at all on the volatility state V . In such a case we can not speak of a truestochastic volatility model, and it will be beneficial to avoid these degenerate cases.We are now ready to give our definition of an affine stochastic volatility model: Definition 2.8.
The process ( X t , V t ) t ≥ is called an affine stochastic volatilitymodel, if it satisfies assumptions A1 – A4.A simple consequence of this definition, that will often be used is the following: Lemma 2.9.
Let ( X t , V t ) t ≥ be an affine stochastic volatility model. Then R ( u, is a strictly convex function, satisfying R (0 ,
0) = R (1 ,
0) = 0 .Proof.
From assumption A3 and Theorem 2.5 it follows that R (0 ,
0) = R (1 ,
0) =0. Lemma 2.2 implies that R ( u,
0) is either strictly convex or an affine function.Assume it is affine. Then R ( u,
0) = 0 for all u ∈ R . This contradicts A4, such thatwe conclude that R ( u,
0) is a strictly convex function. (cid:3) Long-term asymptotics
In this section we study the behavior of an affine stochastic volatility model as t → ∞ . We focus first on the stochastic variance process ( V t ) t ≥ . Under mildassumptions this process will converge in law to its invariant distribution:3.1. Stationarity of the variance process.Proposition 3.1.
Suppose that A1 and A2 hold, that χ (0) < and the L´evymeasure m satisfies the logarithmic moment condition Z y> (log y ) m ( dx, dy ) < ∞ . Then ( V t ) t ≥ converges in law to its unique invariant distribution L , which has thecumulant generating function (3.1) l ( w ) = Z w F (0 , η ) R (0 , η ) dη ( w ≤ . Keller-Ressel and Steiner [2008] show that under the given conditions the pro-cess ( V t ) t ≥ converges in law to a limit distribution L , whose cumulant generatingfunction can be represented by (3.1). A short argument at the end of this para-graph shows that the limit distribution is also the unique invariant distributionof ( V t ) t ≥ . First we make the following definition: Given some affine stochasticvolatility model ( X t , V t ) t ≥ , we introduce the process ( e X t , e V t ) t ≥ , defined as theMarkov process with the same transition probabilities as ( X t , V t ) t ≥ , but startedwith X = 0 and V distributed according to L . We will refer to ( e X t , e V t ) t ≥ as thestochastic volatility model ( X t , V t ) t ≥ ‘in the stationary variance regime’. We alsodefine the associated price process e S t := exp( rt + e X t ). As we discuss in Section 5the process ( e X t , e V t ) t ≥ can be related to the pricing of forward-starting options, MARTIN KELLER-RESSEL when the time until the start of the contract is large.The cumulant generating function of ( e X t , e V t ) is given by(3.2)log E [ e u e X t + w e V t ] = log E h exp (cid:16) φ ( t, u, w ) + e V ψ ( t, u, w ) (cid:17)i = φ ( t, u, w )+ l ( ψ ( t, u, w )) . We verify now that L is indeed an invariant distribution of ( V t ) t ≥ :(3.3) E h exp (cid:16) w e V t (cid:17)i = exp ( φ ( t, , w ) + l ( ψ ( t, , w ))) == exp Z t F (0 , ψ ( s, , w )) ds + Z ψ ( t, ,w ) F (0 , η ) R (0 , η ) dη ! == exp Z ψ ( t, ,w ) w F (0 , η ) R (0 , η ) dη + Z ψ ( t, ,w ) F (0 , η ) R (0 , η ) dη ! = exp( l ( w )) , where we have used that under the conditions of the Proposition above, ψ ( t, , w ) isa strictly monotone function converging to 0 as t → ∞ . (cf. Keller-Ressel and Steiner[2008]). To see that L is unique, assume that there exists another invariant dis-tribution L ′ , and let ( V ′ t ) t ≥ be the variance process started with V ′ distributedaccording to L ′ . Again we use that φ ( t, u, w ) → l ( w ) and ψ ( t, , w ) → t → ∞ (see Keller-Ressel and Steiner [2008]), and get thatlim t →∞ E [exp( wV ′ t )] = E h lim t →∞ exp ( φ ( t, , w ) + V ′ ψ ( t, , w )) i = E [exp( l ( w ))] = e l ( w ) , for all w ≤
0, in contradiction to the invariance of L ′ .3.2. Long-term behavior of the log-price process.
We have seen that ( V t ) t ≥ converges to a limit distribution, but we do not expect the same for the log-priceprocess ( X t ) t ≥ . Nevertheless, it can be shown that the rescaled cumulant gener-ating function t log E (cid:2) e X t u (cid:3) converges under suitable conditions to a limit h ( u ),that is again the cumulant generating function of some infinitely divisible randomvariable. This result can be interpreted such, that for large t the marginal distri-butions of ( X t ) t ≥ are ‘close’ to the marginal distributions of a L´evy process withcharacteristic exponent h ( u ). Furthermore, h ( u ) can be directly obtained from thefunctions F and R , without knowledge of the explicit forms of φ and ψ . We startwith a preparatory Lemma: Lemma 3.2.
Let ( X t , V t ) t ≥ be an affine stochastic volatility model and supposethat χ (0) < and χ (1) < . Then there exist a maximal interval I and a uniquefunction w ∈ C ( I ) ∩ C ( I ◦ ) , such that R ( u, w ( u )) = 0 for all u ∈ I and w (0) = w (1) = 0 .Moreover it holds that [0 , ⊆ I , w ( u ) < for all u ∈ (0 , ; w ( u ) > for all u ∈ I \ [0 , ; and (3.4) ∂R∂w ( u, w ( u )) < for all u ∈ I ◦ . We show Lemma 3.2 together with the next result, which makes the connectionto the qualitative properties of the generalized Riccati equations.
FFINE STOCHASTIC VOLATILITY MODELS 9
Lemma 3.3. (a) For each u ∈ I ◦ , w ( u ) is an asymptotically stable equilibriumpoint of the generalized Riccati equation (2.3b) .(b) For u ∈ I ◦ , there exists at most one other equilibrium point e w ( u ) = w ( u ) , andif it exists, it is necessarily unstable and satisfies e w ( u ) > max(0 , w ( u )) .(c) For u ∈ R \ I , no equilibrium point exists.Proof. Define L = { ( u, w ) : R ( u, w ) ≤ } . As the level set of the closed con-vex function R , it is a closed and convex set. For all u ∈ R , define w ( u ) =inf { w : ( u, w ) ∈ L } , and I = { u ∈ R : w ( u ) < ∞} . Clearly w ( u ) is a continuousconvex function, and I a subinterval of R . We will now show that w ( u ) and I satisfy all properties stated in Lemma 3.2. By assumption A3 and Theorem 2.5, R (0 ,
0) = R (1 ,
0) = 0; together with Lemma 2.2 it follows that the set [0 , × ( −∞ , R . Since R ( u,
0) is by Lemma 2.9 strictly convex, and also χ ( u ) is convex, we deduce that R ( u, < ∂R∂w ( u,
0) = χ ( u ) < u ∈ (0 , R ( u, w ), as a function of w , is either affine or strictly convex,such that there exists a unique point w ( u ), where R ( u, w ( u )) = 0, and necessarily ∂R∂w ( u, w ( u )) <
0. It is clear that for u ∈ (0 , w ( u ) coincides with the functiondefined above, and that w ( u ) <
0. At u = 0 we have that R (0 ,
0) = 0 and χ (0) < w (0) = 0. A symmetrical argument at u = 1 shows that w (1) = 0,and thus that [0 , ⊆ I .We show next that w ( u ) ∈ C ( I ◦ ): Define u + = sup I , and w + = lim u ↑ u + w ( u ); u − , w − are defined symmetrically at the left boundary of I . Note that u ± and w ± can take infinite values. Define the open set K := { ( λu − + (1 − λ ) u + , w ) : λ ∈ (0 , , w < λw − + (1 − λ ) w + } . Lemma 2.2 implies that K is contained in the interior of dom R . On the otherhand, the graph of w , restricted to I ◦ , i.e. the set { ( u, w ( u )) : u ∈ I ◦ } , is clearlycontained in K . Since R is by Lemma 2.2 an analytic function in the interior ofits effective domain, the implicit function theorem implies that w ( u ) ∈ C ( I ◦ ).In addition it follows that ∂R∂w ( u, w ( u )) = 0 for all u ∈ I ◦ , such that the asser-tion ∂R∂w ( u, w ( u )) <
0, which we have shown above for u ∈ (0 , I ◦ . The claim that w ( u ) > u ∈ I \ [0 ,
1] can easily be derived fromthe convexity of w ( u ), and the fact that w ( u ) < ,
1) and w (0) = w (1) = 0.We have now proved most part of Lemma 3.2 (except for the uniqueness), andturn towards Lemma 3.3: Since R ( u, w ( u )) = 0 and ∂R∂w ( u, w ( u )) < u ∈ I ◦ , w ( u ) must be an asymptotically stable equilibrium point of the generalized Riccatiequation 2.3b, showing 3.3a. Assume now that for some u ∈ I ◦ there exists a point e w ( u ) = w ( u ) such that R ( u, e w ( u )) = 0. By Lemma 2.2, R ( u, w ) is, as a functionof w , either strictly convex or affine. If it is affine, it has a unique root, and e w ( u )cannot exist. If it is strictly convex, there can exist a single point e w ( u ) other than w ( u ), such that R ( u, e w ( u )) = 0. Necessarily e w ( u ) > w ( u ) and ∂R∂w ( u, e w ( u )) > e w ( u ) is an unstable equilibrium point of the generalized Riccatiequation for ψ . In addition e w ( u ) > w ( u ), and in particular the fact that e w (0) > e w (1) > w ( u ) in the sense of Lemma 3.2. To see that e w ( u ) > max(0 , w ( u )), note that we only have to show that e w ( u ) >
0, whenever w ( u ) <
0. This is the case only for u ∈ (0 , e w ( u ) ≤ u ∈ (0 , R and ∂R∂w ( u, e w ( u )) > R ( u, ≥ u ∈ (0 , w ( u ) as w ( u ) = inf { w : ( u, w ) ∈ L } and I as the effective domain of w ( u ). (cid:3) We are now ready to show our main result on the long-term properties of thelog-price process ( X t ) t ≥ . Theorem 3.4.
Let ( X t , V t ) t ≥ be an affine stochastic volatility model and supposethat χ (0) < and χ (1) < . Let w ( u ) be given by Lemma 3.2 and define h ( u ) = F ( u, w ( u )) , J = { u ∈ I : F ( u, w ( u )) < ∞} . Then [0 , ⊆ J ⊆ I ; w ( u ) and h ( u ) are cumulant generating functions of infinitelydivisible random variables and lim t →∞ ψ ( t, u,
0) = w ( u ) for all u ∈ I ;(3.5a) lim t →∞ t φ ( t, u,
0) = h ( u ) for all u ∈ J . (3.5b)
Corollary 3.5.
Under the conditions of Theorem 3.4, the following holds: sup u ∈ [0 , | ψ ( t, u, − w ( u ) | ≤ C exp( − X · T ) ;(3.6a) sup u ∈ [0 , (cid:12)(cid:12)(cid:12)(cid:12) t φ ( t, u, − h ( u ) (cid:12)(cid:12)(cid:12)(cid:12) ≤ Ω C exp( − X · T ) ;(3.6b) for some constant C , and with X = inf u ∈ [0 , | χ ( u ) | and Ω = sup u ∈ [0 , ∂∂w F ( u, w ) (cid:12)(cid:12)(cid:12)(cid:12) w =0 Proof.
Let u ∈ [0 , u, w ( u )) ∈ [0 , × ( −∞ , F (0 ,
0) = F (1 ,
0) = 0, such that Lemma 2.2 guarantees that [0 , × ( −∞ , ⊆ dom F . It follows that [0 , ⊆ J . Define z ( t, u ) = ψ ( t, u, − w ( u ) . Inserting into the generalized Riccati equation 2.3b, ∂∂t z ( t, u ) = R ( u, ψ ( t, u, R ( u, ψ ( t, u, − R ( u, w ( u )) , and z (0 , u ) = w ( u ) . If ψ ( t, u, ≤ R ( u, ψ ( t, u, − R ( u, w ( u )) ≤ z ( t, u ) ∂R∂w ( u,
0) = z ( t, u ) χ ( u ) , using convexity of R . By Gronwall’s inequality z ( t, u ) ≤ | w ( u ) | exp ( χ ( u ) t ) . Since χ is convex, χ (0) < χ (1) <
0, we have shown (3.6a). The estimate | φ ( t, u ) − h ( u ) | == (cid:12)(cid:12)(cid:12)(cid:12) t Z t ( F ( u, φ ( s, u )) − F ( u, w ( u ))) ds (cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:12)(cid:12)(cid:12)(cid:12) ∂F∂w ( u, (cid:12)(cid:12)(cid:12)(cid:12) · | ψ ( t, u ) − w ( u ) | yields (3.6b) and we have shown Corollary 3.5.Let now u ∈ I ◦ \ [0 , R ( u, w ) > w ∈ [0 , w ( u )), and R ( u, w ( u )) = 0. It follows that the initialvalue ψ (0 , u,
0) = 0 is in the basin of attraction of the stable equilibrium point w ( u ) FFINE STOCHASTIC VOLATILITY MODELS 11 and thus that ψ ( t, u,
0) is strictly increasing and converging to w ( u ). An additionalargument may be needed at the boundary of I : Let u + = sup I and assume that u + ∈ I (i.e. I is right-closed). Since ( u + , w ) ∈ dom R for all w ≤ w ( u + ), wecan define ∂R∂w ( u + , w ( u + )) at least as a limit for w ↑ w ( u + ). By Lemma 3.2 either ∂R∂w ( u + , w ( u + )) < ∂R∂w ( u + , w ( u + )) = 0. In the first case we can argue as in theinterior of I that w ( u + ) is an asymptotically stable equilibrium point. In the secondcase we use once more that by Lemma 2.2 R ( u + , w ) is, as a function of w , eitherstrictly convex or affine. If it is affine, it must be equal to 0, and thus R ( u + ,
0) = 0,in contradiction to Lemma 2.9. Hence it is strictly convex, and attains its minimumat w ( u + ). This implies that R ( u + , w ) > w ∈ [0 , w ( u + )) and we concludethat ψ ( t, u + ,
0) converges to w ( u + ). For u − = inf I , a symmetrical argumentapplies.Assertion (3.5b) follows immediately from the representation (2.4), andlim t →∞ t φ ( t, u,
0) = lim t →∞ t Z t F ( u, ψ ( s, u, ds = F ( u, w ( u ))for all u ∈ J .We have shown that the sequence of infinitely divisible cumulant generating func-tions ψ ( t, u,
0) converges on I to a function w ( u ) that is continuous in a rightneighborhood of 0. This is sufficient to imply that w ( u ) is again the cumulant gen-erating function of an infinitely divisible random variable (See Feller [1971, VIII.1,Example (e)] for the convergence part, and Sato [1999, Lemma 7.8] for the infinitedivisibility.). The same argument can be applied to φ and h ( u ), and we have shownTheorem 3.4. (cid:3) Moment explosions
In this section we continue to study the time evolution of moments E [ S ut ] = E [ e X t u ] of the price process in an affine stochastic volatility model. We are in-terested in the phenomenon that in a stochastic volatility model, moments of theprice process can explode (become infinite) in finite time. For stochastic volatilitymodels of the CEV-type – a class including the Heston model, but no models withjumps – moment explosions have been studied by Andersen and Piterbarg [2007]and Lions and Musiela [2007]. In the context of option pricing, an interesting resultof Lee [2004] connects the existence of moments of the stock price process to thesteepness of the smile for deep in-the-money or out-of-the-money options. Our firstresult shows that in an affine stochastic volatility model a simple explicit expressionfor the time of moment explosion can be given:4.1. Moment explosions.
By definition, the u -th moment of S t , i.e. E [ S ut ] is givenby S u exp ( φ ( t, u,
0) + V ψ ( t, u, time of moment explosion forthe moment of order u by T ∗ ( u ) = sup { t : E [ S ut ] < ∞} . It is obvious from the Markov property that E [ S ut ] is finite for all t < T ∗ ( u ) andinfinite for all t > T ∗ ( u ). As in the previous section, the main result follows from aqualitative analysis of the generalized Riccati equations (2.3). Even though ψ ( t, u + ,
0) converges to w ( u + ), note that w ( u + ) is not a stable equilibrium point inthe usual sense. This is due to the fact that solutions from a right- neighborhood N ∩ ( w ( u + ) , ∞ )will diverge from w ( u + ) to + ∞ . Theorem 4.1.
Suppose the conditions of Theorem 3.4 hold. Define J = { u ∈ I : F ( u, w ( u )) < ∞} , f + ( u ) := sup { w ≥ F ( u, w ) < ∞} ,r + ( u ) := sup { w ≥ R ( u, w ) < ∞} , and suppose that F ( u, < ∞ , R ( u, < ∞ and χ ( u ) < ∞ .(a) If u ∈ J , then T ∗ ( u ) = + ∞ . (b) If u ∈ R \ J , then T ∗ ( u ) = Z min( f + ( u ) ,r + ( u ))0 dηR ( u, η ) . If F ( u,
0) = ∞ , R ( u,
0) = ∞ or χ ( u ) = ∞ then(c) T ∗ ( u ) = 0 . Proof.
Suppose that u ∈ J . Then Theorem 3.4 implies that both ψ ( t, u,
0) and φ ( t, u,
0) are finite for all t ≥
0. This proves (a). Let now u ∈ R \ J , F ( u, < ∞ , R ( u, < ∞ and χ ( u ) < ∞ . To prove (b) we start by analyzing the maximallifetime of solutions to the generalized Riccati equation(4.1) ∂∂t ψ ( t, u,
0) = R ( u, ψ ( t, u, , ψ (0 , u,
0) = 0 . Define M = [0 , r + ( u )) and note that R ( u, . ) ∈ C ( M ). Since u [0 , R ( u, >
0. It is clear, that at least a local solution ψ ( t, u,
0) to theODE exists, which satisfies 0 ≤ ψ ( t, u, ≤ r + ( u ) and is an increasing function of t as long as it can be continued. Using a standard extension theorem (e.g. Hartman[1982, Lem. I.3.1]) the local solution ψ ( t, u,
0) has a maximal extension to an interval[0 , T ( u )), such that one of the following holds:(i) T ( u ) = ∞ , or(ii) T ( u ) < ∞ and ψ ( t, u,
0) comes arbitrarily close to the boundary of M , i.e.lim sup t → T ( u ) ψ ( t, u,
0) = r + ( u ) . Consider case (i). Since ψ is increasing, its limit for t → ∞ exists, but can beinfinite. Suppose lim t →∞ ψ ( t ) = α < ∞ . Then α must be a stationary point, i.e. R ( u, α ) = 0, but this is impossible by Lemma 3.3. The case that α = ∞ is onlypossible if r + ( u ) = ∞ , such that in this case lim t → T ( u ) ψ ( t, u,
0) = r + ( u ). Considercase (ii). Since ψ is increasing the limes superior can be replaced by an ordinarylimit and we get lim t → T ( u ) ψ ( t, u,
0) = r + ( u ) as before.Let now T n be a sequence such that T n ↑ T ( u ). By (2.5) it holds that(4.2) Z ψ ( T n ,u, dηR ( u, η ) ds = T n . Letting n → ∞ we obtain that T ( u ) = R r + ( u )0 dηR ( u,η ) ds . FFINE STOCHASTIC VOLATILITY MODELS 13
We can write the time of moment explosion T ∗ ( u ) as the maximum joint lifetimeof φ ( t, u,
0) and ψ ( t, u, T ∗ ( u ) = sup { t ≥ φ ( t, u, < ∞ ∧ ψ ( t, u, < ∞} .By the integral representation (2.4) it is clear that if f + ( u ) ≥ r + ( u ), φ ( t, u, ψ ( t, u,
0) is finite and T ∗ ( u ) = T ( u ). If f + ( u ) < r + ( u ) then ψ ( T ∗ ( u ) , u,
0) = f + ( u ). Inserting into the representation (4.2) yields (b).For assertion (c), let F ( u,
0) = ∞ , R ( u,
0) = ∞ , or χ ( u ) = ∞ . In the first case, φ ( t, u,
0) does not exist beyond t = 0. In the other cases no local solution to thegeneralized Riccati equation (4.1) exists, such that ψ ( t, u,
0) explodes immediately. (cid:3)
Moment explosions in the stationary variance regime.
In Section 3.1we have introduced ( e X t , e V t ) t ≥ as the model in the stationary variance regime. Themoment explosions of this process can be analyzed in a similar manner as above.We define the time of moment explosion in the stationary variance regime by T S ∗ ( u ) := sup n T ≥ E [ e S uT ] < ∞ o ;the superscript ‘S’ stands for ‘stationary’.The analogue to Theorem 4.1 is the following result: Theorem 4.2.
Suppose the conditions of Theorem 3.4 hold. Define f + ( u ) , r + ( u ) as in Theorem 4.1, and in addition l + := sup { w > l ( w ) < ∞} . Suppose that F ( u, < ∞ , R ( u, < ∞ and χ (0) < ∞ .(a) If u ∈ J and w ( u ) ≤ l + , then T S ∗ ( u ) = + ∞ . (b) If u ∈ R \ J or w ( u ) > l + , then T S ∗ ( u ) = Z min( f + ( u ) ,r + ( u ) ,l + )0 dηR ( u, η ) . If F ( u,
0) = ∞ , R ( u,
0) = ∞ or χ (0) = ∞ , then(c) T S ∗ ( u ) = 0 . Corollary 4.3.
Under the conditions of Theorem 4.2, T S ∗ ( u ) ≤ T ∗ ( u ) , for all u ∈ R Proof.
By equation (3.2), the moment E [ e S ut ] is given by E [ e S ut ] = exp ( φ ( t, u,
0) + l ( ψ ( t, u, . This expression is finite, if φ ( t, u,
0) and ψ ( t, u,
0) are finite, and if ψ ( t, u, < l + .It is infinite if φ ( t, u,
0) or ψ ( t, u,
0) are infinite, or if ψ ( t, u, > l + . The rest ofthe proof can be carried out as for Theorem 4.1. Note, that now even for u ∈ J ,the moment can explode, if l + is reached by ψ ( t, u,
0) before the stationary point w ( u ). Corollary 4.3 follows easily by comparing the range of integration and theconditions for case (a) and (b) between Theorem 4.1 and Theorem 4.2. (cid:3) Applications
Smile behavior at extreme strikes.
In the preceding section, we have kept u fixed, and looked at the first time T ∗ ( u ) that the moment E [ S ut ] becomes infinite.It will now be more convenient to reverse the roles of T and u , and for a given time t to define the upper critical moment by u + ( t ) = sup { u ≥ E [ S ut ] < ∞} = sup { u ≥ T ∗ ( u ) < t } , and the lower critical moment by u − ( t ) = inf { u ≤ E [ S ut ] < ∞} = inf { u ≤ T ∗ ( u ) < t } . It is seen that u − ( T ) and u + ( T ) can be defined as the generalized inverse of T ∗ ( u )on ( −∞ ,
0] and [1 , ∞ ) respectively. In addition it is easily derived from Jensen’sinequality, that E [ S ut ] < ∞ for all u ∈ ( u − ( t ) , u + ( t )) , and E [ S ut ] = ∞ for all u ∈ R \ [ u − ( t ) , u + ( t )] . The results of Lee [2004] relate the explosion of moments to the ’wing behavior’of the implied volatility smile, i.e. the shape of the smile for strikes that are deepin-the-money or out-of-the-money. To give a precise statement, let ξ be the log-moneyness, which for a European option with time-to-maturity T and strike K isgiven by ξ = log (cid:16) Ke rT S (cid:17) . Proposition 5.1 (Lee’s moment formula) . Let V ( T, ξ ) be the implied Black-Scholes-Variance of a European call with time-to-maturity T and log-moneyness ξ . Then lim sup ξ →−∞ V ( T, ξ ) | ξ | = ς ( − u − ( T )) T and lim sup ξ →∞ V ( T, ξ ) | ξ | = ς ( u + ( T ) − T where ς ( x ) = 2 − (cid:0) √ x + x − x (cid:1) and u ± ( T ) are the critical moment functions. The function ς is strictly decreasing on R > , mapping 0 to 2, and ∞ to 0.Thus for fixed time-to-maturity T , the steepness of the smile is decreasing with | u ± ( T ) | . A finite critical moment u ± ( T ) implies asymptotically linear behavior of V ( T, ξ ) in ξ , and an infinite critical moment implies sublinear behavior of V ( T, ξ ).It is also evident that u − ( T ) determines the ’left’ side of the volatility smile, alsoknown as small-strike, in-the-money-call or out-of-the-money-put side; u + ( T ) de-termines the ’right’ side, or large-strike, out-of-the-money-call, in-the-money-putside. Finally we mention that Lee’s result has been extended and strengthened byBenaim and Friz [2006] from a ‘lim sup’ to a genuine limit under conditions relatedto regular variation of the underlying distribution function.5.2. Forward-smile behavior.
The forward smile is derived from the prices offorward-start options. For a forward-start call option – all options we considerare European – a start date τ , a strike date T + τ and a moneyness ratio M areagreed upon today (at time t = 0). The option then yields at time T + τ a payoffof (cid:16) S T + τ S τ − M (cid:17) + , i.e. the relative return over the time period from τ to τ + T , FFINE STOCHASTIC VOLATILITY MODELS 15 reduced by M and floored at 0. Under the pricing measure the value of such anoption at t = 0 is given by(5.1) e − r ( T + τ ) E "(cid:18) S T + τ S τ − M (cid:19) + = e − τr E h(cid:0) e X T + τ − X τ − e ξ (cid:1) + i , where we define the log-moneyness ξ of a forward-start option as ξ = log M + rT .Forward-start options are not just interesting in their own right, but are used asbuilding blocks of more complex derivatives, such as Cliquet options (see Gatheral[2006, Chapter 10]).Analogously to plain vanilla options, we can define the implied forward volatility σ ( τ, T, ξ ), by comparing the forward option price to the price of an option withidentical payoff in the Black-Scholes model. Note that the implied forward volatilitydepends also on τ , the starting time of the contract. For τ = 0, the implied volatilityof a plain vanilla option is retrieved. More interesting is the behavior for τ > τ in a stochastic volatility model, since the uncertainty of the variance V τ at thestarting date of the option has to be priced in. In an affine stochastic volatilitymodel, it will be seen that under mild conditions, the implied forward volatilities σ ( τ, T, ξ ) actually converge to a limit as τ → ∞ . Not surprisingly, this behavioris related to the convergence of ( V t ) t ≥ to its invariant distribution. In the limit τ → ∞ , the pricing of a forward-start option is equivalent to the pricing of a plainvanilla option in the stationary variance regime (cf. Section 3.1). Proposition 5.2.
Let ( X t , V t ) t ≥ be an affine stochastic volatility model, satisfyingthe conditions of Proposition 3.1. Let σ ( τ, T, ξ ) be the implied forward volatility inthis model. Then lim τ →∞ σ ( τ, T, ξ ) = e σ ( T, ξ ) , where e σ ( T, ξ ) is the implied volatility of a European call with payoff (cid:16) e e X T − e ξ (cid:17) + ,and e X T is the log-price process of the model in the stationary variance regime.Proof. We can write the price of a forward-start call as C ( τ, T, ξ ) = e − τr E h(cid:0) e X T + τ − X τ − e ξ (cid:1) + i = e − rτ E h E (0 ,V τ ) h(cid:0) e X T − e ξ (cid:1) + ii . Denote by C BS ( T, ξ, σ ) the (plain vanilla) call price in a Black-Scholes model withvolatility σ and the normalization S = 1. It is easy to see that the price of aforward-start option in the Black-Scholes model is just the discounted plain vanillaprice, i.e. C BS ( τ, T, ξ, σ ) = e − rτ C BS ( T, ξ, σ ). By definition, the implied forwardvolatility of the call C ( τ, T, ξ ) satisfies C BS ( T, ξ, σ ( τ, T, ξ )) = e rτ C ( τ, T, ξ ) = E h E (0 ,V τ ) h(cid:0) e X T − e ξ (cid:1) + ii . Taking the limit τ → ∞ on both sides we obtain C BS ( T, ξ, lim τ →∞ σ ( τ, T, ξ )) = E h lim τ →∞ E (0 ,V τ ) h(cid:0) e X T − e ξ (cid:1) + ii = E (cid:20)(cid:16) e e X T − e ξ (cid:17) + (cid:21) , using dominated convergence. It is well known that the above equation allows aunique solution in terms of the Black-Scholes implied volatility, and we get e σ ( T, ξ ) =lim τ →∞ σ ( τ, T, ξ ). (cid:3) Combining Lee’s moment formula with our results on moment explosions underthe stationary variance regime (Theorem 4.2), asymptotics of e σ ( T, ξ ) for ξ → ±∞ can be derived. 6. Examples
The Heston model with and without jumps.
In the model of Heston[1993], the log-price ( X t ) t ≥ and the corresponding variance process ( V t ) t ≥ aregiven under the risk-neutral measure by the SDE dX t = − V t dt + p V t dW t dV t = − λ ( V t − θ ) dt + ζ p V t dW t where W t , W t are Brownian motions with correlation parameter ρ , and ζ, λ, θ > F ( u, w ) = λθw (6.1a) R ( u, w ) = 12 ( u − u ) + ζ w − λw + uwρζ . (6.1b)It is easily calculated that χ is given by χ ( u ) = ρζu − λ . We will first analyzethe long term behavior of ( X t ) t ≥ , with the help of Theorem 3.4. To satisfy thecondition χ (1) < λ > ζρ . Note that this condition is always satisfied if ρ ≤
0, the case that is typical for applications. Solving a quadratic equation wefind that w ( u ) = ( λ − uρζ ) − p ( λ − uρζ ) − ζ ( u − u ) ζ , and h ( u ) = λθw ( u ) . Denoting the term under the square root by ∆( u ), we see that w ( u ) and h ( u ) areboth defined on J = I = { u : ∆( u ) ≥ } . Since R is a second order polynomial inthe Heston model, the equilibrium points of the generalized Riccati equation for ψ form an ellipse in the ( u, w )-plane, and w ( u ) is given by its lower part – see Fig-ure 1 for an illustration. Interestingly, w ( u ), and also h ( u ), are cumulant generatingfunctions of a Normal Inverse Gaussian distribution (cf. Barndorff-Nielsen [1997,Eq. (2.4)]). Thus, for large t , the price process of the Heston model is, in termsof its marginal distributions, close to a Normal-Inverse-Gaussian exponential-L´evymodel.Next we consider moment explosions in the Heston model. As mentioned above,moment explosions in the Heston model (and other models) have already beenstudied by Andersen and Piterbarg [2007]. Nevertheless this will provide a firsttest of Theorem 4.1: In the case of the Heston model it is easily determined from(6.1) that f + ( u ) = r + ( u ) = ∞ . Calculating the integral in case (b) of Theorem 4.1,we obtain(6.2) T ∗ ( u ) = + ∞ ∆( u ) ≥ √ − ∆( u ) (cid:18) arctan √ − ∆( u ) χ ( u ) + π { χ ( u ) < } (cid:19) ∆( u ) < . In Figure 2 a plot of this function for typical parameter values is shown. Note thatAndersen and Piterbarg [2007] distinguish an additional case where χ ( u ) <
0, but∆( u ) >
0. A little calculation shows that this can only happen if χ (1) ≥
0, a case
FFINE STOCHASTIC VOLATILITY MODELS 17 ( u ) )Unstable equilibria y ( t , u ) for t = 0.1, 0.5,2 Figure 1.
This plot shows the stable and unstable equilibria of the gen-eralized Riccati equation of a Heston model with parameters ρ = − . ζ = 0 . λ = 1 . θ = 0 . ψ ( t, u ) converge to the stableequilibrium points, which form the lower boundary of an ellipse in the ( u, w )-plane. that is precluded by our assumptions in Theorem 3.4, and never occurs when ρ ≤ J t ) t ≥ be a pure-jump L´evy process, independent of ( W , t ) t ≥ and define theHeston-with-jumps model by dX t = (cid:18) δ − V t (cid:19) dt + p V t dW t + dJ t dV t = − λ ( V t − θ ) dt + ζ p V t dW t . The drift δ is determined by the martingale condition for ( S t ) t ≥ . To make theexample simpler, we assume that ( J t ) t ≥ jumps only downwards. This is equivalentto saying that the L´evy measure m ( dx ) of ( J t ) t ≥ is supported on ( −∞ , F ( u, w ) = λθw + e κ ( u )(6.3) R ( u, w ) = 12 ( u − u ) + ζ w − λw + uwρζ , (6.4) where e κ ( u ) is the compensated cumulant generating function of the jump part, i.e. e κ ( u ) = Z ( −∞ , ( e xu − m ( dx ) − u Z ( −∞ , ( e x − m ( dx ) . Let κ − < e κ ( u ) is finite on ( κ − , ∞ ) and infinite outside.For example, if the absolute jump heights are exponentially distributed with anexpected jump size of 1 /α , then κ − = − α .To analyze the explosion times of this model, note that R , and thus χ ( u ), w ( u ), I and r + ( u ) have not changed compared to the Heston model. As long as u > κ − ,the explosion time T ∗ ( u ) is the same as in the Heston model. However, if u ≤ κ − , F ( u,
0) = ∞ and by Theorem 4.1, T ∗ ( u ) = 0. Thus, the addition of jumps tothe Heston model has the effect of truncating the explosion time to zero, whenever u ≤ κ − .From the viewpoint of the critical moment functions, u + ( t ) does not change com-pared to the Heston model, but u − ( t ) does; in the model with jumps it is givenby u Jump − ( t ) = u Heston − ( t ) ∨ κ − . Since u − is increasing with t , it makes sense to define a cutoff time T ♯ by T ♯ = sup (cid:8) t ≥ u Heston − ( t ) = κ − (cid:9) = T ∗ ( κ − ) , such that u Heston − ( t ) < u Jump − ( t ) , if t < T ♯ u Heston − ( t ) = u Jump − ( t ) , if t ≥ T ♯ . In Figure 2 a comparison of the critical moment functions in the Heston model withand without jumps can be seen. By Lee’s moment formula, the critical moment u − ( t ) moving closer to 0 will cause the left side of the implied volatility smile tobecome steeper. Thus the net effect of adding the jump component ( J t ) t ≥ to theHeston model, is a steepening of the left side of the smile for maturities smallerthan T ♯ . For times larger than T ♯ , the asymptotic behavior of the smile (in thesense of Lee’s formula) is exactly the same as in the Heston model without jumps.This corresponds well to the frequently made observation (see e.g. Gatheral [2006,Chapter 5]) that a Heston model with jumps can be fitted well by first fitting a(jump-free) Heston model to long maturities, and then calibrating only the ad-ditional parameters to the full smile. In fact Gatheral proposes (on heuristicalgrounds) the concept of a ‘critical time’ T , after which the influence of an inde-pendent jump component on the implied volatility smile can be neglected. Theanalysis of the Heston model with jumps is of course easily extended to the casethat ( J t ) t ≥ is not one-sided. In that case the effects discussed above will be seento affect also the right side of the implied volatility smile.6.2. A model of Bates.
We consider now the model given by dX t = (cid:18) δ − V t (cid:19) dt + p V t dW t + Z D x e N ( V t , dt, dx ) dV t = − λ ( V t − θ ) dt + ζ p V t dW t . where as before λ, θ, ζ > ρ . The jump component is given by e N ( V t , dt, dx ) = N ( V t , dt, dx ) − n ( V t , dt, dx ),where N ( V t , dt, dx ) is a Poisson random measure, and its compensator n ( V t , dt, dx ) FFINE STOCHASTIC VOLATILITY MODELS 19 is of the state-dependent form V t µ ( dx ) dt , with µ ( dx ) the L´evy measure given in(2.6). A model of this kind has been proposed by Bates [2000] to explain the time-variation of jump-risk implicit in observed option prices. Bates also proposes asecond variance factor, which we omit in this example, in order to remain in thescope of Definition 2.8. It would however not be difficult to extend our approach tothe two-factor Bates model, since the two proposed variance-factors are mutuallyindependent, causing the corresponding generalized Riccati equations to decouple.Since it is affine, the above model can be characterized in terms of the functions F and R : F ( u, w ) = λθw (6.5) R ( u, w ) = 12 ( u − u ) + ζ w − λw + uwρζ + e κ ( u ) . (6.6)where e κ ( u ) is the compensated cumulant generating function of the L´evy measure µ . As in the Heston model we can obtain w ( u ) and h ( u ) explicitly, and get h ( u ) = − χ ( u ) − p ∆( u ) ζ , and h ( u ) = λθw ( u ) , where χ ( u ) = ρζu − λ and ∆( u ) = χ ( u ) − ζ ( u − u + 2 e κ ( u )). Both w ( u ) and h ( u )are defined on I = J = { u : ∆( u ) ≥ } . The time of moment explosion can againbe calculated explicitly, and is given by(6.7) T ∗ ( u ) = + ∞ ∆( u ) > √ − ∆( u ) (cid:18) arctan √ − ∆( u ) χ ( u ) + π { χ ( u ) < } (cid:19) −∞ < ∆( u ) <
00 ∆( u ) = −∞ . The Barndorff-Nielsen-Shephard model.
The Barndorff-Nielsen-Shephard(BNS) model was introduced by Barndorff-Nielsen and Shephard [2001] as a modelfor asset pricing. In SDE form it is given in the risk-neutral case by dX t = ( δ − V t ) dt + p V t dW t + ρ dJ λt dV t = − λV t dt + dJ λt where λ > ρ < J t ) t ≥ is a L´evy subordinator, i.e. a pure jump L´evyprocess that increases a.s. The drift δ is determined by the martingale conditionfor ( S t ) t ≥ . The time-scaling J λt is introduced by Barndorff-Nielsen and Shephardto make the invariant distribution of the variance process independent of λ . Thedistinctive features of the BNS model are that the variance process has no diffusioncomponent, i.e. moves purely by jumps and that the negative correlation betweenvariance and price movements is achieved by simultaneous jumps in ( V t ) t ≥ and( X t ) t ≥ . The BNS model is an affine stochastic volatility model, and F and R aregiven by F ( u, w ) = λκ ( w + ρu ) − uλκ ( ρ )(6.8) R ( u, w ) = 12 ( u − u ) − λw (6.9) where κ ( u ) is the cumulant generating function of ( J t ) t ≥ .We simply have χ ( u ) = − λ and w ( u ) from Lemma 3.2 is given by w ( u ) = 12 λ ( u − u ) . It follows that h ( u ) = λκ (cid:18) u λ + u (cid:18) ρ − λ (cid:19)(cid:19) − uλκ ( ρ ) . This expression can be interpreted as cumulant generating function of a Brownianmotion with variance λ and drift ρ − λ , subordinated by the L´evy process J λt andthen mean-corrected to satisfy the martingale condition.To analyze moment explosions in the BNS model, let κ + := sup { u > κ ( u ) < ∞} .It is easy to see that f + is given by f + = max( κ + − ρu, r + = ∞ , we havethat the explosion time for the moment of order u is given by T ∗ ( u ) = Z f + dηR ( u, η ) = − λ log (cid:18) − λ (max( κ + − ρu, u ( u − (cid:19) . The critical moment functions u ± ( T ) can be obtained explicitly by solving a qua-dratic equation, and are given by u ± ( t ) = 12 − ρλ − e − λt ± s
14 + (2 κ + − ρ ) λ − e − λt + ρ λ (1 − e − λt ) . The large-strike asymptotics for the implied volatility smile in the sense of Lee canbe explicitly calculated by inserting u ± into Proposition 5.1.6.4. The Heston model in the stationary variance regime.
In the Hestonmodel the limit distribution of the variance process ( V t ) t ≥ is a Gamma distribu-tion with parameters ( − λθζ , λζ ). This is well-known, but can also be obtained byapplying Proposition 3.1. The cumulant generating function l ( w ) is thus given by l ( w ) = − λθζ log (cid:18) − ζ λ w (cid:19) , defined on ( −∞ , λζ ), such that l + = λζ . As before we have that χ ( u ) = ρζu − λ , andwe assume that χ (1) <
0. In addition we define χ + ( u ) = ρζu + λ . By Theorem 4.2,the explosion time in the stationary regime is given by(6.10) T S ∗ ( u ) = Z λ/ζ dηR ( u, η ) == ∞ p ∆( u ) > − χ + ( u ) , √ ∆ log (cid:12)(cid:12)(cid:12) χ + χ +2 λ √ ∆ − ∆ χ + χ − λ √ ∆ − ∆ (cid:12)(cid:12)(cid:12) < p ∆( u ) < − χ + ( u ) , √− ∆ arctan (cid:16) λ √− ∆ χ + χ − ∆ + π { χ + χ< ∆ } (cid:17) ∆( u ) < . In Figure 2 T S ∗ ( u ) is plotted together with T ∗ ( u ) for the Heston model. FFINE STOCHASTIC VOLATILITY MODELS 21 – ( t ) Heston modelHeston in the stationary variance regimeHeston with jumps
Figure 2.
This plot shows the critical moment functions u ± ( t ) for a Hestonmodel with the same parameters as in Figure 1. Also shown are u S ± ( t ) forthe model in the stationary variance regime, and u Jump ± ( t ) for the Hestonmodel with an independent jump component, whose negative jump heights areexponentially distributed with mean α = − .
1. Note that u Jump ± ( t ) coincideswith u ± ( t ) everywhere except in the lower left corner of the plot. The BNS model in the stationary variance regime.
In the BNS model,the cumulant generating function of the limit distribution L of the variance processis given by Proposition 3.1 by l ( w ) = Z w κ ( η ) η dη , provided the log-moment condition R y> (log y ) µ ( dy ) < ∞ holds for the L´evy mea-sure of ( J t ) t ≥ . The above integral is finite as long as w ∈ ( −∞ , κ + ), and infiniteoutside. Thus l + = κ + . In Section 6.3 we obtained that f + ( u ) = κ + − ρu , suchthat the time of moment explosion under stationary variance is given by T S ∗ ( u ) = Z min ( f + ( u ) ,l + )0 dηR ( u, η ) = − λ log (cid:18) − λk ( u ) u ( u − (cid:19) , where k ( u ) = κ + for u ≥ k ( u ) = max( κ + − ρu,
0) for u ≤
0. Again, thisexpression can be inverted to give the critical moment functions in the stationary variance case. By definition ρ ≤
0, such that we obtain u S − ( T ) = 12 − ρλ − e − λT − s
14 + (2 κ + − ρ ) λ − e − λT + ρ λ (1 − e − λT ) u S + ( T ) = 12 + r
14 + 2 κ + λ − e − λT . Appendix A. Additional proofs
Proof of Theorem 2.1.
Let t ≤ τ . By the flow equation we can write φ ( τ, u, η ) = φ ( t, u, η ) + φ ( τ − t, u, ψ ( t, u, η )) ψ ( τ, u, η ) = ψ ( τ − t, u, ψ ( t, u, η )) . Since the left sides are finite by assumption, it follows that also φ ( t, u, η ) and ψ ( t, u, η ) are. V t is non-negative, such that | E [exp ( uX t + wV t )] | ≤ | E [exp ( uX t + ηV t )] | , whenever Re w ≤ Re η . Thus φ ( t, u, w ) and ψ ( t, u, w ) exist for all w ∈ C withRe w ≤ Re η . As a particular case we can conclude that φ ( t, u, w ) and ψ ( t, u, w )exist for all ( u, w ) in U := (cid:8) ( u, w ) ∈ C : Re u = 0 , Re w ≤ (cid:9) .We also define U ◦ := (cid:8) ( u, w ) ∈ C : Re u = 0 , Re w < (cid:9) , and show next that φ ( t, u, w )and ψ ( t, u, w ) are (right-)differentiable at t = 0 for all ( u, w ) ∈ U ◦ . The key ideaof our proof is originally due to Montgomery and Zippin [1955], and has also beenpresented in Filipovi´c and Teichmann [2003] and Dawson and Li [2006]. First notethat the identity E (cid:2) wV t e uX t + wV t (cid:3) = (cid:18) ∂∂w φ ( t, u, w ) + V ∂∂w ψ ( t, u, w ) (cid:19) exp ( φ ( t, u, w ) + V ψ ( t, u, w ) + X u )shows that ∂∂w φ ( t, u, w ) and ∂∂w ψ ( t, u, w ) exist, and are continuous for all t ≤ τ and( u, w ) ∈ U ◦ . By Taylor expansion it holds that Z s ψ ( r, u, ψ ( t, u, w )) dr − Z s ψ ( r, u, w ) dr = Z s ∂∂w ψ ( r, u, w ) dr ( ψ ( t, u, w ) − w )+ o (cid:0) | ψ ( t, u, w ) − w | (cid:1) . (A.1)On the other side, using the flow property, we calculate Z s ψ ( r, u, ψ ( t, u, w )) dr − Z s ψ ( r, u, w ) dr = Z s ψ ( r + t, u, w ) dr − Z s ψ ( r, u, w ) dr == Z s + tt ψ ( r, u, w ) dr − Z s ψ ( r, u, w ) dr = Z t ψ ( r + s, u, w ) dr − Z t ψ ( r, u, w ) dr . (A.2)Denoting the last expression by I ( s, t ), and putting (A.1) and (A.2) together, weobtain lim t → (cid:12)(cid:12) s I ( s, t ) (cid:12)(cid:12) | ψ ( t, u, w ) − w | = (cid:12)(cid:12)(cid:12)(cid:12) s Z s ∂∂w ψ ( t, u, w ) dr (cid:12)(cid:12)(cid:12)(cid:12) . Thus, writing M s = s R s ∂∂w ψ ( t, u, w ) dr , we havelim t → t | ψ ( t, u, w ) − w | = (cid:12)(cid:12)(cid:12)(cid:12) lim t → I ( s, t ) st (cid:12)(cid:12)(cid:12)(cid:12) · | M s | − = (cid:12)(cid:12)(cid:12)(cid:12) ψ ( s, u, w ) − ws (cid:12)(cid:12)(cid:12)(cid:12) | M s | − . FFINE STOCHASTIC VOLATILITY MODELS 23
But M s is a continuous function of s , and lim s → M s = ∂∂w ψ (0 , u, w ) = 1, such thatfor s small enough M s = 0. We conclude that the left hand side is finite, and using(A.1) we obtain thatlim t → ψ ( t, u, w ) − wt = (cid:18) ψ ( s, u, w ) − ws (cid:19) · (cid:18) s Z s ∂∂w ψ ( r, u, w ) dr (cid:19) − . The finiteness of the right hand side implies the existence of the limit on the left.In addition the right hand side is continuous for ( u, w ) ∈ U ◦ , showing that also theleft hand side is. A similar calculation for φ ( t, u, w ) shows thatlim t → φ ( t, u, w ) t = φ ( s, u, w ) s − lim t → (cid:18) ψ ( t, u, w ) − wt (cid:19) · (cid:18) s Z s ∂∂w φ ( r, u, w ) dr (cid:19) , allowing the same conclusions for φ ( t, u, w ). We have thus shown that the time-derivatives of φ ( t, u, w ) and ψ ( t, u, w ) at t = 0 exist, and are continuous in U ◦ .Combining Duffie et al. [2003, Proposition 7.2] and Duffie et al. [2003, Proposi-tion 6.4] the differentiability can be extended from U ◦ to U , and we have shownthat ( X t , V t ) t ≥ is a regular affine process. The rest of Theorem 2.1 follows now asin Duffie et al. [2003, Theorem 2.7] (cid:3) Proof of Lemma 2.2.
We prove the assertions of Lemma 2.2 for F ; they followanalogously for R . By the L´evy-Khintchine representation (2.6), F ( u, w ) + c is thecumulant generating functions of some infinitely divisible random variables, say X .Writing z = ( u, w ) ∈ R , and using H¨older’s inequality it holds for any λ ∈ [0 , F ( λz + (1 − λ ) z ) = log E h e λ h z ,X i e (1 − λ ) h z ,X i i − c ≤≤ λ log E h e h z ,X i i + (1 − λ ) E h e h z ,X i i − c = λF ( z ) + (1 − λ ) F ( z ) , showing convexity of F . In addition equality in (A.3) holds if and only if ke h z ,X i = e h z ,X i a.s. for some k >
0. This in turn is equivalent to h z − z , X i being constanta.s. Choosing now z and z = z from some one-dimensional affine subspace U = { p + h q, x i : x ∈ R } of R , we see that either h q, X i is constant a.s. in whichcase F | U is affine, or it is not constant, in which case strict inequality holds in (A.3)for all z , z ∈ U , showing (c).Let L α = { z : F ( z ) ≤ α } be a level set of F , and z n ∈ L α a sequence converging to z . Then by Fatou’s Lemmalog E [ e h z,X i ] − c ≤ lim inf n →∞ log E [ e h z n ,X i ] − c ≤ α , showing that z ∈ L α and thus that F is a closed convex function. Finally F isproper, because F (0 ,
0) = c > −∞ , showing (a).Next we show analyticity: Consider the random variables X n := X {| X |≤ n } . Sincethey are bounded, their Laplace transforms, and hence also their cumulant gener-ating functions are entire functions on C , and thus analytic on R . As a uniformlimit of analytic functions F ( u, w ) is analytic in the interior of dom F , showing (b).Assertion (d) follows directly from Theorem 2.1. (cid:3) References
Leif B. G. Andersen and Vladimir V. Piterbarg. Moment explosions in stochasticvolatility models.
Finance and Stochastics , 11:29–50, 2007.G. Bakshi, C. Cao, and Z. Chen. Empirical performance of alternative optionpricing models.
Journal of Finance , 52:2003–2049, 1997.Ole E. Barndorff-Nielsen. Processes of normal inverse Gaussian type.
Finance andStochastics , 2:41 – 68, 1997.Ole E. Barndorff-Nielsen and Neil Shephard. Non-Gaussian Ornstein–Uhlenbeck-based models and some of their uses in financial economics.
Journal of the RoyalStatistical Society B , 63:167–241, 2001.David S. Bates. Post-’87 crash fears in the S&P 500 futures option market.
Journalof Econometrics , 94:181–238, 2000.David S. Bates. Jump and stochastic volatility: exchange rate processes implicit inDeutsche Mark options.
The Review of Financial Studies , 9:69–107, 1996.Shalom Benaim and Peter Friz. Regular variation and smile asymptotics.arXiv:math/0603146v2, 2006.Peter Carr and Liuren Wu. Time-changed L´evy processes and option pricing.
Jour-nal of Financial Economics , 71(1):113–141, 2004.D. A. Dawson and Zenghu Li. Skew convolution semigroups and affine Markovprocesses.
The Annals of Probability , 34(3):1103–1142, 2006.D. Duffie, D. Filipovic, and W. Schachermayer. Affine processes and applicationsin finance.
The Annals of Applied Probability , 13(3):984–1053, 2003.Darrell Duffie, Jun Pan, and Kenneth Singleton. Transform analysis and assetpricing for affine jump-diffusions.
Econometrica , 68(6):1343 – 1376, 2000.William Feller.
An introduction to probability theory and its applications . JohnWiley & Sons, 2nd edition, 1971.Damir Filipovi´c. A general characterization of one factor affine term structuremodels.
Finance and Stochastics , 5:389–412, 2001.Damir Filipovi´c and Josef Teichmann. Regularity of finite-dimensional realizationsfor evolution equations.
Journal of Functional Analyis , 197:433–446, 2003.Jim Gatheral.
The Volatilty Surface . Wiley Finance, 2006.Philip Hartman.
Ordinary Differential Equations . Birkh¨auser, 1982.S. Heston. A closed-form solution of options with stochastic volatility with applica-tions to bond and currency options.
The Review of Financial Studies , 6:327–343,1993.Martin Keller-Ressel and Thomas Steiner. Yield curve shapes and the asymptoticshort rate distribution in affine one-factor models.
Finance and Stochastics , 12:149 – 172, 2008.Roger Lee. The moment formula for implied volatility at extreme strikes.
Mathe-matical Finance , 14(3):469–480, 2004.Alan L. Lewis.
Option Valuation under Stochastic Volatility . Finance Press, 2000.P. Lions and M. Musiela. Correlations and bounds for stochastic volatility models.
Annales de l’Institut Henri Poincar´e , 24:1–16, 2007.Deane Montgomery and Leo Zippin.
Topological Transformation Groups . Inter-science Publishers, Inc., 1955.W. F. Osgood. Beweis der Existenz einer L¨osung der Differentialgleichung dydx = f ( x, y ) ohne Hinzunahme der Cauchy-Lipschitz’schen Bedingung. Monatsheftef¨ur Mathematik und Physik , 9:331–345, 1898.
FFINE STOCHASTIC VOLATILITY MODELS 25
Ken-Iti Sato.
L´evy processes and infinitely divisible distributions . Cambridge Uni-versity Press, 1999.
Vienna University of Technology, Wiedner Hauptstrasse 8–10, A-1040 Wien, Austria
E-mail address ::