Martingale transport with homogeneous stock movements
MMartingale transport with homogeneous stock movements
Stephan Eckstein ∗ Michael Kupper † August 28, 2019
Abstract
We study a variant of the martingale optimal transport problem in a multi-periodsetting to derive robust price bounds of a financial derivative. On top of marginal andmartingale constraints, we introduce a time-homogeneity assumption, which restrictsthe variability of the forward-looking transitions of the martingale across time. Weprovide a dual formulation in terms of superhedging and discuss relaxations of the time-homogeneity assumption by adding market frictions. In financial terms, the introducedtime-homogeneity corresponds to time-consistent call prices, given the state of the stock.The time homogeneity assumption leads to improved price bounds and the possibilityto utilize more market data. The approach is illustrated with two numerical examples.
Keywords : Robust pricing, martingale optimal transport, superhedging, market informa-tion, transaction costs
We consider a discrete stock process S , ..., S T . The goal is to find a fair price of a financialinstrument f ( S , ..., S T ) depending on this stock. We follow the robust pricing idea ofmartingale optimal transport [4, 5], in that we determine the highest and lowest possibleprice for this instrument under pricing rules which are consistent with European call andput prices observed on the market (which determine the risk-neutral one-period marginaldistributions of S , ..., S T ) and the assumption that the process S , ..., S T is a martingale.In addition, this paper adds a notion of time-homogeneity for the process S , ..., S T , madeprecise in Section 3. The reason we introduce this assumption is twofold:(1) While the martingale optimal transport approach is very robust, for practical purposesthe obtained range of prices is often too wide, see also [16, 19, 22]. ∗ Department of Mathematics, University of Konstanz, Universitätsstraße 10, 78464 Konstanz, Germany,[email protected] † Department of Mathematics, University of Konstanz, Universitätsstraße 10, 78464 Konstanz, Germany,[email protected] a r X i v : . [ q -f i n . M F ] A ug
2) In the martingale optimal transport setting, information obtained through optionprices with maturities t , ..., t k has little relevance for pricing an instrument dependingon different time-points t k +1 , ..., t K . Hence, market information is used inefficiently.The notion of time-homogeneity of the stock process mainly aims at overcoming the issueraised in point (2) and thus narrow the range of possible prices to improve on point (1). Inthe spirit of robust pricing the introduced assumption of time-homogeneity is as weak aspossible, while still achieving its purpose.Three key features of the approach are worth pointing out: First, the homogeneousmartingale optimal transport problem is as numerically tractable as the martingale optimaltransport problem without time-homogeneity, in that the discretized version reduces to alinear program and the dual formulation is well suited for various approaches, see e.g. [11,14, 16]. Second, the dual formulation can be interpreted in terms of trading strategiesand superhedging. And third, market frictions and relaxations of the introduced time-homogeneity assumption can be incorporated naturally.In the recent literature, different methods have been studied to improve on point (1)and hence make the martingale optimal transport approach more practicable. In [19, 22]the authors study additional variance and Markovianity constraints on the underlying stockprocess. In [15] additional information from options written on the stock’s volatility isincorporated.The rest of the paper is structured as follows: In Section 2, we give the relevant notationand recall basic facts about martingale optimal transport. In Section 3, the notion oftime-homogeneity is introduced and we state basic properties and duality for the time-homogeneous version of the martingale optimal transport problem. In Section 4, extensionslike market frictions, relaxed assumptions and higher dimensional markets are discussed.Section 5 gives two short numerical examples. All proofs are postponed to Section 6. S = ( S , ..., S T ) denotes the value of a stock at time points t = 1 , ..., T . For simplicity,we assume no risk-free rate and no dividends. We model the asset prices as the canonicalprocess on R T , i.e. S t ( ω ) = ω t for ω ∈ R T . Hereby, R T is endowed with the Borel σ -algebra B ( R T ) and euclidean norm | · | . We denote by C lin ( R T ) (resp. C b ( R T ) ) the set of allcontinuous functions f : R T → R such that | f ( · ) | / (1 + | · | ) is bounded (resp. f is bounded)and by P ( R T ) the set of all probability measures Q on B ( R T ) .Let µ , ..., µ T ∈ P ( R ) have finite first moments. The measures µ , ..., µ T model the risk-neutral marginal distributions of S , ..., S T inferred from option prices, see [6, 17]. Further,fix f ∈ C lin ( R T ) , which defines the financial instrument f ( S ) to be priced. For arbitrary Q ∈ P ( R T ) and a sub-tuple I = ( t , ..., t | I | ) of (1 , ..., T ) let Q I := Q ◦ S − I , where S I : R T → R | I | is given by S I ( ω ) = ( ω t , ..., ω t | I | ) . Denote by Q t := Q ( t ) the t -th marginal of Q , and Π( µ , ..., µ T ) := { Q ∈ P ( R T ) : Q t = µ t for t = 1 , ..., T } , M ( µ , ..., µ T ) := { Q ∈ Π( µ , ..., µ T ) : E Q [ S t +1 | S , ..., S t ] = S t for all t = 1 , ..., T − } .
2e call Π( µ , ..., µ T ) the set of all couplings between µ , ..., µ T and M ( µ , ..., µ T ) the setof all martingale couplings. The martingale optimal transport problem is to find the lowestand highest possible price of the financial instrument f ( S ) among models in M ( µ , ..., µ T ) :Without loss of generality, we focus on the problem to find the highest price: sup Q ∈M ( µ ,...,µ T ) E Q [ f ( S )] (MOT)In contrast, the usual (multi-marginal) optimal transport problem is stated over all couplings sup Q ∈ Π( µ ,...,µ T ) E Q [ f ( S )] . (OT)Both problems allow for a dual formulation, which can be interpreted in terms of trading.For the (OT) problem, the dual formulation reads inf h ,...,h T ∈ C lin ( R ): (cid:80) Tt =1 h t ( S t ) ≥ f ( S ) T (cid:88) t =1 (cid:90) R h t dµ t . (OT-Dual)Thereby, h , ..., h T are trading strategies for a single time-period, which corresponds totrading freely into European call options. Indeed, one can restrict each h t to be a linearcombination of European call options, i.e. h t ( S t ) = (cid:80) Ni =1 α i ( S t − k i ) + for different strikeprices k , ..., k N ∈ R , see [4]. For the (MOT) problem, the martingale condition correspondsto the assumption that one can additionally trade dynamically in the underlying, leading to inf h ,...,h T ∈ C lin ( R ) , ϑ ∈ C b ( R ) ,...,ϑ T − ∈ C b ( R T − ): (cid:80) Tt =1 h t ( S t )+ (cid:80) T − t =1 ϑ t ( S ,...,S t ) ( S t +1 − S t ) ≥ f ( S ) T (cid:88) t =1 (cid:90) R h t dµ t . (MOT-Dual)Hereby, ϑ t ( S , ..., S t ) is the (positive or negative) amount invested into the stock at time t . The purpose of this Section is the introduction and analysis of the notion of time-homogeneityadded to the martingale optimal transport setting, which restricts the variability of the for-ward looking transitions of the martingales across time. The formal condition is introducedin Definition 3.1 and illustrated in Figure 1. Basic properties are stated in Remark 3.2, andthe duality for the time-homogeneous version of the martingale optimal transport problemis given in Theorem 3.3. The duality stated in Theorem 3.3 is shortly discussed in Remark3.4 in terms of swap contracts.We first recall the following: Any π ∈ P ( R ) can be disintegrated as π = π ⊗ K where π is the first marginal of π and K : R → P ( R ) is a (Borel measurable) stochastic kernel, whichis π -a.s. unique. Second, for two measures µ, ν there is a unique Lebesgue decomposition µ = µ ν, abs + µ ν, singular , where µ ν, abs (cid:28) ν and µ ν, singular ⊥ ν . Note that µ ν, abs and ν µ, abs have the same null-sets.The notation used in the following definition is fixed throughout the paper.3 imestock price sx S s = x s + τ ( S s + τ − k ) + | ( S s = x ) price: p s ,τ, k ( x ) tS t = x t + τ ( S t + τ − k ) + | ( S t = x ) price: p t ,τ, k ( x ) Assumption: p s ,τ, k ( · ) = p t ,τ, k ( · ) µ t ∧ µ s -a.s.(Homogeneity) for all ( s , t , τ ) ∈ ∆ and k ∈ R Figure 1: Illustration of Definition 3.1(ii) and Remark 3.2(i). Homogeneity for ( S , ..., S T ) states that the forward looking option pricing rules are independent of the time t or s , giventhat all available information is that the stock is in the same state x for both time points. Definition 3.1.
Let ∆ = { ( s, t, τ ) ∈ { , ..., T } : s < t, t + τ ≤ T } .(i) For µ, ν ∈ P ( R ) we say that an event holds µ ∧ ν -almost surely, if it holds almostsurely with respect to µ ν, abs , which is the absolutely continuous part of µ with respectto ν given by Lebesgue’s decomposition theorem.(ii) We say that Q ∈ P ( R T ) is homogeneous, if K s,s + τ = K t,t + τ Q s ∧ Q t -a.s. for all ( s, t, τ ) ∈ ∆ , where K s,s + τ denotes the stochastic kernel given by Q ( s,s + τ ) = Q s ⊗ K s,s + τ .(iii) We set P hom ( R T ) := { Q ∈ P ( R T ) : Q is homogeneous } , Π hom ( µ , ..., µ T ) := Π( µ , ..., µ T ) ∩ P hom , M hom ( µ , ..., µ T ) := M ( µ , ..., µ T ) ∩ P hom . Remark 3.2.
The proof of the following statements is given in Section 6.(i) For Q ∈ Π( µ , ..., µ T ) define the pricing rule p s,τ,k ( x ) := (cid:90) R ( y − k ) + K s,s + τ ( x, dy )= E Q [( S t + τ − k ) + | S t = x ] . Q is homogeneous if and only if p s,τ,k = p t,τ,k holds µ s ∧ µ t -a.s. for all ( s, t, τ ) ∈ ∆ and k ∈ R .(ii) P hom ( R T ) is not convex, but Π hom ( µ , ..., µ T ) and M hom ( µ , ..., µ T ) are convex.(iii) Let P HM ( R T ) be the set of measures Q ∈ P ( R T ) such that the canonical process ( S , ..., S T ) is a homogeneous Markov chain under Q . It holds conv( P HM ( R T )) ⊂P hom ( R T ) , which is strict for T ≥ .(iv) It is Π hom ( µ , ..., µ T ) (cid:54) = ∅ if and only if ( µ , ..., µ T − ) dominates ( µ , ..., µ T ) in hetero-geneity (see [23, Definition 3.4.]). To state duality, we make the following assumption: (A)
For all ( s, t ) ∈ { , ..., T } there exists a finite Borel measure θ s,t on R which is equivalentto µ µ s , abs t such that dθ s,t dµ t and dθ s,t dµ s are continuous and bounded.This is satisfied in a large number of cases, for example if all marginals µ , ..., µ t are discrete,or if all marginals have a continuous and strictly positive Lebesgue density. We furthernote that properties (like continuity) of densities are always understood in the sense thatthere exists a representative among the almost sure equivalence class satisfying the property.
Theorem 3.3.
Let µ , ..., µ T ∈ P ( R ) have finite first moment, and f ∈ C lin ( R T ) . Assume M hom ( µ , ..., µ T ) (cid:54) = ∅ and ( A ) holds. Then max Q ∈M hom ( µ ,...,µ T ) E Q [ f ( S )] (HMOT) = inf (cid:26) T (cid:88) t =1 (cid:90) R h t dµ t : h t ∈ C lin ( R ) , t ∈ { , ..., T } ,ϑ t ∈ C b ( R t ) , t ∈ { , ..., T − } ,g s,t,τ ∈ C b ( R ) , ( s, t, τ ) ∈ ∆ , such that f ( S ) ≤ T (cid:88) t =1 h t ( S t ) + T − (cid:88) t =1 ϑ t ( S , . . . , S t ) (cid:0) S t +1 − S t (cid:1) + (cid:88) ( s,t,τ ) ∈ ∆ (cid:16) g s,t,τ ( S s , S s + τ ) dθ s,t dµ s ( S s ) − g s,t,τ ( S t , S t + τ ) dθ s,t dµ t ( S t ) (cid:17)(cid:27) The proof of Theorem 3.3 is given in Section 6. We thank Ruodu Wang for pointing this relation out to us. In fact, the latter generalizes to the case where µ , ..., µ T all have a continuous and strictly positivedensity with respect to any reference measure θ . Then, one can define θ s,t by dθ s,t dθ := min { dµ t dθ , dµ s dθ } andfinds that dθ s,t dµ t and dθ s,t dµ s are continuous and bounded by 1. emark 3.4. Compared to the usual martingale optimal transport, the additional tradingterm arising from the homogeneity condition in the dual formulation is the sum (cid:88) ( s,t,τ ) ∈ ∆ (cid:16) g s,t,τ ( S s , S s + τ ) dθ s,t dµ s ( S s ) − g s,t,τ ( S t , S t + τ ) dθ s,t dµ t ( S t ) (cid:17) . Each individual summand can be interpreted as a swap contract which, under the assump-tion of time-homogeneity, has fair price 0.To simplify, say g s,t,τ ( S s , S s + τ ) = V ( S s ) · ( S s + τ − k ) + . This means, one buys V ( S s ) many call options at time s which expire at time s + τ . Under homogeneity, see Remark3.2 (i), conditioned on S s = x , the price of such a financial instrument is the same whenreplacing time point s with some other time point t . If two traders were to agree that such aninstrument is equally valuable for time points s and t , it has to be taken into considerationhow likely the events S s = x and S t = x are. The fair weighting to take this into account isachieved by the terms dθ s,t dµ s ( S s ) and dθ s,t dµ t ( S t ) . This section aims at discussing the following extensions and variations of the approach:(i) Non-equally spaced time-grids(ii) Variations of the time-homogeneity assumption(iii) Market frictions(iv) Extension to several assets (high-dimensional market)The first two points are specific to the setting at hand, while the latter two points arereoccurring themes in robust pricing. We hence go briefly over the latter two issues, whilereferencing related work.
Non-equally spaced time-grids.
The notion of time-homogeneity introduced in Section3 makes sense when subsequent time steps are equally far apart, i.e. if there exists a constant C such that time points t and s are | t − s | · C many trading days apart. Available data onoption prices is not always equally spaced. The framework can account for this by modelingthe time steps as t < t < ... < t N with t i ∈ N (instead of t = 1 , ..., T ), where | t i − t j | measures the number of trading days between time points t i and t j . Then one can set ˜∆ := { ( i, j, τ i , τ j ) ∈ { , ..., N } : i < j, i + τ i ≤ N, j + τ j ≤ N, | t i + τ i − t i | = | t j + τ j − t j |} and Definition 3.1 (i) changes to equality of K i,τ i = K j,τ j for all ( i, j, τ i , τ j ) ∈ ˜∆ , where now Q ( t i ,t i + τi ) = µ t i ⊗ K i,τ i , etc. 6 ariations of the time-homogeneity assumption. A natural stronger version of time-homogeneity is the extension from one-period transitions to many-period transitions. Fortwo-period transitions for instance, Definition 3.1 can be extended via the condition K s, ( s + τ ,s + τ ) = K t, ( t + τ ,t + τ ) Q s ∧ Q t -a.s. for all ( s, t, τ ) , ( s, t, τ ) ∈ ∆ with τ < τ , where Q ( s,s + τ ,s + τ ) = Q s ⊗ K s, ( s + τ ,s + τ ) and K s, ( s + τ ,s + τ ) : R → P ( R ) . With such anextension, all relevant properties like convexity of Π hom ( µ , ..., µ T ) remain unchanged.Weakening the notion of time-homogeneity can be done in various ways. First, note that K s,τ = K t,τ µ s ∧ µ t -a.s. ⇔ θ s,t ⊗ K s,τ = θ s,t ⊗ K t,τ with θ s,t as in condition (A) stated before Theorem 3.3. So time-homogeneity can simply bestated as equalities of measures. A natural relaxation is to instead assume that the measuresare close in a suitable distance D ( · , · ) , like Wasserstein-distance or relative entropy. Therelaxation from homogeneity to r -homogeneity takes the form θ s,t ⊗ K s,τ = θ s,t ⊗ K t,τ Relaxation −→ D ( θ s,t ⊗ K s,τ , θ s,t ⊗ K t,τ ) ≤ r s,t,τ , where r s,t,τ ≥ for all ( s, t, τ ) ∈ ∆ . If the mapping ( µ, ν ) (cid:55)→ D ( µ, ν ) is convex, the set of r -homogeneous couplings between µ , ..., µ T remains convex.Alternatively, one can directly penalize the distance between θ s,t ⊗ K s,τ and θ s,t ⊗ K t,τ in the statement of the optimization problem, which leads to sup Q ∈M ( µ ,...,µ T ) E Q [ f ( S )] − (cid:88) ( s,t,τ ) ∈ ∆ r s,t,τ D ( θ s,t ⊗ K s,τ , θ s,t ⊗ K t,τ ) . (Pen-HMOT)For appropriately chosen D ( · , · ) , this penalization corresponds to the inclusion of transactioncosts in the dual formulation, which is discussed below. Market frictions.
The most flexible notion of market frictions that can be incorporatedin the framework is that of transaction costs. Transaction costs result in more costly hedgingstrategies on the dual side, and in relaxed constraints for the considered models Q on theprimal side. Proportional transaction costs correspond to an enlargement of the set offeasible models, see e.g. [7, 9, 14]. With superlinear transaction costs on the other hand, theconstraint is completely removed, and instead a penalization term is added to the objectivefunction, see e.g. [2, 7].An instance of such a penalized primal formulation resulting from superlinear transactioncosts is the above defined (Pen-HMOT) for appropriately chosen penalization term. If D = G is the Gini index in (Pen-HMOT), this corresponds to the use of quadratic transactioncosts in the dual formulation, i.e. the term g s,t,τ ( S s , S s + τ ) dθ s,t dµ s ( S s ) − g s,t,τ ( S t , S t + τ ) dθ s,t dµ t ( S t ) For two measures ν, µ the Gini index G is defined as G ( ν, µ ) = (cid:82) (cid:16) dνdµ (cid:17) dµ − , if ν (cid:28) µ and G ( ν, µ ) = ∞ ,else. See also [20]. r s,t,τ | g s,t,τ ( S s , S s + τ ) | dθ s,t dµ s ( S s ) . So it holds
Corollary 4.1.
Under the assumptions of Theorem 3.3, it holds max Q ∈M ( µ ,...,µ T ) E Q [ f ( S )] − (cid:88) ( s,t,τ ) ∈ ∆ r s,t,τ G ( θ s,t ⊗ K s,τ , θ s,t ⊗ K t,τ ) ( G -Pen-HMOT) = inf (cid:26) T (cid:88) t =1 (cid:90) R h t dµ t : h t ∈ C lin ( R ) , t ∈ { , ..., T } ,ϑ t ∈ C b ( R t ) , t ∈ { , ..., T − } ,g s,t,τ ∈ C b ( R ) , ( s, t, τ ) ∈ ∆ , such that f ( S ) ≤ T (cid:88) t =1 h t ( S t ) + T − (cid:88) t = ϑ t ( S , . . . , S t ) (cid:0) S t +1 − S t (cid:1) + (cid:88) ( s,t,τ ) ∈ ∆ (cid:16) g s,t,τ ( S s , S s + τ ) dθ s,t dµ s ( S s ) − g s,t,τ ( S t , S t + τ ) dθ s,t dµ t ( S t ) − r s,t,τ | g s,t,τ ( S s , S s + τ ) | dθ s,t dµ s ( S s ) (cid:17)(cid:27) The proof is sketched in Section 6.3. In general, the time-homogeneity assumptionbehaves quite similarly to the martingale or marginal assumptions in terms of transactioncosts, and hence many different modeling approaches can be applied.In one respect, the notion of time-homogeneity is however more restrictive: When includ-ing the assumption of time-homogeneity, one has to take care with relaxing the assumptionof precisely knowing the marginal laws. The reason is that P hom ( R T ) is not convex, andthe optimization problem only becomes feasible by adding constraints so that the resultingset of models Q is convex (this is achieved by specifying marginal distributions, see Remark3.2(ii)), which is crucial for the tractability of the resulting optimization problem. Extension to several assets.
The martingale optimal transport setting generalizes asfollows to higher dimensions: One specifies µ t,i of each stock S t,i for time points t = 1 , ..., T and dimensions i = 1 , ..., d individually. The dimensions are coupled through a joint mar-tingale constraint E [ S t +1 ,i | S , ..., S t ] = S t,i where S t = ( S t, , ..., S t,d ) . In terms of homogeneity, one has two choices: First, one can define the notion of ho-mogeneity as in Definition 3.1 (ii) for each dimension i = 1 , ..., d individually. This is See also [10, 18]. Some papers [8, 13, 21] extend the MOT problem in a different way, where it isassumed that for each time point, the d -dimensional marginal distribution is known. While it leads to aninteresting mathematical problem, this assumption is less well justified from a financial viewpoint, as onecan only infer each individual one-dimensional marginal distribution from market data. f ( S ) = ( S − S ) + are depicted. For both figures, time steps used indicates howmany marginal distributions are known, i.e. how much market data is used. We see thatfor the martingale optimal transport approach alone, using more data does not improve theobtained price bounds. Incorporating homogeneity however leads to improved bounds whenadding data.straightforward and sensible, and the resulting optimization problem remains convex. Analternative to take into consideration is to specify homogeneity jointly across dimensions,similarly to the martingale constraint. Then, Definition 3.1 is stated for Q ∈ P (( R d ) T ) , and Q t ∈ P ( R d ) , etc. In this case however, convexity of the set of models Q becomes an issue.If only the individual one-dimensional marginals of S t,i are known, the set of models Q forthe time-homogeneous multi-dimensional martingale optimal transport problem will not beconvex. Hence, for financial purposes, the first alternative is more suitable. In this section, we present two short numerical examples that showcase the potential of theintroduced setting.
Consider a discrete model where µ t is the uniform distribution on the set { − t, − t +2 , ...,
100 + t } for t = 1 , ..., . So the support of the marginals is the same as in a binomialmodel where the stock starts at at time point and can either go up or down by eachperiod. The financial instrument is a forward start option, f ( S ) = ( S − S ) + . First, wesolve the model using just the data (i.e. marginal distributions) from time points t = 8 , (time steps used = 2 ). Then, we gradually increase the information that is used, by addingthe marginal information from t = 7 (time steps used = 3 ), t = 6 , etc. until all marginals9 , ..., µ are included (time steps used = 9 ). The results are reported in Figure 2. Onthe left, we see that without the homogeneity assumption, the bounds do not get sharperwith additional information used. With the added assumption of homogeneity however, thebounds tighten drastically. Let T = 3 and µ t ∼ X t for t = 1 , , , where X t = X exp( σW t − σ t ) for σ > is a geometricBrownian motion. Following [1, 22] we consider the option f ( S , S , S ) := ( S − S + S ) + .Set X = 1 , σ = 0 . . The model price for the Black-Scholdes model Q BS is given by E Q BS [ f ( S )] ≈ . . Compared to the previous example, where the (homogeneous) martingale transport problemis a linear program, the current example has to be solved approximately. Discretization isnon-trivial even with just the martingale condition, see [1, 14]. Homogeneity, which cruciallydepends on the given marginals’ support, adds difficulty for a discretization scheme. Hence,we instead calculate this example using the dual formulation and the penalization approachof [11], i.e. we approximate each trading strategy h t , ϑ t and g s,t,τ by a neural network. Without the homogeneity, this leads to inf Q ∈M ( µ ,µ ,µ ) E Q [ f ( S )] ≈ . and sup Q ∈M ( µ ,µ ,µ ) E Q [ f ( S )] ≈ . . On the other hand, incorporating homogeneity improves the bounds slightly but notably to inf Q ∈M hom ( µ ,µ ,µ ) E Q [ f ( S )] ≈ . and sup Q ∈M hom ( µ ,µ ,µ ) E Q [ f ( S )] ≈ . . While the strengths of the homogeneity assumption certainly lie with cases where more timesteps are involved, even in this example the bounds are narrowed by around .Assuming homogeneity of the underlying process also becomes more restrictive when themarginals µ , µ , µ are less homogeneously evolving. As an extreme case, if in the abovewe instead set µ ∼ X , then the interval of possible prices for the MOT is [0 . , . while for the homogeneous MOT it is [0 . , . , which is drastically more narrow. To approximate a trading strategy with d inputs, we use a network structure with 5 layers, hiddendimension · d and ReLu activation function. For the penalization as in [11], we use the product measure θ = µ × µ × µ and β γ ( x ) = 10000 max { , x } . For training, we use batch size 8192, learning rate . (after the first 60000 iterations, learning rate is decreased by a factor of 0.98 each 250 iterations for another60000 iterations), and the Adam optimizer with default parameters. The reported values are primal values,as described at the start of Section 4 in [11]. In comparison, using the discretization scheme from [1, Subsection 6.4.] with 100 samples for eachmarginal, we get inf Q ∈M ( µ ,µ ,µ ) E Q [ f ( S )] ≈ . and sup Q ∈M ( µ ,µ ,µ ) E Q [ f ( S )] ≈ . . Proofs
Proof of (i): If Q is homogeneous, then by definition p s,τ,k = p t,τ,k holds µ s ∧ µ t -a.s.. Thereverse follows since the function class { h ( x ) = ( x − k ) + : k ∈ R } is measure determining,see e.g. [4, Footnote 2]. Proof of (ii):
First, we show that P hom ( R T ) is not convex. Consider T = 3 and thetwo homogeneous Markov chains Q a = 0 . δ (0 , , + 0 . δ (1 , , and Q b = 0 . δ (0 , , +0 . δ (1 , , . Both Markov chains start in state with probability . and state 1 withprobability . . Chain a always switches states and chain b always stays in the same state.Obviously Q a , Q b ∈ P hom ( R ) . But Q := 0 . Q a + 0 . Q b (cid:54)∈ P hom ( R ) . Indeed, at time the Markov chain transitions from state to each state with equal probability, i.e. withthe notation as in Definition 3.1, it is K , (0) = 0 . δ + 0 . δ . However, at time it is K , (0) = 0 . δ + 0 . δ .Next, we show that Π hom ( µ , ..., µ T ) is convex, which implies that M hom ( µ , ..., µ T ) is convex too. Let Q a , Q b ∈ Π hom ( µ , ..., µ T ) , λ ∈ (0 , and Q := λ Q a + (1 − λ ) Q b .Take ( s, t, τ ) ∈ ∆ . We have to show K s,s + τ = K t,t + τ with notation as in Definition 3.1,i.e. Q ( s,s + τ ) = Q s ⊗ K s,s + τ . Denote by K as,s + τ the stochastic kernel satisfying Q a ( s,s + τ ) = µ s ⊗ K as,s + τ (same for K bs,s + τ ). By the general formula K ( s,s + τ ) = λ d Q as d Q s K as,s + τ + (1 − λ ) d Q bs d Q s K bs,s + τ , and since all measures have the same marginals, it follows K s,s + τ = λK as,s + τ +(1 − λ ) K bs,s + τ ,which yields the claim. Proof of (iii):
The inclusion is trivial: Indeed, if K : R → P ( R ) is the transition ker-nel of the homogeneous Markov chain, then with the notation as in Definition 3.1, it is K s,s + τ = K τ , which is independent of s . (Hereby, K s is defined as usual by K s +1 ( x, A ) := (cid:82) K ( y, A ) K s ( x, dy ) .)That the inclusion is strict, consider the following example: Let Q a := 0 . δ (0 , , +0 . δ (1 , , and Q b := 0 . δ (0 , , + 0 . δ (1 , , . It is Q := 0 . Q a + 0 . Q b ∈ P hom ( R ) . However,straightforward calculation shows that Q cannot be written as a convex combination ofhomogeneous Markov chains, so Q (cid:54)∈ conv( P HM ( R )) . Proof of (iv):
First, we note the following: If Π HM ( µ , ..., µ T ) = Π( µ , ..., µ T ) ∩ P HM ( R T ) ,then Π hom ( µ , ..., µ T ) (cid:54) = ∅ if and only if Π HM ( µ , ..., µ T ) (cid:54) = ∅ . Indeed, by inclusion, the’if’ direction is clear. On the other hand, if Q ∈ Π hom ( µ , ..., µ T ) , then one can define Q HM ∈ Π HM ( µ , ..., µ T ) as the Markov chain with initial distribution µ and transition kernel K , , where Q (1 , = µ ⊗ K , . Now, by [24, Theorem 9.7.3, (iii) ⇔ (iv’)], it follows that Π HM ( µ , ..., µ T ) (cid:54) = ∅ if and only if ( µ , ..., µ T − ) dominates ( µ , ..., µ T ) in heterogeneity.11 .2 Proof of Theorem 3.3 Define φ ( f ) as the infimum term in the statement of the proposition for f ∈ C lin ( R T ) . For π ∈ P ( R T ) , the convex conjugate φ ∗ is given by φ ∗ ( π ) := sup f ∈ C lin ( R T ) (cid:16) (cid:90) f dπ − φ ( f ) (cid:17) . We show φ ( f ) = sup π ∈P ( R T ) (cid:82) f dπ − φ ∗ ( π ) using [3, Theorem 2.2.] and calculate φ ∗ ( π ) sothat the proposition follows. Dual representation:
To apply [3, Theorem 2.2.], we show that φ ( f ) is real-valued on C lin ( R T ) and condition (R1) stated within the Theorem holds, i.e. for C lin ( R T ) (cid:51) f n ↓ it holds φ ( f n ) ↓ φ (0) for n → ∞ . Regarding φ ( f ) ∈ R , φ ( f ) < ∞ is obvious. Onthe other hand, φ ( f ) > −∞ will follow by calculation of φ ∗ ( π ) and the assumption that M hom ( µ , ..., µ T ) is non-empty, since for π ∈ M hom ( µ , ..., µ T ) then φ ∗ ( π ) ≥ (cid:82) f dπ − φ ( f ) and since (cid:82) f dπ < ∞ (all marginals have first moments and f ∈ C lin ( R T ) ), it holds φ ( f ) > −∞ . Regarding condition (R1), note that φ (0) = 0 and φ ( f ) ≤ φ ot ( f ) where φ ot isthe optimal transport functional φ ot ( f ) := inf h ,...,h T ∈ C lin ( R ): ∀ x ∈ R T : (cid:80) Tt =1 h t ( x t ) ≥ f ( x ) T (cid:88) t =1 (cid:90) R h t dµ t . Since φ ot is continuous from above on C lin ( R T ) (see e.g. [12, Proof of Theorem 1]), φ is aswell. Computation of the convex conjugate:
We show φ ∗ ( π ) = 0 , if π ∈ M hom ( µ , ..., µ T ) ,and φ ∗ ( π ) = ∞ , else. First, by plugging in the definition of φ ( f ) , exchanging supremaand plugging in the maximal f (note that to choose f maximally, the condition that θ s,t isbounded and continuous is used), one obtains: φ ∗ ( π ) = sup h t ∈ C lin ( R ) T (cid:88) t =1 (cid:90) R h t dπ t − (cid:90) R h t dµ t (a) + sup ϑ t ∈ C b ( R t ) T − (cid:88) t =1 (cid:90) R T ϑ t ( x , ..., x t ) · ( x t +1 − x t ) π ( dx , ..., dx T ) (b) + sup g s,t,τ ∈ C b ( R ) (cid:88) ( s,t,τ ) ∈ ∆ (cid:90) R T (cid:16) g s,t,τ ( x s , x s + τ ) dθ s,t dµ s ( x s ) − g s,t,τ ( x t , x t + τ ) dθ s,t dµ t ( x t ) (cid:17) π ( dx , ..., dx T ) (c)By martingale optimal transport duality, we have: Term (a) is zero if π t = µ t for all t = 1 , ..., T , and else infinity. Term (b) is zero if the canonical process is a martingale under π , and else infinity. It only remains to show that term (c) is zero if π is homogeneous, and12lse infinity. This is done already under the assumption that π t = µ t for all t = 1 , ..., T . Wewrite π ( t,t + τ ) = π t ⊗ K t,t + τ . Then one calculates for ( s, t, τ ) ∈ ∆ and g s,t,τ ∈ C b ( R ) (cid:90) R T (cid:16) g s,t,τ ( x s , x s + τ ) dθ s,t dµ s ( x s ) − g s,t,τ ( x t , x t + τ ) dθ s,t dµ t ( x t ) (cid:17) π ( dx , ..., dx T )= (cid:90) R g s,t,τ ( x s , x s + τ ) dθ s,t dµ s ( x s ) π ( s,s + τ ) ( dx s , dx s + τ ) − (cid:90) R g s,t,τ ( x t , x t + τ ) dθ s,t dµ t ( x t ) π ( t,t + τ ) ( dx t , dx t + τ )= (cid:90) R g s,t,τ ( x s , x s + τ ) θ s,t ⊗ K s,s + τ ( dx s , dx s + τ ) − (cid:90) R g s,t,τ ( x t , x t + τ ) θ s,t ⊗ K t,t + τ ( dx t , dx t + τ )= (cid:90) R g s,t,τ dθ s,t ⊗ K s,s + τ − (cid:90) R g s,t,τ dθ s,t ⊗ K t,t + τ And hence, term (c) is zero if θ s,t ⊗ K s,s + τ = θ s,t ⊗ K t,t + τ for all ( s, t, τ ) ∈ ∆ , and elseinfinity. So it is zero if and only if K t,t + τ = K s,s + τ holds θ s,t -a.s. for all ( s, t, τ ) ∈ ∆ . This,by choice of θ s,t , corresponds to µ s ∧ µ t -a.s. equality, and hence term (c) is zero if and onlyif π is homogeneous. We sketch the proof for the duality formula shown for ( G -Pen-HMOT) from Section 4. Theargument works the same as for Theorem 3.3. The only difference is term (c), which for thecase of transaction costs reads sup g s,t,τ ∈ C b ( R ) (cid:88) ( s,t,τ ) ∈ ∆ (cid:90) R T (cid:16) g s,t,τ ( x s , x s + τ ) dθ s,t dµ s ( x s ) − g s,t,τ ( x t , x t + τ ) dθ s,t dµ t ( x t ) − r s,t,τ | g s,t,τ ( x s , x s + τ ) | dθ s,t dµ s ( x s ) (cid:17) π ( dx , ..., dx T ) . (c)Similar to the proof of Theorem 3.3, using π t = µ t for t = 1 , ..., T , this simplifies to (cid:88) ( s,t,τ ) ∈ ∆ sup g s,t,τ ∈ C b ( R ) (cid:90) R g s,t,τ dθ s,t ⊗ K s,s + τ − (cid:90) R g s,t,τ dθ s,t ⊗ K t,t + τ − r s,t,τ (cid:90) R | g s,t,τ | dθ s,t ⊗ K s,s + τ and finally the terms inside the sum are precisely the dual representation for the Gini index G ( θ s,t ⊗ K s,s + τ , θ s,t ⊗ K t,t + τ ) as shown in [20], which yields the claim. References [1] A. Alfonsi, J. Corbetta, and B. Jourdain. Sampling of probability measures in the convex orderand computation of robust option price bounds. hal-01963507 , 2018.
2] P. Bank, Y. Dolinsky, S. Gökay, et al. Super-replication with nonlinear transaction costs andvolatility uncertainty.
The Annals of Applied Probability , 26(3):1698–1726, 2016.[3] D. Bartl, P. Cheridito, and M. Kupper. Robust expected utility maximization with mediallimits.
Journal of Mathematical Analysis and Applications , 471(1-2):752–775, 2019.[4] M. Beiglböck, P. Henry-Labordère, and F. Penkner. Model-independent bounds for optionprices—a mass transport approach.
Finance and Stochastics , 17(3):477–501, 2013.[5] M. Beiglböck, N. Juillet, et al. On a problem of optimal transport under marginal martingaleconstraints.
The Annals of Probability , 44(1):42–106, 2016.[6] D. T. Breeden and R. H. Litzenberger. Prices of state-contingent claims implicit in optionprices.
Journal of business , pages 621–651, 1978.[7] P. Cheridito, M. Kupper, and L. Tangpi. Duality formulas for robust pricing and hedging indiscrete time.
SIAM Journal on Financial Mathematics , 8(1):738–765, 2017.[8] H. De March, N. Touzi, et al. Irreducible convex paving for decomposition of multidimensionalmartingale transport plans.
The Annals of Probability , 47(3):1726–1774, 2019.[9] Y. Dolinsky and H. M. Soner. Robust hedging with proportional transaction costs.
Financeand Stochastics , 18(2):327–347, 2014.[10] S. Eckstein, G. Guo, T. Lim, and J. Obłój. Multi-martingale optimal transport: Structure andnumerics.
In preparation , 2019.[11] S. Eckstein and M. Kupper. Computation of optimal transport and related hedging problemsvia penalization and neural networks.
Applied Mathematics & Optimization , pages 1–29, 2018.[12] S. Eckstein, M. Kupper, and M. Pohl. Robust risk aggregation with neural networks. arXivpreprint arXiv:1811.00304 , 2018.[13] N. Ghoussoub, Y.-H. Kim, and T. Lim. Structure of optimal martingale transport plans ingeneral dimensions.
The Annals of Probability , 47(1):109–164, 2019.[14] G. Guo and J. Obloj. Computational methods for martingale optimal transport problems.
Annals of Applied Probability , 2019.[15] J. Guyon. The joint s&p 500/vix smile calibration puzzle solved.
Available at SSRN 3397382 ,2019.[16] P. Henry-Labordère. Automated option pricing: Numerical methods.
International Journal ofTheoretical and Applied Finance , 16(08):1350042, 2013.[17] J. C. Jackwerth and M. Rubinstein. Recovering probability distributions from option prices.
The Journal of Finance , 51(5):1611–1631, 1996.[18] T. Lim. Multi-martingale optimal transport. arXiv preprint arXiv:1611.01496 , 2016.[19] E. Lütkebohmert and J. Sester. Tightening robust price bounds for exotic derivatives.
Quan-titative Finance , pages 1–19, 2019.[20] F. Maccheroni, M. Marinacci, and A. Rustichini. A variational formula for the relative Giniconcentration index.
CiteSeer, doi=10.1.1.561.6568 , 2004.[21] J. Obłój and P. Siorpaes. Structure of martingale transports in finite dimensions. arXiv preprint rXiv:1702.08433 , 2017.[22] J. Sester. Robust price bounds for derivative prices in markovian models. Available at SSRN:https://ssrn.com/abstract=3297018 or http://dx.doi.org/10.2139/ssrn.3297018 , 2018.[23] J. Shen, Y. Shen, B. Wang, and R. Wang. Distributional compatibility for change of measures.
Finance and Stochastics , pages 1–34, 2019.[24] E. Torgersen.
Comparison of Statistical Experiments , volume 36. Cambridge University Press,1991., volume 36. Cambridge University Press,1991.