Good Deal Hedging and Valuation under Combined Uncertainty about Drift and Volatility

We study robust notions of good-deal hedging and valuation under combined uncertainty about the drifts and volatilities of asset prices. Good-deal bounds are determined by a subset of risk-neutral pricing measures such that not only opportunities for arbitrage are excluded but also deals that are too good, by restricting instantaneous Sharpe ratios. A non-dominated multiple priors approach to model uncertainty (ambiguity) leads to worst-case good-deal bounds. Corresponding hedging strategies arise as minimizers of a suitable coherent risk measure. Good-deal bounds and hedges for measurable claims are characterized by solutions to second-order backward stochastic differential equations whose generators are non-convex in the volatility. These hedging strategies are robust with respect to uncertainty in the sense that their tracking errors satisfy a supermartingale property under all a-priori valuation measures, uniformly over all priors.


Introduction
Hedging and valuation under model uncertainty (ambiguity) about volatility has been a seminal problem in the topical area of robust finance. In mathematics, it has motivated to no small extend recent advances on subjects such as second order backward stochastic differential equations (2BSDEs), G-expectations and related stochastic calculus, sub-linear conditional expectations and control of non-linear kernels, using a variety of different methods from stochastic control, quasi-sure analysis and capacity theory, or expectation-spaces and PDE-theory, see Denis and Martini (2006); Denis et al. (2011);Nutz (2012b); Nutz and Soner (2012); ; Nutz and van Handel (2013); Hu et al. (2014a);  and many more references therein. The research has been challenging (and fruitful) since one has to deal (in a probabilistic setup) with families of non-dominated probability measures, also called multiple priors, that can be mutually singular. In contrast, uncertainty solely about drifts in a continuous time setting of stochastic Itô-processes could be dealt with in a dominated framework of measures which are absolutely continuous with respect to a single reference probability measure.
The main contributions of the current paper are twofold. For incomplete markets in continuous time, we solve the problem of robust hedging and valuation under combined uncertainty about both the drifts and the volatilities of the Itô processes which describe the evolution of the tradeable asset prices in an underlying non-Markovian model for the financial market. Further, we investigate to this end the no-good-deal approach to hedging and valuation, that is much cited in the finance literature (cf. Cochrane and Saá-Requejo (2000); Cerný and Hodges (2002); Björk and Slinko (2006)) and provides more narrow valuation bounds and less extreme hedges than the more fundamental approach of almost-sure-hedging by superreplication with its corresponding no-arbitrage valuation bounds.
Concerning the influential application of hedging under volatility uncertainty in continuous-time models which has been stipulated, at least, by Avellaneda et al. (1995); Lyons (1995), the literature so far has almost entirely been concerned with the superreplication approach, for uncertainty being restricted solely to volatility as ambiguity about the drift of the asset prices has no effect there. While the notion of superreplication is fundamental to the theory of stochastic processes and for applications, being related to the optional decomposition and excluding the possibility of losses, it is also known from a practical point of view that superreplication could be overly expensive already in the absence of uncertainty for incomplete market models. This calls for an adaption of other, less conservative, concepts for partial (not almost-sure) hedging in incomplete markets to a framework that is robust with respect to model ambiguity. Under combined uncertainties on drift and volatility, which are going to be relevant to such alternative approaches where the (ambiguous) distributions of hedging errors matter, however new mathematical challenges have to be overcome, as the non-dominated family of measures will not just consist of local martingale measures, as e.g. in . Such has been noted and addressed just recently in Nutz (2012b); Epstein and Ji (2014); . Likewise, we are aware of only few recent articles on the related problem of expected utility maximization under uncertainty about both drifts and volatilities Tevzadze et al. (2013); Biagini and Pınar (2017); Neufeld and Nutz (2016), some of which achieve quite explicit results for models with specific parametric structure. Among many interesting contributions on utility optimization under only one type of uncertainty, see for instance Chen and Epstein (2002); Quenez (2004); Garlappi et al. (2007); Schied (2007); Øksendal and Sulem (2014) for solely (dominated) uncertainty about drifts, or Matoussi et al. (2015); Hu et al. (2014b) for ambiguity solely about volatilities but not about drifts. For equilibrium prices of a representative agent under ambiguity about the volatility, see (Epstein and Ji, 2013, Sect.3.3). To the best of our knowledge, there appear to be hardly any studies on hedging approaches for (generically) incomplete markets under combined ambiguity about drifts and volatilities -apart from superreplication.
We are going to investigate a robust extension of the no-good-deal hedging approach in continuous time under combined ambiguity about the volatilities and drifts. Without model uncertainty, gooddeal bounds have been introduced as valuation bounds in incomplete markets which do not only prevent opportunities for arbitrage but also for deals with an overly attractive risk-to-reward ratio. The most cited reference in the finance literature appears to be Cochrane and Saá-Requejo (2000). We refer to Björk and Slinko (2006); Cerný and Hodges (2002); Klöppel and Schweizer (2007) for mathematical and conceptual ideas and many more references. By using only a suitable subset of "no-good-deal" risk neutral prices, the resulting valuation bounds are tighter than the classical no-arbitrage bounds (which are often too wide) but still have economic meaning. Good-deal bounds have been defined predominantly by constraints on the instantaneous Sharpe ratios in (any) extension of the financial market by additional derivatives' price processes, see Björk and Slinko (2006). For model without jumps, such is equivalent to imposing constraints on the optimal expected growth rates, see Becherer (2009). Although good-deal theory has been initiated merely as a valuation approach (cf. the conclusions in Björk and Slinko, 2006), a corresponding approach to hedging has been proposed, cf. Becherer (2009), where (good-deal) hedging strategies are defined as minimizers of a certain dynamic coherent risk measure, in the spirit of Barrieu and El Karoui (2009), such that the good-deal bounds appear as a market consistent risk measures. Naturally, results on valuations and hedges in good-deal theory, like in Cochrane and Saá-Requejo (2000); Björk and Slinko (2006); Becherer (2009), are sensitive to the assumptions of the probability model on the drifts and volatilities of the asset prices. Since the objective real world measure is not precisely known and financial models can, at best, be useful but idealized simplifications of reality, robust approaches to model ambiguity are relevant to good-deal theory.
As far as we know, a robust approach to good-deal hedging in continuous time under nondominated uncertainty has not been available so far. For drift uncertainty and more references see Becherer and Kentia (2016). Robust results on valuation and hedging will be obtained by 2BSDE theory, building on recent advances by  whose analysis provides a general wellposedness result that fits well for the present application under combined uncertainty, cf. Remark 4.3. Indeed, their result neither requires convexity nor uniform continuity of the generator, and it can deal with general (measurable) contingent claims without assuming further regularity (like e.g. uniform continuity) or a Markovian framework.
The organization of the present paper is as follows. The setup and preliminaries are explained in Section 2, with a brief summary of key results on 2BSDEs. Then we begin Section 3 by a description of the financial market and the (non-dominated) confidence set of (uncertain) priors that captures the combined ambiguity about drifts and volatilities. Let us note that, in comparison to most literature on hedging under ambiguous volatility, like Avellaneda et al. (1995);Lyons (1995), we are going to consider a model for asset prices that constitutes a generically incomplete market, even if seen just under (any) one individual prior. That means not only that there exists in general no perfect hedging (i.e. replicating) strategy which is robust with respect to ambiguity on priors, but that there does not even exists a replicating strategy in general in the model for (any) one given probability prior, without ambiguity. Section 3 then proceeds by taking drift and volatility to be known at first, in order to explain the idea for the good-deal approach as simply as possible. Following classical good-deal theory, good-deal restrictions are defined by constraints on the instantaneous Shape ratios, i.e. by radial bounds on the Girsanov kernels of pricing measures, and standard BSDE descriptions of valuation bounds and hedges are summarized. Adopting a multiple-priors approach, like e.g. in Gilboa and Schmeidler (1989); Chen and Epstein (2002), while accommodating for the fact that priors here are non-dominated, Section 4 starts by defining the good-deal bounds under uncertainty as the worst-case bounds over all priors. For hedging purposes, we define good-deal hedging strategies as minimizers of suitable a-priori risk measures under optimal risk sharing with the market. We derive 2BSDE characterizations for the dynamic valuation bounds and the hedging strategies. We show that tracking errors from good-deal hedging satisfy a supermartingale property under all a-priori valuation measures uniformly for all priors. The proof relies on saddle-point arguments to identify the robust good-deal hedging strategy through a minmax identity. Finally, we finish Section 4 with a simple but instructive example about hedging a put option on a non-traded (but correlated) asset in an incomplete market. This allows for an elementary closed form solution, offering intuition for the general but abstract main Theorem 4.6. It illustrates for instance that the good-deal hedging strategy generally is very different from the super-replicating strategy, which has been studied in e.g. Avellaneda et al. (1995); Lyons (1995); Denis and Martini (2006); Nutz and Soner (2012); Neufeld and Nutz (2013); Vorbrink (2014). The concrete case study also illustrates, how already in an elementary Markovian example additional complications arise from combined uncertainty.

Mathematical framework and preliminaries
We consider filtered probability space (Ω, F = F T , P 0 , F) where Ω is the canonical space {ω ∈ C([0, T ], R n ) : ω(0) = 0} of continuous paths starting at 0 endowed with the norm ω ∞ := sup t∈[0,T ] |ω(t)|. The filtration F = (F t ) t∈[0,T ] is generated by the canonical process B t (ω) := ω(t), ω ∈ Ω and P 0 is the Wiener measure. We denote by F + = (F + t ) t∈[0,T ] the right-limit of F, with F + t = F t+ := ∩ s>t F s . For a probability measure Q, the conditional expectation given F t will be denoted by E Q t [·]. A probability measure P is called a local martingale measure if B is a local martingale w.r.t. (F, P ). One can, cf. Karandikar (1995), construct the quadratic variation process B pathwise such that it coincides with B P P -a.s. for all local martingale measures P . In particular this yields a pathwise definition of the density a of B w.r.t. the Lebesgue measure as We denote by P W the set of all local martingale measures P for which a is well-defined and takes values P -almost surely in the space S >0 n ⊂ R n×n of positive definite symmetric n × n-matrices. Note that the measures in P W can be mutually singular, as illustrated e.g. in Soner et al. (2011).
For any P ∈ P W , the process W P := (P ) · 0 a − 1 2 s dB s is a Brownian motion under P . To formulate volatility uncertainty, we concentrate only on the subclass P S ⊂ P W of measures s.. A benefit of restricting to the subclass P S is the following aggregation property (cf. Soner et al., 2011, Lem.8.1, Lem.8.2).
Lemma 2.1. For P ∈ P W , let F P denote the P -augmentation of the filtration F and F W P P that of the natural filtration F W P of W P . Then B has the martingale representation property w.r.t. (F P , P ) for all P ∈ P S , and P S = P ∈ P W : F P = F W P P . Moreover, (P, F) satisfies the Blumenthal zero-one law for any P ∈ P S . Remark 2.2. For any P ∈ P S , Lemma 2.1 implies E P t [X] = E P [X | F + t ] P -a.s. for any X in L 1 (P ), t ∈ [0, T ]. In particular, any F + t -measurable random variable has a F t -measurable P -version. Let a, a ∈ S >0 n . We will work with the subclass P [a,a] of P S defined by P [a,a] := P ∈ P S : a ≤ a ≤ a, P ⊗ dt-a.e.
(2.1) and assumed to be non-empty. We use the language of quasi-sure analysis as it appears in the framework of capacities of Denis and Martini (2006) as follows.
Definition 2.3. A property is said to hold Q-quasi-everywhere (Q-q.e. for short) for a family Q of measures on the same measurable space if it holds outside of a set, which is a nullset under each element of Q.
Unless stated otherwise, inequalities between random variables will be meant in a P [a,a] -quasisure sense (written P [a,a] -q.s. for short), while inequalities between F + -progressive processes will be in the P [a,a] ⊗ dt-q.e. sense, for P [a,a] ⊗ dt := P ⊗ dt, P ∈ P [a,a] . We now introduce spaces and norms of interest for the paper. Some of these spaces are already quite classical, and have been modified here to account for the possible mutual singularity of measures in P [a,a] . For a filtration a] , we consider the following function spaces: (resp. P -a.s.), and satisfying for the set of measures P [a,a] (t, P, X) := P ∈ P [a,a] : P = P on X t , e) I 2 (X, P ) the space of X-predictable processes K with càdlàg and non-decreasing paths P -a.s., K 0 = 0 P -a.s., and K 2 I 2 (P ) := E P [K 2 T ] < ∞. In particular we denote by I 2 X P P ∈P [a,a] the family of tuples (K P ) P ∈P [a,a] s.t. K P ∈ I 2 (X P , P ) for any P ∈ P [a,a] and sup P ∈P [a,a] K P I 2 (P ) < ∞. The reader will note the analogy with the spaces and norms defined in  (though with slightly different notations) for the specific family of collection of measures P(t, ω) := P [a,a] for any (t, ω) ∈ [0, T ] × Ω. A filtration which might in the sequel play the role of X in the 2.1. Second order backward stochastic differential equations. Following , we summarize an existence and uniqueness result for Lipschitz 2BSDEs and state a representation of solutions that will be key to characterize the good-deal bounds and hedging strategies under combined drift and volatility uncertainties: See Proposition 2.8. The generator for a 2BSDE is a function F : [0, T ] × Ω × R × R n × S >0 n → R for which we denote F t (ω, y, z) := F t (ω ·∧t , y, z, a t (ω)) and F 0 t := F t (0, 0). For wellposedness we will require generators F and terminal conditions X that satisfy the following combination of Assumption 2.1.(i)-(ii) and Assumption 3.1. in  (for κ = p = 2).
(ii) F is jointly Borel measurable, and F-progressive in (t, ω) for each (y, z, a), (iii) ∃ C > 0 such that for all (t, ω, a) ∈ [0, T ] × Ω × S >0 n , y, y ∈ R, z, z ∈ R n , F t (ω, y, z, a) − F t (ω, y , z , a) ≤ C |y − y | + |z − z | , Remark 2.5. Assumption 2.4-(iv) is satisfied for F such that Assumption 2.4-(ii) holds and F 0 is bounded P [a,a] -q.s.. It implies sup A second-order BSDE is a stochastic integral equation of the type In comparison to , because the canonical process B satisfies the martingale representation property simultaneously under all measures in P [a,a] (cf. Lemma 2.1), we do not have the orthogonal martingale components in the formulation of 2BSDEs as (2.2). The same formulation can be used in the more general framework with semimartingale laws for the canonical process, but working under a saturation property for the set of priors (cf. Possamaï et al., 2015, Def.5.1). a] is called solution (triple) to the 2BSDE (2.2) if it satisfies the required dynamics P [a,a] -q.s. and the family K P , P ∈ P [a,a] satisfies the minimum condition If the family {K P , P ∈ P [a,a] } can be aggregated into a single process K, i.e. K P = K, P -a.s. for all P ∈ P [a,a] , then (Y, Z, K) is said to solve the 2BSDE.
Remark 2.7. Note in the 2BSDE dynamics (2.2) the dependence of the stochastic integrals (P ) · 0 Z tr s dB s on the probability measures P ∈ P [a,a] . Indeed since the measures in P [a,a] may be non-dominated, it might be that these integrals do not aggregate (see Soner et al., 2011, for more on aggregation). However under additional set theoretical assumptions (for instance continuum hypothesis plus the axiom of choice) a method by Nutz (2012a) can be used to construct the stochastic integral · 0 Z tr s dB s pathwise for any predictable process Z. As a by-product, the family {K P , P ∈ P [a,a] } for a 2BSDE solution (Y, Z, (K P ) P ∈P [a,a] ) would automatically aggregate into a single process K : Reciprocally for a solution (Y, Z, K), the family (P ) · 0 Z tr s dB s , P ∈ P [a,a] also aggregates. The pair (F, X) will be called the parameters of the 2BSDE (2.2). We will refer to Y as the value process and Z as the control process. The following proposition provides the wellposedness result of interest in this paper, as well as a representation of the value process in terms of solutions to standard BSDEs. The proof relies on an application of (Possamaï et al., 2015, Thm.4.1, Thm.4.2) for the specific family of measures P [a,a] , the details being deferred to the appendix. We employ the classical sign convention for standard BSDE generators according to which the generator of the BSDE (2.5) below is − F (i.e. with a minus sign). The convention for associated 2BSDE generator however remains as already introduced, e.g. the 2BSDE (2.2) has the generator F . Proposition 2.8. If X ∈ L 2 (F + ) and Assumption 2.4, then the 2BSDE (2.2) has a 1. unique solution (Y, Z, (K P ) P ∈P [a,a] a] 2. and for any P ∈ P [a,a] , the Y -part of the solution has the representation where (Y P (τ, H), Z P (τ, H)) denotes the unique solution to the standard BSDE with parameters (− F , H), for an F P -stopping time τ and H ∈ L 2 (F P τ , P ). Remark 2.9. Note in comparison to  that uniform continuity (in (ω)) and convexity (in a) of the generator function F are not required for the more general wellposedness results of  summarized here in Proposition 2.8. The latter is what we need in Section 4 for applications to valuation and hedging under combined drift and volatility uncertainty, as the results of  may not be applicable in that situation; cf. Part 1 of Remark 4.3 for a detailed justification. Note that the generalized theory also works for terminal conditions which are merely Borel measurable and do not need to be in the closure of uniformly continuous functions as required in .

Financial market model and good-deal constraints
We apply the 2BSDE theory of Section 2.1 to good-deal valuation and hedging of contingent claims in incomplete financial markets under drift and volatility uncertainty. In comparison to standard BSDEs which are used in Becherer and Kentia (2016) in the presence of solely drift uncertainty, 2BSDEs are an appropriate tool for describing worst-case valuations in the presence of volatility uncertainty. As in Cochrane and Saá-Requejo (2000); Björk and Slinko (2006), we consider good-deal constraints imposed as bounds on the Sharpe ratios (equivalently bounds on the optimal growth rates as in Becherer, 2009) in the financial market extended by further (derivative) asset price processes. But first we specify the model for the market with uncertainty about the volatility and the market price of risk.
3.1. Financial market with combined uncertainty about drift and volatility. The financial market consists of d tradeable stocks (d ≤ n) with discounted price processes (S i where b (resp. σ) is a R d -valued (resp. R d×n -valued) F-predictable uniformly bounded process, with σ being such that the family (P ) · 0 σ s dB s , P ∈ P [a,a] aggregates into a single process · 0 σ s dB s . (3.1) We assume in addition that σσ tr is uniformly elliptic in the sense that there exists Υ, Λ ∈ (0, ∞) such that Υ I d×d ≤ σσ tr ≤ Λ I d×d , P [a,a] ⊗ dt-q.e., where I d×d denotes the d × d identity matrix. In particular σ a 1/2 is P [a,a] ⊗ dt-q.e. of maximal rank d ≤ n, since σ aσ tr is uniformly elliptic and bounded (by (2.1),(3.2)).
Remark 3.1. For economic interpretation, one should clearly have aggregation of S, which should be quasi-surely defined as a single process. The latter is ensured here by the aggregation condition (3.1) which might seem restrictive at first sight, but is ensured for instance if σ is càdlàg in which case · 0 σ s dB s can even be constructed pathwise as in Karandikar (1995). The market model captures uncertainty about the volatility of the stock prices S which is σ a 1/2 under each measure P ∈ P [a,a] . Since dB t = a 1/2 t dW P t P -a.s. for a P -Brownian motion W P , then the dynamics of (S t denoting the market price of risk in each reference model P ∈ P [a,a] . Note that ξ is R n -valued, F-predictable and uniformly bounded by a constant depending only on a, a, Λ, Υ, and the uniform bound on b. The financial market described is thus typically incomplete for any reference measure P ∈ P [a,a] for the volatility σ a 1/2 if d < n. In practice, the bounds a,ā and the uniform bounds on σσ tr can be viewed as describing some confidence region for future volatility values, which might be set e.g. according to expert opinion about the range of historical or future (implied) volatility scenarios.
To incorporate also uncertainty about the drift, we admit for market prices of risk ξ θ being from a radial set of which ξ is the center, that is we consider , and δ is a fixed non-negative bounded F-predictable process. The set (3.3) of market prices of risk corresponds to an ellipsoidal confidence region of drift uncertainty on risky asset prices S, such that ambiguous drifts could attain values in ellipsoids Ellipsoidal specifications of uncertainty are common in the literature, and appear naturally in the context of uncertainty about the drifts of tradeable asset prices in multivariate Gaussian settings, cf. e.g. Garlappi et al. (2007); Biagini and Pınar (2017).
We denote by Θ : with radial values. The notation " " emphasizes that Θ is set-valued. F-predictability of δ implies that Θ is F-predictable in the sense of Rockafellar (1976) Hence by measurable selection arguments (e.g. Rockafellar, 1976, Cor.1 Combined Knightian uncertainty (i.e. ambiguity) about drift and volatility scenarios is then captured by a (typically non-dominated) set R := Q : Q ∼ P, dQ = (P ) E θ · W P dP for some θ ∈ Θ and P ∈ P [a,a] of candidate reference probability measures (priors), where (P ) E(M ) := exp M − M 0 − 1 2 M P denotes the stochastic exponential of a local martingale M under P . For any Q ∈ R, there are P Q ∈ P [a,a] , θ Q ∈ Θ such that the canonical process B is a Q-semimartingale with decomposition We will simply denote by Q P,θ a reference measure Q ∈ R associated to P ∈ P [a,a] and θ ∈ Θ, and we take note that the specification range for P and θ, respectively, account for uncertainty about volatilities and drifts, respectively. S evolves under Q P,θ ∈ R, P -a.s., as where W P,θ := W P − · 0 θ s ds is a Q P,θ -Brownian motion. Hence ξ θ is the market price of risk in the model Q P,θ , with volatility σ a 1/2 and a satisfying a ≤ a ≤ a, P ⊗ dt-a.e.. Let M e (Q) := M e (S, Q) denote the set of equivalent local martingale measures for S in a model Q. Then standard arguments (analogously to Becherer and Kentia, 2016, Prop.4.1) easily lead to the following Lemma 3.2. For any P ∈ P [a,a] and θ ∈ Θ, the set M e (Q P,θ ) is equal to Remark 3.3. Note by Lemma 3.2 that for each P ∈ P [a,a] and θ ∈ Θ, the minimal martingale measure Q P,θ given by d Q P,θ = (P ) E(− ξ θ · W P,θ ) dQ P,θ is in M e (P ) since ξ and δ are uniformly bounded. This implies that M e (Q) = ∅ for any Q ∈ R. Thus, the market satisfies the no-free lunch with vanishing risk condition (Delbaen and Schachermayer, 1994) under any (uncertain) prior Q ∈ R. We will interpret this as a notion for no-arbitrage under drift and volatility uncertainty (as in e.g. Biagini et al., 2015).
We parametrize trading strategies ϕ = (ϕ i ) d i=1 in terms of amount ϕ i of wealth invested in the stock with price process S i , such that ϕ is a F P [a,a] -predictable process satisfying suitable integrability properties that will be made precise. In this respect, the wealth process V ϕ associated to a trading strategy ϕ with initial capital V 0 (so that (V 0 , ϕ) quasi-surely satisfies the self-financing requirement) would have dynamics, on [0, T ], Re-parameterizing trading strategies in terms of integrands φ := σ tr ϕ ∈ Im σ tr w.r.t. B + · 0 a 1/2 s ξ s ds yields as dynamics for the wealth process for any P ∈ P [a,a] . The set Φ(P ) of permitted strategies in the model P (shortly P -permitted), P ∈ P [a,a] , is defined as Under uncertainty and i.p. for non-dominated priors P ∈ P [a,a] , we want the wealth process for a trading strategy to be defined (q.e.) as single process, and not to vary with the prior. This requires as additional condition on strategies φ that the family (P ) · 0 φ tr s dB s , P ∈ P [a,a] of "profit&loss"-processes aggregates (cf. p.5) into a single process, denoted · 0 φ tr s dB s . To this end, we make the × Ω and such that the family (P ) · 0 φ tr s dB s , P ∈ P [a,a] aggregates into a single process · 0 φ tr s dB s . The trading strategies in Φ will be termed as P [a,a] -permitted (in short: permitted) strategies. By definition, one has Φ ⊆ Φ(P ) for all P ∈ P [a,a] . Hence V φ is a martingale under some measure equivalent to P for any φ ∈ Φ and P ∈ P [a,a] . This excludes arbitrage strategies from Φ for any scenario σ a 1/2 of the volatility.
Remark 3.5. Note that by a result of Karandikar (1995) the aggregation property of Definition 3.4 would be satisfied for φ being càdlàg and F-adapted (like in Remark 3.1 for the volatility σ). Alternatively, aggregation would automatically hold under additional (non-standard) axioms for set-theory as in Nutz (2012a); cf. Remark 2.7.

3.2.
No-good-deal restriction and implications to dynamic valuation and hedging. In the absence of uncertainty, we consider a (classical) no-good-deal restriction defined as a bound on the instantaneous Sharpe ratios, for any extension of the market by additional derivatives' prices computed by the no-good-deal pricing measures, following Björk and Slinko (2006); Cochrane and Saá-Requejo (2000). In a setup without jump, such a constraint is equivalent to a bound on the optimal expected growth rate of returns in any market extension, as in e.g. Becherer (2009). It is known that such is ensured by imposing a bound on the norm of the Girsanov kernels from the risk-neutral pricing measures. As slightly more detailed elaboration on the features of this valuation approach is given in Remark 3.6. Under solely drift uncertainty, robust good-deal hedging has been studied by Becherer and Kentia (2016). To extend the theory by including moreover ambiguity on volatility, we impose as no-good-deal restriction under combined drift and volatility uncertainty now the same Sharpe ratio bound but require it under each model Q ∈ R separately. By doing so, we obtain for each Q ∈ R a set of no-good-deal measures Q ngd (Q) ⊆ M e (Q). Then following a worst-case approach to good-deal valuation under uncertainty we define the robust good-deal bound under drift and volatility uncertainty as the supremum over all no-good-deal prices for all reference priors Q ∈ R.
To explain the idea for the good-deal approach, we will next do so at first without addressing model uncertainty. To this end, let us assume just for the remaining of this subsection that we neither have ambiguity about drifts nor about volatility and let us consider, just as a starting point, one arbitrary probability reference measure Q P,θ from R to be the objective real world measure for the standard (non-robust) good deal problem of valuation and hedging. This means that we take one specification P and θ for volatility and drift to be given. Having explained the problem without ambiguity, this will set the stage to explain the robust problem with model uncertainty thereafter in the main Section 4.
3.2.1. The no-good-deal problem under a given model Q P,θ without uncertainty. Let h be a fixed non-negative bounded F-predictable process. For a given probability measure Q P,θ in R (with P ∈ P [a,a] and θ ∈ Θ), we consider the set Q ngd (Q P,θ ) of no-good-deal measures in the model Q P,θ as the subset of M e (Q P,θ ) consisting of all equivalent local martingale measures Q, whose Girsanov kernels λ w.r.t. the Q P,θ -Brownian motion W P,θ are bounded by h in the sense that |λ t (ω)| ≤ h t (ω) for all (t, ω) ∈ [0, T ] × Ω. More precisely using Lemma 3.2, we define In the sequel, any selection θ (not necessarily measurable) of Θ is taken to satisfy Clearly, (3.5) implies that the minimal martingale measure Q P,θ is in Q ngd (Q P,θ ) = ∅ for any P ∈ P [a,a] , θ ∈ Θ. Beyond this (3.5) will also be used to prove wellposedness of a 2BSDE whose solution will describe the robust good-deal valuation bound and hedging strategy (see Theorems 4.2 and 4.6, respectively). As in (Becherer and Kentia, 2016, Lem.2.1, part b)), one can show for any P ∈ P [a,a] and θ ∈ Θ that the set Q ngd (Q P,θ ) is convex and multiplicatively stable (in short m-stable). The latter property, also referred to as rectangularity in the economic literature Chen and Epstein (2002), is usually required for time-consistency of essential suprema of conditional expectations over priors; see Delbaen (2006) for a general study. For fixed P ∈ P [a,a] and θ ∈ Θ, the upper good-deal bound for a claim X ∈ L 2 (F P T , P ) in the model Q P,θ , for given drift and volatility specification, is defined as Remark 3.6. According to classical good-deal theory in continuous time with continuous price dynamics, the set Q ngd (Q P,θ ) of risk-neutral measures Q is specified in such a way, that in any extension (S, S ) of the financial market by additional derivatives' price processes S , which are (local) Q-martingales, the extended market is not just arbitrage-free but moreover does not permit opportunities for dynamic trading with overly attractive reward to risk ratios. More precisely, the choice of Q ngd (Q P,θ ) is such that the extended market does not offer trading opportunities that yield instantaneous Sharpe ratios which exceed the bound h (see Björk and Slinko, 2006, Section 3 and Appendix A, for details); Equivalently, one can show that the extended market does not admit for portfolio strategies that offer (conditional) expected growth rates larger than h (see Becherer, 2009, Section 3), and the respective good-deal restrictions are sharp. In this sense, the dynamic upper π u,P,θ t (X) and lower −π u,P,θ t (−X) good-deal bounds determine at any time t a sub-interval of the arbitrage-free prices, that is determined by those valuation measures Q which restrict the opportunities for overly good deals.
The hedging objective for the seller of a claim X ∈ L 2 (P ) who believes in the model Q P,θ is to find a trading strategyφ P,θ ∈ Φ(P ) that minimizes the residual risk under a risk measure ρ P,θ from holding X and trading dynamically in the market. As the seller charges a premium π u,P,θ · (X) for X, she would like this premium to be the minimal capital that dynamically makes her position acceptable, so that π u,P,θ becomes the market consistent risk measure corresponding to ρ P,θ , in the spirit of Barrieu and El Karoui (2009). Following Becherer (2009), define ρ P,θ as a dynamic coherent risk measure by Hence P ngd (Q P,θ ) is the set of a-priori valuation measures equivalent to Q P,θ , which satisfy the no-good-deal restriction under Q P,θ but might not be local martingale measures for the stock price process S (yet they are w.r.t. the market with only the riskless asset S 0 ≡ 1). This implies Q ngd (Q P,θ ) = P ngd (Q P,θ ) ∩ M e (Q P,θ ), and thus ρ P,θ t (X) ≥ π u,P,θ t (X), t ∈ [0, T ], P -a.s. for all X ∈ L 2 (P ). Moreover as Q ngd (Q P,θ ) is m-stable and convex, the set P ngd (Q P,θ ) also is, for any P ∈ P [a,a] and θ ∈ Θ. Therefore by (Becherer and Kentia, 2016, Lem.2 For a fixed reference measure Q P,θ , the hedging problem of the seller for X ∈ L 2 (F P T , P ) is then to findφ P,θ ∈ Φ(P ) such that P -almost surely for all t ∈ [0, T ] holds (3.8) One can show (cf. Cor.5.6 in Becherer (2009) (2016)) that for each P in P [a,a] and θ in Θ, the tracking (or hedging) error of the good-deal hedging strategyφ P,θ is a supermartingale under any measure Q ∈ P ngd (Q P,θ ). This supermartingale property of tracking errors could be viewed as a robustness property of the hedging strategy with respect to the family P ngd (Q P,θ ) of 'valuation' probability measures as generalized scenarios (cf. Artzner et al., 1999).

3.2.2.
Standard BSDEs for valuation and hedging under Q P,θ without uncertainty. The solution to the valuation and hedging problem described by (3.6) and (3.8) can be obtained in terms of standard BSDEs under P ∈ P [a,a] . In order to be more precise, let us introduce some notations that will also be used throughout the sequel. For a ∈ S >0 n , we denote by Π a (t,ω) (·) and Π a,⊥ (t,ω) (·) respectively the orthogonal projections onto the subspaces Im σ t (ω)a 1 2 tr and Ker σ t (ω)a 1 2 of R n , (t, ω) ∈ [0, T ] × Ω. Explicitly, for each a ∈ S >0 n and t ∈ [0, T ] (omitting ω-symbols for simplicity), the projections for z ∈ R n are given by In particular we define Π (t,ω) (·) := Π at(ω) (t,ω) (·) and Π ⊥ (t,ω) (·) := Π at(ω),⊥ (t,ω) (·). For P ∈ P [a,a] and θ ∈ Θ, let us consider the standard BSDE under P , for data X ∈ L 2 (F P T , P ) and generator − F θ t (ω, ·) = −F θ (t, ω ·∧t , ·, a t (ω)) with for all (t, ω, z, a) ∈ [0, T ] × Ω × R n × S >0 n . By (Becherer, 2009, Thm.5.4), the upper valuation bound π u,P,θ · (X) and good-deal hedging strategyφ P,θ in the model Q P,θ are identified in terms of the solution (Y P,θ , Z P,θ ) ∈ D 2 (F P , P ) × H 2 (F P , P ) to the standard BSDE (3.9) as π u,P,θ · (X) = Y P,θ and a 1 2 tφ P,θ t As drift uncertainty can be incorporated in a setup with one dominating reference measure, the above standard BSDE description of good-deal bounds and hedging strategies has been extended by Becherer and Kentia (2016) to the presence of uncertainty solely about the drift. Under uncertainty about the volatility, however, a dominating probability measure would fail to exist and hence standard BSDE techniques will not be applicable anymore. Instead, we will rely on 2BSDEs to provide in Section 4, under combined uncertainty about drifts and volatilities, an infinitesimal characterization of (robust) good-deal bounds and hedging strategies, after suitably defining the latter in such a typically non-dominated setup.

Good-deal hedging and valuation under combined uncertainty
We describe good-deal bounds in the market model of Section 3.1 using 2BSDEs and study a corresponding notion of hedging that is robust w.r.t. combined drift and volatility uncertainty. We first define the good-deal valuation bounds from the no-good-deal restriction of Section 3.2, but taking into account the investor's aversion towards drift and volatility uncertainty. Furthermore, still as in Section 3.2, hedging strategies are defined as minimizers of some dynamic coherent risk measure ρ under uncertainty, so that the good-deal bound arises as the market consistent risk measure for ρ, in the spirit of Barrieu and El Karoui (2009), allowing for optimal risk sharing with the market under uncertainty.
For valuation, it seems natural to view uncertainty aversion as a penalization to the no-good-deal restriction in the sense that it implies good-deal bounds under uncertainty that are wider than in the absence of uncertainty. To formalize this idea, we rely on a classical worst-case approach to uncertainty in the spirit of e.g. Gilboa and Schmeidler (1989); Hansen and Sargent (2001). The idea is that an agent who is averse towards ambiguity about the actual drifts and volatilities may opt, to be conservative, for a worst-case approach to valuation in order to compensate for losses that could occur due to the wrong choice of model parameters. Acting this way, she would sell (resp. buy) financial risks at the largest upper (resp. smallest lower) good-deal bounds over all plausible priors in her confidence set R capturing drift and volatility uncertainty. Given the technical difficulties that may arise from R being non-dominated, in particular for writing essential suprema, we define the worst-case upper good-deal bound π u · (X) for a financial risk X ∈ L 2 P [a,a] as the unique process π u · (X) ∈ D 2 F P [a,a] (if it exists) that satisfies π u t (X) = P ess sup P ∈P [a,a] (t,P,F+) P ess sup θ∈Θ π u,P ,θ t (X), t ≤ T, P -a.s., P ∈ P [a,a] , (4.1) with π u,P ,θ · (X) defined in (3.6). The lower bound π l · (X) = −π u · (−X) is defined analogously, replacing essential suprema in (4.1) by essential infima, and for this reason we focus only on studying the upper bound. For X ∈ L 2 (F + ), we shown in Section 4.1 that (4.1) indeed defines a single process π u · (X) that can be identified as the Y -component of the solution of a specific 2BSDE. For robust good-deal hedging, we define the hedging strategy similarly as in Section 3.2, but taking into account model uncertainty. Indeed, under combined drift and volatility uncertainty, the hedging objective of the investor for a liability X is to find a P [a,a] -permitted trading strategy that dynamically minimizes the residual risk (under a suitable worst-case risk measure ρ) from holding X and trading in the market. As a seller charging the premium π u · (X) for X, she would like the upper good-deal bound to be the minimal capital to make her position ρ-acceptable at all times so that π u · (·) becomes the market consistent risk measure corresponding to ρ · (·). The second objective of the investor being robustness (of hedges and valuations) w.r.t. ambiguity about both drifts and volatilities, ρ should be compatible with the no-good-deal restriction in the market and should also capture the investor's aversion towards uncertainty. From the definition of π u · (X) (for X ∈ L 2 (F + )) in (4.1) and the hedging problem (3.8) in the absence of uncertainty, we define ρ · (X) for ρ P,θ · (·) defined in (3.7). The above considerations then lead to a good-deal hedging problem that is analogous to (3.8) in the case without uncertainty and, mathematically, for a F T -measurable contingent claim X ∈ L 2 (F + ) writes: Findφ ∈ Φ such that for all P ∈ P [a,a] holds P -almost surely for all t ∈ [0, T ], Analogously to Becherer and Kentia (2016) we relate robustness of the hedging strategy (w.r.t. uncertainty) to a supermartingale property of its tracking (hedging) error under a class P ngd of a-priori valuation measures containing all P ngd (Q), uniformly over all reference models Q ∈ R.
To introduce in more precise terms the notion of robustness w.r.t. ambiguity about drifts and volatilities, we define the tracking error R φ (X) of a strategy φ ∈ Φ for a claim X ∈ L 2 (F + ) as In words, R φ t (X) is the difference between the dynamic variations in the (monetary) capital requirement for X and the profit/loss from trading (hedging) according to φ up to time t. Note that also π u is a dynamic coherent risk measure. Subsequently, we will say that that a good-deal hedging strategyφ(X) for a claim X is robust w.r.t. (drift and volatility) uncertainty if Rφ(X) is a (F P , Q)-supermartingale for every Q ∈ P ngd (Q P,θ ) uniformly for all P ∈ P [a,a] and θ ∈ Θ.
Remark 4.1. If the tracking error Rφ(X) is a (F P , Q)-supermartingale for any Q in P ngd (Q P,θ ), then the strategyφ is said to be "at least mean-self-financing" in the model Q P,θ , analogously to the property of being mean-self-financing (like risk-minimizing strategies studied in Schweizer, 2001, Sect.2, with valuations taken under the minimal martingale measure) which would correspond to a martingale property (under Q P,θ ) of the tracking errors. Holding uniformly over all P ∈ P [a,a] and θ ∈ Θ, such can be interpreted as a robustness property ofφ (cf. Becherer and Kentia, 2016) w.r.t. ambiguity about drifts and volatilities.
We show in Section 4.2 that the robust good-deal hedging strategyφ can be obtained in terms of the control process of a 2BSDE describing the good-deal valuation bound π u , and that it is robust with respect to combined uncertainty. This strategy will be shown to be quite different from the almost-sure hedging (i.e. superreplicating) strategy in general, in Section 4.3.

Good-deal valuation bounds.
For each P ∈ P [a,a] , t ∈ [0, T ] and P ∈ P [a,a] (t, P, F + ), the worst-case good-deal bound under drift uncertainty in the model P is defined for X ∈ L 2 (F P T , P ) by π u,P t (X) := P ess sup θ∈Θ π u,P ,θ t (X), P -a.s., so that (4.1) rewrites for F T -measurable X ∈ L 2 (F + ) as π u t (X) = P ess sup P ∈P [a,a] (t,P,F+) π u,P t (X), t ∈ [0, T ], P -a.s., P ∈ P [a,a] . (4.5) Note from (Becherer and Kentia, 2016, Thm.4.11) that the worst-case good-deal bound under drift uncertainty π u,P · (X) for P ∈ P [a,a] is the value process of the standard BSDE under P with terminal condition X and generator − F (t, ω, ·) = −F (t, ω ·∧t , ·, a t (ω)), where F (t, ω, z, a) := inf θ∈Θ F θ (t, ω, z, a), for all (t, ω, z, a) ∈ [0, T ] × Ω × R n × S >0 n , (4.6) and F θ is given by (3.10) for each selection θ of Θ. With (4.5) and the above BSDE representation of π u,P · (X), P ∈ P [a,a] , the robust good-deal bound π u · (X) can alternatively be formulated like in (Soner et al., 2013, eq.(4.12) and Prop.4.10) (or alternatively Possamaï et al., 2015, eq.(2.6) and Lem.3.5) as the value process of a stochastic control problem of nonlinear kernels. Roughly, this means to write, up to taking right-limits in time over rationals, where π u,P,t,ω · (X) are given by solutions to standard BSDEs under P ∈ P t [a,a] , with terminal value X and generator as in (4.6), but defined on the canonical space Ω t := {ω ∈ C([t, T ], R n ) :ω(t) = 0} of (shifted) paths with (shifted) canonical process B t , (shifted) natural filtration F t , and associated (shifted) set of priors P t [a,a] , t ∈ [0, T ]. Note that by a zero-one law as in Lemma 2.1, π u,P,t,ω t (X) is indeed constant for any (t, ω) and P ∈ P t [a,a] (like in Remark 4.3, Part 2.), and hence the pointwise supremum (4.7) is well-defined. Although such a pathwise description of the worst-case good-deal bound π u t (X) reflects better the economic intuition behind the latter and, as a stochastic control problem, would be more classical for the literature, it might be less suitable for approaching the hedging problem (4.3) for which the essential supremum formulation (4.5) seems more appropriate.
It is known Soner et al. (2012);  that under suitable assumptions (on terminal condition and generator) the value process of a control problem like (4.7) can be described in terms of the solution to a 2BSDE. We provide such a description for π u · (X) by considering the 2BSDE with generator F given by (4.6) and terminal condition X. We have the following Theorem 4.2 (Good-deal valuation under combined uncertainty). Let X be a F T -measurable claim in L 2 (F + ). Then a unique solution (Y, Z, a] to the 2BSDE (4.8). Moreover the process Y uniquely satisfies (4.1) and any P ∈ P [a,a] holds π u t (X) = Y t , t ≤ T, P -a.s.. Proof. We aim to apply Part 1. of Proposition 2.8 to show existence and uniqueness of the solution to the 2BSDE (4.8). To this end and since X is F T -measurable and in L 2 (F + ) it suffices to check that X and the function F satisfy conditions (ii) to (iv) of Assumption 2.4. Because F 0 ≡ 0, then (iv) obviously holds. As for (iii) about the uniform Lipschitz continuity of F in z, it is enough to show that the functions F θ , for θ ∈ Θ, are equi-Lipschitz in z uniformly for all (t, ω, a) ∈ [0, T ] × Ω × S >0 n . The latter holds since by Minkowski and Cauchy-Schwarz inequalities one has for all θ ∈ Θ, for C ∈ (0, ∞), making use of (3.5) and boundedness of the functions δ and h. To infer the first claim of the theorem, it remains to show (ii) about F being jointly and Fprogressive. We first show that F is F-progressive in (t, ω). For each z ∈ R n , a ∈ S >0 n , the map [0, T ] × Ω × R n (t, ω, ϑ) → F ϑ (t, ω, z, a) defined from (3.10), after straightforward calculations by (omitting the dependence on ω for notational simplicity) is continuous in ϑ and F-predictable in (t, ω) since σ, h, ξ are F-predictable. Since the correspondence Θ is F-predictable then by measurable maximum and measurable selection results (Rockafellar, 1976, Thms 2K and 1.C) one has for each (z, a) ∈ R n × S >0 n that F (·, ·, z, a) is F-predictable (hence F-progressive) and there exists θ z,a ∈ Θ such that F (t, ω, z, a) = F θ z,a (t,ω) (t, ω, z, a) = inf ϑ∈Θt(ω) F ϑ (t, ω, z, a) for all (t, ω) ∈ [0, T ] × Ω. Now consider Θ : [0, T ] × Ω × R n × S >0 n R n , a correspondence that is constant in its last two arguments and defined by Θ t (ω, z, a) := Θ t (ω) for all (t, ω, z, a) ∈ [0, T ] × Ω × R n × S >0 n . Then Θ is a P(F) ⊗ B(R n ) ⊗ B(S >0 n )-measurable correspondence because Θ is F-predictable, with P(F) denoting the predictable sigma-field w.r.t. F. To prove joint measurability of F , it suffices by measurable selection arguments analogous to the ones above to show that the map (t, ω, z, a) → inf ϑ∈ Θt(ω,z,a) F ϑ (t, ω, z, a) Showing that (t, ω, z, a, ϑ) → F ϑ (t, ω, z, a) is a Carathéodory function would imply the result by (Rockafellar, 1976, Thm.2K). As ϑ → F ϑ (t, ω, z, a) is continuous for each (t, ω, z, a), it remains only to show that (t, ω, z, a) → F ϑ (t, ω, z, a) is jointly measurable for each ϑ. First note that F ϑ (·, ·, z, a) is B([0, T ]) ⊗ F T -measurable for each (z, a, ϑ) ∈ R n × S >0 n × R n since it is F-predictable. Moreover for any x ∈ R n×k , y ∈ R n×p (with k, p ∈ N) the function R n×n a → x tr ay ∈ R k×p is continuous. Likewise a → a −1 and a → a 1/2 are continuous over S >0 n , the former being a consequence of the continuity of the determinant operator. Hence from (4.9) one infers that the map S >0 n a → F ϑ (t, ω, z, a) is continuous. Overall because F ϑ is Lipschitz in z uniformly in (t, ω, a) and continuous in a, it is continuous in (z, a). As a Carathéodory function, (t, ω, z, a) → F ϑ (t, ω, z, a) is jointly measurable and this concludes joint measurability of F . Overall we have shown that F satisfies Assumption 2.4. Hence the first claim of the theorem follows by Part 1. of Proposition 2.8 while the second claim is obtained from Part 2. of Proposition 2.8, recalling the expression (4.5) for the good-deal bound and using (Becherer and Kentia, 2016, Thm.4.11).
Remark 4.3. 1. In a situation with only volatility uncertainty and zero drift, one could apply instead of  an earlier but less general wellposedness result by  to the 2BSDE (4.8), as noted in (Kentia, 2015, Thm.4.20). The latter requires (besides some continuity of X) the generator function F to be convex (in a) and uniformly continuous (UC) in ω, which would hold under UC-assumptions on σ, h. However as soon as one considers non-zero drift (let alone drift uncertainty), convexity of F from our application is no longer clear. Uniform continuity for F given by (4.6) in the present situation with combined uncertainties is neither. E.g., a sufficient condition for it would be that the family (F θ ) θ∈Θ is equicontinuous (in ω), what however seems restrictive.
2. Theorem 4.2 shows in particular that the family of essential suprema in (4.1) indexed by the measures P ∈ P [a,a] effectively aggregates into a single process π u · (X) ∈ D 2 F P [a,a] for Borelmeasurable claims X ∈ L 2 (F + ). In this case π u · (X) is F P [a,a] -progressive, and therefore π u 0 (X) is F P [a,a] 0 −measurable. Hence π u 0 (X) is deterministic constant by the zero-one law of Lemma 2.1, with π u 0 (X) = sup P ∈P [a,a] π u,P 0 (X). 3. By Proposition 2.8, the good-deal bound π u · (·) satisfies a dynamic programing principle for each Borel-measurable claim X in L 2 (F + ): for any P ∈ P [a,a] holds π u s (X) = P ess sup P ∈P [a,a] (s,P,F+) π u,P s (π u t (X)) = π u s (π u t (X)), s ≤ t ≤ T, P -a.s.. (Klöppel and Schweizer, 2007, Thm.2 1, Part a)) one can show that (t, X) → π u t (X) has the properties of a time-consistent dynamic coherent risk measure.
As the set P ngd (Q P,θ ) is m-stable and convex for each P ∈ P [a,a] , θ ∈ Θ, then by (Becherer and Kentia, 2016, Lem.4.9) the union θ∈Θ P ngd (Q P,θ ) is also m-stable and convex. Thus the dynamic coherent risk measure ρ P : L 2 (P ) → L 2 (P, F t ) is time-consistent (see e.g. Becherer and Kentia, 2016, Lem.2.1), with ρ P t (X) ≥ π u,P t (X), t ∈ [0, T ], P -a.s. for all X ∈ L 2 (P, F T ). Consider the generator function F : (4.11) and the associated 2BSDE Proposition 4.4 below gives a 2BSDE description of ρ · (X) that, similar to parts 3.-4. of Remark 4.3 for π u · (·), implies that ρ defines a dynamic risk measure (analogous to π u · (·)) that is time-consistent over all F T -measurable X in L 2 (F + ). The proof is analogous to that of Theorem 4.2, and its details are included in the appendix.
Proposition 4.4. Let X be a F T -measurable claim in L 2 (F + ). Then there exists a unique solution a] to the 2BSDE (4.12). Moreover the process Y uniquely satisfies (4.10) and for all P ∈ P [a,a] holds ρ t (X) = Y t , t ∈ [0, T ], P -a.s..
Note from results on good-deal valuation and hedging in the presence solely of drift uncertainty (cf. e.g. Becherer and Kentia, 2016, Prop.4.12 and Thm.4.13), that for any P ∈ P [a,a] , P ∈ P [a,a] (t, P, F + ) and X ∈ L 2 (F P T , P ) one has, P -a.s. on t ∈ [0, T ], Hence from (4.5) and for F T -measurable claims X in L 2 (F + ) holds for all P ∈ P [a,a] , P -almost surely, (4.13) Furthermore one can infer from (Becherer and Kentia, 2016, Thms. 4.11, 4.13) the existence of a family φP ∈ Φ(P ), P ∈ P [a,a] of trading strategies satisfying for each P ∈ P [a,a] , π u t (X) = P ess sup P ∈P [a,a] (t,P,F+) T t (φ P s ) tr a 1/2 s ξ s ds + dB s , t ∈ [0, T ], P -a.s.. (4.14) In additionφ P is given for each P ∈ P [a,a] by a 1 2 where (Y P,X , Z P,X ) with Y P,X = π u,P · (X) is the solution to the standard BSDE under P ∈ P [a,a] with data (− F ( a 1 2 ·), X) for F defined in (4.6), andθ P is a F P -predictable selection of Θ satisfying with Fθ P given as in (3.10) for θ =θ P .
Remark 4.5. In the case of no volatility uncertainty (as studied in Becherer and Kentia, 2016), i.e. for P [a,a] = {P 0 } with a = a = I n×n , the strategyφ P 0 (X) for a claim X in L 2 P [a,a] (F T ) = L 2 (F + ) = L 2 (F T , P 0 ) would be P [a,a] -permitted and hence already the robust good-deal hedging strategy (w.r.t. drift uncertainty solely) for robust valuation π u · (X) = π u,P 0 · (X) and risk measure ρ = ρ P 0 . In the presence of volatility uncertainty however, the situation is more complicated because each strategyφ P and risk measure ρ P may be defined only up to a null-set of each (non-dominated) measure P ∈ P [a,a] .
Since we are looking for a single processφ ∈ Φ solution to the hedging problem (4.3), one way is to investigate appropriate conditions under which the family {φ P , P ∈ P [a,a] } can be aggregated into a single strategyφ ∈ Φ, i.e. withφ =φ P P ⊗ dt-a.e., for all P ∈ P [a,a] . Were this possible, (4.14) would be written for any P ∈ P [a,a] as π u t (X) = P ess sup P ∈P [a,a] (t,P,F+) such that the aggregateφ would readily satisfy (4.3) since π u · (·) ≤ ρ · (·). However, since general conditions for aggregation are somewhat restrictive and highly technical (see e.g. Soner et al., 2011), we shall abstain from following this path and will directly show that a suitable strategy described in terms of the Z-component of a 2BSDE (cf. (4.15) below) solves the good-deal hedging problem. In the simpler framework of drift uncertainty only, there exists a worst-case measureP ∈ R such that π u · (·) = π u,P · (·) holds and the good-deal hedging strategyφP in the modelP is robust w.r.t. (drift) uncertainty. This has been shown in Becherer and Kentia (2016) by first considering an auxiliary (larger) valuation bound for which the associated hedging strategy automatically satisfies the robustness property and then by a saddle point result (cf. Becherer and Kentia, 2016, Thm.4.13) by identifying the auxiliary bound and hedging strategy with the standard good-deal bound and hedging strategy respectively. In the present non-dominated framework, however, it is questionable but indeed rather unlikely that such a worst-case measureP exists in R for general claims. Instead, we shall follow a slightly different approach and show in a more straightforward manner using Lemma 5.1 that the candidate strategy in (4.15) indeed satisfies the required robustness property w.r.t. drift and volatility uncertainty, which will then be used to recover its optimality. Note that our proof does not use any comparison theorem for 2BSDEs (as e.g. in Kentia, 2015, Sect.4

.1.3). Such would need the terminal wealths
T 0 φ tr s dB s as possible terminal conditions for 2BSDEs to be in L 2 (F + ), which might not be the case for φ ∈ Φ.
For a Borel-measurable claim X in L 2 (F + ), consider the processφ :=φ(X) defined by a 1 2 tφt := Π t a 1 2 for the unique solution (Y, Z, (K P ) P ∈P [a,a] ) to the 2BSDE (4.8),θ being a F P [a,a] -predictable process with |θ t (ω)| ≤ δ t (ω) for all (t, ω) ∈ [0, T ] × Ω and satisfying (4.16) Existence ofθ ∈ Θ satisfying (4.16) easily follows by measurable selection arguments similar to those in the proof of Theorem 4.2, using the fact that Z is F P [a,a] -predictable, Θ is F-predictable and P(F) ⊆ P F P [a,a] . The following result shows thatφ given by (4.15) is indeed a robust good-deal hedging strategy if the family (P ) · 0 Z tr s dB s , P ∈ P [a,a] of "profit & loss" processes aggregates. Theorem 4.6 (Robust good-deal hedging under combined uncertainty). For an F T -measurable X in a] to the 2BSDE (4.8) and assume that the family (P ) · 0 Z tr s dB s , P ∈ P [a,a] aggregates into one process · 0 Z tr s dB s . Then: 1. The processφ =φ(X) from (4.15) is in Φ and solves the good-deal hedging problem (4.3). 2. The tracking error Rφ(X) of the good-deal hedging strategyφ (cf. (4.4)) is a (F P , Q)supermartingale under any Q ∈ P ngd Q P,θ , for all P ∈ P [a,a] , θ ∈ Θ.
Note that, apart from measurability and some integrability, no regularity conditions (like e.g. uniform continuity) are imposed on the contingent claim X. As the setup is non-Markovian, the contingent claim could clearly be path-dependent.
Proof. We first prove the second claim, and then use it to imply the first. Note that by the condition on the integral Z, the strategyφ given by (4.15) clearly belongs to Φ. By Theorem 4.2, we know that π u · (X) = Y for (Y, Z, (K P ) P ∈P [a,a] ) solution to the 2BSDE (4.8). Let P ∈ P [a,a] , θ ∈ Θ and Hence since K P is non-decreasing, then combining (4.17) and (4.18) imply that the finite variation part of the Q-semimartingale Rφ is non-increasing. Furthermore one has Rφ ∈ D 2 (F P , P ) because π u · (X) ∈ D 2 F P [a,a] ⊆ D 2 (F P , P ) andφ ∈ Φ ⊆ Φ(P ). Now since λ + θ is bounded then dQ dP F T is in L p (F T , P ) for any p < ∞ and by Hölder's inequality it follows that Rφ ∈ D 2− (F P , Q) holds for some > 0. Thus Rφ is a (F P , Q)-supermartingale. We turn to the proof of the first claim of the theorem. To show thatφ solves the hedging problem (4.3), let P ∈ P [a,a] . Then by (4.13) and Φ ⊆ Φ(P ) for all P ∈ P [a,a] one has P -a.s. To conclude that someφ ∈ Φ is a good-deal hedging strategy satisfying (4.3), it suffices to show for all θ ∈ Θ, P ∈ P [a,a] that P -a.s. for all t ∈ [0, T ], P ∈ P [a,a] (t, P, F + ) and Q ∈ P ngd (Q P ,θ ) holds To this end, let θ ∈ Θ, P ∈ P [a,a] . By the first claim of the theorem, the tracking error Rφ := Rφ · (X) ofφ is a (F P , Q)-supermartingale for every Q ∈ P ngd (Q P ,θ ), P ∈ P [a,a] . This implies by Lemma tr s a 1/2 s ξ s ds+dB s , P -a.s., for all Q ∈ P ngd (Q P ,θ ), P ∈ P [a,a] (t, P, F + ), for any t ≤ T . Reorganizing that inequality yields the claim.
Remark 4.7. 1. By a result of Karandikar (1995) it is possible to define the stochastic integral · 0 Z tr t dB t pathwise if the process Z is càdlàg and F-adapted, such that in particular it satisfies the hypothesis of Theorem 4.6. Although the Z-component of a 2BSDE solution is not guaranteed in general to be càdlàg, we emphasize that Theorem 4.6 may still be applicable in some specific situations. For instance in a Markovian diffusion setting, one may be able to use partial differential equation (PDE) arguments to show that the Z-component is even continuous. An example in such a setting is provided in Section 4.3, where beyond the continuity of Z we can even obtain explicit solutions to the 2BSDE (4.8), for some bounded contingent claims. In the most general situation, however, the result of Nutz (2012a) can be used under additional set-theoretical assumptions to get rid of the aggregation condition in Φ and in the statement of Theorem 4.6; cf. Remarks 2.7 and 3.5.
2. A consequence of Theorem 4.6 is a minmax identity: For P ∈ P [a,a] holds a.s.
4.3. A case study: Hedging a put on non-traded but correlated asset. In this subsection we investigate a simple Markovian example to provide more intuition and to illustrate the general but abstract main theorems, giving closed-form formulas for robust good-deal valuations and hedges.
To this end, we investigate here the particular application of a vanilla put option on a non-traded asset in concrete detail. The latter is a typical problem for optimal partial hedging in incomplete markets. The example is elementary enough to permit even for closed-form solutions, in some parameter settings, and it can be solved also by more standard optimal control methods, exploiting the Markovian structure and certain properties of the claim's payoff function. We will make use of this fact to demonstrate clear differences to hedging by superreplication, and further to elucidate the difficulties arising from combined drift and volatility uncertainty, compared to only one type of ambiguity. More precisely, we show how under combined uncertainty the optimal control problem of robust good-deal valuation becomes effectively one over a non-rectangular domain of control variables, making it more complex to identify optimizers and worst-case priors (or parameters), even for monotone convex payoff functions for which intuition from examples in other related valuation approaches (e.g. robust superreplication) might at first suggest otherwise. Of course, one cannot expect to get similarly elementary solutions for measurable contingent claims in general models: For the general problem formulation, the solution has been fully characterized by means of 2BSDEs in the main Theorems 4.2 and 4.6. Let us consider a financial market where only one risky asset (a stock) with discounted price process S is tradeable, apart from the riskless asset (with unit price). In addition, there is another asset that is not tradeable but whose value L is correlated with S. The processes S and L are, P [a,a] -q.s., given by where B = (B 1 , B 2 ) is the canonical process, P [a,a] is the set defined as in (2.1) for diagonal matrices a = diag(a 1 , a 2 ) and a = diag(a 1 , a 2 ) in S >0 2 , with S 0 , L 0 , σ S , β in (0, ∞), for drifts b, γ in R, volatility matrix σ := (σ S , 0) ∈ R 1×2 and a P 0 -correlation coefficient ρ ∈ [−1, 1]. For a constant bound h ∈ [0, ∞) on the instantaneous Sharpe ratios, we are going to to derive closed-form expressions for robust good-deal valuations and hedges, for a European put option X := (K − L T ) + on the non-traded asset L with strike K ∈ (0, ∞) and maturity T , and to identify the corresponding worst-case drifts and volatilities. At first we assume the drift rate b of S to be zero, so that the center market price of risk ξ vanishes quasi-surely, and consider the case of uncertainty solely on volatility (i.e. for δ ≡ 0). For this case we will identify a worst-case volatility for the robust valuation bounds in closed form, and compare robust good-deal hedging with classical robust superreplication under volatility uncertainty. After this, we discuss the more complex case with combined drift and volatility uncertainties and the difficulties that arise in identifying worst-case drifts and volatilities. Finally, we investigate sensitivities of the derived robust good-deal bound under volatility uncertainty with respect to variations of the drift parameter γ ∈ R for the non-traded asset. implying Im (σ a 1 2 ) tr = z ∈ R 2 : c 12 z 1 − c 11 z 2 = 0 and Ker (σ a 1 2 ) = z ∈ R 2 : c 11 z 1 + c 12 z 2 = 0 . Hence for z = (z 1 , z 2 ) tr ∈ R 2 one gets Π(z) = 1 a 11 ( c 11 ) 2 z 1 + c 11 c 12 z 2 c 11 c 12 z 1 + ( c 12 ) 2 z 2 and Π ⊥ (z) = 1 a 11 ( c 12 ) 2 z 1 − c 11 c 12 z 2 ( c 11 ) 2 z 2 − c 11 c 12 z 1 . (4.20) In the case δ ≡ 0, the 2BSDE (4.8) thus rewrites here as for z = (z 1 , z 2 ) tr ∈ R 2 . Clearly L T is F T -measurable, and since the put option payoff function x → (K − x) + is bounded and continuous it follows that X is F T -measurable and in L 2 (F + ). Therefore the conditions of Theorem 4.2 are satisfied, yielding that the worst-case good-deal bound π u · (X) coincides with the Y -component of the solution of the 2BSDE (4.21). In Lemma 4.8 below we will express the solution to the 2BSDE (4.21) in terms of the classical solution To this end, let P a = P 0 • (a 1/2 B) −1 ∈ P [a,a] such that B t = at P a ⊗ dt-a.s.. The process L under P a is a geometric Brownian motion with dynamics dL t = L t γdt +β ρdW 1,P a t + 1 −ρ 2 dW 2,P a t , t ∈ [0, T ], P a -a.s., where W P a = (W 1,P a , W 2,P a ) := (a) −1/2 B is a P a -Brownian motion and Hence a closed-form expression for v(t, L t ), with v being the solution to the PDE (4.23), is given by the Black-Scholes formula for the price of the put option X = (K − L T ) + in the model P a . By arguments analogous to the derivations of the formulas in (Becherer and Kentia, 2016, Sect.3.2.1), v(t, L t ) coincides P a -a.s. with the valuation bound π u,P a t X for all t ≤ T , and is given in closed form as where "B/S-put-price" denotes the standard Black-Scholes put pricing formula with zero interest rate, "vol" abbreviating volatility, N denoting the cdf of the standard normal law, m := γ − hβ 1 −ρ 2 = γ − hβ 1 − ρ 2 √ a 2 , and d ± := ln The details of proof for the following lemma can be found in the appendix.
Worst-case model for valuation and hedging: Using Lemma 4.8, Theorem 4.2 implies by (4.24) that the robust good-deal bound for a put option X = (K − L T ) + is given in closed-form by a.s.. (4.25) Hence the robust good-deal bound π u · (X) for a put option X = (K − L T ) + is attained at the "maximal" volatility matrix a, and can be computed as in the absence of uncertainty, but in a worst-case model P a ∈ P [a,a] in which B t = at holds P a ⊗ dt-a.e., yielding π u t X = π u,P a t X for all t ∈ [0, T ], P a -a.s.. In addition, π u t X is given explicitly in terms of a Black-Scholes' type formula, for modified strike K exp(−m(T − t)) and volatilityβ = β ρ 2 a 1 + (1 − ρ 2 )a 2 1/2 . Similarly Theorem 4.6 and Lemma 4.8 imply by (5.1) that the robust good-deal hedging strategyφ :=φ(X) for X is given bȳ -a.e., (4.26) where we have used the fact that a −1/2 Π a 1/2 z = z 1 + a 12 a 11 z 2 , 0 tr for z = (z 1 , z 2 ) tr ∈ R 2 , which is straightforward by (4.20) and (4.19). Analogously the lower good-deal bound π l · X and corresponding hedging strategy can also be computed, but under the worst-case measure P a ∈ P [a,a] for "minimal" volatility matrix a. Comparison with robust superreplication under volatility uncertainty: Intuitively as the bound h on the Sharpe ratios increases to infinity, the (upper) good-deal bound π u · (X) should increase towards the robust upper no-arbitrage valuation bound, studied in Avellaneda et al. (1995);Lyons (1995); Denis and Martini (2006); Vorbrink (2014); Nutz and Soner (2012); Neufeld and Nutz (2013). Our result in this example is in accordance with Avellaneda et al. (1995); Lyons (1995);El Karoui et al. (1998); Epstein and Ji (2013); Vorbrink (2014) who showed that under ambiguous volatility, Black-Scholes-valuation and hedging of vanilla put (or call) options under maximal (resp. minimal) volatility corresponds to the worst-case for the seller (resp. for the buyer). Instead of superreplication we focus here on robust good-deal hedging under volatility uncertainty for valuation at the worst-case good-deal bound. Beyond some simplifications (particular payoff function) in comparison to Avellaneda et al. (1995);El Karoui et al. (1998);Vorbrink (2014), our setup also includes some more original aspects: We study an option on a non-tradable asset L with possibly non-perfect correlation with S, and there is incompleteness in the sense that the option X is already clearly non-replicable by dynamic trading under any individual prior (if |ρ| < 1), even without model uncertainty. Further, we now show that the robust good-deal hedging strategyφ(X) is very different from (the risky asset component of) the super-replicating strategy, in general. Indeed since 0 ≤ X ≤ K holds pathwise, the (upper) no-arbitrage bound (or superreplication price) process V (X) under P a defined by satisfies π u,P a t (X; h) ≤ V t (X) ≤ K, P a -a.s., for π u,P a t (X; h) denoting the good-deal bound in the model P a with Sharpe ratio constraint h ∈ [0, ∞) and being given by (4.24) for arbitrary but fixed h. If |ρ| < 1 then π u,P a t (X; h) for t < T increases to K as h tends to +∞ (since m → −∞, d ± → −∞), and hence we get, in striking difference to the good deal bound from (4.25), that (4.27) The superreplication price process V (X) has the optional decomposition (cf. Kramkov, 1996, Thm.3.2) where · 0 φ s dB 1 s and C are unique (see Kramkov, 1996, Thm.2.1 and Lem.2.1). One obtains by (4.27) that · 0 φ s dB 1 s = 0 and C = (K − X)1 {T } . Note from (4.26) thatφ = (Z 1 , 0) tr P a ⊗ dt-a.e. since a = a P a ⊗ dt-a.e.. For ρ = 0, the process Z 1 is non-trivial under P a ⊗ dt. Overall for 0 < |ρ| < 1, · 0φ tr s dB s = · 0 Z 1 s dB 1 s cannot be equal to · 0 φ s dB 1 s ≡ 0 P a ⊗ dt-almost everywhere. This means that for 0 < |ρ| < 1 the good-deal hedging strategyφ for the put option on the non-traded asset is not the traded asset component of the option's super-replicating strategy for the model P a . Therefore the robust good-deal hedging strategyφ is in general not a super-replicating strategy under volatility uncertainty for the set P [a,a] of reference priors and π u t (X) + T tφ tr s dB s does not dominate the claim X P a -almost surely, let alone P [a,a] -quasi-surely. 4.3.2. Case with combined drift and volatility uncertainty. The approach of Section 4.3.1 to derive closed-form valuations and hedges may not work in general when there is drift uncertainty in addition to volatility uncertainty. Indeed one would need, by the preceding valuation and hedging Theorems 4.2 and 4.6, first to identify a candidate worst-case drift parameterθ (and then possibly a worst-case volatility) satisfying (4.16). This requires finding a minimizer over θ ∈ Θ ≡ {x ∈ R 2 : |x| ≤ δ} of the quantities F θ ( a 1/2 Z) given by (3.10) for the Z-component of the solution to the 2BSDE (4.8). However, it is not clear whatsoever how this could be done in general (let alone explicitly), given that the expression for F θ is quite complex by non-triviality of the kernel Ker (σ a 1/2 ). Recall that the latter is due to the possible singularity of the volatility matrices σ t ∈ R d×n when d < n, i.e. under market incompleteness under each prior. Furthermore, even if one could identifyθ, it would still be questionable what the corresponding worst-case volatility should be as Fθ( a 1/2 Z) = F ( a 1/2 Z) is still a very complicated function of the coefficients of a ∈ S >0 2 . This issue does not appear in the less general setting of Section 4.3.1 where, thanks to the zero drift assumption b = 0 for the traded asset S, the expression for F ( a 1/2 Z) greatly simplifies to (4.22) and this allows by direct comparison to obtain a as the corresponding worst-case volatility. If market incompleteness is mainly due to the presence of volatility uncertainty (i.e. market is complete under every prior, and hence Ker (σ a 1/2 ) is trivial), then drift uncertainty is redundant as it does not have influence on (essentially superreplication) valuation bounds (and respective strategies) for any contingent claim, as F θ (z) = F (z) = ξ tr z for any θ ∈ Θ, z ∈ R n . This has been argued in more detail by (Epstein and Ji, 2013, Example 3) and is perhaps not surprising. Indeed, it is well-known that the Black-Scholes price does not depend on the drift of the underlying and consequently remains unaffected by drift uncertainty. Super-replication (q.s.) appears as a natural notion for robust hedging under market completeness for every prior and is well studied in the literature, (cf. Avellaneda et al., 1995;Lyons, 1995;Nutz and Soner, 2012;Neufeld and Nutz, 2013;Epstein and Ji, 2013;Vorbrink, 2014), where for the above reason it is standard to assume zero drift. However in the case of incomplete markets (i.e. for d < n), combined uncertainty on drifts and volatilities becomes relevant for related approaches to valuation and hedging, that are less expensive than superreplication. One might wonder whether one could identify worst-case drifts and volatilities explicitly for certain examples. Yet, even for the vanilla put option of Section 4.3.1, we are not aware of a closed-form solution for this. E.g. for non-zero drift b, which is not even subject to uncertainty, we admit that we are not able to state a worst-case volatility in closed form (e.g. by identifying it asā like we did on p.20). To better demonstrate why this case is effectively already less tractable, let us assume for simplicity the following values for the model parameters: σ S = 1, γ = 0, β = 1, δ = 0 and ρ = 0. The dynamics for S and L are then dS t = S t (bdt + dB 1 t ) and dL t = L t dB 2 t , P [a,a] -q.s..
Here ξ = b c 11 / a 11 , b c 12 / a 11 tr , Ker (σ a 1/2 ) = Span η with η := 1, − c 11 / c 12 tr . By (3.4) for P in P [a,a] , the set Q ngd (P ) consists of all measures Q ε with dQ ε = (P ) E (− ξ + ε η) · W P dP, Pa.s., for R-valued processes ε satisfying |ε| ≤ c 12 ( a 11 ) −1/2 h 2 − b 2 / a 11 1/2 everywhere. Since B 2 = ( c 12 , c 22 ) · W P and ( c 12 ) 2 + ( c 22 ) 2 = a 22 , then a change of measure from P to Q ε yields where W ε is a one-dimensional Q ε -Brownian motion. As the put payoff function (K − ·) + is non-increasing, the good-deal bound under P at time zero π u,P 0 ((K − L T ) + ) = sup ε E Q ε [(K − L T ) + ] is then attained at the largest possible ε which is c 12 ( a 11 ) −1/2 h 2 − b 2 / a 11 1/2 . After replacing in the dynamics of L and taking the supremum over all P ∈ P [a,a] we obtain by Part 2 of Remark 4.3 that the robust good-deal bound at time t = 0 is (4.28) with β(a) := a 22 and γ(a) := −b (a 12 /a 11 ) − h 2 − b 2 /a 11 1 2 a 11 a 22 − (a 12 ) 2 1 2 (a 11 ) − 1 2 . Moreover L γ,β has dynamics dL γ,β t = L γ,β t γ t dt + β t dW t ) , t ∈ [0, T ], for some one-dimensional P 0 -Brownian motion W . Hence, we recognize that (4.28) is given by the value of a standard (Markovian) optimal control problem with state process L γ,β but with (dynamic) control variables (γ t , β t ) t∈[0,T ] taking values in a generally non-rectangular set (γ, β) ∈ R 2 β = β(a) and γ = γ(a) for some a ∈ S >0 2 ∩ [a, a] , what clearly is a complication for the derivation of optimal controls. Because the payoff function x → (K − x) + is non-increasing and convex, then clearly a volatility matrix a * which simultaneously maximizes a → β(a) and minimizes a → γ(a) would be an optimizer for (4.28). In the particular case where b = 0, one can easily check that a * equals a, yielding (again) the simple closed-form solution of Section 4.3.1. Yet, when b = 0, it becomes less simple to describe a * . 4.3.3. Sensitivity with respect to drift parameter of non-traded asset. Despite the lack of closed-form expression for valuations under combined uncertainties, still, some further insight can be obtained by investigating simply the sensitivity of the robust good-deal bound π u · (X) =: π u · (X; γ) with respect to a (fixed constant) drift γ of the non-traded asset L, solely under volatility uncertainty. This is straightforward, by exploiting the explicit formulas obtained in Section 4.3.1. By (4.25) one gets for any t ∈ [0, T ] and γ ∈ R that π u ≤ 0. Therefore we obtain, what intuition suggests: The robust good-deal bound π u · (X; γ) for the European put option X = (K−L T ) + is non-increasing in γ for the model P a . This implies that for γ, γ ∈ R specifying an interval range [γ, γ] for the (constant) drift parameter of L, the worst-case drift corresponds to γ, and is that for which the supremum ess sup P a γ∈[γ,γ] π u t (X; γ) is attained for any t ∈ [0, T ]. Note that this supremum may be different from the worst-case good-deal bound (defined in (4.1) and characterized by Theorem 4.2) under combined drift and volatility uncertainties, as the latter parametrizes drift uncertainty instead in terms of stochastic drifts γ that satisfy γ ≤ γ ≤ γ pointwise and there is no apparent reason why the worst-case volatility for every fixed (stochastic) drift γ should be a.

Appendix
This section contains some proofs and details omitted in the main body of the paper. This includes the Lemma 5.1 which is used in the proof of Theorem 4.6.
Proof of Proposition 2.8. Consider the family (P(t, ω)) (t,ω)∈[0,T ]×Ω of sets probability measures given by P(t, ω) := P [a,a] for all (t, ω) ∈ [0, T ] × Ω and the constant set-valued process D : [0, T ] × Ω → S >0 n with D t (ω) := [a, a] for all (t, ω) ∈ [0, T ] × Ω. Clearly D has the properties required in (Neufeld and Nutz, 2013, Example 2.1) from which the first claim of (Neufeld and Nutz, 2013, Cor.2.6) implies that the constant family P(t, ω) ≡ P [a,a] satisfies Condition A therein, hence in particular the measurability and stability conditions of Assumption 2.1, (iii)-(v) of , whereby (iii) in particular follows from (Neufeld and Nutz, 2013, Condition (A1)) since a countable product of analytic sets is again analytic (Bertsekas and Shreve, 1978, Prop.7.38). This together with our Assumption 2.4 and X ∈ L 2 (F + ) imply Assumptions 2.1 and 3.1 of  from which a direct application of Theorems 4.1 and 4.2 therein yields existence and uniqueness of a 2BSDE solution (Y, Z, (K P ) P ∈P [a,a]    vanishes in this case thanks to the martingale representation property of P [a,a] in Lemma 2.1. Moreover since by Lemma 2.1 the filtration F P is actually right continuous for every P ∈ P [a,a] , it follows that F P t+ = F P t for t ∈ [0, T ], P ∈ P [a,a] which implies that the minimum condition in the definition of a 2BSDE solution in (Possamaï et al., 2015, Def.4.1) is equivalent to (2.3). In particular, we have F P [a,a] + = F P [a,a] .
Proof of Proposition 4.4. By the classical comparison theorem for standard BSDEs, one easily sees for every P ∈ P [a,a] that Y P,X = ρ P · (X), where ( Y P,X , Z P,X ) is the unique solution to the standard Lipschitz BSDE under P with data (− F ( a 1 2 ·), X) for F given by (4.11). In addition, one can verify as in the proof of Theorem 4.2 for (F, X) that (F , X) satisfies Assumption 2.4 as well. The required result then follows from an application of Proposition 2.8.
Thus the process K is P -a.s. non-decreasing because the delta of the put option in the Black-Scholes model is non-positive and the gamma is non-negative, i.e. ∂v ∂x (t, L t ) ≤ 0 and ∂ 2 v ∂x 2 (t, L t ) ≥ 0 for all t ∈ [0, T ] using (4.24). Moreover it can be shown following arguments analogous to those in the proof of (Soner et al., 2012, Thm.5.3) that the process K satisfies the minimum condition (2.3); we omit the essentially technical details which we refer to (Kentia, 2015, Sect.4.3.3, proof of Lem.4.25).
It remains to show that v(·, L · ) ∈ D 2 F P [a,a] , Z ∈ H 2 F P [a,a] and that the stochastic integral · 0 Z tr s dB s can be constructed pathwise. This will conclude by uniqueness of the solution to the 2BSDE that (v(·, L · ), Z, K) is the unique solution to the 2BSDE (4.21) for Z given as in (5.1) and K as in (5.2). Since v is C 1,2 and L is P [a,a] -q.s. continuous and F P [a,a] -adapted, then v(·, L · ) is P [a,a] -q.s. càdàg and F P [a,a] -progressive and Z is F P [a,a] -predictable. That v(·, L · ) is in D 2 F P [a,a] now follows from (4.24) which indeed implies 0 ≤ v(t, L t ) ≤ K pathwise. By (5.1) and since a ≤ a ≤ a holds P ⊗ dt-a.e. for any P ∈ P [a,a] , one has a 1/2 t Z t 2 ≤ max(a 1 , a 2 )β 2 e 2|m|T L 2 t P ⊗ dt-a.e. for any P ∈ P [a,a] . Hence to conclude that Z ∈ H 2 F P [a,a] it suffices to show that sup P ∈P [a,a] E P T 0 L 2 t dt < ∞. To this end, note that for any P ∈ P [a,a] hold T 0 L 2 t dt ≤ β −2 (min(a 1 , a 2 )) −1 L T and L 2 T ≤ L 2 0 e 2|γ|+β 2 max(a1,a2) TL T (5.3) P -almost surely, forL = 1 + · 0 2L s β ρdB 1 s + 1 − ρ 2 dB 2 s P [a,a] -q.s.. Clearly E P [L T ] ≤ 1 for every P ∈ P [a,a] . Thus taking expectations in (5.3) gives E P T 0 L 2 t dt ≤ β −2 (min(a 1 , a 2 )) −1 L 2 0 e 2|γ|+β 2 max(a1,a2) T , for all P ∈ P [a,a] . Now taking the supremum over all P ∈ P [a,a] implies that Z ∈ H 2 F P [a,a] . As a consequence (v(·, L · ), Z, K) is the unique solution to the 2BSDE (4.21). Finally · 0 Z tr s dB s can be constructed pathwise by Karandikar (1995), since Z is continuous and F-adapted.