Robust Trading of Implied Skew
Abstract
In this paper, we present a method for constructing a (static) portfolio of comaturing European options whose price sign is determined by the skewness level of the associated implied volatility. This property holds regardless of the validity of a specific model – i.e. the method is robust. The strategy is given explicitly and depends only on one’s beliefs about the future values of implied skewness, which is an observable market indicator. As such, our method allows to use the existing statistical tools to formulate the beliefs, providing a practical interpretation of the more abstract mathematical setting, in which the belies are understood as a family of probability measures. One of the applications of the results established herein is a method for trading one’s views on the future changes in implied skew, largely independently of other market factors. Another application of our results provides a concrete improvement of the modelindependent super and sub replication strategies for barrier options proposed in [10], which exploits the given beliefs on the implied skew. Our theoretical results are tested empirically, using the historical prices of S&P 500 options.
1 Introduction
In this paper, we present a method for constructing a (static) portfolio of comaturing European options whose price sign is determined by the skewness level of the associated implied volatility. We define the implied skewness (or implied skew) as a ratio of the implied volatilities of outofthemoney (OTM) calls to those of comaturing OTM puts. The target portfolio is constructed so that its price has a specified sign, whenever the level of implied skew satisfies a given upper or lower bound, regardless of the exact values of implied volatilities (e.g. regardless of the overall level of volatility).
The existing literature on implied skew can be split into three main categories, depending on the interpretation of what this term actually means. The first interpretation of implied skew is as a measure of asymmetry of the riskneutral distribution of the underlying value at a fixed future time (i.e. maturity). This measure is defined through the first three moments of the associated distribution, and each of these moments, in turn, can be computed as a price of a corresponding (static) portfolio of European options (with the same maturity and multiple strikes). Such definition has been adopted by CBOE in the construction of its Skew Index, and its empirical performance is analyzed in [3]. Nevertheless, it is worth mentioning that the CBOE Skew Index does not correspond to a price of a tradable portfolio: its value is given by a nonlinear function of the prices of several static portfolios of European options. The second approach is to measure the implied skew through the correlation between the underlying returns and the returns of its (spot) volatility. Indeed, in a classical stochastic volatility model, the derivative of a shortterm atthemoney implied volatility, with respect to the logmoneyness variable, is determined by the aforementioned correlation. Such a relation may fail to hold in practice, leading to potential trading opportunities: e.g. [19] proposes a trading strategy that benefits from such deviations. Finally, another interpretation of implied skew is the one adopted in the present paper: namely, the implied skew is defined as a ratio of OTM implied volatilities (calls to puts). It is a popular approach, both among practitioners and among academics (cf. [4], [11]), to attempt to profit from changes in such implied skew by opening a long or short position in a “risk reversal” portfolio: i.e. a static portfolio consisting of a long position in an OTM put and a short position in (several shares of) a comaturing OTM call. This portfolio can be combined with a static position in the underlying. Then, choosing the weights of the three instruments appropriately, one can ensure that the price of the portfolio is locally (i.e. asymptotically, for an infinitesimal time period) insensitive to changes in the underlying and to parallel shifts in the implied volatility surface (or its logarithm). At the same time, the price of such a portfolio is locally (for an infinitesimal time period) monotone in the implied skew. Note that such a portfolio can be constructed for any choice of the two strikes: one for an OTM call and one for an OTM put. In the present paper, we show how to construct a static risk reversal portfolio whose price sign is insensitive to arbitrary changes in the implied volatility surface, as long as the implied skew stays above (below) a chosen lower (upper) bound. These properties of the portfolio hold for a sufficiently small, but not infinitesimal, time period. It turns out that, in order to achieve the desired properties of a risk reversal portfolio, in addition to fixing the weights of the two OTM options, one has to restrict the choice of the two strikes. In particular, given an arbitrary OTM put (call) strike, our method shows how to choose the appropriate strike of an OTM call (put).
The results presented herein do not require any assumptions on the distribution of the underlying or the implied volatility surface. As such, they present an explicit example of robust, or modelindependent, methods in Financial Mathematics. Many researchers (cf. the seminal work of [30], as well as e.g. [28], for a broader discussion) have been interested in properties of option prices which are independent of any model specification. In particular, many of such works aim to incorporate existing market constraints or information in order to obtain interesting properties without specific assumptions. Chief examples of such constraints are given by market prices of liquidly traded options on the underlying. The resulting properties may include constraints on the prices of illiquid options, such as barrier options, and the associated sub and superreplicating strategies: see e.g. [9], [27], [10], [20], [21], [6], [25], [1], [12]. The works on modelindependent pricing and hedging mentioned above, while of fundamental interest, have been criticized for the lack of practical relevance due to typically wide interval of noarbitrage prices they produce. And so, even if the modelindependent hedging strategies were shown to be competitive to classical deltavega hedging, see e.g. [33], it is of vital interest to develop robust framework which is capable of interpolating between the modelindependent and the modelspecific setups. A classical probabilistic way to address this problem is by considering families of models (i.e. probability measures), which may vary from a single model to the class of all possible models: see e.g. [24], [7]. Another approach is to focus not on any probabilistic models but rather on the set of possible paths of observed quantities (e..g prices of financial assets). The latter may vary between a very large (all paths) and a small (the support of some fixed model) set of paths. The crucial insight here is that the feasible paths should be defined using market observable and meaningful quantities. This should allow one to apply existing statistical methods to the time series data, together with expert views, to restrict the universe of paths: see e.g. [29], [2], [31, 32], [35]. Herein, we consider a framework in which the set of tradable instruments includes the underlying and a set of European options written on it. This has been previously done in works on the socalled market models, see e.g. [34], [14], [16], [15], [13]. While these works focused on a fixed probabilistic setting, herein, we pursue the robust approach. Instead of postulating any specific probabilistic features, we prescribe constraints on the price paths of the assets. More specifically, we consider beliefs on the implied skew, as the latter is a market observable quantity of interest, and it is reasonable to expect that many practitioners have a view on its future behavior. The resulting theoretical framework is very parsimonious: it allows for a continuous interpolation between an entirely model–free and (the classical) modelspecific setting.
As indicated above, one of the applications of the results established herein is a method for trading one’s views on the future changes in implied skew, largely independently of other market factors. Another application of our results provides a concrete improvement of the modelindependent super and sub replication strategies for barrier options proposed in [10], which exploits a given set of beliefs on the implied skew. Our theoretical results are tested empirically: herein, we implement the proposed risk reversal portfolio using the historical prices of S&P 500 options, and verify that its price possesses the desired properties.
Finally, in order to achieve our goal, we introduce and analyze a new family of models, called Piecewise Constant Local Variance Gamma models (see [18] for related results). These turn out to be rather remarkable. First, they allow for explicit computation of European options’ prices and of the shortterm implied volatility. Second, these models allow for skew in the implied volatility they generate, which is flexible enough to provide a good fit, as well as the upper and lower bounds, on the empirically observed implied skew. Third, any model from this family has an explicit static hedging strategy, constructed via the weak reflection principle (cf. [17], [5]). We believe that these results are of independent interest.
The remainder of this paper is organized as follows. In Section 2, we discuss an example which provides a rationale behind the proposed trading strategies and presents one of their applications. Section 3 formalizes the notion of beliefs on implied skewness. Section 4 introduces the class of auxiliary models, needed to construct the desired portfolios, for given beliefs. These portfolios are constructed in Section 5, where it is also shown how to use them for trading the implied skew. Section 6 presents another application of the proposed portfolios, concerned with super and subreplication of barrier options. Section 7 contains the empirical analysis. Finally, we summarize our results in Section 8.
2 Case study: superhedging of UpandOut Put
We start with a motivating example which illustrates how the beliefs on implied skewness can be used in practice. We consider a trader who wants to hedge a short position in an upandout put (UOP) option with the payoff
(1) 
where , and . We suppose the trader does not want to make any specific probabilistic assumptions about the dynamics of the risky assets, but believes asset prices are continuous and that the sign of the implied skew will remain constant through time. We show here how such beliefs can be used to improve the mode–independent approach of [10]. For simplicity, assume that European options with just two strikes and are liquidly traded at all times and denote their BlackScholes implied volatilities at time as , . The model–independent superreplication strategy proposed in [10] is based on the following inequality:
(2) 
which is satisfied as long as . This strategy is discussed in more detail in Section 6. The first term in the right hand side of (2) is a static position in shares of a put option with strike and maturity . The second term on the RHS in (2) is a static position in shares of a forward struck at , with maturity and which is liquidated at time . This is done at no cost since . The initial capital needed to set up this strategy is
It is easy to see from (2) that the above strategy superreplicates the payoff of the UOP option, provided the paths of the underlying are continuous. Hence, it allows one to hedge the risk of a short position in a UOP option.
Let us now construct an improvement of this strategy by specifying some rather weak beliefs about future implied skew. Namely, we assume that the trader believes that, whenever , , the market implied volatility will exhibit a nonnegative skew:
(3) 
Recall that the well known static hedging formula in the BlackScholes model (cf. [8]) implies:
(4) 
where denote the put and call prices in the BlackScholes model. Combining this with the beliefs above, instantly yields
which has to hold for any with and where we take, e.g., . Then, the improvement of the superreplicating strategy proposed in [10] consists in shortselling, initially, additional shares of comaturing calls struck at , and closing the position at (if ). If the underlying does not hit before time (i.e. ), then the payoff of the additional calls is zero. If the underlying does hit before (i.e. ), then, at the time , the superreplication portfolio is closed at a profit:
Clearly, the initial value of the new strategy is smaller than the initial capital of the original one, as we shorted call options. The strategy exploits the beliefs (3) to improve the modelindependent approach of [10]. In hindsight, the explicit construction was possible due to two factors. First, we could identify the boundary case of beliefs in (3), which in some sense “dominates” all the other consistent market scenarios. Second, this boundary case corresponded to a flat implied volatility surface, i.e. the BlackScholes model, which allows for an explicit static hedging formula (4). These two observations turn out to be crucial. We will see below that they can be extended to allow for more general beliefs, leading, in particular, to a more general superreplication algorithm described in Section 6. In order to develop a general approach extending the case study above, we need to describe a convenient, flexible family of martingale models for the underlying, which, on the one hand, can produce a dominating surface for general beliefs, and, on the other hand, allow for explicit static hedging formulas in the spirit of (4). We develop such a family of models in Section 4.
Finally, we emphasize that (3), and also the superhedging strategy, were invariant with respect to the scaling of the implied volatilities and only depended on the implied skew. We show later in the paper that the same phenomenon occurs in general: the superreplicating strategy depends on the skewness of the dominating surface, but it is robust with respect to scaling of this surface. Thus, a superreplication strategy for a UOP option is an example of trading the implied skew.
3 The Market Setup and Beliefs on Implied Skewness
3.1 The Market Setup
We consider a market consisting of an underlying tradable asset, whose price process we denote by and a family of European call and put options, with available strikes, , and with arbitrary maturities, whose price processes are denoted
Namely, at any time , the underlying can be purchased or sold in any quantity at the price , and a European call or put option with any strike and any maturity can purchased or sold in any quantity at the price or , respectively. There is no specific probabilistic model underlying these processes. Instead, we work directly with the set of their admissible paths. In fact, for convenience, we express the prices of European options at time through their implied volatility surface:
(5) 
For simplicity, we assume that the carrying costs are zero (or, equivalently, that all prices are in units of a numeraire). Then, the call and put prices at time are defined through the implied volatility as follows:
for all and , where and are the BlackScholes prices of the European call and put options, respectively, with the current level of underlying , strike , time to maturity , volatility , and zero interest and dividend rates. Note that the above definition in particular implies that the putcall parity is satisfied:
Clearly, the above represents the time price of a forward contract, which depends only on and but not on . We say that a call option with strike is inthemoney (ITM) at time if , it is atthemoney (ATM) if , and it is outofthemoney (OTM) if . The terminology is similar for the put options, except that the inequalities are reversed.
We now impose conditions on each market value to ensure it does not admit any static arbitrage.
Definition 1.
The market value , where is a nonnegative number and is as given in (5), is admissible if the following conditions hold.

For any , there exists a strictly positive limit .

For any , we have: , for all .

For any and any , we have: .

For any , we have:
where . An admissible path of is a function from into the set of all admissible market values, .
Remark 1 (On static arbitrages).
Remark 2 (On dynamic arbitrages).
Note that, with the definition of admissible paths of above, we are allowing paths which may admit dynamic arbitrage opportunities. We could further restrict the set of admissible paths to, essentially, those which are supported by some martingale measure (for and the options) and satisfy additional constraints, e.g. pathwise beliefs, introduced further in the paper. All of the statements in Sections 5 and 6 would hold with such a modified definition of admissibility. In some abstract setups such an approach is necessary to have the pricing–hedging duality, see e.g. [12]. However in our setup, similarly to e.g. [28], this is not necessary, as we do not analyze the optimality of the proposed trading strategies in the classical probabilistic sense.
We work with the sets of admissible paths of and, in many cases, introduce additional continuity assumption on . Every path of is a realization of the future states of the market: it determines the future prices of all traded instruments. Crucially, later on we further restrict the set of possible market values, so that they are admissible and satisfy some additional constraints stated via beliefs on the implied volatility. The beliefs are expected to arise from statistical observations of the implied volatility and may e.g. take a form of confidence intervals. They are discussed in detail in the following subsection.
The traded instruments can be used to construct portfolios, which arise as the linear combinations of the European options and the underlying, with the time varying (adapted) weights. A static portfolio has constant weights and we restrict ourselves to portfolios which trade finitely many times. This is both realistic and allows to circumvent issues with pathwise definitions of a stochastic integral, seen in e.g. [29, 23]. We assume that the pricing operator is linear: the price of any portfolio at time is the associated linear combination of the market prices of its elements. Definition 1 and Remark 1 imply that the price of any static portfolio of coterminal calls, puts and underlying, whose terminal payoff function (i.e. the associated linear combination of the payoffs of its elements) is nonnegative, has a nonnegative price at any time before the maturity. We will use this observation implicitly in the proofs in Sections 5 and 6.
3.2 Beliefs on Implied Skewness
We now turn to the specific type of beliefs we want to study in this paper. In order to define the beliefs, we consider a family of functions:
(6) 
which determine the bounds for the skewness of the market implied volatility. Typically, we have , but there is no need to enforce this inequality.
Definition 2.
Given a set of functions , as in (6), and a barrier , we define the beliefs , for any , as the set of all implied volatility surfaces , which admit a constant (depending on ), satisfying
(7) 
Notice that it is natural to make beliefs about the values of implied volatility in the moneyness variable, , as opposed to the strike variable, . This is why the beliefs are formulated w.r.t. a given barrier : it is often convenient to assume that the beliefs are expected to be satisfied when . Even though it is implicitly assumed that is close to , so that the above inequalities have a correct interpretation, in some cases, we need to consider the above beliefs for . The latter is stated explicitly in the main results of the paper.
Notice also that, for , we can estimate the implied skewness from below:
(8) 
Moreover, for any , the beliefs generated by , are the same as the beliefs generated by . In other words, the beliefs restrict the skewness of a member implied volatility w.r.t. the barrier (defined as the left hand side of (8)), via the skewness of the input functions. However, these beliefs are invariant w.r.t. scaling of the input functions.
The beliefs provide lower bounds on the implied skewness. Similarly, we introduce the beliefs , which provide upper bounds on the implied skew.
Definition 3.
Given a set of functions , as in (6), and a barrier , we define the beliefs , for any , as the set of all implied volatility surfaces , which admit a constant (depending on ), satisfying
Clearly, any , satisfies the upper bound on its skewness:
Notice that the beliefs and can be defined with a single set of functions – either or . The reason we introduce both sets is that, if , the interval can be interpreted as a confidence interval for the values of . Then, when approach zero and grow to infinity, the beliefs and include more and more surfaces, converging to the set of all admissible implied volatilities, with arbitrary skewness. On the other hand, when , for all , the set is either empty or includes implied volatilities with the same shortterm skewness (i.e. as ). In this sense, the methodology we develop herein interpolates between a specific market setting (where the shortterm implied skewness is determined uniquely) and the modelindependent setting (where the implied skewness may take arbitrary values).
Whenever or is invoked, we assume that it is created with some barrier and some input functions , as in (6). The latter functions may be left unspecified if it causes no ambiguity. We say that the market implied volatility satisfies the beliefs or at time , if or , respectively.
Remark 3.
It may often be convenient to specify the beliefs on implied skewness for a smaller set of strikes than the one that is actually available in the market. For example, the results in Sections 5 and 6 are formulated for a fixed strike , and they describe trading strategies which never include any options with other strikes below the barrier. In this case, the input functions (and, hence, the beliefs) may be constructed for the set of strikes that includes all and only one .
It is often convenient to extend the set or , in order to make it more tractable. This is done via the dominating implied volatility surfaces, or, the dominating surfaces, for short.
Definition 4.
An implied volatility is a lower dominating surface for beliefs , if is an admissible market value and
Similarly, an implied volatility is an upper dominating surface for beliefs , if is an admissible market value and
In particular, the lower dominating surface has a lower shortterm implied skew than any implied volatility in . Similarly, the upper dominating surface has a higher shortterm implied skew than any implied volatility in . Corollary 1 below shows that a dominating surface always exists, provided and is sufficiently small. A dominating surface, in general, does not belong to or . However, if it does so for arbitrarily large , we say that it generates the beliefs or .
Definition 5.
A lower dominating surface generates the beliefs if
Similarly, an upper dominating surface generates the beliefs if
It turns out that it is very convenient to analyze the beliefs generated by dominating surfaces. This is due to the fact that a dominating surface can be constructed via a specific martingale model, and this is what allows us to use the classical probabilistic tools of Financial Mathematics in the present robust (pathwise) analysis. We achieve this in Section 4.1 below, where we construct a convenient parametric family of dominating surfaces, produced by a specific class of martingale models for the underlying. The main results are, then, formulated in Sections 5 and 6 for the beliefs generated by such dominating surfaces.
4 Piecewise Constant Local Variance Gamma (PCLVG) Models
In this section, we introduce and analyze the relevant family of timechanged local volatility models, which posses three important features. First, they allow for explicit computation of European options’ prices and of the shortterm implied volatility. Second, the models allow for skew in the implied volatility they generate, which is flexible enough to bound the market implied skew. Third, any model form this family has an explicit static hedging formula, in the spirit of (4). We believe this is a remarkable set of properties and the family of models is of independent interest. The proofs of the results in this section are technical and are all relegated to Appendix A to allow for a more concise presentation.
The models presented herein are closely related to the Local Variance Gamma (LVG) models, introduced in [18]. Let the process , taking values in , be defined for as the unique weak solution to
(9) 
absorbed at zero. In the above, is a Brownian motion, and the function is piecewise constant, of the form
with some constants . The existence and uniqueness of the solution to (9) is discussed in [18]. Finally, we define the stochastic process as a time change of . Consider a random variable , such that is independent of and has an exponential distribution with mean one. Then, we set
(10) 
It is clear that is a continuous nonnegative martingale. Equation (10), parameterized by , describes a plausible riskneutral evolution of the underlying (recall that the carrying costs, including interest and dividend rates, are assumed to be zero). We will refer to the parametric family of models given by (10) as the PCLVG (Piecewise Constant Local Variance Gamma) models. The name is motivated by the similarity of the above construction and the LVG models introduced in [18]. Indeed, the main difference between the two is the different choice of a time change: the latter is assumed to be a Gamma process in [18]. However, it is also mentioned in [18] that any other independent time change will produce models with similar features, as long as the marginal distributions of the time change are exponential. Herein, we choose as the desired time change process, so that it has exponential distribution at any time . This particular choice is motivated by the desire to have a nontrivial shortterm implied volatility produced by the models, and it is explained by the results of the next subsection.
4.1 Dominating PCLVG Surfaces
In this subsection, we compute the implied volatility in a PCLVG model and show that it can be used to generate a dominating surface for general beliefs, provided the barrier does not coincide with any of the strikes and the maturity is sufficiently small. Importantly, our proof is constructive – it provides a method for computing the dominating surfaces numerically.
Denote the time zero price of a European call option produced by a PCLVG model with parameter by
where is the initial level of underlying, denotes the strike, and is the time to maturity. Similarly, we define the price of a European put, denoted .
Lemma 1.
For any and any , we have:
(11) 
For any and any , we have:
(12) 
For any and , denote by the BlackScholes implied volatility of . It is easy to check, using equations (11)–(12), that the associated call prices always lie in the interval , hence, the implied volatility is always well defined. In addition, it is not difficult to check that the call prices, given by (11)–(12), satisfy all the static noarbitrage conditions of Definition 1 except the first condition on the existence of a shortterm limit of implied volatility. The following proposition verifies the first condition of Definition 1, thus, showing that is an admissible market value, and it provides an explicit expression for the shortterm implied volatility.
Proposition 1.
For any and any , we have
For any and any , we have
Several graphs of , for small but strictly positive , are given in Figure 4. The above proposition has an important corollary which shows that, by choosing , we can construct a dominating surface for any beliefs, provided the maturity is sufficiently small.
Corollary 1.
Consider any , s.t. , and any as in (6). Then, there exist and , such that
(13) 
In particular, is an upper dominating model for the beliefs generated by and . Similarly, there exists and , such that
(14) 
In particular, is a lower dominating model for the beliefs generated by and .
Proof.
For brevity, when the level of the barrier is fixed, we simply refer to as a dominating surface instead of writing .
4.2 Exact Static Hedge in PCLVG Models
As follows from the results of [17] (further analyzed and extended in [5]), the static hedge of a UOP with strike and barrier , in a timehomogeneous diffusion model with coefficient , and with zero carrying costs, is given by a European option with the payoff , where
(15) 
and , for . In the above, is an arbitrary, sufficiently large, constant, and ’s are the two fundamental solutions of the following ODE, normalized to take value at :
It is easy to verify that, in the present case, we have:
The static hedging property means that, with denoting the aforementioned timehomogeneous diffusion, we have:
where and is the filtration w.r.t. which is defined. The left hand side of the above is the price of a UOP, and the right hand side is the price of a comaturing European option, evaluated at or before the underlying hits the barrier (if the barrier is not hit, the payoffs of the two options coincide). The above result also applies to the processes obtained as an independent continuous time change of a timehomogeneous diffusion, such as the PCLVG processes. However, strictly speaking, the PCLVG process does not satisfy the assumptions made in [17]. Indeed, the coefficient is discontinuous: it is piecewise constant, taking values , for , and , for . Nevertheless, we can still use (15) to compute a candidate , which is expected to produce a static hedge in the PCLVG model. Assuming and , we obtain
and, in turn,
By closing the contour of integration on the right, we conclude that the second integral in the above expression is zero. Thus,
By closing the contour of integration on the right, we conclude that the integrand inside the th integral in the first sum is zero when . For the other values of , we close the contour of integration on the left and compute the integral via the residue calculus. Similarly, we proceed with the second sum. As a result, we obtain
(16) 
A graph of is given in Figure 1. Having an explicit candidate, we can use Markov property and explicit option prices in (11)–(12) to show that this is indeed the correct function. The latter is stated precisely in the following proposition, whose proof is given in Appendix A.
Proposition 2.
Assume that follows a PCLVG model, with , , . Then, for any , any , and all , we have:
(17) 
where is given by (16) and .
Assuming , taking expectations in (17) and using (16), we obtain the following formula, which holds for all , any , and any :
(18) 
The above formula is the analogue of (4), but it holds in a model in which the underlying is not symmetric w.r.t. the barrier (and, hence, the implied volatility may have a skew).
Notice also that
(19) 
where
(20) 
The function is the largest convex function dominating from below. The function is one of the minimal convex functions dominating from above: i.e. there exists no other convex function which dominates from above and is dominated by from above. The graphs of , , and , are given in Figure 1. We can see that provides a good approximation of for the values of the argument that re not too far from the barrier (which are the most important values, if the maturity is small and the underlying is close to ). Moreover, the right hand side of Figure 1 shows that, if is close to (this graph uses , which is the value used in our empirical analysis), is very close to the payoff of a call struck at , and, in particular, both , and provide a very good approximation of in such a regime.
Thus, in a PCLVG model, conditional on the level of underlying, we can bound the price of an OTM put uniformly, from above and from below, by the prices of respective static portfolios, consisting of a single OTM call each:
(21) 
Note that the above inequalities hold for all (not only asymptotically, for ) and, hence, for all times until the maturity, whenever the underlying is at . Considering a symmetric problem of static hedging a downandoutcall (DOC), we can derive the following, equivalent, system:
(22) 
Remark 4.
It is easy to deduce from Lemma 1 that
In particular, the second inequality in (21) is “infinitely tighter” than the first one, for small . The above equivalence, in principle, is well known in a general diffusion model (cf. [26]) and can be obtained without establishing the exact static hedge: one simply needs to notice that and are symmetric w.r.t. the geodesic distance given by . However, such an equivalence, alone, does not yield the second inequality in (21), even if it is restricted to small , which explains the need for Proposition 2. In addition, Proposition 2 yields the first inequality in (21) and shows that both of them hold for all .
5 Trading the Deviations of Implied Skew
Let us show how (21) can be used to construct a portfolio for robust trading of the implied skew – further referred to as RTIS. Buying such a portfolio, whenever the implied skewness deviates from its typical range of values, and selling it when the implied skewness comes back to its normal range, should yield a positive return. More precisely, for a given barrier and a given set of beliefs, or , on the implied skewness, the RTIS portfolio is a static portfolio of vanilla options, which satisfies the following two properties:

if the implied skewness satisfies the beliefs, and if the underlying is on the appropriate side of the barrier, the portfolio has a positive price;

if the implied skewness deviates sufficiently far from the beliefs, and if the underlying is not too far from the barrier, the portfolio has a negative price.
Note that, whenever the underlying is on the appropriate side of the barrier, the price of an RTIS portfolio is guaranteed to have a positive sign, depending on the level of implied skew, but regardless of the overall level of implied volatility, and independent of the exact value of the underlying. Of course, the exact price of an RTIS portfolio may depend on the entire shape of the implied volatility and on the underlying value. The precise meaning of the two properties of an RTIS portfolio is given below, in Propositions 3 and 4.
However, the above formulation is sufficient to see the practical benefit of RTIS. For example, assume that, at time , the implied skewness measured relative to the current level of underlying (i.e. using as a barrier) deviates sufficiently far from the beliefs (constructed w.r.t. ). Then, one can choose a new barrier and construct the corresponding beliefs w.r.t. . The new barrier needs to be close to , so that the implied skewness measured relative to is also sufficiently far from the beliefs, and, in turn, the price of the associated RTIS portfolio is negative at time . On the other hand, the barrier should not be too close to , so that the underlying is likely to remain on the same side of for some time (the choice of optimal is the “art” of implementing such a strategy successfully). One can, then, open a long position in the RTIS portfolio at time . If, at some future time, the skewness returns to a level consistent with the beliefs (one has to assume that this will occur eventually, as this is the principle by which the beliefs are constructed), and the underlying remains on the appropriate side of the barrier, the price of the RTIS portfolio becomes positive, yielding a positive return. Whenever this strategy is implementable, it yields a positive return. Its implementability, in turn, depends on the behavior of the implied skew and the underlying, but not on any other changes in the implied volatility. The sensitivity of the strategy with respect to changes in the underlying can be mitigated by opening an additional static position in a forward contract struck at (which has initial price zero and is insensitive to changes in the implied volatility), so that the initial BlackScholes delta of the portfolio is zero.
Let us show how to construct an RTIS portfolio satisfying the two defining properties stated above. The choice of a portfolio depends on whether we are given beliefs or . Herein, we consider the case of , with the other case discussed in Remark 9. Consider a barrier , and beliefs , generated by , with some (cf. Definition 4). Consider any time , s.t. is admissible, , with some , and . Notice that, scaling with a positive constant , scales the in the natural way: