Calibrating the SABR Model to Noisy FX Data
Kellogg College University of Oxford
A thesis submitted in partial fulfillment of the MSc in Mathematical Finance
Hilary 2018 Abstract
We consider the problem of fitting the SABR model to an FX volatility smile. It is demonstrated that the model parameter β cannot be deter- mined from a log-log plot of σATM against F . It is also shown that, in an FX setting, the SABR model has a single state variable. A new method is proposed for fitting the SABR model to observed quotes. In contrast to the fitting techniques proposed in the literature, the new method allows all the SABR parameters to be retrieved and does not require prior beliefs about the market. The effect of noise on the new fitting technique is also investigated. Acknowledgements
I would like to thank both of my supervisors Dr Daniel Jones and Guil- laume Bigonzi for their guidance and support throughout this project. They provided the direction for this work and offered invaluable insight and advice along the way. I also gratefully acknowledge the financial support from Bank Julius Baer and Co. Ltd. Finally I thank Dr Beate Solleder for her love and tireless support and for tolerating me during the time it has taken me to complete this project. 1 Introduction
The work presented here is concerned with fitting the SABR stochastic volatility model to foreign exchange (FX) data. Specifically, we are interested in the implied volatility of an option, σ, as a function of the strike of the option, K. This relation- ship between σ and K is known as the volatility smile. For specified SABR model parameters, the volatility smile is given by the well known equation of Hagan et al. [1]. Here we focus on the inverse problem, i.e. given a volatility smile which was generated using the SABR model, how can we obtain the parameters of the underlying SABR model? In section 2 we introduce the quoting conventions used in the FX market and define the three options which are commonly used to describe the FX volatility smile. A method for calibrating a volatility smile to market quotes is also described. The SABR model is presented in section 3 and the equations which will be used throughout this work are stated. Section 4 reviews previous works related to fitting the SABR model to market data and section 5 gives details of the Monte Carlo method which was used to generate simulated market data. In this work we have focused on fitting simulated data since this removes any uncertainty regarding whether the SABR model accurately describes the market and what the true model parameters are. A key topic when fitting the SABR model to market data is the determination of the parameter β. In section 6 we explain why an approach to fitting β that is often described in the literature does not produce reliable results. Section 7 examines whether variance-covariance matching can be used to estimate the SABR parameters from time series of σATM and F . The relationship between the three main FX options quotes that would be predicted by the SABR model is investigated in section 8. These predictions are also compared to sample market data. An approximation for the correlation between the implied volatility at-the-money and the forward price is considered in section 9. This correlation is important because it allows vega exposure to be partially hedged with delta. A new method for fitting the SABR model to FX data is proposed in section 10 and the ability of this method to retrieve the parameters of the underlying SABR model is investigated. We begin with the case that the quotes are free from noise and then systematically introduce noise on each of the three main option quotes. Section 11 considers the case of pricing a digital option when the volatility smile is described using the SABR model. Conclusions are drawn in section 12, which also includes suggestions for further work.
1 2 Introduction to FX Market Conventions
The foreign exchange (FX) market is one of the most liquid and competitive mar- kets in the world. Because many of the FX market conventions are unique to this market, this section provides a brief introduction to these conventions. In the FX market participants agree to exchange one currency for another on a specified day at a specified FX rate. An FX rate is the price of one currency expressed in terms of another currency. Consider the currency pair XXXYYY. The tag XXX represents the “foreign” currency, while YYY represents the “domestic” currency. The FX rate XXXYYY specifies the price of the foreign currency in terms of the domestic currency. For example; EURUSD specifies the price of one Euro in US dollars. This section begins with an explanation of the delta conventions which are used for quoting options in the FX market. Thereafter we introduce three commonly traded options structures: the at-the-money straddle, risk reversal and vega-weighted butterfly. These three structures are particularly important because they are often used to define the volatility smile in the FX market. The section concludes with a description of a method for calibrating a volatility smile to observed market prices.
2.1 Option Quotes in the FX Market
Options in the FX market are not typically quoted in terms of strike, K, but as the delta of the option, assuming a Black-Scholes (BS) model. The delta of an option in the BS model is given by ∂V ∆ = ∂S R T f − rs ds = we t Φ(wd1), (1) where F (t,T ) σ2 ln( K ) + 2 (T − t) d1 = √ (2) σ T − t and Φ is the normal cumulative distribution function. Here V is the value of the option and S is the current (spot) exchange rate. F (t, T ) is the forward price of the exchange rate at time t expiring at time T and w takes the value of w = 1 for a call f and w = −1 for a put option. rt is the risk free rate of the foreign currency at time t. Although FX options are quoted in terms of ∆, options are actually written with a specified strike. Therefore we need to be able to convert ∆ into the corresponding
2 strike. Re-arranging equation (1) leads to the following expression for K
√ |∆| σ2(T − t) K = F (t, T ) exp − wσ T − tΦ−1 + . (3) − R T rf ds e t s 2 In the context of this work, equation (3) is important because the SABR model gives the implied volatility as a function of K, rather than ∆. FX option quotes are further complicated by the use of different definitions of delta depending on the market convention of the currency pair being traded. For some currency pairs the market convention is to quote the premium in the foreign currency, e.g. a vanilla option on the USDJPY pair is quoted in USD. Since the premium is in foreign currency, the premium itself should be hedged. Therefore the market convention is to use the premium included delta, which is given by V ∆PI = ∆ − . (4) S Consider the case that we write a call option on the USDJPY pair. At expiry this gives the buyer the right to purchase USD at a price specified by the strike K, which is in JPY. To make our position (instantaneously) risk free with respect to S we should hold ∆ USD given by equation (1). However, as the option writer, we receive the option premium given by V/S, where V is the value of the option in JPY and V/S is the premium in USD. The premium included delta is the amount of USD that we need to hold in addition to the premium, which leads to equation (4). Finding the value of K which corresponds to a specified value of ∆PI is more involved because both ∆ and V depend on K. Castagna [2] proposed the following method based on Newton’s method to calculate K:
1. Calculate an initial estimate of K using equation (3)
2. Calculate ∆PI for the current value of Ki using equation (4)
3. Estimate the derivative of ∆PI with respect to Ki by “bumping” Ki by a small amount (e.g. 1%) and re-evaluating ∆PI for this new value of K
4. Calculate Ki+1 as ∆PI − ∆¯ Ki+1 = Ki − (5) ∂∆PI ∂K where ∆¯ is the target value of ∆PI.
5. Iterate until |Ki+1 − Ki| < , where is a tolerance parameter.
3 2.2 At-the-money Straddle
The most liquid FX option is the at-the-money (ATM) straddle. This structure consists of a call and put both struck at the “at-the-money” level. The definition of the ATM strike depends on market conventions. One choice is the zero delta ATM strike, which is defined as the strike that leads to the call and the put having the same delta (but with opposite sign). Another possible definition of the ATM strike is the ATM forward. Under this convention the ATM strike is set equal to the forward price of the underlying currency pair with the same expiry as the option. By no-arbitrage the forward price is given by
− R T rf ds e t s F (t, T ) = St , (6) − R T rdds e t s
d where rt is the risk free rate of the domestic currency at time t. The final definition of the ATM strike is the at-the-money spot, where the ATM strike is defined to be S, the current spot rate of the underlying pair. The at-the-money straddle describes the level of the implied volatility surface: changing the ATM volatility results in a parallel shift of the implied volatility surface along the implied volatility axis.
2.3 Risk Reversal
A risk reversal is a highly-traded structure consisting of a long call and a short put. The call and put are symmetric in that they are chosen to have the same delta (but with opposite sign). The most commonly traded risk reversal contract is the 25 delta contract, where the call and put are stuck such that they have deltas of 0.25 and -0.25, respectively. In the market, the risk reversal is quoted as the difference between the implied volatilities of the call and the put, i.e.
σ25RR(t, T ) = σ25C (t, T ) − σ25P (t, T ). (7)
The risk reversal can be either positive or negative and describes the skew of the implied volatility surface. A positive risk reversal indicates that there is more demand for calls than puts, whereas a negative risk reversal suggests that puts are favoured over calls.
4 2.4 Butterfly
A vega-weighted butterfly (VWB) is a highly-traded structure consisting of a long call, a long put and a short ATM straddle. The long call and long put are again symmetric in delta and together form a strangle. For the most commonly traded butterfly, delta is again chosen to be 0.25 for the call and -0.25 for the put. This is referred to as the 25 delta butterfly. The vega of the strangle is larger than that of the ATM straddle meaning that the quantity of the straddle needs to be larger than the quantity of the strangle in order that the structure is vega neutral. The market quote for the VWB is defined as the difference between the volatility of the strangle and the volatility of the ATM straddle (σATM). Market quotes for the vega-weighted butterfly are complicated by the existence of two conventions for the strangle. The most straight forward definition of the strangle is to use the same put and call options which were used for the risk reversal. This results in the ‘two-vol’ butterfly, which is defined as 1 σ (t, T ) = σ (t, T ) + σ (t, T ) − σ (t, T ). (8) 25BF 2 25C 25P ATM Under this convention the volatility of the strangle is defined as the mean of the volatilities of the put and the call. However, the most common market quote for the VWB is not the two-vol butterfly, but the single-vol butterfly. In this case the volatilities of the put and the call are chosen to be equal to one another. Define σVWB to be the volatility of the put and the call for the single-vol strangle. The market quote for the single-vol butterfly is then
σ1−vol−25BF(t, T ) = σVWB(t, T ) − σATM(t, T ). (9)
For σ25RR = 0, equation (8) reduces to equation (9) and the two conventions for VWB are equivalent. In general, however, the two definitions are not equivalent and the discrepancy between σ25BF and σ1−vol−25BF tends to increase as the magnitude of σ25RR increases. Either σ25BF or σ1−vol−25BF can be used to construct a volatility smile. What is important is to understand which convention is being used and how to interpret the market quotes in term of the constraints that they place on the volatility smile. For simplicity the majority of this work has been performed using the two-vol butterfly, σ25BF. This choice means that fewer strikes are required in the calibration process. Section 2.5 describes how to calibrate a volatility smile using market quotes for σ1−vol−25BF, which is the case that will be most frequently encountered in practice.
5 The butterfly describes the curvature of the implied volatility surface; a high value of σ25BF(t, T ) implies that the implied volatility in the wings is large compared to the implied volatility at-the-money.
2.5 Building the Volatility Smile from Market Data
The volatility smile is a mapping between strike, K, and implied volatility:
K 7→ σ(K). (10)
In this section it is assumed that we have a functional form for σ(K) which we wish to fit to market quotes for σATM, σ25RR and σ1−vol−25BF. This is the case that is most frequently encountered in practice. It is assumed further that, given three points on the volatility smile, we can fit the function σ(K) such that we can obtain the volatility for any K ≥ 0. Although this work focuses on fitting the SABR model, the method described below can be applied to any functional form which meets these criteria. For example, the vanna-vega interpolation method proposed by Castagna [2] or the simplified parabolic interpolation method introduced by Reiswich [3]. Constructing the volatility smile from market data is achieved by recognising the three constraints placed on the smile by the three options quotes discussed above (σATM, σ25RR and σ1−vol−25BF). These constraints are described in detail by
Reiswich [3] and Castagna [2]. The market quote for σATM provides the constraint
σ(KATM) = σATM, (11) where KATM is determined by market conventions. To ensure that σ25RR is priced correctly by the volatility smile we have
σ(K25C) − σ(K25P) = σ25RR. (12)
Here the strikes K25C and K25P fulfil
∗ ∆ (K25C, σ(K25C)) = 0.25, ∗ ∆ (K25P, σ(K25P)) = −0.25. (13)
The function ∆∗(K, σ) is either the standard delta or the premium included delta and is determined by market conventions. The final constraint is that the value of a VWB priced by the volatility smile should match the price quoted in the market. Here we
6 assume that the market quote for the VWB uses the single volatility convention. The put and call that make up the VWB have a volatility given by
σVWB(t, T ) = σ1−vol−25BF(t, T ) + σATM(t, T ). (14)
The strikes of these options can be found by solving the equations
∗ ∆ (K25C, σVWB) = 0.25, ∗ ∆ (K25P, σVWB) = −0.25 (15) for K25C and K25P. The value of the strangle component of the VWB is:
C(K25C, σVWB) + P (K25P, σVWB). (16)
Here C(K, σ) and P (K, σ) are, respectively, the Black-Scholes price of a call (put) option with strike K and volatility σ. The volatility smile must be able to reproduce the price of this strangle, which leads to