Calibration of the Volatility Surface

Erik Nilsson erinil@kth.se 840428-0292 June 12, 2008

Abstract
This thesis consists of two parts, one concerning implied volatility and one concerning local volatility. The SABR model and SVI model are investigated to model implied volatility. The performance of the two models were tested on the Eurcap market in March 2008. Two ways of extracting local volatility are reviewed by a test performed on data from European options based on the S&P 500 index. The rst method is a way of solving regularized Dupire's equation and the other one is based on nding the most likely path.

Acknowledgements
I would like to thank Robert Thorén and his coworkers at Algorithmica Research AB for guiding me during my work. I am also grateful to Professor Anders Forsgren at KTH, who has been my tutor.

. . . . . 3. . .2 Calibration . . . . . . 1. . . . . . . . . . . . . . . . . . . . . . . . . . .1 Procedure . . . . . . . . . . . . . . . . . . . .2 Volatility . . . . . . . . . . . . . . . . . . . .2. . . . . . . .2 Results . . . . . 5 5 5 5 2 Stochastic volatility 8 8 9 3 Local volatility 12 12 14 14 17 4 Test of the implied volatility models 4. . . . . .1 Theory . . . . . . . .2. . . . . .1 The regularized Dupire equation 5. . . . . . . . . . . . . 3. . . . . .2 The most likely path calibration . .2. . . . . . . . . . . . .2 Local volatility by most likely path calibration . . . . . . . . . . . . . . . . .1 Theory . . . . . . 2. . . . . . . . . .2 Calibration . . . 3. . . . . .3 Introduction to arbitrage-free option pricing . . . . . 5. . . .1 Procedure . . . . . . . .2 Results . . . 3. . . . . . . . . . . . . . . . . . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . .Contents 1 Introduction 1. 19 5. . . . . . . . . . . 1. . . . . . . 22 19 5 Test of the local volatility models 22 24 24 26 6 Discussion 30 . . . . . . . . . . . . . . . . . . . . .1 Local volatility by solving Dupire's equation . . . . . . . . 19 4. . . . . . . . . . . . . . .2. . . . . . . . . . . . . . .1 Purpose and background . . . . . . . . . . . .

an interest rate or an equity. There are a lot of dierent stochastic volatility models which will be covered in a later section. developed by Algorithmica Research AB. Local volatility is the volatility used in the general Black-Scholes model and it is a deterministic function of expiration time and price of the underlying.2 Volatility Volatility is the standard deviation of the time series for a stock. The contracts can be a lot more complex. The second goal is to investigate whether there is a method which can recover a plausible local volatility surface from a market implied volatility surface. interest rate etc.1 Introduction Purpose and background The purpose of this thesis is to give a theoretical background of which methods that are used to model volatility. Stochastic volatility is an extension to the Black-Scholes model where the volatility itself is a stochastic process. The reader is expected to have basic knowledge in optimization and stochastic calculus. Mathematically this means that at the maturity T. Quantlab has been the tool for implementation. stable and fast on the option interest rate market. for example exotic options where the payo at maturity not only depends on the value of the underlying at maturity but its value several times during the contract's life or it could depend on more than one underlying. and one modeling local volatility. specialized for nance. 1. There are three dierent types of volatility which will be treated in this thesis: Implied volatility. stochastic volatility and local volatility. The implied volatility is the volatility used in Black-Scholes formula to generate a given option price. The underlying is usually a stock. The second part is implemented in Matlab. A European put option is the same. The thesis can be divided into two sections. Volatility has traditionally been used in nance as a measure of risk. ST − K } for a European call option. An easy example of an option is the European call option which gives the buyer the right but not the obligation to buy a stock for the strike price K at maturity T.1 INTRODUCTION 1 1. For the rst section. but this time the buyer has the right but not the obligation to sell the underlying. 5 . one about modeling implied volatility. It is important to understand the dierences between these. The rst goal is to nd an implied volatility method which is robust. 1. the value is max{0. It is a powerful numerical computing environment.3 Introduction to arbitrage-free option pricing An option is a nancial contract dependent on an underlying.

They believed that a price is determined relative to other prices quoted in the market in such a manner as to preclude any arbitrage opportunities. Later on. S − K } under the risk neutral measure Q given the information today Ω. σBS ) = Sφ(d1 ) − Ke−rT φ(d2 ). which gives a better understanding for the meaning of arbitrage-free. i. (1. Black and Scholes postulated that the price S of an underlying behaves like a diusion process. dS = rdt + σdWt S (1. risk neutral pricing and more in the theory of stochastic calculus. the option price is given by the discounted conditional expectation of the payo function max{0. A cap can be treated as a sum of caplets. This theory is called Arbitrage-Free Option pricing and was among others developed by Fischer Black and Myron Scholes. Therefore a mathematical theory was needed to nd some kind of consensus in option pricing. For interested readers. A caplet is a european option with an interest rate as an underlying and having the expiration time equal to the period of payments in the cap contract. The payment is compensating the excess rate. methods for extracting the implied volatility surface on the cap market will be tested. The diculty with options is to determine a fair price. hence the buyer never has to pay more interest rate than the strike rate. Arbitrage opportunity means that there is a nancial strategy which guarantees no loss and has positive probability of a prot. where just a few contracts are settled. The more complex a contract is. especially for ill-liquid markets.2) i.1. (1. This expression has a closed form formula called the Black-Scholes formula. "There is no such thing as a free lunch".e. A one year cap with a 6 month payment period is the sum of one 6 month caplet and another 6 month caplet starting in 6 months. K. To derive the remainder of the Black-Scholes formula. An example of a cap would be an agreement to receive a payment for each 6 months the LIBOR rate exceeds 2. so therefore a huge step is made to obtain the following option price formula CBS = e−rT E Q [max{0. the more dicult it is to actually know the "right" price for it. An interest rate cap is a derivative in which the buyer receives payments at the end of each period in which the interest rate exceeds the agreed strike rate. r. we give the explicit form CBS (S. σ is the volatility and W is a Brownian motion.e. S − K }|Ω]. see [2].3 Introduction to arbitrage-free option pricing 1 INTRODUCTION A product which will be treated in this thesis is a so called cap. This thesis cannot cover all of this. There is a famous expression often used in nancial literature.3) 6 . the readers have to be familiar with martingale theory.5 %.1) where r is the risk-free rate. T. 6 months in the case mentioned above.

that the model assumes log-normally distributed returns for stocks. see [2]. Henceforth just CBS (S. r.1 INTRODUCTION 1. This is due to the fact. but since then. T. A lot of research has been done to handle the weaknesses of the Black-Scholes model. K. which not always agrees with observed returns.3 Introduction to arbitrage-free option pricing where d1 = d2 = 2 /2)T ln(S/K ) + (r + σBS σBS σBS (T ) (T ) . (1. it has been more popular to just model implied volatility and then keeping the BlackScholes analytical formula as a tool to quote option prices from implied volatilities. Riccardo Rebonato's view of implied volatility is "the wrong number put in the wrong formula to obtain the right price". σBS ) will be used as the option price given by the Black-Scholes formula. 7 .4) (1. The Black-Scholes formula worked well before the huge crashes 1987 and 1989. Since Black-Scholes is widely used across the world by banks. see [12].5) 2 /2)T ln(S/K ) + (r + σBS and φ is the standard normal cumulative distribution function. it has been observed that implied volatility has a skewness and smile structure. which contradicts an assumption of constant volatility. These observations usually give a skewed distribution with fatter tails. .

11) 8 .5) (2. v ¯ long term mean. This is an extension to the dynamics of the Black and Scholes model. √ (2. the more dicult they are to calibrate.1 Stochastic volatility Theory A stochastic volatility model is a model where the volatility itself is a stochastic process.8) Three new parameters λ. Ft (0) = f. The Bates model has nine parameters and SVJJ even more.3) where St is the price of the underlying. Mean reversion means that the process strives to a long term mean value.9) (2. E (dW1 dW2 ) = ρdt. Xt ∼ P o(λt). Despite the fact that Stochastic volatility models are dicult to calibrate in general. This serves as a better explanation for discontinuous jumps in the market. which has simultaneous jumps in the underlying and the volatility. Another popular model is the Bates model which is an extension to the Heston model. The dynamic model from [8] has the following mathematical representation vt St dW1 .6) (2. dαt = ναt dW2 . µ is the constant drift. αt (0) = α. 2 J (2. The dierence lies within the price process where a Poisson process is added. A further extended model is the SVJJ. vt is the variance of the underlying. σ2 . the SABR model given by dFt = σt Ftβ dW1 .2) (2. √ dvt = κ(¯ v − vt )dt + η vt dW2 . there is one stochastic model that diers.4) (2. One popular model is the Heston model. κ speed of reversion.7) where r is the risk-free rate and the random variable Jt determines the jump size and follows the distribution log(1 + Jt ) ∼ N log(1 + µJ ) − 2 σJ .1) (2. where the price of the underlying is a geometric brownian motion and the volatility is a geometric brownian motion with mean reversion. The mathematical representation from [1] is dSt = (r − λµJ )St dt + vt St dW1 + Jt St dXt .2 STOCHASTIC VOLATILITY 2 2. dSt = µSt dt + √ (2. (2. √ dvt = κ(¯ v − vt )dt + η vt dW2 . η volatility of volatility and ρ the correlation between the two brownian motions W1 and W2 .10) (2. E (dW1 dW2 ) = ρdt. However the more complex these models are. µJ and σJ are added to this model. E (dW1 dW2 ) = ρdt.

2 Calibration where Ft is the forward value. ρ. From empirical observations. ν and has the analytical closed form formula σBS (K. σ. translates the graph. Therefore. However as the expiration time τ → 0. Due to the analytical formula. b. m) = a + b(1 + ρ)(k − m) k → ∞. This model was created by Paul Hagan et al.14) In this section. a. β. which agrees with the assumption of linear wings. ρ. The SABR model has four parameters α. 2. It is a clever parameterization of the implied volatility surface. Calibration of the other stochastic volatility models are beyond the scope of this thesis. m) = a − b(1 − ρ)(k − m) k → −∞.2 STOCHASTIC VOLATILITY 2. an exact expression for the implied volatility can be obtained by using singular perturbation techniques.15) ex ex 24 (f K )1−β 4 (f K )(1−β )/2 24 « z · x(z ) 9 . determines how smooth the vertex is.13) (2. (2. a. b. b. the implied variance is always linear in the wings and curved in the middle when plotted against the logarithmic moneyness. the model is easy to calibrate. Another easy calibrated model is the SVI (Stochastic Volatility Inspired) model found by Jim Gatheral [5]. The forward value and volatility are under the forward measure and the two processes are correlated with ρ. The variance has the left and right asymptotes varL (k. m) = a + b ρ(k − m) + where a b σ ρ m (k − m)2 + σ 2 (2. f ) = α n (1−β )2 (1−β )4 „ o f f f (f K )(1−β )/2 1 + 24 log2 K + 1920 log4 K + O(log6 K )  » – ff 2 2 2 (1 − β ) α 1 ρβνα 2 − 3ρ 2 2 · 1+ + + ν t + O ( t ) .2 Calibration (2. α is the volatility. ρ.12) gives the overall level of variance. σ. varR (k. The forward value is Ft = Sert and r is the rate. ν is the volatility of volatility and W1 and W2 are Brownian motions. the calibration of the SABR and SVI model is described. σ. ρ. the author suggested the following parameterization var(k. see [6]. gives the angle between the left and right asymptotes. a. determines the orientation of the graph. The dynamics of this model is similar to the ones shown above but do not have the volatility mean reversion property and is therefore only good for short expirations theoretically.

For well behaved functions and with good initial guesses. which has an algebraic solution. a certain preprocess is needed.2 Calibration 2 STOCHASTIC VOLATILITY Notice that this formula is expressed in strike K and the forward value. The LMA interpolates between the Gauss-Newton algorithm and the gradient descent method. Only two parameters ρ and ν have to be calibrated. This decision is made on the basis of market experience. a linear interpolation from the nearest neighbors is used. therefore it is important to try dierent starting points.ν (K. f ) = f (1  » – ff (1 − β )2 α2 1 ρβαν 2 − 3ρ2 2 α 1+ + + ν tex + O(t2 ex ) . As explained in section 1. f = Sertex . or an assumption.T )∈Ω ||σSABR (K. 2 − 2 β 1 − β 24 f 4f 24 − β) (2.3.18) There are many local minima of this function. a. The L2 norm of a vector a. The method used to solve the problem is LevenbergMarquardt (LMA). Same optimization procedure as above is used to calibrate the SVI model. see. A special case of the formula above is used when the strike and forward value are equal. The calibration procedure leads to a non-convex and non-linear optimization problem.ρ. The scalar β is determined either from a loglog-plot with historical data. Since β controls the distribution function. When extracting implied volatility surfaces from the CAP-market. β = 0 gives raise to a normally distributed change of the underlying. [11]. with elements ai 2 is i ai . i. The LMA is more robust than the Gauss-Newton. where S is the value today. T )||2 (K.. and β = 1 a lognormally distributed change of the underlying. T ) − σ ∗ (K. The SVI model has ve parameters and all of them have to be calibrated. When σAT M is not given. r is the rate and tex is the expiration time.b.g. β = 1 2 can also be used. e. an a priori view of the distribution function could be a way to set the β .σ. α is calculated whenever needed on the y by inserting values of ρ and ν . T ) − σ ∗ (K. a Cap is a sum of caplets and it is the caplet volatilities that 10 .m min ||σSV I (K. LMA tends to be a bit slower than GaussNewton. which means that in many cases it nds a solution even if it starts very far o the nal minimum. Another method is to relax α completely and then receive a least square t to the SABR model.17) where || · ||2 is the L2 norm. It is a useful algorithm for non-linear least square tting problems.e. It is called Caplet-stripping. [10].2. The idea is to choose α so that σAT M is xed at the level given from the market. min ρ. It is a third order polynomial equation.T )∈Ω (2. T )||2 (2.16) from this equation. σAT M = σBS (f.

Within each section there are two 6 month caplets. This function draws a straight line through the last and rst volatility on that section and calculates the sum of the caplet values which are evaluated with BS-formula. Newton's method is used to nd the exact volatility which will be consistent with the CAP-value. 11 . Consider this like two sections.2 Calibration are needed to generate the correct surface. then the 6 month caplets in-between can be extracted in the following way. 0 to 1 and 1 to 2. a at line is calculated.e. CAPLET0→ 1 + CAPLET 1 →1 = CAP0→1 . since there is no start-value in that section. Create a function dependent on the last implied volatility in the section. The LMA algorithm and the caplet stripping procedure were already implemented in Quantlab. the linearly interpolated caplet volatilities are extracted. CAP0→2 are given. For the rst section. i. if the values of the one year cap CAP0→1 and the two years cap. For example. therefore it was favorable to implement SABR and SVI in Quantlab. In this way. For the second section a linear interpolation is made according to the procedure.2 STOCHASTIC VOLATILITY 2. The caplet-stripper uses the Cap-values to implicitly calculate the corresponding caplet volatilities.CAP0→1 . The sum of the 2 2 two remaining caplets is CAP0→2 .

4) where ϕ(ST . who extended the volatility to be a state-dependent function of the price of the underlying and the time to expiration. is the local volatility function. S0 ) is the probability density of the nal spot at time T. T ) of a European option with strike K and expiration T is given by ∞ C (S0 . K. therefore Monte Carlo simulation has to be used. T ) = K dST ϕ(ST .1) whis is also called the general Black-Scholes dynamics. S0 )(ST − K ).3 LOCAL VOLATILITY 3 3. (3. when an option depends on an underlying several times during the life time of the contract.1 Local volatility Theory Local volatility is needed when pricing exotic options. S (3. K. K. consistent with the given European option prices. One of the developers of local volatility theory was Bruno Dupire. t)dWt . The main idea of nding the local volatility is to derive the risk-neutral density from market prices of European options. t) according to the equation dS = µt dt + σL (St .5) 12 . (3.2) where Dt is the dividend yield and C is the European option price C (S0 . The new dynamics becomes dS = rdt + σL (S. There is no closed form formula for these contracts. t)dW. Suppose the stock price diuses with a risk-neutral drift µt = rt − Dt and local volatility σL (S.3) The undiscounted risk-neutral value C (S0 . and the preferable volatility is the local volatility. T . S Proof from [4]. The unique local volatility function is the solution to Dupire's equation σ2 K 2 δ2C δC ∂C = L + (rt − Dt ) C − K 2 δT 2 δK δK (3. The corresponding unique statedependent diusion coecient σL (S. 2 2 ∂ST ∂ST ∂T (3. It evolves according to the Fokker-Planck equation 1 ∂2 ϕ ∂ϕ 2 2 σL ST ϕ − (µST ϕ) = . T . t). T ). The great breakthrough was when Dupire showed that under risk-neutrality there was a unique diusion process consistent with these risk-neutral distributions.

gives ∞ dx = [f g ]b a − g dx.12) The second term of (3. K (3.4). S0 ). dierentiating (3.14) 13 . the term in brackets vanishes.7) ∂2C = ϕ(K.8) ∂C = δT ∞ dST K 1 ∂2 ∂ 2 2 (µST ϕ) (ST − K ).12) nally gives σ2 K 2 ∂ 2C ∂C ∂C = L + µ( T ) C − K .11) In the same way as before.3 LOCAL VOLATILITY 3. T . T .9) by parts yields ∞ − K dST ∂ (µST ϕ) (ST − K ) = ∂ST ∞ ∞ − µST ϕ(ST − K ) K + K dST (µST ϕ). ∂K 2 Now. S0 ) (ST − K ). σL ST ϕ − 2 2 ∂ST ∂ST b a fg (3.12) may be written as ∞ dST µST ϕ = µ K K dST ϕ(ST − K ) + K K dST ϕ .5) in (3.1 Theory Dierentiating (3.7) to the rst term of (3.13) Notice that the rst term of the right hand side above is exactly the undiscounted option value from (3.8) By using (3. S0 ). (3. (3. Integrating the second term of (3. K (3.6) to the second term of (3. (3.10) 2 ∂ST L T Using that limK →∞ ST = 0 the term in brackets vanishes and the second 2 K2 term becomes σL2 ϕ. T . ∂T (3.9) by parts.4) with respect to K gives ∂C =− ∂K ∞ dST ϕ(ST . By using (3. 2 ∂T 2 ∂K ∂K (3.13) and (3.9) b a f Integrating the rst term of (3. This leads to the following equation σ2 K 2 ∂C = L ϕ+ ∂T 2 ∞ ∞ ∞ dST µST ϕ.4) with respect to time T gives ∂C = δT ∞ dST K ∂ ϕ(ST . dST K 1 ∂2 2 2 2 (σL ST ϕ) (ST − K ) = 2 ∂ST ∞ 1 ∂ 2 − (σL ST ϕ)(ST − K ) 2 ∂ST K ∞ dST K 1 ∂ (σ 2 S 2 ϕ).6) (3.

the smoothest. The a priori guess is of course crucial for the calibration and needs to be chosen carefully on the basis of market information and gut feeling. The entropy regularization generates the solution. One way to solve the Dupire equation is a method described in [7].3. that the theory assumes a well dened European option price space.t) for which market prices are given. i. Ill-posed problems like this one need to be reformulated for numerical treatment. (3. In this case the regularization is based on penalizing the discretized gradient. It uses natural cubic splines and Tikhonov regularization. such as an assumption of the smoothness or a bound on the norm. For the inverse problem above. That is why calibration is considered as an art by some authors [9]. This is a famous inverse problem in computational nance. whose corresponding probability distribution has the shortest entropy distance to an a priori probability distribution. the one closest to some a priori guess or some other. There is no consensus of which solution is the best. This is done by introducing some additional information of the solution. This process is called regularization. This is due to the fact.1 Calibration Local volatility by solving Dupire's equation Though the theory ensures a unique local volatility it is a non-trivial problem to recover it from real option data. The natural cubic spline is dened as the minimizer g ˆ of the optimization 14 . t)||2 2 (3. The idea is to interpolate the given option prices by a natural cubic spline per time slice.16) (3.17) where Cm is the given market prices. t) − C (S. Tikhonov regularization penalizes either the rst or the second derivatives of the surface. One shortcoming of this procedure is that one penalty factor has to be determined a priori. It is known for being very sensitive to noisy input data.15) subject to σ2 K 2 ∂ 2C ∂C ∂C = L + µ(T ) C − K ∂T 2 ∂K 2 ∂K C (S0 .2 3. the solution changes dramatically to small changes in the data.e. which is not the case on real markets. which of course makes the problem severely underdetermined. 0) = max(K − S0 . 0). for example the Frobenius norm of the hessian.2 Calibration 3 LOCAL VOLATILITY 3. Ωm is the set of pairs (S. This results in an ill-posed optimization problem min σL (S.2.t)∈Ωm ||Cm (S. As a matter of fact there are only a few dozens of option prices available. This reduces the noise and generates the smoothest solution to the problem above. K. Tikhonov regularization and entropy regularization are used by dierent authors.

2 Calibration problem below n min g i=1 ∗ ||yi − g (ui )||2 2+λ un g (v )dv. j ) ∈ {(i. n − 1 1 ri. n − 1 3 1 ri. ui the strike values and λ is a penalty parameter.23) is dened in the following way Q is a n × (n − 2)-matrix and hi = ui+1 − ui . The matrix In is the n × n identity matrix. n − 1 qi.i+1 = ri+1. . j )| |i − j | ≥ 2} the (n-2)-vector ξ is the corresponding second derivatives for the interior nodes of the natural cubic spline. . j ) ∈ {(i. Q (3. .18) where yi∗ are the given option prices.j = h− j −1 . .j = 0 ∀ (i. The optimization problem (3.22) B= In 0 0 λR .j = −h− j −1 − hj . j )| |i − j | ≥ 2} R is (n − 2) × (n − 2) dened by its elements ri. u1 (3. .j ∀ i. Therefore CK .i = hi ∀ i = 2.j = 0 ∀ (i. . 1 −1 qi. . . n − 2 6 ri. . y= y∗ 0 . . .19) (3. then 1 qj −1. . j = 2.21) A= Q −RT . It is convenient to use natural cubic splines. .18) may be written as a quadratic program 1 min −y T x + xT Bx x 2 subject to AT x = 0 (3. 1 qj. 15 .j +1 = h− j ∀ j = 2. (3. since they are known to be smooth.20) where x= g ξ . continuous and twice dierentiable everywhere. Note that the second derivatives at the boundary nodes are zero by denition.j = (hi−1 + hi ) ∀ i = 2. .3 LOCAL VOLATILITY 3. . (3. .

Therefore the local volatility surface can be obtained by solving a linear σL equation system. Euler forward and Euler backwards are used at the end-points. . the Dupire equation can be written like 2 σL = 2(CT + rKCK ) K 2 CKK (3. − τ1 m 1 τm     by increasing γ suciently. which could yield complex volatilities. The optimization problem above can be written as a linear equation system (D2 + γ LT L)z = Db L is dened as L= Im+1 ⊗ LK LT ⊗ In (3.25).. CT (Ki . By z is σL applying the regularization. hence this method is a lot faster than other methods.25) as its elements.. the problem becomes 2 min ||Dz − b||2 2 + γ ||Lz ||2 z (3. the local variance 2 . . Tj ) = τj τj +1 τj τj +1 1 gj +1 (Ki ) + ( − )gj (Ki ) − gj −1 (Ki ) τj + τj +1 τj +1 τj τj +1 τj (3.2 Calibration 3 LOCAL VOLATILITY CKK can be calculated by just dierentiating g ˆ.27) where γ > 0 is a regularization parameter and L is a discrete gradient operator. Unfortunately there is no guarantee that the solution is positive.3. m − 1. LT =    .29) and LK ∈ Rn−1.26) where D is a diagonal matrix with the denominator of (3.28) (3. The rst key approach was to solve Dz ≈ b (3.. −1 1 1      . CT is obtained using a derivative approximation. a positive solution is guaranteed.. .. and with zero dividend. For discounted option prices. there is only one unknown.. 16 .m+1  LK   =  −1 1 −1   − τ1 1 1 τ1 − τ1 2  . 1 τ2 . . . 2 andb is the corresponding vector of the enumerator of (3. j = 2. A priori guesses has to be used when choosing the parameters λ for the splines and γ for the regularization.n .24) with τj = Tj − Tj −1 . Therefore the problem needs to be regularized. which will be the case in the test later on.25) Having all these quantities... LT ∈ Rm.

This guess is used in the equation above to obtain the next guess of the most likely path. The procedure is repeated until it converges.34) The procedure continues until the change between two iterations is suciently small. t) = σmarket (S. When having the most likely path. dSt = µt dt + σ (E [St |ST = K ])dWt = µt dt + σ∗ (t)dWt St (3. 2) Calculate the new local volatility. It is based on a concept called most likely path rst found by Jim Gatheral [4]. according to [9]. where the initial guess is set to the forward price path. t) (3. i+1 σBS (S.33) ¯t is the most likely path. t). The procedure to nd the most likely path is itself an iterative process. This calculation is complicated when the normal dynamics dSt St = µt dt + σ (t.2.T = t 2 0 σ (s)ds T 2 0 σ (s)ds (3. the i +1th implied volatility surface can be calculated by i+1 σBS = 1 T T 0 i ¯ σL (St .32) is a regression coecient between 0 and 1.T K αt. there is a more robust way of nding the local volatility. where each iteration consists of two steps: 1) Find the most likely path. Now the i + 1th local volatility surface is where S obtained according to the following point-wise adjustment. t)dt.31) where Ft = S0 e Rt 0 µs ds is the forward value and αt. St )dWt is used. t) i σL (S. He shows that the implied variance is well approximated by the time integral from t = 0 to t = T of the local variance along the most likely price path E [St |ST = K ].30) There is a closed form formula to calculate the conditional expected value.2 Calibration 3.3 LOCAL VOLATILITY 3. 17 . then the nal path is the most likely path.2 Local volatility by most likely path calibration Due to the numerical diculties of solving Dupire's equation. Therefore a simpler version is introduced in [9]. The calibration is an iterative process. i+1 σL (S.T 1 RT t σ 2 (s)ds (3. (3. E [St |ST = K ] = Ft FT −αt.T e 2 αt.

38) ∆S ∆t H (˜ σL (S. min σ ˜L (S.2 Calibration 3 LOCAL VOLATILITY As will be seen in the test section 5. tj −1 ) ∂St σ ˜L = . the gradient of the objective function is calculated. The smooth surface is the solution to the following optimization problem. Since that is not expected. The operator H gives 2σ ˜L ∂tS σ ˜L ∂S . Tikhonov regularization with a discrete Hessian is used to eliminate small noise. an additional smoothing process is used. tj ) − σ ˜L (Si−1. t)||2 σL (S. Solving the rst necessary optimality equation for a quadratic program ∇f = 0 is equivalent to solving a linear equation system. Ω is the set of all states and times. t) = Of course these approximations are slightly altered at the boundary points. tj ) − 2˜ σL (Si . the generated surface tends to be very noisy. tj ) 2 ∂S σ ˜L = .37) (∆S )2 σ ˜L (Si + 1. which is solved quickly and yields the global optimal solution.36) 2 ˜L ∂St σ ˜ L ∂t σ σ ˜L (Si+1 . To speed up the optimization. t)||2 2 + ζ ||H (˜ F. The Frobe2 nius norm of a matrix A with elements aij is i.35) where σL is the surface obtained from the most likely path calibration.tj ) − σ ˜L (Si . || · ||F is the Frobenius norm. t) − σL (S. H is an operator calculating an approximation of the Hessian in (S.t) of σ ˜L . (3.j aij . tj ) + σ ˜L (Si−1 . (3. tj +1 ) + σ ˜L (Si .t)∈Ω ||σ ˜L (S. (3.3. (3. 18 .

both SVI and SABR perform very well. Both of them are almost identical and coincide almost perfectly with each data point. 4. since each calibration uses 10 randomly chosen initial points. It is one stochastic volatility model. The dots are the stripped volatilities The gure 2 shows a case were there is a noticeable dierence between the data points and the SABR t. The calibration is done for all caps each day in March 2008. the SVI takes between 1 to 20 seconds to calibrate. SABR and one parametric model SVI. 24/3-2008. Compare to the t in gure 4 with a relaxed α a few pictures below.1 Test of the implied volatility models Procedure Two methods of the ones mentioned in section 2 are tested to model the implied volatility on the Eurcap market in March 2008. The test is to investigate how stable they are. SABR is calibrated according to the article.5 year volatility slice. 19 . Figure 1: SVI and SABR ts to the 1.2 Results As can be seen. Performance-wise the SABR model is calibrated almost instantly. The goal is to nd a method which can be running without any supervision.4 TEST OF THE IMPLIED VOLATILITY MODELS 4 4.

24/3-2008. SABR and SVI will go in dierent directions. 24/3-2008. The dots are the stripped volatilities Figure 3 shows an example where SABR and SVI diers from each other. 20 . Figure 3: SVI and SABR ts to the 25 year volatility slice. SABR is calibrated according to the article. but it is interesting to notice that they may dier.4. The t is very good.2 Results 4 TEST OF THE IMPLIED VOLATILITY MODELS Figure 2: SVI and SABR ts to the 8 year volatility slice. Both curves are pretty close to the data points but when extrapolating these curves. It is far from clear which model to choose and luckily it is not the job of the author to choose. The dots are the stripped volatilities Figure 4 shows the calibration on the same data as Figure 1. SABR is calibrated according to the article. but now the SABR model is calibrated with a relaxed α.

2 Results Figure 4: SVI and SABR ts to the 1. The t is closer to the data than when α is chosen according to the article.5 year volatility slice. 24/3-2008. SABR is calibrated with relaxed α. The dots are the stripped volatilities Figure 5 shows another t which also is very good. SABR is calibrated with relaxed α. The dots are the stripped volatilities. Figure 5: SVI and SABR ts to the 8 year volatility slice.4 TEST OF THE IMPLIED VOLATILITY MODELS 4. 24/3-2008. 21 .

For the Dupire calibration. The scalar N. 22 . we consider that the local volatility is correct. the splines for each time slice are evaluated for values S · [0. and the data is displayed in table 1. The market implied volatility surface is shown in gure 6.85 : 0. the rst calculation is done on a grid S · [0. Below is the original implied volatility surface from the S&P500 and a table with the data. represents the number of simulations. If the prices can be regenerated within a 95 % condence interval. The monte carlo simulation consists of simulating Sn (ti+1 ) = Sn (ti ) + r∆ti Sn (ti ) + σL (Sn .2 : 5]. (5. which in this test is set to 10 000.05 : 5] using Matlab's griddata. n=1 (5.85 : 0. The option price is obtained from S (Kj . These two surfaces are then linearly interpolated to a ner grid S · [0.06.005 : 1.1 Test of the local volatility models Procedure This test is done on the European S&P 500 index European call options of October 1995.4] × [0.025 : 1. For the likely path calibration.005 : 1. 1).4]×[0. Sn (ti ) − Kj }.1) where n is the index of simulation and i is the index of time.85 : 0.2) An approximate 95 % condence interval is calculated as well. ti ) = e−rti N N max{0.4]. The real market prices are obtained by Black-Scholes formula from the given implied volatilities. The reason to this is that Monte Carlo simulation will be used to verify that the option prices can be regenerated from the local volatility surface obtained from the calibration.2 : 0.5 TEST OF THE LOCAL VOLATILITY MODELS 5 5. where the present value is S=$590. and interest rate r=0. ti )Sn (ti ) ∆ti N (0.2 : 0.

124 0.119 0.149 0.127 0.171 0.107 0.940 1.175 0.5 TEST OF THE LOCAL VOLATILITY MODELS 5.200 0.153 0.114 0.171 0.149 0.168 0.124 0.151 0.115 0.169 0.110 0.142 0.143 0. 0.123 0.161 0.104 0.000 3.119 0. 23 .113 0.000 5.136 1.18 0.100 0.128 0.130 0.159 1.160 0.137 0.108 0.159 0.100 0.128 0.300 0.138 0.1 5 4 0.161 0.157 0.14 0.400 0.142 0.2 0.103 0.133 0.108 0.145 0.148 1.137 0.103 0.100 0.150 0.169 0.102 0.111 0.850 0.097 0.143 0.154 1.168 0.168 0.120 0.172 0.900 0.159 0.190 0.139 0.150 0.137 0.132 Table 1: Implied volatility data from S&P500 index European call options of October 1995.177 0.126 0.118 0.425 0.109 0.151 1.08 850 3 800 750 2 700 650 600 1 550 500 0 Expiration time Strike price Figure 6: Implied volatility from European S&P500 index European call options of October 1995.169 0.133 0.135 0.130 0.200 0.150 0.106 0.144 0.124 0.155 0.16 Implied volatility 0.1 Procedure 0.107 0.695 0.000 1.125 0.128 0.157 0.950 0.168 0.138 0.115 0.000 4.162 0.164 0.155 0.113 0.133 0.099 0.103 0.500 2.050 0.101 0.100 0.133 0.12 0.130 0.148 0.000 S T\S 0 0.140 1.000 0.113 0.102 0.152 0.144 1.

but for short expirations it is probably not correct.2 Results 5 TEST OF THE LOCAL VOLATILITY MODELS 5. Table 2 displays interpolated market prices for a few strike prices and expiration times. Compare the values of the two tables.28 0.08 850 2 800 750 700 1 650 600 550 500 0 5 Expiration time Strike price Figure 7: The linearly interpolated local volatility surface.24 0.2.12 0.2 Results 5. they nally were set to 10000 and 500000 respectively. Table 3 displays the corresponding Monte Carlo prices from the Dupire calibrated volatility surface.14 0. since they seem to generate a plausible surface. The surface seem to be very good at expiration times longer than a few months.18 0. Values in parenthesis are out side the 95 % condence interval. Figure 8 is a way of displaying which values are within the 95% condence interval. 24 . By investigating dierent values of the parameters λ and α.2.5. 0.1.26 0.1 The regularized Dupire equation Figure 7 shows the local volatility surface obtained by the calibration used in section 3.2 Local volatility 0.22 0.1 4 3 0. The slope at the right side of the gure is probably not correct and may depend on numerical diculties. since the denominator of Dupire's equation for low strike prices is very small. A dot means that the corresponding value is signicant.16 0.

803 128.213 66.701 1.630 43.600 90.25 2.75 1.324 96.25 1.580 64.27 0.131 159.825 153.983 171.253 8.25 4.460 0.566 68.785 52.836 174.671 1.112 47.21 0.629 139.778 104.110 99.328 30.75 3.25 3.489 157.774 85.696 145.117 41.39 0.793 119.211 13.554 13.97 28.806 1.461 82.919 32.273 209.552 54.706 80. S T\S 0 0.351 54.738 40.33 0.780 43.156 134.794 27.85 94.056 71.15 6.518 0.461 27.871 15.435 54.749 115.301 185.667 184.849 22.002 0.008 0.935 4.75 2.712 1.193 143.885 Table 2: Market prices in $ 25 .830 54.089 56.75 4.520 98.804 111.998 29.75 0.084 71.952 0.09 1.928 42.91 60.366 1.027 0.082 197.147 67.495 75.304 130.679 42.25 0.098 0.239 126.775 80.051 0.003 0.728 111.260 2.883 0.982 1.085 133.280 80.906 6.2 Results 0 10 20 30 Expiration time index 40 50 60 70 80 90 0 20 40 60 Strike index 80 100 Figure 8: Dots show values within the approximate 95 % condence interval.500 88.729 197.620 17.085 28.349 95.781 3.309 220.984 120.874 5.229 20.562 148.010 113.043 172.940 59.232 11.080 104.5 TEST OF THE LOCAL VOLATILITY MODELS 5.425 161.03 1.095 10.803 27.

909 70.25 3.047 63.75 0.039 0.924 53.75 3. Figure 10 shows the signicance of the most likely path calibrated surface. As mentioned before.864 22.088 98.2 Results 5 TEST OF THE LOCAL VOLATILITY MODELS S T\S 0 0. Figure 9: The linearly interpolated local volatility surface.782) (0.402 13.27 1.756 14.832 5.343 219.09 1.2 The most likely path calibration Figure 9 shows the surface obtained from the most likely path calibration.701 59.130) (7.854 54.970 160.642) (28.2.781) 0.770 42.050 103.172 65.309 79.91 0.429 138.39 94.962) (0. 26 .312 (59.000) (0.390 158.144) 171.571 8.703 79.837 125. it is very noisy.908 134.373 111. Values in parenthesis are not signicant.041) (0.839 196.15 1.351 114.844 147.111 11.455 41.932) (3. Because of the noise.097 66.220 (0.788 27.194 104.463 94.221 56.929 18.002 117.017) 126.426 1.405 4.565 20.000) (0.21 1.058) (0.490 80.241 151.334) 157.25 1.974 43.963 97.000) (0.856 70.818 27.105 15.33 1. 5.505 111.884 84.75 2. It takes about 30 seconds in matlab to generate the surface.528 (7.450 53.111) 27.441 (2.5.992 54.492 88.577 75.25 0.960 43.161 120.986 196.03 1.97 1.603 (51.327 184.010 81.316 27.951 0.25 4.75 1.189 32.301 29. This is due to that the iterative process is point-wise and therefore does not take its neighbors into account.282 183.194 130.317 173.600 41.000) 110. a smoothing process was tested.639 41.827 (11.225 30.648 171.764 46.25 2.474 142.103 96.732 67.564 Table 3: Monte carlo prices in $ evaluated with the Dupire calibrated surface.317 145.85 0.222 89.976 2.850 208.75 4.190 130.

633 (42. Then the surface becomes like gure 11.980 68.000) (0.824 134.25 1.15 1.292 152.505 28.24 4.503) (0.364) 23.644 105. Since the rst surface was pretty noisy.029 100.586 33.109 (12.724 170.97 28.314 51.029) (0.27 1.357) 0.714) (21.830) 1.673) (14. 0.880 57.794) (0.693 72.594 (17.85 94.502) (30. 27 .678 0.728 132.701 139.158) (4.003 127.828 142.969 183.094) (9. 0 10 20 30 Expiration time index 40 50 60 70 80 90 0 20 40 60 Strike index 80 100 Figure 10: Dots show values within the approximated 95% condence interval.051 208.555 43.346 184.2 Results This surface is not as good as the Dupire-calibrated surface.900 111.192 56.749 75. according to (3.373 44.35) in the calibration section 3.282 81.381 83.000) (0.647 60.776 114.236 97.1.000) 28.192) (0.293 119.565 S T\S 0 0.331 171.636 98.21 1.463) (12.2.25 3.614 144.806) (2.5 TEST OF THE LOCAL VOLATILITY MODELS 5.161 79.881 125.25 0.939) (32.037 47.458 64.241) (7.899 55.008 1.383 105.437 55.071 86.25 2. Values in parenthesis are not signicant.441) (15.063) (1.110 0.662 44.192 88.159) 81.445 90.674 158.002) (0.09 1.406 114.984) (2.094 0.014) (6.420 69. The values for low strikes are good but for high strikes a lot of values are not signicant.03 1.824 161.872 157.75 Table 4: Monte carlo prices in $ evaluated with the most likely path calibrated surface.75 2.478) (4.75 1.671) (19.866 43. It is much smoother but still has the overall shape of the original one.131 147.314) 96.393 56.047 174.91 60.685 130.75 3.859) 111.005.75 4.33 1.056 196.652 196.409 0.066 219.133 119.409 (29.39 (7.488 (28.262 71.468 68.090 (42. a Tikhonov smoother was applied with ζ = 0.

18 Local volatility 0. it unfortunately does not give the correct prices. Figure 12 shows the signicance of the smoothed surface. Values greater than 0 is within the approximated 95 % condence interval. 0 10 20 30 Expiration time index 40 50 60 70 80 90 0 20 40 60 Strike index 80 100 Figure 12: Dierence between the condence bounds and the market price.2 Results 5 TEST OF THE LOCAL VOLATILITY MODELS 0.14 0.22 0.5.08 850 800 1 750 700 650 600 550 0 500 Expiration time Strike price Figure 11: The linearly interpolated local volatility surface. 28 . The surface seem equally good as the unsmoothed surface.2 0.1 3 2 0.12 5 4 0.16 0. Even though the local volatility surface looks smoother.

868) (1.15 1.33 1.330 4.278 11.244 55.349) (3.25 0.000) 110.034 (126. 29 .2 Results S T\S 0 0.25 1.25 3.141) 28.644 74.756 71.456 20.189) (1.277 42.380 198.038 56.054) (0.348 104.446 16.844 161.982 59.781 100.221 133.5 TEST OF THE LOCAL VOLATILITY MODELS 5.148 28.647 171.700) 103.85 0.003) (0.410) 0.018 44.434 143.029 31.029) (0.700) (83.829 (98.499 157.937 182.208 175.863 140.511 169.000) (0.591 41.711 (79.857 9.385 44.339 29.715) (128.599) (4.451 115.668) 97.39 94. Values in parenthesis are not signicant.75 4.562 13.145 110.994 84.505) (57.426 23.119 0.655 208.267 28.424 18.269 79.266 (12.27 1.75 0.022 69.840 154.831) 28.503 134.124 88.75 3.126 195.251) (7.91 0.372) (15.112 69.625 95.802 56.024 64.981 221.75 1.737 184.000) (0.458) Table 5: Monte carlo prices in $ evaluated with the smoothed most likely path calibrated surface.886 2.691 55.440 70.057 147.490 43.117 119.505 33.03 1.25 2.538) 142.471) (1.21 1.97 1.102 114.75 2.367 118.368) (51.763 86.416 42.265) (69.526 89.314 (7.003 (28.425 47.499 (59.530 125.25 4.451 (155.319) (0.09 1.401) (0.187 6.

For each restart. To further give a measure of how good the surfaces are. The second goal was to investigate whether a fairly fast and stable method could extract a local volatility surface from quoted European option prices or not. Therefore it was necessary to use random initial points. the calibration time increases. It is actually often the case that banks in Stockholm rather use stochastic volatility to price exotic options than local volatility. Points for greater expiration time index than 40 are not displayed. This is probably due to that the SABR model just has a few parameters. especially when the α is chosen according to the article. The same initial point was used on all test data and a solution was always found. with the parameters obtained from the calibration. but in the worst case it could be 20 seconds. they both perform well. The SVI was slightly more dicult to calibrate. the SABR model might be preferable. but both models could be interesting to use and as was said earlier. which were mentioned earlier. With 10 restarts. in a Monte Carlo pricer. but requires a lot of tweaking to set the parameters right. But from an optimization point of view it was more interesting to work with the local volatility. the calculation time was about a few seconds. since some initial points did not lead to a solution. The SABR model was really stable and fast. The short answer is. The other one is fairly quick to calibrate and does not need any initial guesses but it is doubtful whether the obtained surface is signicant or not. However it is no guarantee that it will always nd a solution with exactly 10 restarts. are a lot more accurate than when using the implied volatility surface. the original implied volatility surface was used as the local volatility in the Monte Carlo pricer. When using 10 initial points it was enough to nd a solution for every day and option from the test data set. The test has shown that both the SABR and SVI serve as good models. Figure 13 shows the signicance of using implied volatility as local volatility.6 DISCUSSION 6 Discussion Our rst goal was to nd a couple of models which in a fast and stable manner can give an implied volatility surface. especially Heston and Bates model. and just two parameters are calibrated. This shows that the calibrated surfaces obtained before. 30 . If one model was to be chosen. Notice that almost all points are insignicant. probably not. A minor point which not has been discussed earlier in the thesis is that the SABR model can be used to price exotic options as well by using the SABR dynamics. so that number of restarts was chosen as a compromise of stability and calculation time. but maybe not good enough. One of them is fast and could generate a plausible surface. since they are all insignicant.

18 Local volatility 0.2 0. This is due to that the second derivative. 0. Perhaps a pde-solver based on nite elements.16 0.12 0.22 0. To illustrate the problem of setting the right parameters. locally is relatively small.14 0. 31 .08 5 4 3 2 1 0 850 800 750 700 650 600 550 500 Expiration time Strike price Figure 14: The local volatility when λ changes from 10 000 to 1 000. look what happens if λ decreases with 0. the variations are treated. By punishing the splines suciently.1 %. which yields a smoother surface. and therefore the denominator of (3.25). It seems that the local volatility is dicult to extract and probably a more rigorous method is needed.26 0. There is not yet a strategy to automatically set these parameters in advance.1 0. Two humps appear in the middle of the surface.6 DISCUSSION 0 5 10 Expiration time index 15 20 25 30 35 40 0 20 40 Strike index 60 80 100 Figure 13: Dots show values within the approximated 95 % condence interval.24 0.

Computational Methods in Finance. The Quarterly of Applied Mathematics 2: 164-168. The volatility surface. & Scholes. P.. S. P. John Wiley & Sons Inc. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bonds and Currency Options. Global Derivatives & Risk Management 2004 [6] Hagan. & Pender. John Wiley & Sons. J. L.REFERENCES REFERENCES References [1] Bates. Fachbereich Mathematik. Johannes Gutenberg-Universität. 1993 [9] Hirsa. 2005 [4] Gatheral. 9: 69-107. Volatility and Correlation. 3: 637-654. & Rösler. Humboldt-Universität zu Berlin.. 1963 [12] Rebonato. S. "The Pricing of Options and Corporate Liabilities. 2006 [10] Levenberg. The Review of Financial Studies. An Algorithm for Least-Squares Estimation of Nonlinear Parameters. F. D. Jump and Stochastic Volatility: Exchange Rate Processes Implicit in Deutsche Mark Options. Inc. SIAM Journal on Applied Mathematics 11:431-441. The Review of Financial Studies. Journal of Political Economy 81. 1999 32 . M. M. Local Volatility Calibration using the Most Likely Path. 2004 [8] Heston. R. 2006 [5] Gatheral. 1973 [3] Fengler. A Method for the Solution of Certain NonLinear Problems in Least Squares. E. A parsimonious arbitrage-free implied volatility parameterization with application to the valuation of volatility derivatives. 2002 [7] Hanke. A Practitioner's Guide. M. Computation of Local Volatilities from Regularized Dupire Equations. A. K. 1996 [2] Black. Managing smile risk. Wilmott Magazine. D. 1944 [11] Marquardt. J. Arbitrage-Free Smoothing of the Implied Volatility Surface.