On the Rate of Convergence of Binomial Greeks San-Lin Chung, Weifeng Hung and Pai-Ta Shih*** Abstract The recent literature indicates that the most efficient binomial models, including binomial Black-Scholes (BBS) model of Broadie and Detemple (1996), the flexible binomial model (FB) of Tian (1999), the smoothed payoff (SPF) approach of Heston and Zhou (2000), and the generalized Cox-Ross-Rubinstein (GCRR) model of Chung and Shih (2007), generally yield binomial option prices with monotonic and smooth convergence. However, their efficiency for the calculation of delta and gamma are not known. To fill the gap of the literature, we investigate the convergence patterns and the rates of convergence of these binomial models for calculating delta and gamma. The method proposed by Hull (1993, hereafter Hull method) and extended binomial tree proposed by Pelsser and Vorst (1994, hereafter extended method) are applied to these binomial models. The numerical results indicate that by using the Hull method or the extended method, these binomial models can generate monotonic convergence with the rate of convergence O(1/ n) for calculating delta and gamma. Moreover, the GCRR-XPC model with an extended tree is the most efficient model to compute deltas and gammas for options when a two-point extrapolation formula is used. Keywords: binomial option pricing model, Greeks, monotonic convergence, rate of convergence JEL Classification: G13 Corresponding author, Department of Finance, National Taiwan University, 85, Section 4, Roosevelt Road, Taipei 106, Taiwan, R.O.C.. Tel: 886-2-3366-1084. Email: chungs@management.ntu.edu.tw. Department of Finance, Feng-Chia University. 100, Wenhwa Rd., Seatwen, Taichung, Taiwan 40724, Taiwan, R.O.C.. Tel: 886-04-24517250#4173. Email: wfhung@fcu.edu.tw. *** Department of Finance, National Taiwan University, 85, Section 4, Roosevelt Road, Taipei 106, Taiwan, R.O.C.. Tel: 886-2-3366-1093. Email: ptshih@management.ntu.edu.tw. Introduction Binomial methods, developed first by Cox, Ross and Rubinstein (1979, CRR thereafter), are well-known for their flexibility and efficiency in calculating option price. One stream of the literature modifies the lattice or tree type to improve the accuracy and computational efficiency of binomial option prices. The pricing errors in the discrete-time models are mainly due to “distribution error” and “nonlinearity error” (see Figlewski and Gao (1999) for thorough discussions). Within the literature, there are many proposed solutions that reduce the distribution error and/or nonlinearity error. For example, Broadie and Detemple (1996) and Heston and Zhou (2000) suggested modified binomial models by replacing the binomial prices prior to the end of the tree by the Black-Scholes values, or by smoothing payoff functions at maturity, and computing the rest of binomial prices as usual. The advantage of their approach is that the resulting binomial option prices converge to the Black-Scholes prices monotonically and smoothly, and thus the standard Richardson extrapolation can be applied to obtain solutions with high accuracy. The other improved lattice approaches including Omberg (1988), Leisen and Reimer (1996), Figlewski and Gao (1999), Widdicks, Andricopoulos, and Newton, and Duck (2002).1 In this paper, we focus on recent binomial models whose binomial option prices converge to the “true” value monotonically and smoothly. In other words, the pricing 1 Omberg (1988) developed a family of efficient multinomial models by applying the highly efficient Gauss-Hermite quadrature to the integration problem (e.g. N (d1 ) in the Black-Scholes formula) presented in the option pricing formulae. Leisen and Reimer (1996) modified the sizes of up- and down-movements by applying various normal approximations (e.g. the Camp-Paulson inversion formula) to the binomial distribution derived in the mathematical literature. Figlewski and Gao (1999) proposed the so-called adaptive mesh method which sharply reduces nonlinearity error by adding one or more small sections of fine high-resolution lattice onto a tree with coarser time and price steps. 1 errors of these binomial option prices are of the same sign and decrease at a known rate as the number of time steps increases.2 These models include the binomial Black-Scholes (BBS) model of Broadie and Detemple (1996), the flexible binomial model (FB) of Tian (1999), the smoothed payoff (SPF) approach of Heston and Zhou (2000), and the GCRR-XPC model of Chung and Shih (2007).3 Although the above models have been widely applied to price options, their convergent patterns and rate of convergence for calculating the hedge ratios are not well known. To fill this gap, we apply the method proposed by Hull (1993) and extended method proposed by Pelsser and Vorst (1994) to calculate Greeks under these models. We find that by using Hull method and extended method, these binomial models can generate monotonic convergence with O 1n rate for calculating deltas and gammas.4 Thus one can apply the extrapolation formula to enhance the computation of the hedge ratios. Among all the binomial models considered in this paper, the GCRR-XPC model is the most efficient one for the evaluation of deltas and gammas for European options when a two-point extrapolation is used. The rest of the paper is organized as follows. Section 2 briefly reviews the binomial models considered in this paper. The rates of convergence of applying 2 Binomial models with the smooth and monotonic convergence property are the most accurate ones in the recent literature because their accuracy can be substantially improved by applying the Richardson extrapolation technique (Chang, Chung, and Stapleton, 2007). 3 In the GCRR-XPC model, a stretch parameter ( ) is used to adjust the lattice so that the strike price is located in the center of the final nodes. See Section 2.1 for details of the GCRR-XPC model. 4 Numerical differentiation formulae can also be applied to these binomial models for the calculations of Greeks. However, by using numerical differentiation formulae, it needs to construct two binomial trees to calculate delta and to construct three binomial trees to calculate gamma. Thus, it takes more time to calculate Greeks. Besides, the numerical results show that the convergent patterns of binomial Greeks using numerical differentiation formulae are not smooth and monotonic. Thus, we do not apply numerical differentiation formulae in this paper. 2 extended binomial tree to CRR model, SPF model and GCRR-XPC model are proved and the extrapolation formula is discussed in Section 3. Section 4 presents the numerical results of various binomial models for the evaluations of deltas and gammas of options. Section 5 concludes the paper. 2. Reviews of the binomial option pricing models We assume that the Black-Scholes economy holds and thus options can be valued as if the investor is risk-neutral. In other words, the options can be priced under the risk-neutral measure where stock price follows a geometric Brownian motion given by dS (r q ) dt dZ , S (1) where S is the stock price, r is the risk-free rate, q is the dividend yield, and is the instantaneous volatility of S . The binomial option pricing model was first developed by Cox, Ross, and Rubinstein (1979) and Rendleman and Bartter (1979). Consider the pricing of an option maturing at time T . In an n-period binomial model, the time to maturity [0, T ] is partitioned into n equal time steps t T / n . If the stock price is S in this period, then it is assumed to jump either upward to uS with probability p or downward to dS with probability 1 p in the next period, where 0 d e( r q ) t u and 0 p 1 . The binomial model is completely determined by the jump sizes u and d and the risk-neutral probability p. 2.1. The binomial models with monotonic and smooth convergence property In the traditional CRR model, the following three conditions are utilized to determine u, d, and p. 3 pSu (1 p) Sd Se( r q ) t , pu 2 (1 p)d 2 pu (1 p)d 2 t , 2 (2) ud 1. The first two conditions are used to match the mean and variance of the stock price in the next period, and the third condition is imposed arbitrarily by CRR. With these three conditions, one can easily determine the binomial parameters as follows: u e t , d e t , and p (e( r q ) t d ) /(u d ) . According to the Central Limit Theorem,5 the discrete distribution of the asset price under the CRR model will converge to its continuous-time limit (i.e. Black-Scholes model). In other words, as t 0 , the price distribution of the CRR model converges to a lognormal distribution, 2 ln ST N ln S0 (r q )T , T . 2 d (3) As a result, the option prices calculated by the binomial model will also converge to the Black-Scholes price. Although the CRR model is a well known and widely used model for valuing options, the option prices calculated from it usually converge to the Black-Scholes price in a wavy erratic fashion. In order to enhance the accuracy of the binomial option prices, the recent literature has developed several ways to overcome the wavy erratic problem embedded in the CRR model. Four important articles are discussed and applied in this paper. These papers developed binomial models with monotonic and smooth convergence property. Monotonic convergence is desirable because more time steps do guarantee more accurate prices. Moreover, smooth convergence is also 5 The following condition should be satisfied to apply the Central Limit Theorem: p u e( r q ) t 1 p d e( r q ) t 3 3 [ p(u e( r q ) t )2 1 p (d e( r q ) t )2 ]1.5 n 0 as t T / n 0 . 4 advantageous because the standard Richardson extrapolation can be used to enhance the accuracy. Broadie and Detemple (1996) proposed a binomial model termed Binomial Black and Scholes (BBS) model to price options. The BBS model is identical to the CRR model except that at one time step before option maturity the Black-Scholes formula replaces the usual “continuation value.” Broadie and Detemple (1996) showed that the option prices obtained from the BBS model converge to the Black-Scholes formula smoothly and monotonically. Thus, they suggested using the Richardson extrapolation to enhance the accuracy. Tian (1999) developed a flexible binomial (FB) model with a “tilt” parameter ( ) that alters the shape and span of the binomial tree. The parameters of the FB binomial model are as follows: t 2 t ue , d e t 2 t e( r q ) t d . , p ud (4) With a positive tilt parameter, > 0, the up movement is larger than the corresponding up move in a standard CRR tree. Consequently, the central nodes depict an upward sloping line. The resulting tree span is thus shifted upward. The exact opposite is true for binomial trees with a negative tilt parameter, < 0. Fig. 1 illustrates the effect of the “tilt” parameter on binomial trees. <Insert Fig. 1. Here> Tian (1999) first proved that his FB model converges to its continuous-time counterparts (i.e. the Black-Scholes model) for any value of the tilt parameter.6 Tian 6 However, the tilt parameter must satisfy the following inequality in order to have “nonnegative probability”: ( r q) 2 1 t . 5 (1999) then suggested the way to select a particular tilt parameter that enhances the rate of convergence of binomial prices to their continuous-time limit. This is done by selecting a tilt parameter such that a node in the tree coincides exactly with the strike price at the maturity of the option. This idea was inspired by Leisen and Reimer (1996). 7 As a result, one can derive the formula for the tilt parameter as the following: ln( SX0 ) (2 j0 n) t n 2 t , (5) ln( SX ) n ln(d0 ) 0 where j0 , [.] denotes the closest integer to its argument, X is the ln( ud00 ) strike price, u0 e t , and d 0 e t . Using the FB model with the above tilt parameter, Tian (1999) showed that monotonic and smooth convergence is possible for pricing standard European and American options, and extrapolation methods are used to enhance the accuracy. Heston and Zhou (2000) showed that the accuracy or rate of convergence of the binomial model depend crucially on the smoothness of the payoff function. They have developed an approach to smooth the payoff function. Intuitively, if the payoff function at singular points can be smoothed, the binomial recursion might be more accurate. They defined the smoothed payoff function G x as follows: G x 1 2 x x x g x y dy , (6) where g x is the payoff function. This is a rectangular smoothing of g x . The 7 Leisen (1996) independently proposed a binomial model that has some similar features as the flexible binomial model. See his SMO model. However, his choice of u and d is rather ad hoc. A systematic way of determining u and d is proposed here through the choice λ. 6 smoothed function, G x , can be easily computed analytically for most payoff functions used in practice. For example, the smoothed payoff function of a European put option with a strike price X can be derived as follows: 0, e t e t , G ( ST )= X ST 2 t X t ln ST X ST e 2 t 2 t ln ST ln X t ln ST ln X t t (7) X , t ln ST ln X t ST Applying the smoothed payoff function G x , instead of g x , to the binomial model yields a rather surprising and interesting result. The associated SPF model converge at the O 1n rate to its continuous-time limit and this convergence is smooth and monotonic as the number of time steps increases. Chung and Shih (2007) proposed a generalized CRR (GCRR) model by adding a stretch parameter into the model specification. In the GCRR model the jump sizes and probability of going up are as follows: u e t , d e 1 t , and p e( r q ) t d , ud (8) where 0 is a stretch parameter which determines the shape (or spanning) of the binomial tree. The CRR model is obviously a special case of our GCRR model with 1 . When t 0 , i.e., the number of time steps n grows to infinity, the GCRR binomial option prices also converge to the Black-Scholes formulae for plain vanilla European options. Generally speaking, the rate of convergence of the GCRR model for pricing 7 options is of order O 1 n when 1 .8 This rate of convergence is inferior to other binomial models such as the BBS model and the FB model. Thus, to best apply the GCRR model, Chung and Shih (2007) suggested using the so-called GCRR-XPC model, where the lattice is set up in a way that the strike price is in the center of the final nodes, to price options. In particular, the binomial option prices of the GCRR-XPC model not only converge at a high order (of order O 1n ) but also converge smoothly and monotonically to the Black-Scholes formulae. Thus in the numerical analysis of this paper, we adopt the GCRR-XPC model to calculate hedge ratios. 2.2. Calculating hedge ratios with binomial models Greeks (or hedge ratios) are the sensitivity of the option price with respect to the change of the underlying risk factors. For the Hull method, the estimated delta is calculated at time t T / n and the estimated gamma is calculated at time 2t in the standard tree where the current underlying asset price and option price are S0,0 and C0,0 , respectively. In Fig. 2, the estimates of delta and gamma can now be obtained as follows: C C1,0 ˆ 1,1 , S1,1 S1,0 C2,2 C2,1 C2,1 C2,0 2 C S2,2 S2,1 S2,1 S2,0 ˆ 2 . 1 S S S 2,2 2,0 2 (9) (10) <Insert Fig. 2. Here> Instead of using the Hull method, we also apply the extended method to calculate delta and gamma. In the extended binomial tree, we build the binomial tree 8 Please refer to Theorem 2 of Chung and Shih (2007). 8 starting from two time steps back from today. Fig. 3 illustrates the extended binomial tree, where the current underlying asset price and option price are S0,1 and C0,1 , respectively. The estimates of delta and gamma can now be obtained as the following: C C0,0 ˆ 0,2 , S0,2 S0,0 (11) C0,2 C0,1 C0,1 C0,0 2 C S0,2 S0,1 S0,1 S0,0 . ˆ 2 1 S S0,2 S0,0 2 (12) <Insert Fig. 3. Here> 3. The rate of convergence of the binomial models and the extrapolation formula When the convergence pattern of the binomial option prices is monotonic and smooth, one can apply the extrapolation technique to enhance its accuracy. However, the extrapolation formula depends on the rate of convergence of the binomial model. Thus we first discuss the rate of convergence of the binomial models discussed in this paper. Leisen and Reimer (1996) proved that the convergence is of order O 1n for the CRR model. Since the FB model of Tian (1999) is a fine tune of the CRR model, it is straightforward to verify that its convergence is also of order O 1n . Moreover, Heston and Zhou (2000) showed that the rate of convergence can be improved by smoothing the option payoff functions and cannot exceed 1 n at some nodes of the tree.9 They showed that the rate of convergence of the binomial model with smoothed payoff functions (e.g. the BBS model and the SPF model) is also of order O 1n . Finally, Corollary 1 of Chung and Shih (2007) proved that the rate of convergence of 9 Please see p. 57 of Heston and Zhou (2000). 9 the GCRR-XPC model is of order O 1n . Thus, all the binomial models considered in this paper have the same rate of convergence, i.e. of order O 1n for calculating option prices. Let e(n) be the pricing error of the n-step binomial model, i.e.: e(n) Cn CBS , (13) where Cn is the n-step binomial price and C BS is the Black-Scholes price. Define the error ratio as: (n) e(n / 2) e(n). (14) The error ratio is a measure of the improvements in accuracy as the number of time steps doubles. Given the fact that all the binomial models considered in this paper have the convergence of order O 1n , we obtain that the error ratio will converge to 2 as the number of time steps increases. Thus the Black-Scholes price can be approximated as follows: CBS (n) 2Cn Cn 2 . (15) It is not difficult to show that the pricing error after applying the extrapolation can be written as: e(n) CBS (n) CBS (2 (n))e(n) . (16) Thus the pricing error after applying the extrapolation critically depends on the difference between 2 and the error ratio. If the error ratio ( n ) is within the range of (1, 3) , the absolute pricing error with extrapolation is smaller than the absolute pricing error without extrapolation. Otherwise, applying the extrapolation method to the binomial prices will increase rather than decrease the pricing errors. Similarly, to derive the extrapolation formula for hedge ratios, one needs to 10 derive the rate of convergence of the Greeks calculated from the binomial models. In the following, the theoretical proofs of the convergent rates of applying the extended method to CRR, SPF, and GCRR-XPC for calculating deltas are offered.10 Similarly, calculating gamma also has the same rate of convergence and the detailed proofs are available on request. Theorem 1. Let ˆ n,CRR be the n-period CRR binomial delta of a standard European put option with extended tree and BS is the true delta. Therefore, n ,CRR = BS d12 2 e f (n) O 1n , n 2 (17) where f (n) = 2d 2 ( (n) (n) 2 ) , N(.) is the cumulative density function of the standard normal distributions, d1 ln XS (r 2 / 2)T T and (n) ln SXx ln du , S x Su m d nm , and m is the largest integer which satisfies Su m d n m X . Proof: It is assumed that Su m d n m X and m is the largest integer. In other words, m ln XS n n ( n) = z , 2 2 2 t (18) ln X ln Su m d nm (n) and 0 (n) 1 . Since where z (n) , we have 2 t 2 t ln X S Su m d nm Xe2 t . (19) On the other hand, 10 Generally, BBS and SPF can be classified as the models with the smoothed payoff function. On the other hand, FB and GCRR-XPC can be classified as the models with the adjusted tree so that the strike price is allocated at one of the final nodes. Therefore, in addition to the convergent rate of applying extended binomial tree method to CRR, only the convergent rates of applying extended binomial tree method to SPF and GCRR-XPC for calculating deltas are offered for simplicity. 11 ert e p where a r t e e t t 1 a t O 1n , 2 (20) 2 2 . 2 In the extended tree, the option value at the node with stock price equal Su 2 at present time is m 1 e rT C nj p j (1 p)n j X Su 2u j d n j , j 0 (21) and the option value at the node with stock price equal Sd 2 at present time is m 1 e rT C nj p j (1 p)n j X Sud 2u j d n j . j 0 The estimated delta is n,CRR (22) (21) (22) . We denote (21)-(22) =A+B+C, where S (u 2 d 2 ) n m n m 1 , A Xe rT Cmn p m 1 p Cmn 1 p m1 1 p B= nm n m1 e rT Cmn p m 1 p Su 2u md nm Cmn 1 p m1 1 p Sd 2u m1d nm1 , m C= e rT C nj p j 1 p j 0 n j Su 2 (23) Sd 2 u j d n j . According to Chung and Shih (2007), m C n j rT e C nj p j 1 p u j d n j 2 2 S (u d ) j 0 0.5 (n) N d1 O 1n . np 1 p Because Cmn 1 Cmn nm and Su m d nm Xe2 m 1 12 t by (19), we have (24) A+B Xe rT Cmn p m 1 p = Xe rT Cmn p m 1 p where D = u 2e2 nm nm t t n m p 2 e m 1 1 p t 1 nm p m 1 1 p (25) D, 1 2nz nm Since m 1 1 2( zn1) 2 2 u e n m p 2 e m 1 1 p t 1 nm nm p . First, let’s check . m 1 m 1 1 p 2 2 z 2( z 1) 4 z 1 1 1 ... , by (18), n n n2 2 ln XS nm 1 O m 1 n t . e1 ( n ) n (26) Besides, we have u 2e2 t 1 2 t 1 (n) 2 2 t 1 (n) O 2 e2 ( n ) n15 , (27) and by (20) p 1 4a t O 1n . 1 p (28) Therefore, D defined in (25) is 2ln XS D 1 O n t e1 ( n ) n 1 4a × 2 t (n) 2 2 t 2 (n) O t O 1n e3 ( n ) n15 + 2 t 1 (n) 2 2 t 1 (n) O 2 e2 ( n ) n15 4 t 0.5 (n) 2 2 t 2 (n) 8at (n) 4ln XS 2 (n) 2 2 t 1 (n) O n e4 ( n ) n1.5 On the other hand, according to Feller’s (1971, P.53~P.54), 13 (29) n! 2 n n 1 2 n e 1 O . 1 n Therefore, Cmn Since 1 1 1 O 1n . 1 1 2 n 1 m nm 2 m m 2 n n m 1 z according to (18), n 2 n 1 O 1n 1 . C n 1 1 1 2 n 1 2 z nm 2 2 z m 2 1 n 1 n 2 n m (30) On the other hand, by (20), p m 1 p nm n m n m 1 1 x 1 x , 2 (31) where x 2a t O 1n . Therefore, by (30) and (31), C p 1 p n m m nm 1 1 1 x 2 n 1 1 2nz 2 nm 1 x 2z 1 n n m 1 1 2z 2 n 1 2z 2 n 1 1 m 1 x 1 x 1 2nz 1 O 1n 2z 2z n 1 n 1 x 1 n 2 4 1 . (32) y 2 y3 According to the Taylor’s series of ln(1 y ) y ... and the 2 3 definitions of z and x in (18) and (31), n m 1 x 1 x 1 2nz nx 2 2z 2 ln 2 z 2z 2 xz O 2 n 1 n 1 x 1 n 2 1 n z = 2 a t n O n 14 . 1 n (33) 2 r 2 ln XS 2 (n) z ( n) 1 d2 Since a t n T , 2 n 2 n 2 T n d (33) 2 2 (nn ) 2 2 O 1 n d22 2d2 (n) O 2 n . 1 n Therefore, d 22 2 d 2 ( n ) 2 n Xe rT Cmn p m 1 p nm e Xe rT n 2 4 1 S 1 O 2 n 4 e d12 2 e 2 d 2 ( n ) n 1 O 1 n 1 n . 1 O 1 O 1 n . 1 O 1 O 1 n 1 n (34) By (29) and (34), A+B in (25) is ( A B) S 1 2 n 4 e d12 2 e 2 d 2 ( n ) n 1 n 4 t 0.5 (n) 2 2 t 2 (n) 8at (n) 4ln XS n (n) 2 2 t 1 (n) O 2 e4 ( n ) n1.5 (35) Therefore, after tedious calculations, we have n,CRR (21) (22) A B C (35) (24) . 2 2 2 2 S (u d ) S (u d ) S (u 2 d 2 ) N d1 d12 2 e f (n) O 1n , n 2 where f (n) = 2d 2 (n) 2 (n) . (36) Q.E.D. According to Theorem 1, although using CRR with an extended tree to calculate delta has the rate of convergence O 1n , the convergent pattern is oscillatory due to the 15 relative allocation of the strike price between two most adjacent nodes. However, SPF and GCRR-XPC models demonstrated that the CRR’s binomial prices monotonically converge to the Black-Scholes formula at the rate of O( 1n ). The rates of convergence of applying extended tree to SPF and GCRR-XPC models in calculating delta are proved in the following theorems. Theorem 2. Let ˆ n, SPF be the n-period SPF binomial delta of a standard European put option with extended tree and BS is the true delta, n , SPF = BS O 1n . (37) Proof: For simplicity, it is assumed that (n) 0.5 . The proofs for the other values of (n) are quite similar, and they are available on request. In the extended tree, the option value at the node with stock price equal Su 2 at present time is m 1 e t e e rT C nj p j (1 p) n j X Su 2u j d n j 2 t j 0 t rT n m n m e Cm p (1 p ) X t ln( Su m1d nm1 ) X e t X m 1 n m 1 × Su d , 2 t 2 t 2 t (38) and the option value at the node with stock price equal Sd 2 at present time is m 1 e rT C nj p j (1 p ) n j ( X Sd 2u j d n j j 0 e rT n m 2 C p m 2 (1 p) n m2 e e 2 t t t ) X t ln Su m1d nm1 X 2 t e t X Su m 1d n m 1 . 2 t 2 t The estimated delta is denoted by n, SPF (39) (38) (39) . For the convenience of proof, S (u 2 d 2 ) 16 We define t ln Su m1d nm1 t X Su m1d nm1 1 e A= . X 2 t 2 t 2 t Xe rT A[Cmn p m (1 p)nm Cmn 2 p m 2 (1 p)n m2 ] (38) (39) Let’s first calculate in . 2 2 S (u 2 d 2 ) S (u d ) Because ln Su m1d nm1 2 t 1 (n) , X t (1 2 ( n )) 1] 2 1 [e A= (1 (n)) t 0.5 (n) O 2 t 2 E1 ( n ) n . (40) In addition, Xe rT [Cmn p m (1 p) n m Cmn 2 p m 2 (1 p) n m 2 ] (n m 1)(n m) p 2 Xe rT Cmn p m (1 p)nm 1 . (m 1)(m 2) (1 p) 2 Since p a r (41) [1 2( zn1) ][1 2nz ] 1 (n m 1)(n m) a t O 1n and where 2 (m 1)(m 2) [1 2( zn1) ][1 2( zn 2) ] 2 2 2 and z ln X S 2 t ( n) , (n m 1)(n m) p 2 4ln XS (r 2 ) 1 4 t O (m 1)(m 2) (1 p) 2 n t 2 4d 2 O n E1 ( n ) n E2 ( n ) n . (42) By (34) and (42), we have (41)= S 1 2 n 4 e d12 2 e 2 d 2 ( n ) n (1 O ) 1 O 4nd 1 n 1 n Therefore, by (40) and (43), 17 2 O E2 ( n ) n . (43) Xe rT A[Cmn p m (1 p)n m Cmn 2 p m 2 (1 p )n m 2 ] S (u 2 d 2 ) 1 2 e n 2 d12 2 (d 2 ) 0.5 (n) O 2 E3 ( n ) n1.5 . 2 1 2 d21 e (d 2 ) (n) 2 (n) O 1n n 2 (44) From the above theorem, we know d12 2 U D 2e N d1 d 2 (n) 2 (n) O 1n , 2 2 Su Sd n 2 (45) where m 1 e t e U = e rT C nj p j (1 p )n j X Su 2u j d n j 2 t j 0 t , m 1 e t e D = e rT C nj p j (1 p )n j X Sd 2u j d n j 2 t j 0 t . Therefore by (44) and (45), the estimated delta n, SPF (40) (41) = (44)+(45) = N (d1 ) O 1n . 2 2 S (u d ) Q.E.D. Theorem 3. Let ˆ n,GCRR XPC be the n-period GCRR-XPC binomial delta of a standard European put option with extended tree and BS is the true delta, n ,GCRR XPC = BS O 1n . (46) Proof: Please refer to Appendix 1. To sum up, in this section, we discuss the rate of convergence of the binomial models and offer the theoretical proofs of the rates of convergence of applying extended method to CRR, SPF and GCRR-XPC to calculate deltas. It is found that applying extended tree to SPF and GCRR-XPC can generate monotonic convergence with O 1n rate for calculating Greeks. 18 4. Numerical results and analysis In the following numerical analysis, we will extensively illustrate the convergent patterns, rate of convergence and numerical efficiency of various binomial models of calculating delta and gamma. First, to show the error ratios of various binomial models defined in (14), the following parameters are chosen: stock price=40, strike price = 45, volatility = 0.2, time to maturity= 6 months, riskless interest rate=6% and dividend yield= 0 so that we can make sure if the extrapolation technique can be applied to enhance the accuracy. For calculating deltas, from Tables 1 and 2, it shows that applying Hull method or applying the extended method to these binomial models result in quite stable error ratios around 2. It means that they converge to the true delta monotonically with O(1/n) rate. Therefore, the extrapolation technique can be applied to enhance the accuracy. We first choose a large set of options (243 options). The 243 parameter sets are drawn from the combinations of X {35, 40, 45}, T {1/12,4/12,7/12}, r {3%, 5%, 7%}, q {2%, 5%, 8%}, {0.2, 0.3, 0.4}, and S 40 . We report the results for different numbers of time steps n {20, 40, 60, 80, 100, 500, 1000} in order to examine the convergence property. The accuracy measure used in this paper is the root mean squared (RMS) absolute error. RMS absolute error is defined as 1 m 2 RMS ei , m i 1 (47) where ei cˆi ci is the absolute error, ci is the “true” value (estimated from the Black-Scholes formula), cˆi is the estimated value from the binomial model. <Insert Tables 1 and 2 Here> 19 To clarify the role played by Richardson extrapolation, we first compare the numerical efficiency of each model on a raw basis without any extrapolation procedure. From Table 3, the extended method on average underperforms Hull method without extrapolation.11 Overall, the SPF with Hull method is the best. However, when Richardson extrapolation is applied, Table 4 shows that the extended method on average outperforms the Hull method in calculating Greeks. 12 The improvement of using Richardson extrapolation for the extended method is because the error ratios of the extended method are more like to be 2 for these four binomial models compared to the Hull method as shown in Table 1 and Table 2. Overall, applying extended method to GCRR-XPC is the best. For calculating gammas, from Table 5 and Table 6, applying Hull method or extended method to these binomial models result in quite stable error ratios around 2, so the extrapolation technique can be applied to enhance the accuracy. Again, to clarify the role played by Richardson extrapolation, we first compare the numerical efficiency of each model on a raw basis without any extrapolation procedure. In Table 7, it shows that the extended method is on average superior to the Hull method.13 Overall, the performance of FB with the extended tree is the best. However, when Richardson extrapolation is used in Table 8, FB performs worst and this is consistent with Table 5 and Table 6 in which the error ratios of using FB are most unstable 11 This is true for all these four binomial models except for GCRR-XPC model. 12 The numerical results of using extended method are better than those of using Hull method for GCRR-XPC model and FB model. Besides, the numerical result of using extended method is almost as good as those of using Hull method for BBS model. 13 The numerical results of using extended method are better than those of using Hull method for CRR model, GCRR-XPC model, and FB model. However, the numerical result of using Hull method is better than those of using extended method only for BBS model and SPF model. 20 around 2. Overall, applying extended tree to GCRR-XPC is the best. To sum up, applying extended tree to GCRR-XPC with extrapolation results in the most efficient estimates of deltas and gammas. <Insert Tables 3 to 8 Here> 5. Conclusion This paper uses various binomial models to calculate Greeks and discuss the convergent pattern, convergent rate, and the numerical efficiency among these models. We contribute to the literature at least three aspects. First, although BBS model of Brodie and Detemple (1996) , SPF of Heston and Zhou (2000), FB of Tian (1999), and GCRR-XPC of Chung and Shih (2007) have been widely applied to price options, their convergent patterns and rate of convergence for calculating the hedge ratios are not well known. To fill this gap, we apply the Hull method and the extended method to these models in order to compute deltas and gammas. Second, we apply the Hull method and extended method to these models and compare numerical efficiency of calculating Greeks. It is found that applying the extended method to GCRR-XPC is the most efficient binomial model in calculating Greeks when a two-point extrapolation formula is used. Finally, we derive the convergence rate of the binomial model in calculating Greeks. We prove that applying the extended method to GCRR-XPC model and SPF model converge to the true delta monotonically with the rate of convergence O(1/n). 21 Appendix 1 (Proof of Theorem 3) At the node with stock price equal to Su1u2 , the option price is n 1 2 e rT C nj p j (1 p)n j ( X Su1u2 u1j d1n j ) , (A1) j 0 and at the node with stock price equal to Sd1d 2 , the option price is n 1 2 e rT C nj p j (1 p)n j ( X Sd1d 2 u1j d1n j ) , (A2) j 0 where u1 e ( n ) t , u2 e t (n) , d1 e t (n) and d 2 e ( n ) t . The estimated delta is (A1) (A2) defined as ˆ n ,GCRR XPC and S (u1u2 d1d 2 ) n 1 n2 1 2 n j n j n j n j (A1) (A2) X C j p (1 p) C j p (1 p) j 0 j 0 n 1 n2 1 2 j n j n j n j j n j n j n j rT + Su1u2 u1 d1 C j p (1 p) Sd1d 2 u1 d1 C j p (1 p ) e j 0 j 0 Because C(nn / 2) 1 n! n n/2 n C(nn / 2) , n2 n n n n n 1 1 ! 1 ! ! ! 2 2 2 2 2 n 1 n2 1 2 n j n j n j n j rT X C j p (1 p) C j p (1 p) e j 0 j 0 n n n n 1 1 X Cnn/2 p 2 (1 p) 2 C(nn /2) 1 p 2 (1 p) 2 e rT n p rT XC(nn /2) p n /2 (1 p)n /2 1 e . n 2 1 p 22 (A3) 1 a t O 1n ,where 2 According Chung and Shih (2007), p 2 ln r T 2 1 r ln XS 1 1 1 . a d2 2 T 2 2 T 2 T X S (A4) Therefore, n 1 1 2 p (1 p) a t O 1n a t O 1n 2 2 n 2 n 2 n 1 2 1 4a 2 t O 4 . n 2 1 n 3 (A5) 2 On the other hand, n 1 n2 1 2 Sd d u j d n j C n p j (1 p ) n j Su u u j d n j C n p j (1 p )n j e rT 1 2 1 1 j 1 2 1 1 j j 0 j 0 n 2 ( Sd1d 2 Su1u2 ) C nj p j (1 p) n j u1j d1n j e rT j 0 n n 1 1 2 2 1 2 1 1 Sd d u d n ( n /2) 1 C n n 2 2 1 1 p n 1 2 Since Su d X , u1 e (1 p ) ( n ) t n 1 rT 2 e , u2 e n 2 1 2 1 n 2 1 Su u u d C t (n) , d1 e n n /2 t (n) n 2 n 2 p (1 p ) e rT . and d 2 e ( n ) t (A6) , n n n n n n n n 1 1 1 1 rT n 2 2 2 2 2 2 n 2 2 Sd d u d C p (1 p ) Su u u d C p (1 p ) 1 2 1 1 ( n /2) 1 e 1 2 1 1 n /2 n n n n ddu n p rT Cnn/2 p 2 (1 p ) 2 Su12 d12 u1u2 1 2 1 e d1 n 2 1 p C n n /2 p n rT p (1 p) X u1u2 e 1 p n 2 n 2 n 2 By (A3), (A6) and (A7), 23 (A7) n 2 n 2 XC p (1 p) [u1u2 1] rT (A1) (A2) C nj p j (1 p)n j u1j d1n j e rT e . S (u1u2 d1d2 ) S (u1u2 d1d 2 ) j 0 n n /2 n /2 (A8) Because 2 1 p(1 p) a t O 1n 2 2 1 [1 x ] 4 where x 4a 2 t O n (A9) , 1 n1.5 n X Cnn/2 p 2 (1 p) 2 [u1u2 1] rT X n 1 e Cn /2 S u1u2 d1d 2 S 4 Since (1 x) e d22 2 e n /2 e ( d1 a T )2 2 ln(1 x ) e n 2 d12 2 e ( x x2 n ...) 2 2 n /2 e2 a tn e 2 O 1 x n /2 u1u2 1 e rT u1u2 d1d 2 (A10) , by (A4) and 1 n S rT e ,we have X (1 x) n /2 e d22 2 S e d2 erT eO . e O 2 1 1 n 1 n X and n d X 1 2 S rT 21 rT O 1n (u1u2 1) e e e e (A10) Cnn/2 S u1u2 d1d 2 4 X n 2 d 1 2 1 O 1 (u1u2 1) Cnn/2 e 2 e n . u1u2 d1d 2 4 2 (A11) According to Feller’s (1971, p.53 ~ p.54), n! 2 n n 1 2 n e If we set n 2k ,we have 24 1 O . 1 n C n n /2 n 2 (2k )! 1 1 k !k ! 2 4 2k 2k 1 2 e 2 k 1 O 21k 1 2k 2 2 2 k 2 k 1e2 k 1 O 1k 2 (2k ) 1 1 O 1k . k Therefore, n n n /2 C 1 1 1 2 1 O 1n 1 O 1n , k n / 2 4 (A12) and (A11) e e d12 2 O 1 u u 1 1 1 O 1n e n 1 2 u1u2 d1d 2 2 (n / 4) d12 2 1 d21 e 2 1 1 O 1n 1 O 2 (n / 4) 12 O 1 n 1 O 1n . 2 (n / 4) 1 n (A13) In (A8), according to Chung and Shih (2007), n /2 n /2 J 0 J 0 C nj p j (1 p) n j u1j d1n j e rt C nj p* j (1 p* )n j 0.5 O 1n . N d1 * * np (1 p ) where p* (A14) pu1 . e r t 0.5 (A14) N (d1 ) N (d 1 ) O 1n np* (1 p* ) 1 d1 N (d1 ) e 2 2 1 O 1n . 2 (n / 4) 25 (A15) Finally, we have (A1) (A2) ˆ n ,GCRR XPC (A13) (A15) N ( d1 ) O 1n . Su1u2 Sd1d 2 (A16) Q.E.D. 26 References Broadie, M., Detemple, J., 1996. American option valuation: New bounds, approximations, and a comparison of existing methods. Review of Financial Studies 9, 1211-1250. Chang, C.C., Chung, S.L., Stapleton, R., 2007. Richardson extrapolation techniques for the pricing of American-style options. Journal of Futures Markets 27, 791-817. Chung, S.L., Shih, P.T., 2007. Generalized Cox-Ross-Rubinstein binomial model. Management Science, 53, 508-520. Cox, J.C., Ross, A., Rubinstein, M., 1979. Option pricing: A simplified approach. Journal of Financial Economics 7, 229-263. Feller, W., 1971. An introduction to probability theory and its applications. NY: John Wiley&Sons, Inc. Figlewski, S., Gao, B., 1999. The adaptive mesh model: A new approach to efficient option pricing. Journal of Financial Economics 53, 313-351. Heston, S., Zhou, G., 2000. On the rate of convergence of discrete-time contingent claims. Mathematical Finance 10, 53-75. Hull, J.C., 1993. Options, futures and other derivatives. NJ: Prentice-Hall International, Inc. Leisen, D., Reimer, M., 1996. Binomial models for option valuation-examining and improving convergence. Applied Mathematical Finance 3, 319-346. Omberg, E., 1988. Efficient discrete time jump process models in option pricing. Journal of Financial and Quantitative Analysis 23, 161-174. Pelsser, A., Vorst, T., 1994. The binomial model and the Greeks. Journal of Derivatives 1, 45-49. 27 Rendleman, R., Batter, B., 1979. Two-state option pricing. Journal of Finance 34, 1093-1110. Tian, Y.S., 1999. A flexible binomial option pricing model. The Journal of Futures Markets 19, 817-843. Widdicks, M., Andricopoulos, A.D., Newton, D.P., Duck, P.W., 2002. On the enhanced convergence of standard lattice methods for option pricing. Journal of Futures Markets 22, 315-338. 28 0 0 0 Fig. 1. The flexible binomial tree for different λ values S 2,2 S1,1 C2,2 C1,1 S 0,0 S 2,1 S1,0 C0,0 S 2,0 C 2,1 C1,0 C2,0 Fig. 2. The Hull method C0,2 S 0,2 S 1,1 C 1,1 S 2,0 S 0,1 C 0,1 C2,0 S 1,0 C1,0 S 0,0 C0,0 -2 t - t 0 -2 t - t Fig. 3. The extended binomial tree 29 0 TABLE 1 Delta and error ratios for European put options by using Hull method GCRR-XPC FB Delta Error Error ratio Delta Error Error ratio 20 -0.719297 -0.010454 — -0.716577 -0.007734 — 40 -0.714714 -0.005871 1.780563 -0.712432 -0.003589 2.154864 80 -0.712009 -0.003166 1.854442 -0.710782 -0.001939 1.850665 160 -0.710508 -0.001665 1.901807 -0.709920 -0.001077 1.800574 320 -0.709704 -0.000861 1.932918 -0.709386 -0.000543 1.983631 640 -0.709284 -0.000441 1.953739 -0.709094 -0.000251 2.165007 1280 -0.709067 -0.000224 1.967875 -0.708973 -0.000130 1.936146 2560 -0.708956 -0.000113 1.977577 -0.708908 -0.000065 1.983873 5120 -0.708900 -0.000057 1.984291 -0.708875 -0.000032 2.016248 Time steps n BBS SPF Delta Error Error ratio Delta Error Error ratio 20 -0.713379 -0.004536 — -0.711156 -0.002313 — 40 -0.711087 -0.002244 2.021622 -0.709940 -0.001097 2.107662 80 -0.709950 -0.001107 2.026989 -0.709392 -0.000549 2.000170 160 -0.709394 -0.000551 2.007205 -0.709103 -0.000260 2.109401 320 -0.709119 -0.000276 1.999329 -0.708974 -0.000131 1.977799 640 -0.708981 -0.000138 2.003059 -0.708909 -0.000066 1.977977 1280 -0.708911 -0.000068 2.015835 -0.708876 -0.000033 1.996055 2560 -0.708877 -0.000034 2.000920 -0.708860 -0.000017 2.012228 5120 -0.708860 -0.000017 1.995844 -0.708851 -0.000008 1.990182 Black-Scholes -0.708843 Time steps n The number of time steps starts at 20 and doubles each time subsequently. The left side of the table reports delta, delta errors, and error ratios. 30 TABLE 2 Delta and error ratios for European put options by using extended method GCRR-XPC Time steps n FB Delta Error Error ratio Delta Error Error ratio 20 -0.701162 0.007682 — -0.700546 0.008297 — 40 -0.704986 0.003857 1.991408 -0.704675 0.004169 1.990367 80 -0.706911 0.001933 1.995742 -0.706751 0.002093 1.992035 160 -0.707876 0.000967 1.997881 -0.707804 0.001039 2.013715 320 -0.708359 0.000484 1.998943 -0.708324 0.000519 2.000775 640 -0.708601 0.000242 1.999472 -0.708581 0.000262 1.981833 1280 -0.708722 0.000121 1.999736 -0.708713 0.000131 2.003349 2560 -0.708783 0.000061 1.999868 -0.708778 0.000065 2.001280 5120 -0.708813 0.000030 1.999934 -0.708810 0.000033 1.998430 BBS Time steps n SPF Delta Error Error ratio Delta Error Error ratio 20 -0.697113 0.011731 — -0.695143 0.013701 — 40 -0.702848 0.005996 1.956450 -0.701771 0.007073 1.937062 80 -0.705803 0.003040 1.972039 -0.705262 0.003582 1.974643 160 -0.707314 0.001529 1.987975 -0.707027 0.001817 1.971769 320 -0.708077 0.000766 1.995427 -0.707934 0.000910 1.996845 640 -0.708460 0.000384 1.996501 -0.708389 0.000455 2.000010 1280 -0.708651 0.000193 1.993193 -0.708616 0.000228 1.998972 2560 -0.708747 0.000096 1.999067 -0.708730 0.000114 1.997423 5120 -0.708795 0.000048 2.001171 -0.708786 0.000057 2.001033 Black-Scholes -0.708843 The number of time steps starts at 20 and doubles each time subsequently. The left side of the table reports delta, delta errors, and error ratios. 31 TABLE 3 RMS errors of delta estimates (without extrapolation) Time steps n 20 40 60 80 100 500 1000 Hull method CRR GCRR-XPC FB BBS SPF 1.97E-03 7.49E-04 9.90E-04 5.24E-04 4.30E-04 6.82E-05 4.04E-05 (0.08) (0.14) (0.20) (0.28) (0.35) (4.13) (15.34) 6.23E-03 3.40E-03 2.35E-03 1.80E-03 1.47E-03 3.17E-04 1.62E-04 (0.13) (0.20) (0.29) (0.38) (0.50) (5.70) (21.16) 4.05E-03 2.07E-03 1.48E-03 1.07E-03 8.43E-04 1.70E-04 8.68E-05 (0.08) (0.12) (0.16) (0.21) (0.28) (4.20) (16.15) 1.77E-03 8.66E-04 5.71E-04 4.28E-04 3.41E-04 6.79E-05 3.38E-05 (3.88) (7.64) (11.40) (15.16) (18.93) (97.36) (203.27) 8.78E-04 4.41E-04 2.91E-04 2.20E-04 1.74E-04 3.50E-05 1.75E-05 (0.13) (0.19) (0.27) (0.36) (0.47) (5.27) (18.56) Extended method CRR GCRR-XPC FB BBS SPF 7.47E-03 3.94E-03 2.29E-03 1.93E-03 1.49E-03 3.24E-04 1.45E-04 (0.10) (0.15) (0.21) (0.29) (0.36) (4.16) (15.40) 5.09E-03 2.55E-03 1.70E-03 1.28E-03 1.02E-03 2.05E-04 1.02E-04 (0.15) (0.22) (0.31) (0.40) (0.52) (5.75) (21.24) 4.95E-03 2.50E-03 1.66E-03 1.24E-03 1.01E-03 2.02E-04 1.00E-04 (0.10) (0.13) (0.17) (0.22) (0.29) (4.23) (16.21) 7.31E-03 3.73E-03 2.51E-03 1.89E-03 1.51E-03 3.05E-04 1.53E-04 (4.27) (8.02) (11.78) (15.54) (19.31) (97.75) (203.68) 8.71E-03 4.45E-03 2.99E-03 2.25E-03 1.80E-03 3.63E-04 1.82E-04 (0.15) (0.21) (0.29) (0.38) (0.49) (5.31) (18.63) The CPU time (in seconds) required to value 243 deltas is given in parentheses. 32 TABLE 4 RMS errors of delta estimates (with extrapolation) Time steps n 20 40 60 80 100 500 1000 Hull method GCRR-XPC FB BBS SPF 2.48E-04 6.74E-05 3.08E-05 1.76E-05 1.14E-05 4.68E-07 1.18E-07 (0.16) (0.26) (0.38) (0.51) (0.66) (7.38) (26.75) 2.89E-03 1.25E-03 6.36E-04 3.49E-04 2.18E-04 2.85E-05 9.44E-06 (0.10) (0.16) (0.21) (0.28) (0.37) (5.43) (20.42) 6.13E-04 1.70E-04 8.17E-05 4.53E-05 3.05E-05 1.01E-06 7.68E-07 (5.93) (11.56) (17.22) (22.88) (29.26) (145.38) (300.24) 4.01E-04 1.27E-04 6.41E-05 3.75E-05 2.26E-05 2.04E-06 6.96E-07 (0.16) (0.25) (0.36) (0.49) (0.62) (6.82) (23.46) Extended method GCRR-XPC FB BBS SPF 1.55E-04 4.29E-05 1.98E-05 1.14E-05 7.38E-06 3.08E-07 7.73E-08 (0.19) (0.29) (0.41) (0.54) (0.69) (7.44) (26.85) 9.31E-04 4.36E-04 2.43E-04 1.40E-04 9.08E-05 9.56E-06 2.94E-06 (0.13) (0.17) (0.22) (0.30) (0.38) (5.47) (20.49) 6.17E-04 1.71E-04 8.20E-05 4.56E-05 3.06E-05 1.01E-06 7.68E-07 (6.53) (12.14) (17.79) (23.45) (29.85) (145.96) (300.85) 9.22E-04 2.41E-04 1.06E-04 6.70E-05 4.05E-05 2.19E-06 9.22E-07 (0.19) (0.28) (0.39) (0.52) (0.66) (6.93) (23.82) Note: The CPU time (in seconds) required to value 243 deltas is given in parentheses. 33 TABLE 5 Gamma and error ratios for European put options by using Hull method GCRR-XPC Time steps n FB Gamma Error Error ratio Gamma Error Error ratio 20 -0.001827 -0.001827 — 0.001526 0.001526 — 40 -0.000883 -0.000883 2.068289 0.000841 0.000841 1.814899 80 -0.000425 -0.000425 2.080183 0.000383 0.000383 2.197365 160 -0.000205 -0.000205 2.076239 0.000149 0.000149 2.562954 320 -0.000099 -0.000099 2.065368 0.000073 0.000073 2.044971 640 -0.000048 -0.000048 2.052693 0.000046 0.000046 1.600160 1280 -0.000024 -0.000024 2.040803 0.000021 0.000021 2.147565 2560 -0.000012 -0.000012 2.030747 0.000010 0.000010 2.041972 5120 -0.000006 -0.000006 2.022736 0.000005 0.000005 1.959054 BBS Time steps n SPF Gamma Error Error ratio Gamma Error Error ratio 20 0.061341 0.000717 — 0.061022 0.000398 — 40 0.060984 0.000360 1.993806 0.060824 0.000200 1.991723 80 0.060803 0.000179 2.003719 0.060727 0.000103 1.935925 160 0.060713 0.000089 2.016477 0.060684 0.000060 1.732304 320 0.060669 0.000045 1.987626 0.060653 0.000029 2.078587 640 0.060646 0.000022 1.997408 0.060638 0.000014 2.054581 1280 0.060635 0.000011 2.012954 0.060631 0.000007 2.000481 2560 0.060630 0.000006 2.002770 0.060628 0.000004 1.969727 5120 0.060627 0.000003 1.993477 0.060626 0.000002 2.028263 Black-Scholes 0.060624 The number of time steps starts at 20 and doubles each time subsequently. The left side of the table reports gamma, gamma errors, and error ratios. 34 TABLE 6 Gamma and error ratios for European put options by using extended method GCRR-XPC Time steps n FB Gamma Error Error ratio 20 0.059092 -0.001532 — 40 0.059851 -0.000773 80 0.060236 160 Gamma Error Error ratio — 0.060874 0.000249 1.981377 0.060814 0.000190 1.315248 -0.000388 1.990661 0.060690 0.000066 2.876399 0.060429 -0.000195 1.995324 0.060631 0.000007 10.114306 320 0.060527 -0.000097 1.997660 0.060627 0.000002 2.762609 640 0.060575 -0.000049 1.998830 0.060631 0.000007 0.351503 1280 0.060600 -0.000024 1.999415 0.060627 0.000002 2.857169 2560 0.060612 -0.000012 1.999707 0.060625 0.000001 2.256359 5120 0.060618 -0.000006 1.999854 0.060625 0.000001 1.770236 BBS Time steps n Gamma Error 20 0.060102 -0.000522 40 0.060373 -0.000251 80 0.060501 160 SPF Error ratio Gamma Error Error ratio 0.059794 -0.000830 — 2.076224 0.060215 -0.000409 2.028231 -0.000124 2.033228 0.060425 -0.000200 2.051288 0.060562 -0.000062 1.998139 0.060533 -0.000091 2.181807 320 0.060594 -0.000030 2.028958 0.060578 -0.000047 1.960983 640 0.060609 -0.000015 2.009468 0.060601 -0.000024 1.972058 1280 0.060617 -0.000008 1.983999 0.060612 -0.000012 2.001634 2560 0.060620 -0.000004 1.997418 0.060618 -0.000006 2.019246 5120 0.060622 -0.000002 2.010265 0.060621 -0.000003 1.983790 Black-Scholes 0.060624 The number of time steps starts at 20 and doubles each time subsequently. The left side of the table reports gamma, gamma errors, and error ratios. 35 TABLE 7 RMS errors of gamma estimates (without extrapolation) Time steps n 20 40 60 80 100 500 1000 Hull method CRR GCRR-XPC FB BBS SPF 2.18E-03 1.05E-03 7.25E-04 5.29E-04 4.16E-04 8.27E-05 4.13E-5 (0.08) (0.14) (0.20) (0.28) (0.35) (4.13) (15.34) 2.25E-03 1.11E-03 7.41E-04 5.56E-04 4.45E-04 8.98E-05 4.50E-05 (0.13) (0.20) (0.29) (0.38) (0.50) (5.70) (21.16) 2.34E-03 1.13E-03 7.58E-04 5.62E-04 4.51E-04 8.96E-05 4.46E-05 (0.08) (0.12) (0.16) (0.21) (0.28) (4.20) (16.15) 1.39E-03 6.72E-04 4.43E-04 3.30E-04 2.64E-04 5.23E-05 2.61E-05 (3.88) (7.64) (11.40) (15.16) (18.93) (97.36) (203.27) 9.23E-04 4.48E-04 2.94E-04 2.20E-04 1.75E-04 3.46E-05 1.74E-05 (0.13) (0.19) (0.27) (0.36) (0.47) (5.27) (18.56) Extended method CRR GCRR-XPC FB BBS SPF 9.90E-04 5.14E-04 2.99E-04 2.30E-04 1.96E-04 3.91E-05 2.02E-05 (0.10) (0.15) (0.21) (0.29) (0.36) (4.16) (15.40) 1.11E-03 5.66E-04 3.80E-04 2.86E-04 2.29E-04 4.62E-05 2.31E-05 (0.15) (0.22) (0.31) (0.40) (0.52) (5.75) (21.24) 7.38E-04 3.69E-04 2.42E-04 1.86E-04 1.48E-04 2.93E-05 1.47E-05 (0.10) (0.13) (0.17) (0.22) (0.29) (4.23) (16.21) 1.56E-03 8.03E-04 5.41E-04 4.08E-04 3.27E-04 6.61E-05 3.31E-05 (4.27) (8.02) (11.78) (15.54) (19.31) (97.75) (203.68) 1.98E-03 1.02E-03 6.90E-04 5.20E-04 4.17E-04 8.44E-05 4.23E-05 (0.15) (0.21) (0.29) (0.38) (0.49) (5.31) (18.63) The CPU time (in seconds) required to value 243 gammas is given in parentheses. 36 TABLE 8 RMS errors of gamma estimates (with extrapolation) Time steps n 20 40 60 80 100 500 1000 Hull method GCRR-XPC FB BBS SPF 3.51E-04 1.27E-04 7.48E-05 5.19E-05 3.91E-05 4.66E-06 1.77E-06 (0.16) (0.26) (0.38) (0.51) (0.66) (7.38) (26.75) 6.83E-04 2.30E-04 9.97E-05 9.15E-05 4.78E-05 4.87E-06 1.76E-06 (0.10) (0.16) (0.21) (0.28) (0.37) (5.43) (20.42) 1.99E-04 4.63E-05 2.03E-05 1.16E-05 6.97E-06 2.79E-07 9.94E-08 (5.93) (11.56) (17.22) (22.88) (29.26) (145.38) (300.24) 2.23E-04 5.15E-05 2.39E-05 1.70E-05 9.63E-06 9.33E-07 2.95E-07 (0.16) (0.25) (0.36) (0.49) (0.62) (6.82) (23.46) Extended method GCRR-XPC FB BBS SPF 1.30E-04 3.67E-05 1.70E-05 9.76E-06 6.32E-06 2.63E-07 6.62E-08 (0.19) (0.29) (0.41) (0.54) (0.69) (7.44) (26.85) 5.38E-04 1.63E-04 6.49E-05 5.89E-05 3.43E-05 3.31E-05 1.26E-06 (0.13) (0.17) (0.22) (0.30) (0.38) (5.47) (20.49) 1.80E-04 5.12E-05 2.35E-05 1.40E-05 8.92E-06 3.99E-07 1.17E-07 (6.53) (12.14) (17.79) (23.45) (29.85) (145.96) (300.85) 2.93E-04 8.07E-05 3.94E-05 2.28E-05 1.48E-05 9.93E-07 3.32E-07 (0.19) (0.28) (0.39) (0.52) (0.66) (6.93) (23.82) The CPU time (in seconds) required to value 243 gammas is given in parentheses. 37