Farup ArrSantia DemandTec in Stanford, CA, fad statstnfordedu LoReNs IMHOF isan asistant profesor at Instat fr Saisie and ‘Wirschafismathematik in Aachen, German. imhot@stoctasthrvth-achende ‘Tze Leu Lat isa profesor of statins a Stanford University hit@satstanordu Pricing and Hedging of American Knock-In Options FariD AITSAHLIA, LORENS IMHOF, AND TZE LEUNG LAI Anerican barier options of the knock-in type involve rnon-Markovian optimal stopping problems for early exercise, They therefore cannot be priced via stan dard methods such as binomial oF tinomial tees and finte-diffrenc schemes for fee-boundary par- tial differential equations. This article provides a modified tree method to rice these options, It also develops fast and accu- rate analytic appnoximations forthe price and hedge parameters. complex derivatives have become accepted instruments to tailor risk coverage for risk managers and investors, Barrier-type options have become important instruments, particu larly for the valuation of structured products (see Banks [1994)). They are also widely used in currency markets The holder ofa barrier option acquires option coverage on only a subset of the risky ‘outcomes for which a plain vanilla option pays off; this reduces the cost of the resulting cov- erage s0 that the holder of the contract does not have to pay for contingencies the holder thinks are unlikely to occur. Because of this flexibility, barrier options were traded over the counter long before the opening of the Chicago Board Options Exchange, and have become some of the most commonly traded derivative contracts. American barrier options offer the added flexibility of early exercise but have to be priced using numerical algorithms, 44, Prucinc aND HlEncING OF AMERICAN KNocK-IN Ormions as they do not have closed-form solutions, unlike their European-style counterparts (see Merton [1973] and Rubinstein and Reiner (1991p Naive application of the Cox-Ross- Rubinstein binomial tree method for barrier options has been shown by Boyle and Lau [1994] to yield inaccurate values, even with many steps. To address this problem, which stems from the position of the barrier relative to the grid, a number of variants of the tree method have been advanced. Ritchken [1995] implements a trinomial tree method. Cheuk and Vorst [1996] develop a time-dependent shift for the trinomial tree, and Figlewski and Gao [1999] introduce an adaptive mesh model that grafts high-resolution lattices around points that cause the inaccuracies in the hinomial model In an alremative to tree methods, Gao, Huang, and Subrahmanyam 2000] and AitSablia, Imhof, and Lai [2003] extend the price decomposition approach orig inally developed for standard American options by Kim [1990], Jacka [1991], and Carr, Jarrow, and Myneni [1992] to derive analytic approx imations for American knock-out prices and hedge parameters. ‘The sum of American knock-in and knock-out prices does not equal the standard American option price, as isthe case for Euro~ pean barrier options. Moreover, the knock- in option value process is non-Markovian, so the classic binomial or trinomial tree methods or numerical partial differential equations for Sonanic 2004standard American options cannot be applied directly to price American knock-in options. We first explain why this is the case, and contrast American knock-in and knock-out options ‘Then we develop a modified binomial tree method to price and hedge American down-and-in puts. An alter~ native method is based on the price decomposition approach for standard American options. We propose an efficient approximation to implement this approach that leads to analytic approximations that have penny accu~ racy. Numerical results for both approaches are provided, and we also address the evaluation of hedge parameters, I, AMERICAN KNOCK-IN AND KNOCK-OUT OPTIONS Consider an underlying asset whose price process S, follows a geometric Brownian motion with volatility & and that pays dividends ar rate gin a market environment where the riskless rate of return is r, Let H be a barrier for either a knock-in ora knock-out option, Let Tbe the expiration date for any option on the asset. Let g(S, K) denote the option’s payoff when exercised at asset price S. Then, 9(S, K) = (S~ K)* is the payoff of a call option with exercise price K, and g(S, K) = (K~ $)* isthe payo on the corresponding put. Define: f (0) of fp) = init <<: $, > HY inf{t <7: Si < 1H} ‘That is, rf) (or x) is the first time the price of the underlying asser falls below (or rises above) the barrier H. ‘Then, for any stopping (early exercise) time 7 $ T: Bs [e705 [eT HS Ags .g)] + = Bs [eS 35] Ww where EX] denotes the expectation ofa random variable X conditional on the initial value S, = S. When 7 = T, Equation (1) expresses the well-known relation between, the price ofa standard European option, on the left, and the prices of corresponding European knock-in and knock-out options, respectively, on the right. Let T,,, denote the class of stopping times taking values between a and b with a H. While the usual binomial ere stats from the root node S and may not include Has a node value, our modified binomial tree uses lattice that includes the barrier. ‘The integral in (6) involves the distribution of the first time that the geometric Brownian motion S, crosses the level H. Because 5, = S exp{(r~ 4 ~ 07/2)t+0B,} under the risk-neutral measure P, where {B,} is a stan~ ard Brownian motion, and because the first-passage den sity of Brownian motion has a simple formula, we use the change of variables: z=logS, y=logH, N=r—q—02/2 o The knock-in time 1/2 = inf{t > 0: S, $ H} can then be expressed as 1/2) = infft2 0: Z, <7}, where: Aart M+oB ® isa Brownian motion with drift A. Hence the probability distribution of 77’ has a density fanction f. given explic~ idly by: 19-(8)o( where n() is the standard normal density function; see Karatzas and Shreve [1988, p. 196]. ‘Using Equation (9), we can express the value (6) of an American down-and-in put [which we will denote P,(S) instead of Viy(S)] a8: a-M ovt ) for #>0 (9) 2 Pols) = fo ervann goat = aS HV (67 te) falta) a9) Here the approximating Riemann sum involves M + 1 equally spaced time steps f, = 0.< <=" 0 and. space increment X, such that P(X shot 1058) =3(14 ee) a Note that X, has mean Ad = E(Z,,5~Z) and vari- ance o°8 = Var(Z,,~Z), and that the barrier 7 belongs to the lattice L, = {y+ VB (0? + 6d): 7= 0, 1, 2, ...}, ‘The backward induction algorithm of dynamic pro- gramming yields: Ve th1) = max{(K —e)*, eM EV (0%, t,)} for re Ls (12) and is initialized at Thy We, TD) = (Ke). Note that log S may not belong to the lattice Ly, in contrast to the usual binomial tree method in which Sis always the root node of the tree but the barrier may not bea node ofthe tree. In AitSahlia, Imhof, and Lai [2003], ‘we use a similar Bernoulli random walk with absorbing, barrier ~y and increments (11) to approximate a Wiener process with the same absorbing barrier 7 to handle the barrier problem for knock-out options (see Boyle and Lau (1994). Il. FAST AND ACCURATE APPROXIMATION Since the price V(H, 1) of a standard American put option can be decomposed as the sum of a European put plus an early exercise premium, we can likewise decom- pose the price (6) of an American down-and-in put as: Pols) = nots) + [erat —o4ae 43) ) where m(H, T~ 1) is the early exercise premium of a stan dard American pu with matutiey Tf, strike price K, and initial stock price H; f,(d) is given in (9); and z = log S. The p,(S) in (13) is the price of a European down- andi pt option with strike price Kand expiration date T. In view of (1), p(S) can be expressed as the difference between a standard European put and a European down- and-out put, yielding the closed-form expression: Sena 2004pol8) = ~Se-PN (ay, H,T) + Ke"™N (—de{S,H,7)) 486-97 (H/S)*(N (d(H, SK,)) ~ N (6ES,7))} Ke“ (H/SPMN (da(H?, SK,T)) ~ N (d(H, $,7))) (a4) where d (2, y, 7) = {logte/y) + (r+ o°/2)1} / (a\F}s 4, (2, 4, 7) = dy (@, y, 7) ~ 0% and N(@) is the standard normal cumulative probability distribution function, At this point, the early exercise premium ™(H, T— 1) is the only piece needed to fully determine the price of an American down-and-in put using Equation (13). Since this quantity is the difference between standard European and American option prices, for which various numerical methods are available, one may choose a meth od according to one’s preference. Nevertheless, once Tis determined, there remains the task of evaluating the inte~ gral in (13). For fistand accurate approximations of the integrand and inegral in (1), first we se Jo’ [1998] or AitSahlia and Lai’ [2001] method to approximate the early exercise pre- mium m. The method involves approximating the early exercise boundary by a piecewise exponential function (Ju) in the original geometric Brownian motion scale or by a piecewise linear function (AitSahlia and Lai) in the Brow- nian motion scale resulting fom a change of variables. A major advantage of this approach is that it leads to closed~ foun approximations of Ue eally excicise premiiuun. Multiplying the discounted American option pre~ mium €°n(H, T 0 by the first-passage density (9 then gives the value of the integrand in (13). For fist valuation, the integral has to be approximated by a sum of m terms with small m. We evaluate it using Gaussian quadrature with m nodes and weight function f., which results in a weighted sum that approximates the integral in (13) (see Press et al. [1992, Section 4.5)} The weights 1,, ..., w, are determined together with modes fy, 05 fy SO that froscoa = Sebi) 0 ied for all polynomials h of degree less than 2m. To find these nodes and weights, a first step is to evaluate the moments af tpt) dt for k=0, 2, 2m=1 srmune 2004 Note that 4 is exactly the probability (under the risk-neutral measure) that the barrier is hit (ie., knock in occurs) during the life of the option. Let a = 2 - y and 6 =A, Recalling that n(z2) denotes the standard normal density and N(a) its cumulative distribution fimetion, we have: Similarly: bey al 3 =o( =) (Mt, — Ma) (16) where M, = 1 and M, = N@a/(o VT) + bVT /o). From (15) and (16): o=m-mton(%A) n=) «7 and G (a/B){2(N, - Ny) ~ Gg) if #0. ‘To compute the higher moments, note that by par- tial integration: e Bee? C4 = A aE ygaeh a + Te? es where pty (HF ) Therefore, in the case b #0: Ja? a Pe ena ~ Pina) + Heke If 6 =0, then: Tue JouRNAL or Dunwvanives 47eae ‘The moments éy «+; Gq €4n therefore be evaluated recursively. Let aoa ea 1 ae et ult) = det * & Ga a Ont pew | em Smt Gn sss Com Note that the polynomials qy(0), «-. qy(t) are orthogonal with respect to f()dt in [0, T]. The nodes thy sos fy ate the zeros of g, (0), and the weights w,, w,, are given by D’ leclimatay OF RA bem Finally, to evaluate the integral in (13) we use the approximation: > - [cent ote) atx Yen th Tt) (18) Fe choosing m to be small for fast computation IV. NUMERICAL ILLUSTRATION Our numerical examples for the two methods con~ sider both a short-maturity put option on a security paying no dividend (Exhibit 1) and a long-maturity pue option, ‘oma security paying a dividend ata constant rate (Evhibit 2). The middle four columns are generated by the mod~ ified binomial algorithm in Equations (10) and (12), for values of N equal to 1,000, 5,000, 10,000 and 20,000. ‘The expected convergence of the algorithm is clearly noticeable, ‘The last columns in the Exhibits present the integral approximation method with m= 2. In this case, the price ofthe American knock-in put is obtained via the approx- imation (18) of the integral in (13). Observe that with N’= 1,000, the modified bino~ ‘ial algorithm generally yields penny accuracy, and that the integral approximation with m = 2 nodes, which is over ten times faster, is even more accurate. The integral approximation involves ¢, given in (17), which is equal to, the probability of ever hitting the barrier during the life of the option, also tabulated in Exhibits 1 and 2. Note that one could improve the accuracy of the integral approximation by increasing the number m of nodes, thus resulting in higher-degree polynomials 4, (0). EXHIBIT 1 American Down-and-In Put Option Prices: Modified Binomial Algorithm and Integral Approximation (with m = 2) Prob. of Modified Binomial Integral S__H [Biting Banier! N= 1000 N=5000 = 10000 N=20000 ppiox. 75 70 | oss | 17.3026 173007173008 17.3008 | 17.3008 | 80 70 | 03001 | sso? 8873, B76 BaTos | B77 8S 70 | 01389 | ops ostt 4.0909 40907 | 4.0906 90 70 | oosss = | iiss 1739 amiss nist | 7136 ss 0 | 06271 | asta 12a362 12361124360 | 12.8386 oo mo | oases | Tesn Toast 7669 | “sso 9s so | oss | 3.6018 36903 3.6901 3.6900 | Rvs 100 80 | ooo 178621. 78S1_— 1.7850 1.7849 | 1.7847 95 90 | 0663 | 68017 68011680106 8DID. | 68050 wo so | oso | diz 4ni93, 92S zat Ws 90 | 02350 | 23402 © 23d 213 23H | 25 no_90 | ores | 1252 12836 12535 1.2532 | 1.2557 ‘Av, CPU time see) 01s 3602 SBI <0. 106, g=0, 0 =02, T=05, K= 100, 48. PRucive AND HEDGING OF AMERICAN KNOCKIN OPTIONS Srranc 2004EXHIBIT 2 American Down-and-In Put Option Prices: Modified Binomial Algorithm and Integral Approximation (with Mexlfied Binomial Icy 1000 N= 5000 N=10000_ N= 20000 | Approx. MATES 244768 24.4766 24.4765 | 24.4760 17960 178953 7.8930 17.8028 | 17-8922 12.7086 12.7022 12-7018 12.7016 | 12.7009 Balle 88081 8.8076 -8.8074_| 8.8070 6.0021 5.9988 5.9984 S9982_| 5.9978 195239195226 195224 19.5223 | 19.8217 H46ti6 14.6097 14.6094 14.6003 | 14.6090 10727 10.7704 — 10.7201 10.7199 | 10.7198 [Tra 773861 188STIseL | 188 | rsosss 1Sog7t 1510869 15.0868 | 15.0861 113987113272 113270 11.3368 | 11.3266 [83962 83946 83044 83942 83902 <0) ‘AN CPU time (sec-) Os 355 S41 «68.37 K 06, ¢=0.08, 0 = 0.2, ae 1 ‘Alternatively, one could achieve the same goal by dividing gy, a aoa the interval [0, T] into several subintervals of the same jz ~ ont {: +e (Yee | length, and then applying Gaussian quadrature to each subinterval, but with corresponding orthogonal polyno- mials of degree 2, resulting in similar formulas. Both approaches will converge to the correct value of the integral as the number of nodes (or subintervals) tends to infinity. The advantage of the second approach is that the nodes in each subinterval are easily calculated as they are zeros of a quadratic function. For the fitst approach, one must numerically determine all the zeros ofan mth-degree polynomial, Our numerical results sug- ‘gest that using only two nodes for the entire interval gives very accurate results in the first plac. V. AMERICAN KNOCK-IN HEDGE PARAMETERS From (6), it follows that the hedge parameter A is given by: aPp, Toe fe A ses)- [ vt. oe (eat (19) Since 2 = log S, the chain rule yields: 18 ae Moreover, from (9), it follows that: Sean 2004 ‘Therefore, as in (10), we can evaluate A via! 3S pont (67,46) 2 (p) A sue ay (e740 E08) 20) Similar expressions can be obtained for the hedge parameters gamma and theta. Note that (20) shows that the Bernoulli walk algorithm to compute A does not involve numerical differentiation. Alternatively, one can use the decomposition formula (13) to compute the hedge parameters. For example: = 2D, 7 cos fora» Here dp,/AS is given in closed form because of (14), and so is OF/8S. To evaluate the integral in (21), wwe can also use Gaussian quadrature. VI. SUMMARY We lve considered American knock-in options, for which the integral defining the early exercise pre~ ‘ium is very different fiom that of an American knock- out option. Despite the non-Markovian nature of the associated optimal stopping problem, we have been able to develop a modified binomial algorithm to price Amer ican knock-in options. We have also given an alternative ‘Tne Journ: oF Denwwarnes 49approach that makes use of the classic decomposition for- mula for a standard American option and computes the early exercise premium by Gaussian quadrature. The methods presented for American down-and-in put options can be modified for up-and-in put options and. for the corresponding call options. REFERENCES AitSahlia, F, L. Imhof, and TL, Lai, “Fast and Accurate Val- uation of American Barrier Options.” Jourtal of Computational Finance, vol. 7 (2003), pp. 129-148. AitSablia, F., and T.L, Lai. “Exercise Boundaries and ficient Approximations to American Option Prices and Hedge Param- eter.” Journal of Computational Finance, vo, 4 (2001), pp. 85-103, Banks, E. Complex: Derivatives. Chicago: Probus Publishing, 1994, Boyle, P., and S.H. Lau. “Bumping Up Against the Barrier with the Binomial Method.” The Joumal of Derivatives, vol. 2 (1994), pp. 6-14. Carr, P., R. Jarrow, and R. Myneni, “Alternative Character= izations of American Put Options." Mathematical Finance, vol 2 (1992), pp. 87-106. Cheuk, TLH.F, and T.C.P. Vorst “Complex Bartier Options” The Joumal of Derivatives, vol. 4 (1996), pp. 8-22. Figlewski,S.,and B. Gao. “The Adaptive Mesh Model: A New Approach to Efficient Option Pricing.” Joumal of Financial Eao- nomics, vol. 53 (1999), pp. 313-351 Go, B., J. Huang, and M.G, Subrahmanyam. “An Analyeical Approach to the Valuation of American Path-Dependent Options.” Journal of Economic Dynamics and Control, vol. 24 (2000), pp. 1783-1827, 50. Pricine AND Hlupeiss oF AMEIUCAN KNOCK-IN Orntons Jacka, S. D. “Optimal Stopping and the American Put.” Math- ‘ematical Finance, vol. 1 (1991), pp. 1-14, Js, N, “Pricing an American Option by Approximating its Early Exercise Boundary as a Multipiece Exponential Fune~ sion.” Review of Financial Studies, vol. 11 (1998), pp. 627-646 Karataas, 1. “On the Pricing of the American Option.” Applied Mathematics and Optimization, vol. 17 (1988), pp. 37-60. Karataas Land S. Shreve. Brownian Motion and Stochastic Cal- cudus. New York: Springer-Verlag, 1988, Kim, IJ. “The Analytic Valuation of American Options.” The Review of Financial Studies, vol. 3 (1990), pp. 547-572. Merton, R. “Theory of Rational Option Pricing,” Bell journal of Economics and Management Science, 4 (1973), pp. 141-183. Press, W.,S, Teukolsky, W. Verterling, and B. Flannery. Numer ‘eal Resipes in C: The Art of Scientific Computing, 2nd ed, New ‘York: Cambridge University Press, 1992, Ritchken, P. “On Pricing Barrier Options." The Joumal of Derivatives, vol. 3 (1995), pp. 19-28, Rubinstein, M,, and E. Reiner, “Breaking Down the Barties.”, Risk, 4 (1991), pp. 28-35, ‘To order reprints of this article, please contact Ajani Mal at malik@ijournalscom or 212-224-3205. Sense 2004