Anda di halaman 1dari 14

in Arbitrage-Free Models Arbitrage Opportunities of Bond Pricing

David BACKUS New York,NY 10012-1126 (dbackus@stern.nyu.edu) Stern School of Business, New YorkUniversity, Silverio FORESI Salomon BrothersInc., New York,NY 10048 Stanley ZIN PA 15213-3890 GraduateSchool of Industrial Administration, Carnegie MellonUniversity, Pittsburgh,
modelsof bondpricingare used by both academicsand Wall Streetpractitioners, Mathematical with practitioners to modelsto selected introducing time-dependent parameters fit "arbitrage-free" assetprices.We show,in a simpleone-factor a setting,thatthe abilityof suchmodelsto reproduce subsetof securitypricesneed not extendto state-contingent claimsmoregenerally. arguethat We of models shouldbe complemented close attention the additional to parameters arbitrage-free by which mightincludemeanreversion, fundamentals, multiplefactors,stochasticvolatility,and/or nonnormal interest-rate distributions. KEY WORDS: Fixed-income claims. derivatives; Options; Pricingkernel;State-contingent

Since Ho and Lee (1986) initiated work on "arbitragefree" models of bond pricing, academics and practitioners have followed increasingly divergent paths. Both groups have the same objective-to extrapolate from prices of a limited range of assets the prices of a broader class of statecontingent claims. Academics study relatively parsimonious models, whose parametersare chosen to approximate "average" or "typical"behavior of interest rates and bond prices. Practitioners, on the other hand, use models with more extensive sets of time-dependent parameters, which they use to match current bond yields, and possibly other asset prices, exactly. To practitioners, the logic of this choice is clear: The parsimonious models used by academics are inadequate for practical use. The four parameters of the one-factor Vasicek (1977) and Cox-Ingersoll-Ross (1985) models, for example, can be chosen to match five points on the yield curve (the four parameters plus the short rate) but do not reproduce the complete yield curve to the degree of accuracy required by market participants. Even more general multifactor models cannot generally approximate bond yields with sufficient accuracy. Instead, practitioners rely almost universally on models in the Ho and Lee (1986) tradition, including those developed by Black, Derman, and Toy (1990), Black and Karasinski (1991), Cooley, LeRoy, and Parke (1992), Heath, Jarrow, and Morton (1992), Hull and White (1990, 1993), and many others. Although analytical approaches vary across firms and even within them, the Black-Derman-Toy model is currently close to an industry standard. Conversely, academics have sometimes expressed worry that the large-parameter sets of arbitrage-free models may mask problems with their structure. A prominent example is Dybvig (1997), who noted that the changes in parameter values required by repeated use of this procedure contra13

dicted the presumption of the theory that the parameters are deterministic functions of time. Black and Karasinski (1991, p. 57) put it more colorfully: "When we value the option, we are assuming that its volatility is known and constant. But a minute later, we start using a new volatility. Similarly, we can value fixed income securities by assuming we know the one-factor short-rate process. A minute later, we start using a new process that is not consistent with the old one." Dybvig argued that these changes in parameter values through time implied that the framework itself was inappropriate. We examine the practitioners' procedure in a relatively simple theoretical setting, a variant of Vasicek's (1977) onefactor Gaussian interest-rate model that we refer to as the benchmark theory. Our thought experiment is to apply models with time-dependent drift and volatility parameters to asset prices generated by this theory. We judge the models to be useful, in this setting, if they are able to reproduce prices of a broad range of state-contingent claims. This experiment cannot tell us how well the models do in practice, but it allows us to study the role of time-dependent parameters in an environment that can be characterized precisely. We find, in this environment, that, if the world exhibits mean reversion, then the use of time-dependent parameters in a model without mean reversion can reproduce prices of a limited set of assets but cannot reproduce the prices of general state-contingent claims. In this sense, these arbitragefree models allow arbitrage opportunities: A trader basing prices on, say, the Black-Derman-Toy model can be exploited by a trader who knows the true structure of the economy.
9 1998 American Statistical Association Journal of Business & Economic Statistics January 1998, Vol. 16, No. 1

14

Journalof Business & EconomicStatistics,January1998

Considera one-periodbond.FromEquation(5) and the initial conditionb? = 1, we see that the price is the conditionalmean of the pricingkernel:b1= Etmt+l. Because the kernelis conditionally we lognormal, needthe following of lognormal randomvariables: log x is normal If property with mean p and variancea2, then log E(x) = p + a2/2. 1. A THEORETICALFRAMEWORK From we see that log mt+l has conditional We use a one-factorGaussianinterest-rate model as a mean Equation(3) and conditionalvariance(Aa)2. The one-period -zt in procedure bond laboratory whichto examinethe practitioners' price satisfies log b1 = -zt + (Aa)2/2 and the short to of choosingtime-dependent parameters fit a bond-pricing rate is assetprices.The modelis a close relative modelto observed of one describedby Vasicek(1977). Althoughin some rert = - log b1 = zt - (Aa)2/2. (6) its spectsit is simplerthanthoseused by practitioners, logis linear structure extremelyuseful in clarifyingthe roles Thus, the shortrate r is the state z with a shift of origin. The meanshortrate is 6 - (Aa)2/2, which we denoteby p by its variousparameters. played To fix the notation,let bVbe the price at date t of a in the rest of the article. bond of maturity the claim to one dollarat n, The stochastic zero-coupon (2), processfor z, Equation impliessimilar date t + n. By conventionb - 1 (one dollartoday costs behaviorfor the shortrate:
one dollar). Bond yields are y7 = -n-1 log bZ andforward rates are ft" = log(bt"/bn+ ). We label the short rate rt = yl f,.
rt+ = rt + (1 - P)(p - rt) + Et+l,
(7)

A strikingexampleof mispricing this settinginvolves in options on long bonds. Options of this type vary across two dimensionsof time, the expirationdate of the option andthe maturity the bondon whichthe optionis written. of modelhas a one-dimensional The Black-Derman-Toy array that can be chosen to reproduce of volatility parameters dateon oneeitherthe pricesof optionswith anyexpiration periodbonds or of optionswith commonexpirationdates on bonds of any maturity. But, if the world exhibitsmean this the reversion, vectorcannotreproduce two-dimensional arrayof pricesof bondoptions.If the volatilityparameters are chosento matchpricesof optionson one-period bonds, then the model overpricesoptionson long bonds. Althoughwe think that mean reversionis a useful feature in a model, it illustratesa more generalpoint-that of whatevertheir form, misrepresentations fundamentals, cannot generallybe overcomeby adding arraysof timeFrom this perspective,the role of dependentparameters. academicresearchis to identifyappropriate fundamentals, which might includemultiplefactors,stochasticvolatility, We movements. conjecture interest-rate and/or nonnormal that arbitrage-free models with inaccuracies along any of these dimensionswill mispricesome assets as a result.

-log mt+l = zt +/ACt+l.

(3)

The parameter which we refer to as the price of risk, A, determines covariance the betweeninnovations thekernel to andthe state and thus the risk characteristics bondsand of relatedassets. Given a pricingkernel,we deriveprices of assets from the pricingrelation 1 = Et(mt+xRt+x), (4)

which holds for the gross returnRt+l on any tradedasset. Because the one-periodreturnon an n + 1-periodbond is b 1/b"+1,the pricingrelationgives us i+
bn+1 = Et(mt+lbtn,+),

(5)

whichallowsus to computebondpricesrecursively, starting


with the initial condition bt = 1.

(1)

a discrete-time versionof Vasicek's(1977) short-rate diffusion. Futurevalues of the shortrate are

asset pricesin ourbenchmark We characterize theory,or with a pricing kernel,a stochasticprocessgovlaboratory, claims.Existenceof such erningprices of state-contingent

a process is guaranteed in any arbitrage-free environment. for n _ 1, which yields conditional first and second moWe describe the kernel for our theoretical environment in ments of two steps. The first step involves an abstract state variable z, whose dynamics follow: (8) Et(rt+n) = rt + (1 - (P)(p - rt)
Zt+i =

pzt + (1 - p)6

+et+l
(2)

and

= zt + (1 - cp)(6- zt) +et+1,

with {et} distributed normally and independently with mean 0 and variance a2. The parameter controls mean reversion: With = 1, the state follows a random walk, but with values between 0 and 1, the conditional mean of future values of z converges to the unconditional mean 6. Step 2 is the pricing kernel m, which satisfies

= vart(rt+n) 2

j=1

) 2(nj-

2 1- 2

(9)

We return to these formulas later. Prices of long bonds follow from (5); details are provided in Appendix A.1. Their properties are conveniently summarized by forward rates, which are linear functions of

in Modelsof BondPricing Backus,Foresi,and Zin:Arbitrage Opportunities Arbitrage-Free

15

the shortrate: fth= rt + (1-

Table1. Propertiesof U.S. Government Bond Yields Maturity


pn)(p

Mean 7.483 7.915 8.190 8.372 8.563 9.012 9.253 9.405 9.524 9.716 9.802

Standarddeviation 1.828 1.797 1.894 1.918 1.958 1.986 1.990 1.983 1.979 1.956 1.864

Autocorrelation .906 .920 .926 .928 .932 .940 .943 .946 .948 .952 .950

- rt) + [A2

(A?

2/2

(10) for all n > 0. Given forwardrates, we can computebond prices and yields from their definitions.The right side of If Equation(10) has a relativelysimple interpretation. we compareit to (8), we see that the first two terms are the expectedshortraten periodsin the future.We referto the last term as a risk premiumand note that it dependson three parameters-the magnitudeof risk (a), the price of risk (A), and meanreversion(po). Both the Ho and Lee (1986) and Black, Derman,and an Toy (1990) models are capableof reproducing arbitrary forwardratecurve (equivalently, yields or bondprices),inmodel like this one. cludingone generated a theoretical by An issue we addresslateris whetherthis capability extends to more complexassets. With this in mind,considera Eudate t + 7 and ropeancall option at date t, with expiration strike price k, on a zero-couponbond with maturityn at expiration.Given the lognormalityof bond prices in this setting,the call price is given by the Black-Scholes(1973) formula,
cn" = br+nN(dx) - kb[N(d2),

1 month 3 months 6 months 9 months 12 months 24 months 36 months 48 months 60 months 84 months 120 months

NOTE: Dataare monthly estimatesof annualized, U.S. compounded, continuously zero-coupon bondyieldscomputed McCulloch Kwon and is government by (1993).Thesampleperiod January 1982 to February 1992.

month.FromEquation we see thatp is the unconditional (7) meanof the shortrate,so we set it equalto the samplemean of the one-month yield in Table1, 7.483/1,200. (The 1,200 convertsan annual rate percentage to a monthlyyield.)The mean-reversion of ?o parameter is the first autocorrelation the shortrate.In Table 1 the autocorrelation .906, so we is set ?pequal to this value. This indicatesa high degree of persistencein the short rate but less than with a random walk. The volatilityparameter is the standard deviation a of innovationsto the short rate, which we estimate with the standard errorof the linearregression(7). The resultis to matchthe mean,standard deviation,and autocorrelation of the shortrate.We choose the final parameter, price the of risk A,to approximate slope of the yield curve.Note the from (10) thatmeanforwardrates,in the theory,are
E(f)
=

a = .6164/1,200. Thus the values of (p,ua, o) are chosen

(11)

where N is the cumulative normaldistribution function,


log[bl+n/(bVk)] + v2 /2 V1-,n
d2 dl vr,n,

p+

A2

A+

a2/2.

and the option volatility is

(log v2,n= vart b

)=

1-2

1-

2.

(12)

See Appendix A.4. Jamshidian (1989) reported a similar formula for a continuous-time version of the Vasicek model. The primary difference from conventional applications of the Black-Scholes formula is the role of the mean-reversion

parameter in (12). cn
2. PARAMETERVALUES

The parametervalues used in the following numerical examples come from an informal moment-matching exercise based on properties of monthly yields for U.S. government securities computed by McCulloch and Kwon (1993). Some of the properties of these yields are reported in Table 1 for the sample period 1982-1991. We choose the parameters to approximate some of the salient features of bond yields using a time interval of one

we see that A = -750 produces theoretical mean yields (the line in the figure) close to their sample means (the stars) for maturities between one month and 10 years. With more negative values, the mean yield curve is steeper, and with less negative (or positive) values, the yield curve is flatter (or downward sloping). Thus we see that all four parameters are required for the theory to imitate the dynamics of interest rates and the average slope of the yield curve. We use these benchmark values later to illustrate differences in prices across bondpricing models. 3. HO AND LEE REVISITED

This tells us that to producean increasingmean forwardrate curve (implyingan increasingmean yield curve) we need A to be negative.The price of risk parameter, other in words, governsthe averageslope of the yield curve. One way of fixing A, then, is to select a value that makes the theoreticalmean yield curve similarto the sample mean yield curve, given our chosen values for the other three An parameters. exampleis picturedin Figure 1, in which

We turn now to the use of time-dependent parameters to fit theoretical models to observed asset prices. We apply,

16

Journalof Business & EconomicStatistics,January1998

rates implied by this model are


n 9.5

rt+n

rt +

Y
j=1

(at+j + rlt+j),

(16)

9-

which implies conditional means of


n

Et (rt+n)
8-

Tt + Zat+j
j=1

(17)

for n > 0. If we compare this to the analogous expression for the benchmark theory, Equation (8), we see that the two are equivalent if we set
n

20

40

80 60 in Maturity Months

100

120

Z
j=1

at+,

(1 - yn)(

Tt).

(18)

Figure 1. Mean Yieldsin Theoryand Data. The line is theory,the stars are data points.

in turn, analogs of the models of Ho and Lee (1986) and Black, Derman, and Toy (1990) to a world governed by the benchmark theory of Section 1. Our first example is a Gaussian analog of Ho and Lee's (1986) binomial interest-rate model. The analog starts with a state equation, zt+l = zt + at+l + rlt+, (13)

Thus, the time-dependent drift parameters of the Ho and Lee model can be chosen to imitate this consequence of mean reversion in the benchmark theory. In practice it is more common to use the drift parameters to fit the model to the current yield curve. To fit forward rates generated by the benchmarktheory, we need [compare (10) and (15)]
n

j=1

Zat+j

rt) - (1- ,9n)4I -

with time-dependent parameters {at} and normally and independently distributed innovations {t } with mean 0 and constant variance 32. The pricing kernel is
- log mt+ = zt + 7r7t+1. (14)

+ [A2- (A+ (1 )/(1 - (p))2]a2/2 - [Y2- ( + n)2132/2. (19)


The drift parameters implied by (18) and (19) are, in general, different. Because

1 - (pn =, n lim 1 We use different letters for the parameters than in the o--1 benchmark theory to indicate that they may (but need not) take on different values. Our Ho and Lee analog differs from we can equate the two expressions when ? = 1 by setting the benchmark theory in two respects. First, the short-rate 3 = a and y = A. But when 0 < p < 1 the two expressions cannot be reconciled. This is evident in Figure 2, in which process does not exhibit mean reversion, which we might think of as setting ? = 1 in Equation (2). Second, the state equation (13) includes time-dependent "drift" parameters {at}. 4Given Equations (13) and (14), the pricing relation (5) Match 3 ,' ExpectedShortRates implies a short rate rt = zt - (y3)2/2 and forward rates = fZZ rt +
j= 1

at+j + [.2 - (y + n)2]/32/2

(15)

for n > 1. See Appendix A.2. Note that (10) differs from (15) of the benchmark in two ways. One is the impact of the short rate on long forward rates. A unit increase in r is associated with increases in fn of 1 in Ho and Lee, but (1p") < 1 in the benchmark. The other is the risk premium, the final term in Equation (15). Despite these differences, time-dependent drift parameters allow the Ho and Lee model to reproduce some of the features of the benchmark theory. One such feature is the conditional mean of future short rates. The future short

-4 -3

Match BondYields Current

-50

15

25 30 20 Time Horizon in Months

35

40

45

50

Figure 2. Two Choices of Ho and Lee DriftParameters. One matches the current bond yields. The other matches expected future short rates.

in Modelsof Bond Pricing Backus,Foresi,and Zin:Arbitrage Opportunities Arbitrage-Free

17

we graphthe two choices of cumulativedrift parameters, n estimatedin Section 2, =, at+j, using the parameters with / = a, -y A, and r = 3.0/1,200. The drift parameters meanconvergerapidlyas the the thatreproduce conditional effects of meanreversion wearoff. But the driftparameters that fit the currentyield curve get steadilysmalleras they offset the impactof maturityon the risk premiumin this in model.This results,for the rangeof maturities the figure, in a decliningtermstructure expectedfutureshortrates, of whereasthe benchmark implies the reverse. The discrepancy the figurebetweenthe two choices of in driftparameters a concreteexampleof the following: is Remark1. The parameters the Ho and Lee model of can be chosento matchthe current yield curveor the conditionalmeansof futureshortrates,butthey cannotgenerally do both. Thisproperty the Ho andLee modelis a hintthattimeof drift parameters not adequatelycapturethe do dependent effects of meanreversionin the benchmark theory.Dybvig A (1997) made a similarobservation. closer look indicates thatthe difficulty in the nonlinear lies in interaction the risk betweenmeanreversionand the price of risk. In premium the benchmark theory,the risk premiumon the n-period forwardrate [see Eq. (10)] is
A2A- A+ -

120
.100

o 8
20

10 15 Timeto Expiration Months in

20

25

of Figure3. The Premium Ho and Lee CallPrices Overthe BenchmarkPrice.

intuitionaboutoptionsis easily verified. Black-Scholes The formula,Equation (11), appliesto the Ho andLee model if we use optionvolatility
v,n vart(1ogb,)
= T(n3)2.

(20)

2/2.

This formulacannotbe reconciledwith that of the benchmarktheory,Equation(12), for all combinations T and of the reproduce option volatilityv1,l of a short option on a shortbond,thenwe overstate volatilitiesof long options the on long bonds. As a result, the model overvaluesoptions with more distantexpirationdates and/or longer underlying bonds. We see in Figure 3 that call prices c,'1 impliedby the Ho and Lee model can be substantially higherthan those The figure expresses generatedby the benchmark theory. the mispricingas a premiumof the Ho and Lee price over the price generatedby the benchmark theory.Benchmark values of Section 2. Ho prices are based on the parameter andLee pricesarebasedon driftparameters matchcurthat rent bond prices,Equation(19), and a volatilityparameter
For = 1 the two models generate n unless p = 1. If we choose the volatility parameter 0 to

In the Ho and Lee model,Equation(15), the analogousexpression is [y2 - (y + n)2]032/2. If we choose - = A and
0 = a, the two expressions are equal for n = 1, but they

move apartas n grows. The discrepancy noted in Remark 1 is a directconsequence. A similarcomparison conditional of variances shows also of strain.The conditionalvariancesof future short signs rates implied by the Ho and Lee model are vart(rt+n) = np2. If we comparethis to the analogousexpressionin the benchmark theory,Equation(9), we see thatthey generally = differ when Wp 1. With p < 1 and 0 = a, the conditionalvariances the samefor n = 1, but for longertime are horizons they are greaterin the Ho and Lee model. We 0 = a that matches the option volatility v1,1 of a one-period summarize discrepancy the following: this in option on a one-periodbond. Both are evaluatedat strike
cannot be chosen to reproduce the conditional variances of future short rates.
T Remark2. The parameters the Ho and Lee model the same call of price, but for options with expirationdates

price k = b/b+n.

12 months in the future the Ho and Lee price is more than 50% higher. 4. BLACK,DERMAN,AND TOY REVISITED

Thus, we see that additional drift parameters allow the Ho and Lee model to imitate some of the effects of mean reversion on bond yields. They cannot, however, reproduce the conditional variances of the benchmark theory. Dybvig (1997) summarized this feature of the Ho and Lee model more aggressively: "[T]he Ho and Lee model starts with an unreasonable implicit assumption about innovations in interest rates, but can obtain a sensible initial yield curve by making an unreasonable assumption about expected interest rates. Unfortunately, while this ... give[s] correct pricing of discount bonds ..., there is every reason to believe that it will give incorrect pricing of interest rate options."Dybvig's

Black, Derman, and Toy (1990) extended the timedependent parametersof Ho and Lee to a second dimension. They based bond pricing on a binomial process for the logarithm of the short rate in which both drift and volatility are time dependent. We build an analog of their model that retains the linear, Gaussian structure of previous sections but includes these two sets of time-dependent parameters. Given a two-parameter distribution like the normal, timedependent drift and volatility can be used to match the conditional distribution of future short rates exactly and thus

18

Journalof Business & EconomicStatistics,January1998

to mitigatethe tendencyof the Ho andLee model to over- This implies conditional of variances The questionis whetherthey also allow pricelong options. us to reproduce prices of other interest-rate-derivative the j) vart(rt+n) o=2EW2 (X-a securities. j=1 Ouranalogof the Black-Derman-Toy model addstimeto the structure Section3: A random the same as Equation(9) of the theory.Similarpatterns of volatility dependent of decliningtime-dependent volatilitiesare commonwhen variablez follows these methodsare used in practice,Black et al.'s numeri(21) cal exampleincluded(see theirtableI). To matchthe conZt+1= Zt+ at+1 + 7?t+l, ditionalmean of the benchmark theory,Equation(8), we choose drift parameters satisfy that the pricingkernelis and
- log mt+ = Zt + yTt+1. (22)
j=1
-Zat+j (1 - p)(p -

rt)

2(t+l

t+n+l)/2.

(28)

Given the critical role played by volatility in pricing derivativeassets, this representsan essential advancebeft = rt + at+j + [72 - ( + n)2]t2+1/2 yond Ho andLee. j=1 the Despite time-dependent volatilityparameters, model the conditional moments cannotsimultaneously 2 n reproduce + + n _ j)2( o2 of the short rate and the forwardrates of the benchmark + + - )2 ;+j - t+j+l)/2 (23) n the rate that The driftparameters reproduce forward theory. j=1 are curve,Equation(10), for n > 1. Equation (23) reducesto the Ho andLee expression, Equation(15), when 3t = /3 for all t. As in the Ho in ratesdifferfromthe benchmark andLee model,forward movementson long forward both the impactof short-rate ratesand the form of the risk premium. We can use both sets of time-dependent to - [T2- (y + n)2]t2+1/2 parameters assetpricesin the benchmark theory.Consider approximate n of distribution futureshortrates.The short the conditional +n -) (It+j 3- t+j+l)/2. S (29) rate follows rr+1 = rt + 2 21+ -t+2)/2 + at+1 j=1 so futureshortratesare Comparing(28) and (29), we see that the two are not n in We note the differencein the fol2 (2 _o2 (24) equivalent, general. rt+n = rt2+t+ (at+j +,7t+j). 1-- t+n+l )/2+ lowing: j=1 4. Remark Given volatilityparameters (27) that repromeanand varianceare Theirconditional the conditional variances futurerates,the driftpaof duce model can be chosen rametersof the Black-Derman-Toy + = + (25) to matchthe currentyield curve or the conditional t+n+ 1)/2 l- at+j rt Et(rt+n) 72-(/+1i means of futureshortrates,but they cannotgenerallydo both. j.=1 and Figure4 plots the differencebetween(29) and (28). The parametervalues are those of Section 2, with y = A,/3t+j =
n

is The new ingredient thateachrqthas time-dependent varidiffersfrom the benchmark ance 3t2.This structure theory in its absenceof meanreversion(the coefficientof 1 on zt drift (the in the state equation)and in its time-dependent a's) and volatility(the /'s). bond pricing as before. We show in ApWe approach pendix A.2 that the short rate is rt = -log(Etmt+1) = ratesare Zt - (y3t+1)2/2 and forward

Thus we see, as Black et al. (1990, p. 33) suggested,that we can fit the firsttwo momentsof the shortrate with two of "arrays" parameters: Remark3. The parameters the Black-Derman-Toy of model can be chosen to reproducethe conditionalmeans and variances futureshortrates. of

those of Figure2 for the Ho andLee modelbutarenonzero j=1 nonetheless. Remark4 is reminiscentof Remark1 for the Ho and This model, in contrastto that of Ho and Lee, is able to in matchthe conditionalvariancesof the benchmark theory, Lee model.Moresurprising, our view, is thatthe Blackthat whichwe do by choosingvolatilityparameters decline Derman-Toymodel's ability to reproducethe conditional varianceof future short rates does not extend to option geometrically: prices. Consider Europeancall options on zero-coupon = pj-lo. (27) bonds. Once more the lognormalstructureof the model =t+j

vart(rt+n) =

3t+j.

(26)

qj-l

, and r = 3.0/1,200. The differences are smaller than

in Modelsof Bond Pricing Backus,Foresi,and Zin:Arbitrage Opportunities Arbitrage-Free


1.4

19

1.2
0 0

> .8 0

C0.60
CD

0.4

0.2

10

15

20 30 35 25 in TimeHorizon Months

40

45

50

DifferenceBetween Black-Derman-ToyDrift Figure4. Cumulative Bond Yieldsand Expected ParametersChosen to Matchthe Current FutureShort Rates.

means that the Black-Scholes formula,Equation(11), apto plies. If we choose drift parameters fit the currentyield betweencall pricesin the model then any difference curve, and the benchmark theorymust lie in theiroptionvolatilianaties. The option volatilityfor the Black-Derman-Toy is log
2 =n2 y
j=1

See AppendixA.4. If we restrictourselvesto options on one-periodbonds (so that n = 1), we can reproducethe volatilitiesof our theoreticalenvironment choosingf's by with time, the same choice that that decline geometrically of variances futureshortrates,Equareplicatesconditional tion (27). The Black-Derman-Toy analogcannot,however,simultaneouslyreproduceprices of options on bonds of longer If maturities. we use the geometrically decliningparameters of Equation(27), the optionvolatilityis
V-rn -=

of Remark5. The parameters the Black-Derman-Toy model cannotreproduce pricesof call optionson bonds the for all maturities expiration and dates. The differencein option volatilities, and hence in call prices, between the two models stems from two distinct roles played by mean reversionin determiningprices of first,in the impactof long bonds.Meanreversionappears, short-rate innovationson future short rates. We see from on (9) that the impactof innovations futureshortrates,in the theory,decays geometricallywith the time horizon T. This featureis easily mimickedin the Black-Derman-Toy model,as we have seen,by using volatilityparameters } {IOt that decay at the same rate. The second role of mean reversion concernsthe impact of short-ratemovementson model, a unit long-bondprices.In the Black-Derman-Toy decreasein the shortrateresultsin an n unitincreasein the logarithmof the price of an n-periodbond;see Appendix A.2. In the benchmark theory,the logarithmof the bond rises by only (1 +p+.. - + 9-l) = (1 -_")/(1price p); see AppendixA.1. This attenuation the impactof shortof rate innovationson long-bond prices is a direct consequence of mean reversion.It is not, however,reproduced by choosing geometricallydeclining volatility parameters and is therefore missing from call prices generatedby the Black-Derman-Toyanalog. Stated somewhat differently, the Black-Derman-Toyanalog does not reproduce the hedge ratiosof the benchmark theory. As a practicalmatter,then, we might expect the Blackto Derman-Toy procedure work well in pricingoptionson short-term instruments, includinginterest-rate caps.Foroptions on long bonds,however,the model overstatesthe option volatilityandhencethe call price.A commonexample of suchaninstrument a callablebond.Thisprocedure is will overstatethe valueof the call provisionto the isgenerally and the suer,in the theoretical environment, thusunderstate value of a callablelong-termbond. The discrepancy betweenoptionprices generated the by benchmark theoryandthe Black-Derman-Toy analogillus150

1--2

n2 r2 ) "n0

From (12) we see thatthe ratioof optionvolatilitiesis


BDT volatility benchmark volatility n2 (1 + p +
...

CD
C100

o-

+ p"n-1)2
U)

This ratio is greater than 1 when 0 < p < 1 and n > 1 and implies that, when the Black-Derman-Toy analog prices options on short bonds correctly, it will overprice options on long bonds. We see in Figure 5 that this mispricing gets worse the longer the maturity of the bond and is greater than 150% for bonds with maturities of two years or more. As in Figure 3, benchmark prices are based on the parameters of Section 2, the drift parameters of the Black-Derman-Toy analog are chosen to reproduce the current yield curve, the strike price is k = bT/b+,t, and the expiration period is 7 = 6. Thus we have the following:

O 50

0o

10

15

20

25

Bond Maturityin Months

of CallPrices Overthe Figure5. The Premium Black-Derman-Toy Benchmark Price. Expiration period7 is 6.

20

Journalof Business & EconomicStatistics,January1998

trates one difficultyof using models with time-dependent 0.05 vector of time-dependent parameters-a one-dimensional cannotreproduce conditional the varivolatilityparameters 0.04 ances of bond prices across the two dimensionsof matuWe rity and time-to-expiration. turn now to a second ex0.03 whose of mispricing,a class of "exotic"derivatives ample returnsdisplaydifferentsensitivitythanbonds to interestrate movements. - 0.02 Consideran asset that deliversthe power 0 of the price of an n-periodbondone periodin the future.This assethas with magnifiedsensitivity 0 0.01 some of the flavorof derivatives movements madepopular BankersTrust to interest-rate by of retainsthe convenient log-linearity bondpricesin our yet 0 the As framework. with options,we compare pricesimplied by the benchmark theory, -0.01, -10
0

-8

-6

-4

log dZ'= -(rt + (Ao)2/2) + 0 log b'

4 -2 0 2 Parameter Theta Sensitivity

10

of Prices of the Exotic Figure6. The Premium Black-Derman-Toy Overthe Benchmark Price.Maturity is 60. n

(A+ 01

)2c

2/2 - 0(1 p1)(/

- rt), (30)
n

with those generated the Black-Derman-Toy analog, by


= log d-n(rt

Mt,t+n -

mt+j,
j=1

+ (7y3t+1)/2)+0 log t bT
+ (,y + On)2); +1/2 - 0(nat+l)
n
-

)2 2

which define n-period-ahead state prices. We show in the thatthe benchmark Appendix theoryimplies discountfactors
- log Mt,t+, 3t+j)2/2]. (31)

0>1 [(n
j=1

-- j)-(t+j+l

--Ot+j)

(-y+ n - j)2(3t2j+1

n6 +

(rt - IL)
+ 1 :)A+ ?0 Et+j

j=11 Both follow from pricingrelation(4) with the appropriate andthe Black-Derman-Toy analogimplies pricingkernel. Suppose we choose the parametersof the Black- - log Mt,t+n = n[rt + (7l3t+1)2/2] n n Derman-Toyanalog to match the conditionalvarianceof future short rates [Eq. (26)], the currentyield curve [Eq. + :(n - j)at+j + (7 + n j)rnt+j. and the price of risk [y = A]. Can this model re(29)], j=1 j=1 producethe prices of our exotic asset for all values of the As long as (o , 1 and a , 0, the secondexpressioncannot 0? sensitivityparameter The answeris generallyno. When to n - 1, the two models generatethe same price d' for all be madeequivalent the first: values of 0, just as we saw that the two sets of cumuladrift and volatility Proposition1. The time-dependent were initiallythe same (see Fig. 4 and parametersof the Black-Derman-Toymodel cannot be tive driftparameters the discussionfollowingRemark4). For longermaturities, chosen to reproduce stochasticdiscountfactorsof the the
benchmark theory. A proof is given in Appendix A.3. The difficulty lies in the innovations + 1 (pl
t+j

however, the prices are generally different. Figure 6 is an example with n = 60 and r = 3.00/1,200. We see that, for values of 8 outside the unit interval, the Black-DermanToy analog overprices the exotic, although the difference between models is smaller than for options on long bonds. Remark 6. The parameters of the Black-Derman-Toy model can be chosen to match both the current yield curve and the term structureof volatility for options on one-period bonds, but they cannot generally replicate the prices of more exotic derivatives.

These examples of mispricing illustrate a more general result-the Black-Derman-Toy analog cannot replicate the state prices of the benchmark theory. This result is stated most clearly using stochastic discount factors,

and (7 + n - j)rlt+j. There is no choice of volatility parameters {f3t} in the Black-Derman-Toy analog that can replicate the nonlinear interaction of mean reversion () and the price of risk () in the benchmark theory. In short, the time-dependent drift and volatility parameters of a model like the Black-Derman-Toy analog cannot replicate the prices of derivative assets generated by a model with mean reversion.

in Backus,Foresi,and Zin:Arbitrage Modelsof Bond Pricing Opportunities Arbitrage-Free

21

5. ARBITRAGE AND PROFIT OPPORTUNITIES In the laboratoryof the benchmark theory,the BlackDerman-Toyanalog systematically mispricessome assets. In this section we consider strategiesthat might be used to profitfrom tradersusing prices indicatedby the BlackDerman-Toymodel. One way to exploit someonetradingat Black-Dermanwith someone tradingat benchToy prices is to arbitrage mark prices. If, as we have seen, a Black-Derman-Toy traderoverpricesoptions on long bonds relativeto those on shortbonds and a benchmark traderdoes not, then we can buy from the latterand sell to the former,thus making a risk-freeprofit.This is as clearan exampleof arbitrage as thereis. But becausesuchprice differencesare so obvious, they are unlikelyto be very common. It is less easy to exploit a Black-Derman-Toytrader when tradestakeplace only at pricesdictated the model. by Because Black-Derman-Toyprices are consistentwith a We pricingkernel,they precluderisklessarbitrage. can often devise dynamicstrategies,however,whose returnsare large relativeto their risk. One such strategyinvolvescall options on long bonds, which (again)the Black-DermanIf Toy modelgenerallyovervalues. we sell the optionto the traderat date t, and liquidateat t + 1, we might expect to makea profit.This seems particularly likely in the option's finalperiodbecausethe optionis overvalued the bonds but on which the optionis writtenare not. we Before we evaluatethis strategy, need to explainhow the Black-Derman-Toy trader time.A pricesassetsthrough traderusing the Black-Derman-Toy modelgenerallyfinds, bothin practiceandin ourtheoretical setting,thatthe timemust be reset each period. In our dependentparameters setting, suppose that the traderchooses volatility parameters at date t to fit implied volatilities from options on shortbonds. As we have seen, this leads the traderto set W2 13t+2, 3t+3,... } equal to {a, cpr, a,...}. If the pa{ft+1, then in the followrameterswere literallytime dependent, ing periodlogic dictatesthat we set {ft+2, /t+3, /t+4, ..
0.55 0.s550.5
0.45
0.4

350

300

Ho-Lee

w250

BDT

a 200
150 -

d 100 2 50 0 0 5 10 15 BondMaturity Months in Benchmark 20

Excess Returns. Figure8. AverageRisk-Adjusted

equalto {scp, p2a,V3a,... }. But if we calibrateonce more to options on one-periodbonds, we would instead use This leads to a predictable upward jump {a,lr, pV2a, ...}. in the volatility parameters from one period to the next, as in Figure7. Similarly,the drift parameters must be adeachperiodto retainthe model'sabilityto reproduce justed the currentyield curve.These changesin parameter values time are an indication the modelis imperfectly that through imitatingthe processgeneratingstate prices,but they neverthelessimproveits performance. this sense the internal In notedby Dybvig (1997) is a symptomof uninconsistency derlyingproblemsbutnot a problemin its own right.Thus, our traderprices call optionsusing drift and volatilityparametersthat match,each period,the yield curve and the term structureof volatilities implied by options on oneperiodbonds [Eqs.(29) and (27), respectively]. Now considera strategyagainstsuch a trader sellinga of on an n-periodbondandbuyingit backone option 7--period returnfrom this strategy periodlater.The gross one-period is
Rt+1 = -c
-1,n/cT

RevisedVolatility Curve

for T > 1, with the conventionthat an option at expiration has value c?'" = (b" - k)+. The profit from this strategy is not risk free, but it can be large relativeto the risk involved. In the world of the benchmark theory, the appropriate adjustment for risk is given by the pricing relation (4). We measure the mean excess risk-adjusted return by a = E(mt+lRt+l) - 1, using the pricing kernel m for the

',

~0.35

Initial Volatility Curve

0.25

0.2

benchmark is to economy.This measure analogous Jensen's alphafor the capitalassetpricingmodelandis the appropriate risk adjustment this setting.For the tradingstrategy in values of Section 2, the just describedand the parameter returnsare in the neighborhood severalhundred of percent per year;see Figure8. The returnsin the figurewere computed by Monte Carlo and concern two-period (7 = 2) opDatein Months

Figure 7. Initial and Revised VolatilityParameters for the Black-Derman-ToyModel.

tions on bondsfor maturities between1 and24 months. (n) The returnsare greaterwith 7 = 1 but decline significantly
as 7 increases. For 7 > 6, the overpricing is difficult to

22

Journalof Business & EconomicStatistics,January1998

marketparticipants. of exploit becausethe overpricing a five-periodoptionthe noredby all but the most sophisticated Financialprofessionalsdo not generallyrevealtheirmethfollowingperiodis almostas great. fromothersources.Onesource ods, buttheycanbe inferred AND 6. MEANREVERSION Tuckman's is bookson fixed-income (1995)book modeling. OTHERFUNDAMENTALS orientedM.B.A. studentsand was aimed at quantitatively drift We have seen thatthe time-dependent andvolatility was used in SalomonBrothers'sales and tradingtraining of analogsof the Ho andLee (1986) andBlack- course,yet he devotedonly fourpagesto modelswith mean parameters (1996) targeteda similar audience the (1990) modelscannotreproduce pricesof reversion.Sundaresan Derman-Toy a anddisregarded meanreversion altogether. Perhaps better claims generatedby a model with mean state-contingent of sourceis Bloomberg(1996), the leadingpurveyor fixedreversion.This thoughtexperimentdoes not tell us how well these modelsperformin practice,but it indicatesthat income informationservices. Calculationson Bloomberg are of derivatives terminals the "fairvalue"of interest-rate are extra parameters not a solutionto all problems: They on a lognormal of the Ho andLee model,and analog with sound based need to be used in the context of a structure of for calculations "option-adjusted spreads" callablebonds fundamentals. the same model as the default(Bloomberg1996),with We focused on meanreversionbecausewe find it an ap- use model, yet it is missing Black et al. (1990) and Hull andWhite (1990) availableas pealing featurein a bond-pricing models. The alternatives. from the most popularbinomialinterest-rate We think, however,that the central issue is not mean in of .906 reported Table 1 is not estimatedautocorrelation differentfrom 1, but a value of 1 has, in the reversionbut whether pricing models-arbitrage-free or substantially a to the fundamentals counterfactual benchmark implica- not-provide good approximation theory,two apparently in firstis thatthe meanyield curve drivingbondprices.Meanreversion, this context,is simtions for bondyields. The that In declines(to minusinfinity)withmaturity. fact, ply an exampleof a fundamental cannotbe mimicked eventually drift A by time-dependent andvolatilityparameters. second upwardsloping,Figure 1 beyield curvesare, on'average, exampleis multiplefactors.Workby, amongothers,Brening a typical example.A second implicationwas pointed (1979),ChenandScott(1993),Duffieand out by Gibbonsand Ramaswamy (1993) and may be more nanandSchwartz (1986),LitterSingleton(in press),Dybvig(1997),Garbade telling:With o close to 1, averageyield curvesexhibitsubmanandScheinkman (1988)clearly (1991),andStambaugh than less stantially curvature we see in the data.An example for indicatesthe benefitsof additional factorsin accounting with ? = .99 is picturedin Figure9, in which we have set = -200 to keep the theoretical10-yearyield close to its the evolutionof bond prices throughtime. Longstaffand A of Schwartz of (1992) arerepresentative a growingnumber samplemean. These two implicationsillustratethe added studiesthat identify a second factorwith volatility,whose powerof combiningtime-seriesandcross-sectioninformain models overstate variationthroughtime is apparent the prices of options tion, and suggest to us that random-walk Yet on fixed-income instruments. another exampleis the fat rate the persistenceof the short-term of interest.For these tails in weekly interest-rate changes documentedby Das reasons,we feel that mean reversionis suggestedby the diffusion(1994), who introduced jumps into the standard of properties bondprices. basedmodels.We wouldguess thatmodelsthatignoreany For similarreasons,Black and Karasinski (1991), Heath or all of these fundamentals will, as a directconsequence, (1992), and Hull and White (1990, 1993) developedways state prices. produceinaccurate models. mean reversioninto arbitrage-free of introducing We argue, in short, that fundamentals matter and that is meanreversion generallyigDespitethesedevelopments, cannotgenerallybe rectified poor choice of fundamentals parameters. by judicioususe of time-dependent 10 can Whether examplesof mispricing be usedto direct our in strategies realisticsettingsdependson the relative trading 9.5 of and approximamagnitudes the mispricing the inevitable models.We wouldnot be tion errorsof moreparsimonious willing, at present,to bet our salarieson our benchmark we theory.Nevertheless, thinkthe exerciseindicatesthatit is important get the fundamentals to right.In our thought 8.5 fundamentals were represented the degree by experiment, of meanreversionin the shortrate,andwe saw thata mistake in this dimensioncould lead to largepricingerrorson in some securities.The extra time-dependent parameters, otherwords,are not a panacea: They allow us to reproduce accurate a subsetof assetpricesbutdo not guarantee prices securities.In for the full range of interest-rate-derivative 120 more generalsettings,we expect a model with n arraysof 60 80 0 40 o100 o20 Maturityin Months to parameters be able to reproducethe term structureof are Mean YieldCurveWhencp = .99. The line is prices of n classes of assets, but if the fundamentals Figure9. Theoretical theory,the stars are data points. wrong,therewill be some assets that are mispriced.

in Modelsof Bond Pricing Backus,Foresi,and Zin:Arbitrage Opportunities Arbitrage-Free

23
1
-

in The difficulty practiceis thateventhe best fundamental to models provideonly a roughapproximation the market Thatleaves us in the underivatives. of fixed-income prices certainterritorydescribedby Black and Karasinski (1991, p. 57): "[One]approachis to search for an interest rate process generalenough that we can assume it is true and

)
S-n-j

6)
+j=1 1

for n > 1. Given the discountfactors,we computebond unchanging. ... While we may reach this goal, we don't Best practice,we prices from bn = EtMt,t+n [a consequence of (5)]: know enoughto use this approach today." think, is to combine currentknowledge of fundamentals c (zt - 6) to with enough extra parameters make the approximation - log b= n + (1 for tradingpurposes. adequate 2 n 1A+ S n--(P) a2/2. 7. FINALTHOUGHTS
j=1

We have examined the practitioners'methodology of choosing time-dependentparametersto fit an arbitrary model to currentasset prices.We showed,in bond-pricing a relativelysimpletheoreticalsetting,that this methodcan systematicallymispricesome assets. Like Ptolemy's geocentric model of the solar system, these models can gento erally be "tuned"to provide good approximations (in our case) pricesof a limitedset of assets,butthey may also for provideextremelypoorapproximations otherassets.As Lochoff(1993, p. 92) putit, "Evenbadmodelscanbe tuned to give good results"for simple assets.
ACKNOWLEDGMENTS

rates [see (1)] are Forward


f n = log bn - log bn+1
=

6+n(6-z)-

(A+

1- -

2o/2,

which includes a short rate of rt = fo = zt - (AU)2/2.

It is conventional expressdiscountfactors,bondprices, to and forwardrates in terms of the observableshort rate r rather thanthe abstract statevariable whichis easily done z, the linearrelationbetween them. Because p = 6 given (AU)2/2, we get discountfactors

= We thank Fischer Black and the associate editor for - log Mt,t+n extensive comments;JenniferCarpenter, VladimirFinkelstein, and Bruce Tuckmanfor guidanceon industrypractice; AnthonyLynchfor pointingout an errorin an earlier this draft;andNed Eltonfor inadvertently initiating project. Backus thanksthe NationalScience Foundation finanfor bond prices cial support.Backus and Zin are ResearchAssociate and of FacultyResearchFellow,respectively, the NationalBureau of EconomicResearch.
APPENDIX: MATHEMATICAL SUMMARY

n6(+

( "- )
(P

(rt - P)
A 1+ +
Et+j,

+
=

(A.1)

We derive many of the formulasused in the text. For both the benchmark theory and our analog of the Blackmodel, we derivethe stochasticdiscountfacDerman-Toy tors implied by the pricing kernels listed in the text and fa s tt + (1Euna)t( rt() the implied bond prices, forwardrates, and prices of call +-Eoptionson discountbonds.The resultis effectivelya mathematical summary of the article. as statedin Equation(10).
A.1 Benchmark Theory We characterize bond-pricing theories with stochastic discount factors A.2 Black, Derman, and Toy

A+ _1-(

a2/21

Mt,t+n

n
j=1

mt+j

We approachour Gaussian analog of the Black-DermanToy model the same way, using Equations (21) and (22) to define the pricing kernel. Our analog of the Ho and Lee model is a special case with it = P for all t. The stochastic discount factors are - log Mt,t+n = nzt + E (n - j)at+j ( + n- j)nt+j.

or - log Mt,t+7 - - Z=l- log mt+j. Given the pricing kernel m of our benchmark theory [Eqs. (2) and (3)], the stochastic discount factors are

+ j=l E

24

Journalof Business & EconomicStatistics,January1998

They imply bondprices of


n n

termsfor the benchmark theoryin an arraylike this: (y + n- j)232


j=1

- log b - nzt +

j=1

(n - j)at+j -

/2

and,for n > 1, forwardratesof


ft=

At+3. Forthe Black-Derman-Toy the analogous termsare model, j=1


)2 2+j

j=2 j=1 n= 1 At+l n= 2 (A + 1)Et+1 AEt+2 n=3 (A+ 1 + Po)t+l (A+1)Et+2.

j=3

Zt +
j=1

n at+3 -(y + n)2 2+1/2 ( - +


j=1

j=2
r]t+2 (7 y+1)rt+2

j=3

t+j+1)/2.

n= 1 n= 2 n= 3

yt+1 ( + 1)rt+i (7 y+2)rt+1

77t+3.

If a = 0, the model is risklessandreplicablewith driftpaBecausethe shortrateis rt = fto = zt - (,yt+1)2/2, we can rametersalone.Alternatively, p = 1 we can replicatethe if rewritethese relationsas benchmark theory by setting y = A and rt+j = Et+j, which - log Mt,t+n = n[rt + (7yt+1)2/2] implies 3t+j = a for all j. But with a $ 0 and ?p 1, it is impossibleto choose the price of risk 7 and volatility n "n parameters { 3t+1, t+2, /it+3} to reproduce the benchmark + (n - j)at+ + E ( + n - j)r/t+j, (A.2) theory.Supposewe try to matchthe termssequentially. To j=1 j=1 match the term (n,j) = (1, 1) we need y77t+l = AEt+l, which requires 7y3t+l = A. Similarly, equivalenceof the (2, 1) terms, (7 + 1)r7t+1= (A + 1)Et+l, tells us (for -log bZ= n[rt + (70t+1)2/2]
n n

nonzero A) that the parametersmust be 7 = A and 3t+l = a.

+
and
n

(n - j)at+j -E (7 + n - j)2+j/2, j=1 j=1

fZ = rt + E at+ + [y2 _ (y + n)2]3+1/2


j=1 n

(7 + n - j)2
j=1

+j+l)/2,

The (3, 1) termnow requires7 + 2 = A+ 1 + p, whichis inchoices when p 1. consistentwith our earlierparameter [WhenA = 0 the same conclusionholds, but the argument startswith the (2, 1) term.]Thus we see that our attempt the to reproduce discountfactorsof the benchmark theory modelhas failed. with those of the Black-Derman-Toy It is easy to imaginesimilarproblemswith extensionsof the model. We could, for example,let y dependon time. But for similarreasons,this one-dimensional arraycannot model. the reproduce discountfactorsof the benchmark A.4 Bond Options Any streamof cash flows {ht} can be valuedby
n

as stated in Equation (23). With /t+j = 0, this reduces to


n

ft=

= rt +

E at+3 +
j=1

[72

(7 + n)2]02/2,

Et 1
j=1

Mt,t+j ht+j

equation(15) of the Ho and Lee model.

discountfactorsM. Weuse thisrelation usingthe stochastic A.3 Nonequivalenceof DiscountFactors bonds. to price European optionson zero-coupon call Considera European optionat date t with expiration the BlackOurexamplesof assets thatare mispricedby
Derman-Toy model indicate that Black-Derman-Toy cannot generally reproduce the stochastic discount factors of our benchmark theory: of Proposition 1. The parameters {7, at, 13t} the Blackmodel cannot be chosen to reproduce the Derman-Toy stochastic discount factors of the benchmark theory. The proof consists of comparing the two discount factors, (A.1) for the benchmark theory and (A.2) for the BlackDerman-Toy model. The deterministic terms of (A.1) and (A.2) are relatively simple. For, say, the first n discount factors, we can equate the conditional means of (A.1) and (A.2) by judicious choice of the n drift parameters The stochastic terms, however, cannot ... , at+n}. {at+l, generally be equated. We can represent the initial stochastic date t + 7. The option gives its owner the right to buy an n-period bond at date t + 7 for the exercise price k, thus generating the cash flow ht+, = (b+, - k)+, where X= = max{O, x} is the nonnegative part of x. The call price is

c"

= Et[Mt,t+,(bt+,

- k)+].

(A.3)

Computing this price involves evaluating (A.3) with the appropriate discount factor and bond price. Both of our theories have lognormal discount factors and bond prices, so to evaluate (A.3) we need two properties of lognormal expectations. Let us say that log x = (log x1, log x2) is bivariate normal with mean vector puand variance matrix E. Formula 1 is = E[xil(x2 - k)] exp(Cti + a1/2)N(d)

in Modelsof Bond Pricing Backus,Foresi,and Zin:Arbitrage Opportunities Arbitrage-Free

25

with /2 d'2

and optionvolatility

d- - log k + 012

V2

= 012

0.2 (1--(fln)~2

where N is the standard normaldistribution functionand I is an indicatorfunctionthat equals 1 if its argument is pn)2 (1-2r)= 20(1 0 otherwise. similarresultwas statedandproved A positive, ' "1 cp 1 -_2 by Rubinstein(1976, appendix).Except for the N term, variance the logarithm the futurebond of of this is the usual expressionfor the mean of a lognormal the conditional randomvariable.Formula2 follows from 1 with a change price. The call price for the Black-Derman-Toy modelfollows of variables(X212 for xl): a similar route with discount factor (A.2). If we choose - k)] the model'sparameters matchbondprices,thenthe only to E[(xl22)I(x2 = exp(Apl+ 2 + a1/2 + a2/2 + a12)N(d), differencein the call formulais the optionvolatility, with
k+ + "12 02 d- d #2 - log
'2 2 2E j=1 2

j=

2(-r--j)

Whetherthis is the same as the benchmark theorydepends Our application these formulasto (A.3) uses Mt,t+, as on the choice of of volatility parameters {tM}. x1 and bn+, as X2. The difficultywith the Black-Derman-Toy model with For the benchmark theory,we use discountfactor (A.1) respect to pricing options in our benchmark economy is and evaluate(A.3) using the expectation formulas.To keep similarto thatwith stochasticdiscountfactors (Proposition the notationmanageable, let 1). Volatilitiesare definedover the two-dimensional array indexedby the lengthr of the optionandthe maturity of n the bond on which the optionis written.This arraycannot 2/2. An -A+ 1-be replicatedby the one-dimensional vector of volatility j=1 Now we look at the discountfactorand futurebondprice. The discountfactoris, from (A.1),
- log Mt,t+, = parameters {3tI}.
[Received October 1995. Revised January 1997.]

1+

(rt - I)

REFERENCES
Modelof Interest Black,F.,Derman, andToy,W.(1990),"AOne-Factor E., Ratesandits Application Treasury to BondOptions," Financial Analysts
Journal, 46, 33-39.

j cj ' l"
j=1

Rates Are Lognormal," Financial Analysts Journal, 47, 52-59. It has conditionalmean given by the first two terms and Black,F., andScholes,M. (1973),"ThePricingof OptionsandCorporate conditional variance 2A,. The futurebondpricecan be repLiabilities," Journal of Political Economy, 81, 637-654. resented for on-line facility, (1996), Bloomberg "Help OSY," help May.

P. and WhenShort Black,F.,andKarasinski, (1991),"Bond Option Pricing

-logb

= n6- An+ +

(r+

E. "A Time Brennan, andSchwartz, (1979), Continuous Approach M., to


the Pricing of Bonds," Journal of Banking and Finance, 3, 133-155.

nj) +

-(p

\(rt+j

Multifactor Modelof the TermStructure Interest of Rates," Equilibrium Journalof FixedIncome,3, 14-31. Cooley,T., LeRoy,S., and Parke,W. (1992), "PricingInterest-Sensitive ClaimsWhenInterestRatesHave Stationary Journalof Components,"
Fixed Income, 2, 64-73.

and L. "Maximum Likelihood Estimation a for Chen, R.-R., Scott, (1993),

"A of Cox,J.,Ingersoll, andRoss, (1985), Theory theTerm J., Structure S.


j=1

of Interest Rates," Econometrica, 53, 385-407.

=-

og

n + log b + (An A/ -A
1

Processesandthe BondMarkets," unDas, S. (1994), "Poisson-Gaussian BusinessSchool. 95-034, Harvard published manuscript TermStructure Interest of Rate SwapYields," Journalof Finance, 52. Dybvig,P. (1997), "BondandBondOptionPricingBasedon the Current
Term Structure,"in Mathematics of Derivative Securities, eds. M. A. H.

K. "An Model the of Duffie, andSingleton, (in press), Econometric D.,

1 -f /n)

U.K.: Cambridge Dempsterand S. R. Pliska, Cambridge, University

Press.

K. of in Bankers Garbade, (1986),"Modes Fluctuation BondYields," Trust,


Topics in Money and Securities Markets, 20, June.

K. "A Gibbons, andRamaswamy, (1993), Testof theCox,Ingersoll, M.,


di Ross Model of the Term Structure,"Review of Financial Studies, 6, =

619-658.

d2 = dl - v,,n

A. Heath,D., Jarrow, andMorton, (1992),"Bond R., Pricingandthe Term Structure InterestRates," of Econometrica, 225-262. 60,

26

Journalof Business & EconomicStatistics,January1998

A General Journalof TermStructure: Two-Factor and Structure Movements PricModel," Ho, T. S. Y., andLee, S.-B. (1986),"Term Equilibrium Journalof Finance,41, 1011Finance,47, 1259-1282. Claims," ing InterestRate Contingent 1029. J. Data," McCulloch, H., and Kwon,H.-C. (1993),"U.S. TermStructure Ohio StateUniversity, and SecuriInterest-Rate-Derivative diskettes, Hull, J., and White,A. (1990), "Pricing Dept. of Fimanuscript computer nance. ties," Review of Financial Studies, 3, 573-592. IncomeStreams and of Interest-Rate Models and the Valuationof Rubinstein, (1976),"TheValuation Uncertain M. (1993), "One-Factor
Interest-Rate Derivative Securities,"Journal of Financial and Quantitative Analysis, 28, 235-254. the Pricing of Options," Bell Journal of Economics, 7, 407-425.

Journalof FiF. Jamshidian, (1989), "AnExact Bond OptionFormula,"


nance, 44, 205-209.

in Rates:Implications R. Stambaugh, (1988), "TheInformation Forward


Journal of Financial Economics, 21, for Models of the Term Structure,"

41-70. FactorsAffecting Sundaresan, S. (1996), Fixed Income Markets and Their Derivatives, J. R., Litterman, and Scheinkman, (1991), "Common Cincinnati: South-Western Bond Returns,"Journal of Fixed Income, 1, 54-61. Publishing. Journal Port- Tuckman, B. (1995), Fixed Income Securities, New York: Wiley. ArmsRace," Lochoff,R. (1993),"TheContingent-Claims of of Characterization the TermStrucVasicek,0. (1977),"AnEquilibrium folio Management, 20, 88-92. Rate Volatilityand the E. ture," Journal of Financial Economics, 5, 177-188. F., and Schwartz, (1992), "Interest Longstaff,