Asset Pricing with Misspecified Models - APJFS
Asset Pricing with Misspecified Models - APJFS
Asset Pricing with Misspecified Models - APJFS
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Asset</strong> <strong>Pricing</strong> <strong>with</strong> <strong>Misspecified</strong> <strong>Models</strong><br />
Jialin Yu<br />
Columbia University<br />
Graduate School of Business<br />
Finance and Economics<br />
3022 Broadway Uris Hall 421<br />
New York NY 10027<br />
+1-212-854-9140<br />
jy2167@columbia.edu<br />
Dongyoup Lee<br />
Columbia University<br />
Graduate School of Business<br />
Finance and Economics<br />
3022 Broadway Uris Hall 4k<br />
New York NY 10027<br />
+1-646-530-1303<br />
dl2135@columbia.edu
<strong>Asset</strong> <strong>Pricing</strong> <strong>with</strong> Misspeci…ed <strong>Models</strong> <br />
Jialin Yu y<br />
Columbia University<br />
Dongyoup Lee z<br />
Columbia University<br />
August 25, 2008<br />
Abstract<br />
This paper provides an asset pricing method to address potential model misspeci…cation. The<br />
resulting price estimator is consistent irrespective of misspeci…cation. The pricing precision is<br />
at least that of nonparametric prices, and automatically converges to parametric precision when<br />
model quality improves.<br />
The method is applicable to multi-dimensional asset pricing and to<br />
sensitivity analysis.<br />
Application to the pricing of CBOT Treasury options suggests that the<br />
cheapest to deliver practice is an important source of misspeci…cation.<br />
Potential equilibrium<br />
implications on bounded rationality are discussed.<br />
We thank Yacine Aït-Sahalia, Robert Hodrick, Bo Honoré, Wei Jiang, José Scheinkman, and seminar participants<br />
at Columbia Business School for helpful discussions. All errors are ours.<br />
y Corresponding author. Address: 421 Uris Hall, 3022 Broadway, New York, NY 10027. Email: jy2167@columbia.edu<br />
z Address: 4K Uris Hall, 3022 Broadway, New York, NY 10027. Email: dl2135@columbia.edu.
1. Introduction<br />
Investors constantly face the challenge of imperfect models. For example, a trader of Treasury options<br />
listed on the Chicago Board of Trade (CBOT) may have learned the state-of-the-art term structure<br />
model which prescribes an option pricing formula. Over time, the trader starts to observe option<br />
prices deviate from those predicted by the pricing formula and suspects the model is misspeci…ed<br />
– just as the Black-Merton-Scholes option pricing formula (Black and Scholes (1973) and Merton<br />
(1973)) is found by many to have di¢ culty explaining the Black Monday in October 1987. Misspeci…cation<br />
can take di¤erent forms: the model may be misspeci…ed only along some dimensions of the<br />
state variables, or the model may be poor along all dimensions. Even in the latter case, the model<br />
can still provide useful restrictions that may be utilized by some investors. For example, the model<br />
may approximate the …rst derivative (e.g., the Greek letter delta) well but not the second derivative<br />
(e.g., the Greek letter gamma). 1 Therefore, misspeci…cation is not a binary concept. Rather, there is<br />
a continuous middle ground between correct speci…cation and the case of a useless model, the middle<br />
ground being the more likely scenario in practice than the two polar cases.<br />
How should the investors deal <strong>with</strong> models that are misspeci…ed but not completely useless, such<br />
as those that provide good approximation in some aspects but misleading information in others?<br />
Further, cautious investors may …nd it desirable to take measures before the latest model is con…rmed<br />
misspeci…ed. How should such investors act against the possibility of misspeci…cation, the true nature<br />
of which is unknown as of yet?<br />
This paper proposes a pricing method (referred to as “robust parametric method” in this paper)<br />
based on a possibly misspeci…ed model so that the resulting price estimator has the following<br />
properties: (i) robustness – the price estimator is consistent and the pricing error is at most that<br />
of the nonparametric rate irrespective of misspeci…cation; (ii) adaptive e¢ ciency –the pricing error<br />
decreases when the model quality improves, and the pricing error approaches the parametric rate in<br />
the limit when the model misspeci…cation disappears. 2 In addition, the estimator does not require an<br />
investor to know the model quality. The appropriate rate of convergence is achieved automatically.<br />
To see the potential magnitude of improvement from adaptive e¢ ciency, recall that the pricing error<br />
1 The Greek letter delta refers to the sensitivity of option value to the change in price of its underlying asset. The<br />
Greek letter gamma measures the rate of change in delta when the underlying asset value changes.<br />
2 See (2.8) on measuring model quality.<br />
1
of parametric method, based on a correct model, is in the order of n 1=2 <strong>with</strong> n being the sample size.<br />
The pricing error of nonparametric method is in the order of n 2/(4+d) where d is the dimension of<br />
the state variables. 3 To improve the pricing precision from $0.1 to $0.01, parametric method requires<br />
100 times the sample size and nonparametric method requires 10,000 times the sample size if d = 4.<br />
Multi-dimensional state variable is not uncommon in asset pricing which reduces the value added<br />
of nonparametric methods. For example, option pricing can involve multiple state variables such as<br />
the underlying asset value, underlying asset volatility, option maturity, strike price, etc. That the<br />
robust parametric method can, depending on model quality, improve the pricing error towards that<br />
of the parametric method is a nontrivial contribution. Its advantage relative to parametric methods<br />
lies in the possibility of model misspeci…cation, in which case the parametric pricing error is di¢ cult<br />
to quantify. Therefore, the robust parametric method is especially suited if the available model is<br />
somewhere in between being correct and disastrous.<br />
To see the intuition of the robust parametric method, consider a pricing model f (X; ), where X<br />
is the state variable and is the model parameter. Misspeci…cation is de…ned as the nonexistence of<br />
a parameter such that f (X = x; ) …ts the true model for all values of x. However, misspeci…cation<br />
does not rule out the existence of a parameter (x) such that f (X = x; (x)) …ts the true model for<br />
one value X = x only. Tracing out (x) for various x makes f (X; (X)) match the true model. That<br />
is, if (X) is appropriately chosen, a misspeci…ed model can be turned into a true model. To say that<br />
the original model is correctly speci…ed amounts to say that (X) is a constant. For example, the<br />
existence of a volatility smile implies that no single implied volatility number can …t option prices at<br />
all strike prices using the Black-Merton-Scholes option pricing formula. However, one can …nd one<br />
implied volatility for each strike so that the Black-Merton-Scholes formula using di¤erent implied<br />
volatilities …ts the option prices at di¤erent strikes. These implied volatilities, when plotted against<br />
strikes, constitute the smile curve which is the equivalent of (X). This is one instance where a<br />
misspeci…ed model is converted into a correct one. Therefore, this paper captures the intuition used<br />
informally in the investment community. Next, if the true model is continuous, f (X; (x)) may<br />
provide a good approximation not only for X = x but also for observations in a neighborhood of x.<br />
The robust parametric method applies parametric method to estimate (x) in this neighborhood of<br />
3 See Newey and McFadden (1994) and Fan (1992) on the parametric and nonparametric rates of convergence.<br />
2
…t. The neighborhood is likely small when the model quality is poor, hence only nearby observations<br />
are used to achieve robustness. When the model quality improves, the neighborhood can be enlarged<br />
to include more observations to enhance e¢ ciency. As a by-product, this neighborhood (referred<br />
to as “region of …t” henceforward in the paper) provides information on the quality of the model<br />
near x. This pricing method is applicable to multi-dimensional pricing problem. For example, in<br />
a two dimensional case, the region of …t may take the shape of a rectangle where the side along<br />
the dimension of good model …t is longer to include more distant observations to improve e¢ ciency,<br />
while the side along the dimension of poor model …t is shorter to utilize only nearby observations to<br />
achieve robustness. The intuition will be formalized in section 2.<br />
The intuition of the proposed pricing method relates to Hansen and Jagannathan (1991) and<br />
Hansen and Jagannathan (1997).<br />
These two papers show how security market data restrict the<br />
admissible region for means and standard deviations of intertemporal marginal rates of substitution<br />
(IMRS) which can be used to assess model speci…cation.<br />
Speci…cally, Hansen and Jagannathan<br />
(1991) calculate the volatility bound which is the greatest lower bound on the standard deviation<br />
of IMRS to price the assets. This bound on the variability of IMRS has a natural connection to<br />
the variability of parameter estimates in this paper. Here, the parameter (x) is constant <strong>with</strong>in a<br />
region of …t surrounding x but may di¤er across di¤erent regions of …t. When the model quality is<br />
poor, regions of …t tend to be smaller to allow greater parameter variability across di¤erent regions<br />
of the state variables in order to price the assets adequately. This operationalizes the Hansen and<br />
Jagannathan (1991) volatility bound for investors who know their model is misspeci…ed but have no<br />
alternative at the time of decision making.<br />
The robust parametric pricing method can add value even if the correct model is known. For<br />
example, when the true model is high dimensional and does not admit closed-form pricing formula, estimations<br />
employing numerical procedures may introduce additional noise when computing resource<br />
is …nite. In this case, It may improve pricing precision by using an elegant model and explicitly<br />
adjusting for possible misspeci…cation using the proposed method, compared to parametric estimation<br />
using the true but complicated model. This echoes the “maxim of parsimony” in Ploberger<br />
and Phillips (2003) and is consistent <strong>with</strong>, for example, the widespread practice of using the Black-<br />
Merton-Scholes option pricing model even when evidence suggests possible misspeci…cation. Section<br />
3
3.3 illustrates this point using simulation under a realistic setting of Treasury option pricing.<br />
We then apply the robust parametric pricing method to the pricing of Treasury options traded on<br />
the CBOT. In both in-sample analysis and out-of-sample performance, the robust parametric method<br />
consistently performs better than the nonparametric method and the parametric method based on<br />
models in which the short rate follows an a¢ ne term structure model. This suggests that option<br />
pricing formulas based on such models are misspeci…ed, though they still contain useful information<br />
so that the robust parametric prices perform better than nonparametric prices. Analysis of the region<br />
of …t indicates that these option pricing formulas have poor …t along the dimensions of short rate<br />
and bond maturity but provide useful economic restriction along the dimension of option maturity.<br />
Such information provided by the region of …t facilitates future development of asset pricing models.<br />
Speci…cally, it suggests that the cheapest to deliver (CTD) practice in the CBOT Treasury options<br />
market is an important source of model misspeci…cation which is typically ignored in bond option<br />
pricing formulas based on term structure models. Jordan and Kuipers (1997) document an interesting<br />
event where CTD a¤ected the pricing of those Treasuries used in futures delivery. Results in this<br />
paper suggest that CTD is also an important feature in day-to-day Treasury options pricing.<br />
In addition to being an important issue in asset pricing, model misspeci…cation is an important<br />
topic in the econometrics literature and has motivated study of speci…cation tests (e.g., Hausman<br />
(1978) and Gibbons, Ross, and Shanken (1989)), nonparametric estimation (e.g., Fan and Gijbels<br />
(1996) and Bandi and Phillips (2003)). Nonparametric estimation alleviates the concern of robustness.<br />
However, a misspeci…ed model may nonetheless contain useful information. Given the increasing<br />
popularity of multi-dimensional models, the e¢ ciency loss of nonparametrics from omitting valid<br />
model restrictions can be nontrivial (the “curse of dimensionality” problem illustrated previously).<br />
To improve e¢ ciency, nonparametric pricing can be conducted under shape restrictions implied by<br />
economic theory (Matzkin (1994), Aït-Sahalia and Duarte (2003)).<br />
There is also a literature on<br />
semiparametric estimation reviewed by Powell (1994).<br />
However, shape restriction and semiparametric<br />
estimation apply only to selected classes of models so far. Further, they do not achieve the<br />
parametric rate of convergence and may in some cases lose their robustness when, for example, the<br />
shape restriction is misspeci…ed. Gozalo and Linton (2000) propose to replace the local polynomial<br />
in nonparametric estimation <strong>with</strong> an economic model and show that the resulting estimator is con-<br />
4
sistent and retains the nonparametric rate of convergence. This paper builds on their insight and<br />
shows that incorporating the economic restrictions of a model also improves estimation e¢ ciency<br />
towards that of the parametric rate when the model quality improves, hence constituting a continuous<br />
middle ground between parametric and nonparametric estimations. This paper focuses on the<br />
estimation of asset price, which is the conditional expectation function of discounted future payo¤<br />
given current state. In the context of likelihood estimation, quasi-maximum likelihood estimator<br />
(White (1982)) and local likelihood estimator (Tibshirani and Hastie (1987)) have been proposed to<br />
address misspeci…cation. When the model is correctly speci…ed, the maximum likelihood estimator<br />
is optimal under fairly general conditions (e.g., Newey and McFadden (1994)). When the model is<br />
misspeci…ed, the quasi-maximum likelihood estimator minimizes the Kullback-Leibler Information<br />
Criterion (KLIC) which is the distance between the misspeci…ed model and the true data-generating<br />
process measured by likelihood ratio. However, minimal distance measured by likelihood ratio does<br />
not translate into minimal distance in price (i.e., conditional expectation function) if the model is<br />
misspeci…ed. This also applies to the local likelihood estimator.<br />
This paper is organized as follows.<br />
Section 2 details the proposed robust parametric pricing<br />
method and its properties. Section 3 uses simulation to demonstrate its performance. Section<br />
4 studies the pricing of Treasury options traded on CBOT using the robust parametric method.<br />
Section 5 concludes. The appendix contains the proofs and collects the various Treasury options and<br />
Treasury futures pricing formulas used in the simulation and empirical analysis.<br />
2. <strong>Asset</strong> pricing <strong>with</strong> misspeci…ed models<br />
Consider an asset whose price is P (X) where X is a d-dimensional state variable. We assume an<br />
investor has an economic model which implies a possibly misspeci…ed pricing formula f (X; ). is a<br />
p-dimensional parameter. The data consist of observations fx i ; y i g n i=1 where y i = P (x i ) + " i . " has<br />
zero mean and can capture the market microstructure e¤ects (see Amihud, Mendelson, and Pedersen<br />
(2005) for a recent review) or sampling errors.<br />
As motivated in the introduction, a misspeci…ed model f (X; ) can be turned into a true model<br />
5
if there exists a function (X) such that<br />
P (X) = f (X; (X)) (2.1)<br />
for all X. For example, the potentially misspeci…ed Black-Merton-Scholes option pricing formula can<br />
be used to …t option prices over di¤erent strikes by using the volatility smile curve which corresponds<br />
to (X). Correct speci…cation is equivalent to (X) being constant. Assuming continuity, a Taylor<br />
expansion implies that for X near x,<br />
<br />
P (X) = f (X; (x)) + b 1 (x) (X x) + (X x) T b 2 (x) (X x) + o kX xk 2 (2.2)<br />
i.e., f (X; (x)) approximates P (X) at X near x. Therefore, we propose to estimate (x) using<br />
observations in a neighborhood of x,<br />
b (x) = argmin<br />
<br />
kx i<br />
X<br />
xkh<br />
[y i f (x i ; )] 2 (2.3)<br />
The reason we include observations at X 6= x in the presence of misspeci…cation is that the<br />
additional observations likely reduce estimation noise as long as the misspeci…cation is not severe.<br />
This creates a trade-o¤ between estimation e¢ ciency and robustness which is represented in the choice<br />
of h in (2.3). We will refer to h as “region of …t” in this paper. When the model misspeci…cation<br />
is minor, one can a¤ord to use a larger region of …t to improve e¢ ciency. On the contrary, if model<br />
misspeci…cation is severe, one might want to use a smaller region of …t to ensure robustness. We will<br />
discuss the optimal choice of region of …t shortly. For now, assuming an estimate b (x) is obtained<br />
using the optimal region of …t, we estimate P (X = x) by<br />
bP (X = x) = f<br />
<br />
x; b <br />
(x) :<br />
The (infeasible) optimal choice of region of …t, denoted h , can be determined by minimizing the<br />
integrated mean squared pricing error<br />
h<br />
h = argmin E P (X)<br />
h<br />
f<br />
<br />
X; b (X)i 2<br />
: (2.4)<br />
6
Equation (2.4) cannot be directly applied because the true expectation is unknown.<br />
In this<br />
paper, we follow a method similar to the crossvalidation in nonparametric bandwidth choice. The<br />
crossvalidation procedure is asymptotically optimal <strong>with</strong> respect to the criterion function in (2.4), see<br />
Härdle and Marron (1985) and Härdle, Hall, and Marron (1988). 4<br />
Speci…cally, the crossvalidation<br />
method involves two steps. First, for a given candidate h, we obtain a …rst-step estimate b i;h (x i )<br />
of (x i ) using all observations less than h away from x i except x i itself, 5<br />
b i;h (x i ) = argmin<br />
<br />
0 0. The lower bound in the order of n 1=(4+d) is the nonparametric rate of bandwidth<br />
choice. The upper bound, when ! is close to zero, is allowed to decrease at a very slow rate in the<br />
case of a good model. The propositions in this paper will be proved for the feasible choice of region<br />
of …t b h instead of for the infeasible h . In general, b h depends on the sample size n. However, the<br />
dependence is not made explicit to simplify notations.<br />
Proposition 1 (Consistency) Under Assumptions 1-5, whether or not the model f is correctly spec-<br />
4 There is a large statistics literature on choosing the optimal smoothing parameter h. See Härdle and Linton (1994)<br />
for a review.<br />
5 If x i itself is included in the crossvalidation, it will result in a mechanical downward bias in the h estimator because<br />
a perfect …t is possible by choosing a very small region of …t so that only x i is included to …t itself.<br />
7
i…ed, when n ! 1, we have<br />
b<br />
p<br />
(x) ! (x)<br />
<br />
f x; b p!<br />
(x)<br />
P (x)<br />
if b h n!1 ! 0 and n b h d n!1 ! 1.<br />
Given the consistency, we next study the asymptotic distribution of the price estimate.<br />
The<br />
asymptotic distribution of b (x) varies <strong>with</strong> the quality of the model. (2.2) implies that any model<br />
can always locally match the level of the true pricing formula. Therefore, in this paper, model quality<br />
is measured by the mismatch between the true model and f (X; (x)) when the state variable X<br />
moves away from x, which relates to how well the derivatives of f (X; (x)) match those of the<br />
true model. f (X; (x)) is capable of matching the true model at X far from x if its derivatives<br />
approximate those of the true model well. We say a model matches the true model up to its 2k-th<br />
derivative if, for any x, (using univariate notation for simplicity)<br />
<br />
P (X) = f (X; (x)) + b 2k+1 (x) (X x) 2k+1 + b 2k+2 (x) (X x) 2k+2 + o kX xk 2k+2 (2.8)<br />
for X near x. Let n x;h denote the number of observations less than h away from x. When X is<br />
d-dimensional, the number of observation is in the order of<br />
n x;h = O p<br />
nh d (2.9)<br />
when n ! 1 and h ! 0. When the model …ts reality well at state variables away from x, we can<br />
a¤ord to use a larger region of …t h. This uses more observations and lowers estimation noise. This<br />
intuition is formalized in proposition 2 below.<br />
Proposition 2 (Bias-variance trade-o¤ ) Under Assumptions 1-5, if the model f matches the true<br />
model up to its 2k-th derivative as in (2.8) for some k 0, when n ! 1, b h ! 0 and n b h d ! 1,<br />
<br />
Bias b (x)<br />
<br />
Var b (x)<br />
<br />
= O bh 2k+2 + n 1 b h<br />
d<br />
(2.10)<br />
<br />
= O n 1 b h<br />
d<br />
:<br />
8
This proposition shows the trade-o¤ between estimation e¢ ciency and robustness. When the<br />
region of …t b h is larger, more observations are used which results in lower variance in the estimate.<br />
However, if the model is misspeci…ed (i.e., k is …nite), increasing the region of …t can lead to a larger<br />
bias. When the model quality improves (k increases), the bias becomes smaller. In the limit when<br />
the model is correctly speci…ed, k ! 1, both bias and variance inversely relate to the region of …t<br />
hence the optimal choice is to use all observations as in parametric estimation. The next proposition<br />
shows the estimator will, depending on model quality, automatically select an appropriate region of<br />
…t b h to balance e¢ ciency and robustness.<br />
Proposition 3 (Model quality) Under Assumptions 1-5, when the model f matches the true model<br />
up to its 2k-th derivative as in (2.8) for some k 0,<br />
b h<br />
1<br />
= O p<br />
n 1=(4+4k+d) (2.11)<br />
<br />
P (x) = f x; b <br />
(x) + O p<br />
n (2+2k)=(4+4k+d)<br />
Note that n (2+2k)=(4+4k+d) ! n 1=2 when k ! 1.<br />
When k = 0 (i.e., if the model can only locally match the level but none of the variations of the<br />
true model <strong>with</strong> respect to the state), the estimator automatically achieves the nonparametric rate<br />
of convergence n 2=(4+d) . 6<br />
When the model gives a better …t in the sense that k increases, the rate of convergence of the<br />
proposed method automatically improves towards that of the parametric rate n 1=2 . Therefore, a<br />
continuous middle ground between nonparametric and parametric estimation is achieved depending<br />
on the quality of the model. The e¢ ciency gain comes from a better economic model. When k<br />
increases, (2.11) implies that the region of …t b h decreases at a slower rate. Recall that (2.7) implies<br />
an upper bound n ! for the region of …t. When the model is so good that the region of …t implied<br />
by (2.11) exceeds the upper bound, further e¢ ciency gains hence full parametric rate of convergence<br />
cannot be achieved. This e¢ ciency loss relative to the full parametric rate of convergence is necessary<br />
because we need h ! 0 to ensure robustness in case the model is misspeci…ed. However, ! can be<br />
6 See Fan (1992) on the nonparametric rate of convergence. This is intuitive because standard nonparametric<br />
estimators (e.g., Nadaraya-Watson kernel estimator) do not place any restrictions on how the true model varies <strong>with</strong><br />
the state variable to achieve robustness.<br />
9
made arbitrarily small to make the rate of convergence arbitrarily close to the parametric rate in<br />
the case of a good model. Further, if one views most models as reasonable approximations (i.e.,<br />
misspeci…ed) rather than literal descriptions of the reality, this e¢ ciency loss associated <strong>with</strong> ! > 0<br />
is likely a small price to pay in practice to ensure robustness.<br />
This improved e¢ ciency is achieved <strong>with</strong>out introducing additional parameters. This contrasts<br />
<strong>with</strong> the Taylor expansion used in local polynomial estimators (see Fan and Gijbels (1996)) in which<br />
smaller bias can be achieved by using a higher-order polynomial to approximate the true model.<br />
However, this leads to increased variance due to increased number of parameters.<br />
For example,<br />
going from a local linear model to a local quadratic model can double the asymptotic variance<br />
for typical kernels (Table 3.3 in Fan and Gijbels (1996)). On the contrary, the improved rate of<br />
convergence in this paper comes from a better economic model.<br />
(2.3) weighs observations equally for ease of illustration and does not explicitly discuss the possibility<br />
of weighting the observations as in, for example, GMM estimation (Hansen (1982)) or LOWESS<br />
nonparametric estimation (Fan and Gijbels (1996)). This is similar to using a uniform kernel in nonparametric<br />
estimation where it is known that the choice of kernel is not crucial (Härdle and Linton<br />
(1994)). Equal weighting is also technically convenient. When the model is correct and the sampling<br />
errors are homoskedastic, we would like the estimator to use all observations <strong>with</strong> equal weight just<br />
like the parametric nonlinear least-squares estimation. To achieve this using a kernel <strong>with</strong> unbounded<br />
support (such as normal), h ! 1 is required which is inconvenient in numerical implementation.<br />
However, weighting implicitly occurs in this paper through the region of …t –observations outside of<br />
the region of …t receive zero weight.<br />
2.1. Sensitivity analysis<br />
Sometimes one may be interested in estimating derivatives of the pricing formula. Examples include<br />
the various Greek letters of the option pricing formula or other sensitivity analyses. Recall that (X)<br />
satis…es<br />
P (X) = f (X; (X))<br />
10
for any X. Taking derivative <strong>with</strong> respect to the state variable implies<br />
P 0 (X) = f X (X; (X)) + f (X; (X)) 0 (X) :<br />
To simplify notation, f X is used to denote<br />
@<br />
@X f, similarly for f .<br />
In order to estimate P 0 (x), 0 (x) needs to be estimated. Otherwise there is a bias if f X (X; (X))<br />
alone is used to estimate sensitivity when the model is misspeci…ed (when the model is correctly<br />
speci…ed, there is no bias because (X) is a constant). To estimate the …rst derivative of (X), we<br />
can use the following augmented model<br />
f (X; 0 (x) + 1 (x) (X<br />
x))<br />
to approximate the true model at X near x and the estimation can then proceed in the same way<br />
<br />
as previously discussed using this augmented model. When the parameter estimates b0 (x) ; b 1 (x)<br />
are obtained, the derivative of the true model P 0 (X) at X = x is estimated as<br />
f X<br />
x; b <br />
0 (x) + f x; b <br />
0 (x) b 1 (x) :<br />
The estimation of higher-order derivative is similar.<br />
Counterparts to Proposition 1 – 3 exist<br />
for derivative estimation. These propositions and their proofs are very similar to Proposition 1 –3,<br />
except that the nonparametric rate of convergence is slightly modi…ed to re‡ect derivative estimation<br />
which is standard in the nonparametric literature. These propositions and proofs are omitted for<br />
brevity and are available from the authors upon request.<br />
2.2. Multivariate pricing models<br />
The robust parametric pricing method is well suited for multivariate models. In fact, proposition 1 - 3<br />
are derived for the general case of d-dimensional state variables. Contrary to nonparametric methods,<br />
there need not be a “curse of dimensionality”problem as long as an investor has a good model. As<br />
shown in Proposition 3, when the model quality improves (i.e., k ! 1 in the proposition), the<br />
estimation e¢ ciency approaches that of the parametric rate which is not a¤ected by dimensionality.<br />
11
In this section, we show that the region of …t can be re…ned in a multivariate model to re‡ect<br />
the possibility that a model may …t well along certain dimensions of the state variables but …t<br />
poorly along other dimensions. In (2.3), the estimator is obtained using observations x i satisfying<br />
kx i<br />
xk h. I.e., we consider the pricing formula to have a good …t for observations less than h<br />
away from x. When the state variable is multi-dimensional, we can apply a separate region of …t for<br />
each dimension.<br />
We illustrate this using a two dimensional example where the state variable can be written as<br />
x = x (1) ; x (2) . (2.3) can be modi…ed so that the parameters are estimated from<br />
b (x) = argmin<br />
<br />
X<br />
<br />
<br />
x (1)<br />
i x (1) h (1)<br />
<br />
<br />
x (2)<br />
i x (2) h (2)<br />
[y i f (x i ; )] 2 : (2.12)<br />
The region of …t now takes the shape of a rectangle. The interpretation is that the pricing formula<br />
is considered to …t well for those observations that are less than h (1) away from x in the …rst<br />
dimension and are less than h (2) away from x in the second dimension.<br />
This re…nement can be<br />
used to re‡ect di¤erent scales of the state variables – e.g., measured in di¤erent currencies, or to<br />
re‡ect the di¤erent degrees of misspeci…cation along various model dimensions – those <strong>with</strong> less<br />
misspeci…cation are associated <strong>with</strong> larger regions of …t. The estimation then proceeds in the same<br />
way and the conclusions in Proposition 1 - 3 remain the same.<br />
2.3. Numerical implementation<br />
The estimation in (2.3) and (2.5) involves nonlinear least squares which is programmed in many<br />
statistical software packages. Nonlinear least squares estimation is fairly quick because it is typically<br />
implemented as iterated linear least squares, see Greene (1997). Nonetheless, when the dataset has<br />
a large number of observations and when the state variable has many dimensions, there is room for<br />
faster implementation of the proposed pricing method.<br />
A potential bottleneck of the robust parametric pricing method lies in the crossvalidation step<br />
(2.5) for …nding the optimal region of …t which, in principal, is estimated for all possible candidates<br />
of region of …t h at all observations x i to evaluate the model quality at various regions. However, this<br />
12
is not necessary –evaluations can be done at fewer h and x i to trade e¢ ciency gain for computational<br />
cost savings.<br />
First, one can restrict the choice of h by searching over the following grid<br />
h 1 = n 1=(4+d) ; h 2 = h 1 + ; h 3 = h 1 + 2; ; h m = n ! (2.13)<br />
where ! is a small positive number discussed in (2.7). The grid size is = (h m h 1 ) = (m 1). m<br />
can be increased when additional computational resource is available. The downside from searching<br />
over fewer candidates is that the chosen region of …t may deviate from the optimal choice b h in<br />
proposition 3, which can reduce (though not eliminate) the e¢ ciency gain associated <strong>with</strong> a good<br />
model. However, the price estimator remains consistent and achieves at least the nonparametric rate<br />
of convergence. For a multivariate model discussed in 2.2, the grid for region of …t can be applied<br />
separately to each dimension.<br />
Next, one can restrict the number of observations x i at which (2.5) is evaluated whose output<br />
is then used in (2.6) as a sample analog of (2.4). For the purpose of consistently estimating the<br />
expectation in (2.4) using its sample analog, the number of evaluations should increase asymptotically<br />
towards in…nity though the rate of increase can be lower than that of the sample size. This can<br />
be implemented, for example, by estimating (2.5) at randomly selected n v observations for some<br />
0 < v 1. When v is bigger, the expectation in (2.4) is estimated more precisely which results in<br />
a more precise choice of region of …t at the cost of computational resource. When (2.5) is evaluated<br />
at selected observations (i.e., v < 1), (2.6) needs to be adjusted to include only these observations.<br />
As long as v > 0, the price estimator remains consistent irrespective of misspeci…cation.<br />
3. Simulation –Treasury options pricing<br />
This section uses simulation to illustrate the proposed robust parametric pricing method in realistic<br />
samples, comparing its performance to parametric and nonparametric methods. When a true model<br />
is complicated, we illustrate that it may improve matters to use a simpler model and explicitly adjust<br />
for misspeci…cation using the robust parametric method.<br />
We illustrate in the context of pricing Treasury options.<br />
Speci…cally, we price call options<br />
13
C (; T; X) on Treasury zero-coupon bonds, where is time to option expiration, T is bond maturity<br />
at option expiration, and X includes other state variables such as the prevailing interest rate,<br />
the strike price, etc. This is a multivariate simulation example in that the option pricing formula<br />
needs to be estimated along the dimensions of option maturity, underlying bond maturity, and other<br />
state variables.<br />
We assume that the true data generating process follows the Cox, Ingersoll, and Ross (1985)<br />
model (CIR model) under the risk-neutral probability<br />
dr t = k ( r t ) dt + p r t dW t (3.1)<br />
where r t is the instantaneous short rate at time t.<br />
The short rate mean reverts to its long-run<br />
mean at a speed governed by k. The standard Brownian motion W drives the random evolution<br />
of the short rate. The instantaneous volatility of the short rate is determined by the parameter <br />
and the square root of the short rate (hence the process is also known as the square root process).<br />
Under the CIR model, the Treasury zero-coupon bond option has a closed-form expression (detailed<br />
in the appendix). Alternatively, the Treasury option price under CIR model can be computed via<br />
numerical integration using the method in Du¢ e, Pan, and Singleton (2000), which we later use in the<br />
simulation to quantify the performance of parametric estimator when the true model is complicated<br />
so that closed-form pricing formula is unavailable and numerical method is used instead to obtain<br />
prices.<br />
To implement the robust parametric method, we assume that an investor is aware that the Vasicek<br />
(1977) model delivers a closed-form Treasury option pricing formula (detailed in the appendix), but<br />
this same investor has yet to adopt the CIR model. In the Vasicek model, the evolution of the short<br />
rate under the risk-neutral probability is assumed to follow<br />
dr t = k ( r t ) dt + dW t : (3.2)<br />
That this investor uses the Vasicek model but not the CIR model may happen if this investor<br />
has studied Vasicek (1977) but the Cox, Ingersoll, and Ross (1985) paper is either not published<br />
yet or has not caught this investor’s attention when he faces the pricing problem.<br />
However, the<br />
14
investor is aware that the Vasicek option pricing formula may be misspeci…ed and addresses such<br />
misspeci…cation using the robust parametric method.<br />
We compare the robust parametric method to four other estimation methods: (i) parametric estimation<br />
using the CIR model (parametric estimation using correct model); (ii) parametric estimation<br />
using the Vasicek model (parametric estimation using misspeci…ed model); (iii) nonparametric estimation;<br />
(iv) parametric estimation using the correct CIR model but applying numerical integration<br />
method to obtain option prices instead of relying on the closed-form CIR option pricing formula,<br />
which allows us to compare estimation using a correct but complicated model which relies on numerical<br />
methods to estimation using a simpler model and explicitly adjusting for misspeci…cation using<br />
the robust parametric method. The estimation performance is measured by the sample analog of the<br />
root integrated mean squared error<br />
v<br />
u<br />
RIMSE t 1 n<br />
nX 2<br />
bCi C i (3.3)<br />
i=1<br />
where C b and C are, respectively, the estimated and the true Treasury option prices in each simulation.<br />
RIMSE captures the average goodness of …t <strong>with</strong> smaller RIMSE indicating better …t.<br />
The simulation is designed to re‡ect the characteristics of the Treasury option contracts traded<br />
on the Chicago Board of Trade (CBOT). The simulation draws 100 sample paths of short rate,<br />
each sample path being equivalent to 5 years of weekly observations. Such samples are commonly<br />
encountered in practice, see for example Du¢ e and Singleton (1997). For each sample path, Treasury<br />
call option prices are generated according to the CIR model for the option maturity =1, 2, 3, 6,<br />
9, 12, 15 months, underlying bond maturity T =2, 5, 10, 30 years. The …rst short rate is drawn<br />
randomly from the stationary distribution of CIR process. To simplify the summary of the simulation<br />
result, we consider only at-the-money options which tend to be the most liquid contracts in practice.<br />
As a result, we study the simulation performance of various estimators for three-dimensional state<br />
variables –option maturity, underlying bond maturity, and short rate. In the simulation, the true<br />
CIR parameters are set to the estimates in Aït-Sahalia (1999)<br />
k = 0:145; = 0:0732; = 0:06521 (3.4)<br />
15
and we add to the true option price a zero-mean normally distributed random variable whose standard<br />
deviation equals to one percent of the true option price to generate the observed option price in the<br />
simulation. This extra random variable is intended to capture market microstructure e¤ects such as<br />
the bid-ask bounce. At the true parameter, the bond option prices average to around $1 hence the<br />
pricing errors can be interpreted either as dollar pricing errors or as proportional pricing errors.<br />
3.1. Simulation result: parametric and nonparametric prices<br />
Table 1 panel A shows the performance of the various option price estimators. When an investor<br />
knows correctly the underlying data generating process, parametric option price estimator performs<br />
the best, generating an average pricing error of only 0.022 cents <strong>with</strong> …ve years of weekly observations.<br />
7<br />
However, the accuracy of the option prices depends crucially on the validity of the model.<br />
When the model is misspeci…ed, the parametric prices result in an error of 4.1 cents which is about<br />
200 times that when the model is correct. This calls for caution in practice <strong>with</strong> parametric prices<br />
when the validity of the pricing model is in question.<br />
Nonparametric prices, on the other hand,<br />
do not depend on any model and avoid misspeci…cation. In the simulation, nonparametric prices<br />
register an average pricing error of 1.3 cents, about 70% less than the parametric prices when the<br />
model is wrong. 8<br />
However, the nonparametric prices ignore all model information (whether correct<br />
or not) and perform much worse than parametric prices from a correctly speci…ed model.<br />
3.2. Simulation result: robust parametric prices<br />
The pricing method proposed in this paper aims to achieve a continuous middle ground between<br />
parametric and nonparametric prices. Table 1 panel A shows that option prices from the robust<br />
parametric method have an average pricing error of 0.15 cents, about 7 times larger than that of the<br />
parametric prices using the correct model yet 27 times smaller than the parametric estimates from a<br />
wrong model. The error of 0.15 cents is also an order of magnitude smaller than the nonparametric<br />
errors.<br />
To see the intuition of improvement, let us turn to panel B in Table 1 and Figures 1 and 2. In<br />
7 We use nonlinear least squares to …nd the parameter estimates.<br />
8 We use the Nadaraya-Watson nonparametric estimator <strong>with</strong> uniform kernel and cross-validation bandwidth selection<br />
in the simulation, see Härdle and Linton (1994) for more details.<br />
16
panel B of Table 1, the regions of …t along the option maturity and bond maturity dimensions are<br />
both zero, indicating that the wrong model provides a poor …t of the true bond option prices along<br />
these two dimensions. 9 Figures 1 and 2 further illustrate this. Figure 1 plots the true and estimated<br />
option prices at various option maturities.<br />
The robust parametric price estimate at one month<br />
maturity, which is estimated using observations whose option maturity is around one month, …ts<br />
the true one-month option price almost exactly. This con…rms that even a wrong model can provide<br />
good local …t (see (2.1)). However, the local …t does not extrapolate well to option maturities slightly<br />
di¤erent from one month. 10<br />
Similarly around the other option maturities shown in Figure 1. This<br />
indicates severe model misspeci…cation along the dimension of option maturity. Figure 2 illustrates<br />
that the Vasicek option price …ts the CIR option price poorly along the dimension of bond maturity,<br />
too.<br />
That the Vasicek option price …ts the CIR option price poorly along option maturity and bond<br />
maturity is the reason why the robust parametric pricing method improves over parametric estimation<br />
using a misspeci…ed model.<br />
When the model …ts poorly along these two dimensions, the robust<br />
parametric method sets the regions of …t to zero along both dimensions and conducts estimation using<br />
only observations <strong>with</strong> the same bond maturity and option maturity to address misspeci…cation.<br />
The result is di¤erent along the dimension of short rate. Panel B of Table 1 shows that the region<br />
of …t is 0.026 along this dimension. I.e., if one is estimating the option price at short rate 7%, the<br />
robust parametric estimator will use all observations whose short rates are between 4.4% and 9.6%.<br />
Figure 3 con…rms that the Vasicek price estimates can be extrapolated to nearby short rates and<br />
still …t the true option prices reasonably well (the two option price curves almost overlap). This<br />
allows the robust parametric method to enlarge the region of …t to include more observations which<br />
improves the estimation e¢ ciency. This is the intuition why the proposed robust parametric method<br />
performs better than nonparametric method –it retains valid model restrictions.<br />
9 To be exact, the region of …t for option maturity averages to 0.01. However, because the observations come in<br />
weekly and the interval between successive observations of option maturity is at least 1=52 0:02, the region of …t for<br />
option maturity is essentially zero.<br />
10 Using the notation in section 2, extrapolation here refers to using f (X; (x)) to estimate option prices at state<br />
variable X di¤erent from x.<br />
17
3.3. Simulation result: comparison <strong>with</strong> numerical methods<br />
The proposed robust parametric method can add value even when the correct model is known. A<br />
true model is likely complicated.<br />
A common situation is that the closed-form pricing formula is<br />
unavailable as a result of the complexity. For example, many continuous-time term structure models<br />
do not render closed-form bond option pricing formula. The Vasicek and CIR models used in the<br />
simulation, along <strong>with</strong> a handful of other models, constitute the exception. Sometimes, numerical<br />
methods can be used to obtain the option prices. For example, Du¢ e, Pan, and Singleton (2000)<br />
show that option prices can be obtained via numerical integration of characteristic functions.<br />
In this section, we compare the performance of the proposed pricing method to the performance of<br />
parametric estimation which uses true model and numerical method to get option prices. Speci…cally,<br />
the robust parametric estimator still uses the closed-form Vasicek option pricing formula which is<br />
misspeci…ed. On the contrary, the parametric estimator uses the true CIR model but pretends that<br />
this is a model so complicated that closed-form option pricing formula is unavailable.<br />
To control the magnitude of the numerical error in this study, we …rst model numerical errors by<br />
C NUM = C (1 + ")<br />
where C is the true option price from the closed-form CIR pricing formula. " is set to be a uniformly<br />
distributed random variable over [<br />
!; !]. C NUM is assumed to be the option prices obtained from<br />
numerical methods. In this simulation, we do not actually use a numerical method but rather we<br />
start from the closed-form option price and let ! proxy the degree of numerical error. When ! = 0,<br />
numerical error disappears and we return to the case of parametric estimation using the closed-form<br />
formula. A larger ! indicates larger numerical error. We use ! = 0:01%, 0:1%, 0:2%, 0:3%, 0:5%, and<br />
1% in the simulation and the results are shown in panel C of Table 1. The proposed robust parametric<br />
method using the misspeci…ed Vasicek model is comparable in performance to parametric estimation<br />
using the true model when the numerical error is between 0.2% and 0.3%.<br />
This is remarkable<br />
because Vasicek option prices are grossly misspeci…ed relative to CIR option prices.<br />
Parametric<br />
estimation using the Vasicek model results in an error that is 200 times that when the model is<br />
correctly speci…ed (see panel A). Panel B further shows that the Vasicek model does not …t CIR<br />
18
model along the dimensions of bond maturity and option maturity at all. Nonetheless, adjusting for<br />
misspeci…cation using the proposed robust parametric method improves the estimation performance<br />
to the equivalent of parametric estimation using true model <strong>with</strong> a numerical error of around 0.25%.<br />
Next, to see what the magnitude of numerical error can be in practice, we follow Du¢ e, Pan,<br />
and Singleton (2000) and generate option prices using the correct CIR model by actual numerical<br />
integration. The estimation RIMSE is shown in panel D of Table 1. The result is comparable to<br />
the 1% error in panel C. In practice, numerical integration precision can be improved at the cost<br />
of longer computing time. Therefore, the result in panel D should be taken <strong>with</strong> caution. However,<br />
even <strong>with</strong> a relatively tractable model like CIR, there are already non-trivial issues <strong>with</strong> numerical<br />
integration. For example, Carr and Madan (1999) point out that poor numerical precision can result<br />
from the highly oscillatory nature of the characteristic function in the integrand. When the true<br />
model becomes more complicated, the numerical errors are likely more di¢ cult to understand and<br />
control, especially when the option prices need to be evaluated at all parameter values searched<br />
in the estimation procedure. This implies that sometimes it may be desirable to use a simple but<br />
misspeci…ed model and explicitly adjust for misspeci…cation using the proposed robust parametric<br />
method.<br />
4. Empirical application –Treasury options pricing<br />
We next apply the robust parametric method to the pricing of Treasury options traded on CBOT,<br />
examine its pricing performance both in terms of in-sample …t and out-of-sample performance, and<br />
see what if any model misspeci…cation it can help investors infer.<br />
We collect weekly call option closing price data from CBOT. The sample period is from May<br />
1990 to December 2006. CBOT lists options on 2-, 5-, 10-, and 30-year Treasuries. These options are<br />
more precisely options on Treasury futures. However, those option maturities <strong>with</strong> the most trading<br />
volume (March, June, September, and December) coincide <strong>with</strong> futures expiration. Therefore, upon<br />
option exercise, the delivery is essentially made in the underlying Treasuries. We focus on the option<br />
maturities of March, June, September, and December and will refer to the options as Treasury<br />
options for simplicity. The 2-year option does not have much trading volume and is excluded from<br />
the analysis. To reduce data error, we eliminate those observations where the recorded option price is<br />
19
less than the intrinsic value, i.e., if C < max(F<br />
K; 0) where C, F , and K are the observed Treasury<br />
call option price, observed Treasury futures price, and option strike, respectively. Further, for each<br />
option contract, we use only data for the at-the-money contract (contract whose F is closest to K)<br />
which tends to have the most trading volume. There are a few instances where CBOT supplies a<br />
closing option price but indicates a trading volume of zero. Such observations are eliminated.<br />
As in section 3, we apply the robust parametric pricing method using the possibly misspeci…ed<br />
Vasicek (1977) model which allows for closed-form bond option pricing formula (see appendix F.2). 11<br />
The Vasicek (1977) option pricing formula assumes that a zero-coupon bond underlies the option.<br />
This di¤ers from the cheapest to deliver practice of CBOT listed options where the delivery may not<br />
be in the form of a zero coupon bond. Because we do not have information on the cheapest Treasury<br />
for delivery, we use the following procedure to adjust for unknown coupon of the bond issue used<br />
for delivery. Speci…cally, we convert the delivery bond into a zero coupon bond by assuming that<br />
all the coupons are paid at bond maturity. This assumption ignores the time value between coupon<br />
payment and bond maturity. It is an imperfect way to model the cheapest to deliver practice and<br />
we will discuss more on this issue later. Now the problem of unknown coupon size is translated to<br />
the new problem of unknown par value which we can back out using the observed Treasury futures<br />
price from CBOT. Speci…cally, let M denote the unknown par value, then M can be computed from<br />
M =<br />
F<br />
F (; T; r)<br />
where F is the observed CBOT Treasury futures price, F (; T; r) is the Vasicek (1977) implied futures<br />
price on a zero coupon bond <strong>with</strong> par=$1 (see appendix F.2). 12 This allows us to construct a pricing<br />
formula for the option<br />
C adj (; T; r; K) = M C(; T; r; K M ) (4.1)<br />
where C adj is the call option price adjusted for the cheapest to delivery practice, C is the Vasicek<br />
(1977) pricing formula for call option on a Treasury zero coupon bond <strong>with</strong> $1 par, is the option<br />
11 We have alternatively estimated a model in which the short rate follows the Cox, Ingersoll, and Ross (1985) process.<br />
The result is similar. It is suppressed for brevity and available from the authors upon request.<br />
12 The CBOT Treasury futures price data are from Datastream.<br />
20
maturity, T is the bond maturity, r is the short rate which is measured by one month Treasury bill<br />
rate, and K is the option strike price.<br />
We compare both the in-sample …t and out-of-sample performance of three pricing methods: the<br />
robust parametric method proposed in this paper, the parametric method, and the nonparametric<br />
method. 13<br />
4.1. Misspeci…cation of Treasury option pricing models<br />
We use the root integrated mean squared error (RIMSE) de…ned in (3.3) to measure the in-sample<br />
performance of various estimators. The result is shown in panel A of Table 2. In the sample, the<br />
model is so misspeci…ed that the nonparametric prices which completely ignore the model provide<br />
better …t than parametric prices.<br />
Nonetheless, the model remains useful because the proposed<br />
robust parametric method which uses the same model but adjusts for possible misspeci…cation does<br />
better than either parametric or nonparametric methods. The robust parametric pricing method<br />
also produces the highest R-square in the regression of observed option prices on …tted option prices<br />
–90.2% versus 49.8% and 74.4% from parametric and nonparametric estimators, respectively. The<br />
improvement in R-square is consistent <strong>with</strong> the scatter plots shown in Figure 4.<br />
The intuition of the improved …t is the same as that observed in section 3: the robust parametric<br />
method imposes the model only in the region where the model provides a good approximation of<br />
reality and discards the model elsewhere. Figure 5 shows the RIMSE for various regions of …t along<br />
the dimensions of option maturity, bond maturity, and short rate. The robust parametric method<br />
selects a region of …t separately for each dimension to minimize the RIMSE (see section 2.2). In the<br />
sample period from May 1990 to December 2006, the Vasicek (1977) model does not …t bond option<br />
prices well along the dimensions of bond maturity and short rate. Therefore, regions of …t are set to<br />
minimum along these two dimensions –0 for bond maturity and 0.5% for short rate (we set minimum<br />
possible region of …t along the short rate to 0.5% to make sure there are su¢ cient observations in the<br />
estimation). This implies that, to estimate option prices at bond maturity T and short rate r, the<br />
robust parametric method uses only those observations at which the options are written on a T -year<br />
13 We use nonlinear least squares in the parametric and robust parametric estimations. We use the Nadaraya-Watson<br />
nonparametric estimator <strong>with</strong> uniform kernel and cross-validation bandwidth selection, see Härdle and Linton (1994)<br />
for more details.<br />
21
ond and the short rates are <strong>with</strong>in [r<br />
0:005; r + 0:005] to minimize the impact of misspeci…cation,<br />
which improves its performance relative to the parametric method.<br />
The improvement of the robust parametric method over nonparametric method is due to the<br />
model providing some useful restrictions along the dimension of option maturity. In Figure 5, the<br />
RIMSE bottoms out when the region of …t is set to 3 weeks for option maturity. This implies that the<br />
Vasicek option pricing formula provides a good approximation for observations <strong>with</strong> adjacent option<br />
maturity –when estimating option prices at option maturity , the robust parametric method will<br />
use observations whose option maturities are <strong>with</strong>in 3 weeks of . 14<br />
The information provided by the regions of …t along various dimensions of the state variable<br />
can be used to triangulate model misspeci…cation which is useful for future development of asset<br />
pricing models. In this case, the model …ts well along the dimension of option maturity but not<br />
along bond maturity and short rate. Pinpointing the exact cause of bond options misspeci…cation<br />
requires a separate study, though the evidence is suggestive that the cheapest-to-deliver (CTD)<br />
practice associated <strong>with</strong> the CBOT Treasury futures/options plays a role. CTD refers to the right of<br />
the short party to deliver any Treasuries designated eligible by CBOT. For example, for the 10 year<br />
contracts, deliverable grades include US Treasury notes maturing at least 6 1/2 years, but no more<br />
than 10 years, from the …rst day of the delivery month. To address the fact that Treasuries vary<br />
in their coupon, maturity, and other features, CBOT uses a system known as the conversion factor<br />
to equalize various bonds. According to CBOT, the conversion factor is the price of the delivered<br />
note ($1 par value) to yield 6 percent and the invoice price equals the futures settlement price times<br />
the conversion factor plus accrued interest. The conversion system usually makes some bonds less<br />
costly to deliver than others, which is not captured by the typical bond option pricing formula based<br />
on term structure models.<br />
The actual cheapest-to-deliver bond varies across contracts involving<br />
di¤erent bond maturities and across di¤erent interest rate environments (see, for example, Kane<br />
and Marcus (1984) and Livingston (1987)) which is consistent <strong>with</strong> the misspeci…cation along the<br />
dimensions of bond maturity and short rate indicated by the regions of …t. The region of …t for option<br />
maturity, on the contrary, shows good …t up to 3 weeks. Observations less than 3 weeks apart are<br />
14 The optimal region of …t along the dimension of option maturity is 2 weeks if the Cox, Ingersoll, and Ross (1985)<br />
process is used instead of the Vasicek (1977) process to model the short rate. The optimal regions of …t along bond<br />
maturity and short rate remain the same. This suggests better …t of Vasicek (1977) process for the purpose of modeling<br />
CBOT Treasury option prices.<br />
22
likely consecutive weekly observations of the same contract for which the cheapest-to-deliver bonds<br />
are likely similar or even identical. Therefore, the evidence suggests that the cheapest to deliver<br />
practice is an important source of misspeci…cation for Treasury option pricing which is ignored by<br />
typical bond option pricing formulas based on term structure models.<br />
4.2. Out-of-sample performance<br />
To con…rm that the improved …t is not due to over…tting and can be extrapolated out of the sample,<br />
panel B of Table 2 shows the out-of-sample comparison of the proposed robust parametric method<br />
to parametric and nonparametric methods. Speci…cally, model parameters are estimated using …ve<br />
years of weekly observations which are then used to compute RIMSE and regression R-square in<br />
the subsequent year. Because the sample period starts in May 1990, the …rst year of out-of-sample<br />
comparison is 1996. Panel B shows the RIMSE and the R-square in the regression of observed option<br />
prices on predicted option prices for each year separately. The robust parametric pricing method has<br />
the lowest out-of-sample error in all years. Overall, the robust parametric method has a reduction<br />
of 46.6 and 33.9 percent in RIMSE, and an increase of 39.6 and 16.5 percentage points in R-square<br />
relative to parametric and nonparametric methods, respectively.<br />
5. Conclusion<br />
Misspeci…ed pricing formula confronts most investors.<br />
This paper proposes a robust parametric<br />
pricing method which utilizes information in a model yet explicitly controls for possible misspeci…-<br />
cation. The resulting price estimator is consistent irrespective of misspeci…cation and it provides a<br />
continuous middle ground between parametric and nonparametric estimators in terms of e¢ ciency.<br />
It can improve pricing precision over parametric methods when the model is misspeci…ed. Because<br />
it retains valid information supplied by a model, the robust parametric pricing method does not<br />
su¤er as much from the “curse of dimensionality” problem faced by nonparametric methods and is<br />
applicable to high dimensional pricing problems or sensitivity analyses.<br />
Model restrictions also help to alleviate the concern of over…tting. As pointed out by Campbell,<br />
Lo, and MacKinlay (1997) (page 524), “... perhaps the most e¤ective means of reducing the impact<br />
of over…tting and data-snooping is to impose some discipline on the speci…cation search by a priori<br />
23
theoretical considerations.” The estimator proposed in this paper does exactly that – it confronts<br />
the data <strong>with</strong> an a priori model. This is con…rmed by the out-of-sample performance in section 4.2.<br />
Using an approximate (i.e., misspeci…ed) model may also provide other advantages. For example,<br />
the true model can be complicated and it may sometimes be preferable to use a simple yet misspeci…ed<br />
model, as pointed out by Fiske and Taylor (1991) (page 13), “...<br />
People adopt strategies that<br />
simplify complex problems; the strategies may not be normatively correct or produce normatively<br />
correct answers, but they emphasize e¢ ciency.” Interestingly, one of the simulations shows that<br />
applying the proposed estimator on a good parsimonious model and explicitly adjusting for possible<br />
misspeci…cation can sometimes do better than fully parametric estimation using a complicated model<br />
even if the complicated model is correctly speci…ed.<br />
This echoes the “maxim of parsimony” in<br />
Ploberger and Phillips (2003) and opens up additional application areas for the estimator when the<br />
true model is complicated and a good parsimonious (though misspeci…ed) model is available.<br />
The robust parametric pricing method can have interesting implications on asset pricing equilibrium<br />
based on bounded rationality. According to Simon (2008), “The term ‘bounded rationality’<br />
is used to designate rational choice that takes into account the cognitive limitations of the decision<br />
maker –limitations of both knowledge and computational capacity. Bounded rationality is a central<br />
theme in the behavioral approach to economics, which is deeply concerned <strong>with</strong> the ways in which<br />
the actual decision-making process in‡uences the decisions that are reached.”If using a simple model<br />
and explicitly adjusting for possible misspeci…cation can achieve good results and if the economy is<br />
populated <strong>with</strong> investors <strong>with</strong> su¢ cient sophistication to act accordingly, it can mitigate the e¤ect<br />
of cognitive costs associated <strong>with</strong> complex reality on the equilibrium outcome. On the other hand,<br />
the robust parametric method itself introduces an additional layer of complexity into the economy<br />
(e.g., the uncertainty on if and when other investors adopt the method) and can amplify the e¤ect<br />
of cognitive limitation. Such implications have interesting potentials and await future studies.<br />
References<br />
Aït-Sahalia, Y., 1999, “Transition Densities for Interest Rate and Other Nonlinear Di¤usions,”Journal<br />
of Finance.<br />
24
Aït-Sahalia, Y., and J. Duarte, 2003, “Nonparametric Option <strong>Pricing</strong> under Shape Restrictions,”<br />
Journal of Econometrics, 116, 9–47.<br />
Amihud, Y., H. Mendelson, and L. H. Pedersen, 2005, “Liquidity and <strong>Asset</strong> Prices,” Foundations<br />
and Trends in Finance, 1, 269–364.<br />
Bandi, F., and P. Phillips, 2003, “Fully Nonparametric Estimation of Scalar Di¤usion <strong>Models</strong>,”<br />
Econometrica, 71, 241–284.<br />
Black, F., and M. Scholes, 1973, “The <strong>Pricing</strong> of Options and Corporate Liabilities,” Journal of<br />
Political Economy, 81, 637–654.<br />
Campbell, J. Y., A. W. Lo, and A. C. MacKinlay, 1997, The Econometrics of Financial Markets,<br />
Princeton University Press.<br />
Carr, P., and D. B. Madan, 1999, “Option valuation using the fast Fourier transform,” Journal of<br />
computational …nance, 3, 463–520.<br />
Chen, R.-R., 1992, “Exact solutions for futures and European futures options on pure discount<br />
bonds,”Journal of Financial and Quantitative Analysis, 27.<br />
Cox, J. C., J. E. Ingersoll, and S. A. Ross, 1985, “A Theory of the Term Structure of Interest Rates,”<br />
Econometrica, 53.<br />
Du¢ e, D., J. Pan, and K. Singleton, 2000, “Transform Analysis and <strong>Asset</strong> <strong>Pricing</strong> for A¢ ne Jump<br />
Di¤usions,”Econometrica, 68, 1343–1376.<br />
Du¢ e, D., and K. J. Singleton, 1997, “An econometric model of the term structure of interest-rate<br />
swap yields,”Journal of Finance.<br />
Fan, J., 1992, “Design-adaptive Nonparametric Regression,” Journal of the American Statistical<br />
Association, 87.<br />
Fan, J., and I. Gijbels, 1996, Local Polynomial Modelling and Its Applications, Chapman Hall,<br />
London, U.K.<br />
Fiske, S. T., and S. E. Taylor, 1991, Social Cognition, McGraw Hill, second edn.<br />
25
Gibbons, M. R., S. A. Ross, and J. Shanken, 1989, “A test of the e¢ ciency of a given portfolio,”<br />
Econometrica, 57, 1121–1152.<br />
Gozalo, P., and O. Linton, 2000, “Local nonlinear least squares: Using parametric information in<br />
nonparametric regression,”Journal of Econometrics, 99, 63–106.<br />
Greene, W. H., 1997, Econometric Analysis, Prentice Hall, third edn.<br />
Hansen, L., and R. Jagannathan, 1991, “Implications of Security Market Data for <strong>Models</strong> of Dynamic<br />
Economies,”Journal of Political Economy, 99, 225–262.<br />
Hansen, L., and R. Jagannathan, 1997, “Assessing speci…cation errors in stochastic discount factor<br />
models,”Journal of …nance.<br />
Hansen, L. P., 1982, “Large Sample Properties of Generalized Method of Moments Estimators,”<br />
Econometrica, 50, 1029–1054.<br />
Hausman, J. A., 1978, “Speci…cation tests in econometrics,”Econometrica, 46.<br />
Härdle, W., P. Hall, and J. S. Marron, 1988, “How Far are Automatically Chosen Regression Smoothing<br />
Parameters From Their Optimum?,”Journal of the American Statistical Association, 83.<br />
Härdle, W., and O. Linton, 1994, “Applied nonparametric methods,” in R.F. Engle, and D.L. Mc-<br />
Fadden (ed.), Handbook of Econometrics, vol. 4, . chap. 38, pp. 2295–2339, Elsevier.<br />
Härdle, W., and J. S. Marron, 1985, “Optimal Bandwidth Selection in Nonparametric Regression<br />
Function Estimation,”Annals of Statistics, 13, 1465–1481.<br />
Jamshidian, F., 1989, “An exact bond option formula,”Journal of Finance, 44, 205–209.<br />
Jordan, B. D., and D. R. Kuipers, 1997, “Negative option values are possible: the impact of Treasury<br />
bond futures on the cash U.S. Treasury market,”Journal of Financial Economics, 46, 67–102.<br />
Kane, A., and A. J. Marcus, 1984, “Conversion factor risk and hedging in the Treasury-bond futures<br />
market,”Journal of futures markets, 4, 55–64.<br />
Livingston, M., 1987, “The e¤ect of coupon level on Treasury bond futures delivery,” Journal of<br />
futures markets, 7, 303–309.<br />
26
Matzkin, R. L., 1994, “Restrictions of economic theory in nonparametric methods,” in R.F. Engle,<br />
and D.L. McFadden (ed.), Handbook of Econometrics, vol. 4, . chap. 42, pp. 2523–2558, Elsevier.<br />
Merton, R., 1973, “Rational theory of option pricing,”Bell Journal of Economics and Management<br />
Science, 4, 141–183.<br />
Newey, W. K., and D. McFadden, 1994, “Large sample estimation and hypothesis testing,” in R.F.<br />
Engle, and D.L. McFadden (ed.), Handbook of Econometrics, vol. 4, . chap. 36, pp. 2111–2245,<br />
Elsevier.<br />
Ploberger, W., and P. C. B. Phillips, 2003, “Empirical Limits for Time Series Econometric <strong>Models</strong>,”<br />
Econometrica, 71, 627–673.<br />
Powell, J. L., 1994, “Estimation of semiparametric models,”in R.F. Engle, and D.L. McFadden (ed.),<br />
Handbook of Econometrics, vol. 4, . chap. 41, pp. 2443–2521, Elsevier.<br />
Simon, H. A., 2008, “Bounded rationality,”in Steven N. Durlauf, and Lawrence E. Blume (ed.), The<br />
new palgrave dictionary of economics. Palgrave Macmillan.<br />
Tibshirani, R., and T. Hastie, 1987, “Local Likelihood Estimation,”Journal of the American Statistical<br />
Association, 82.<br />
Vasicek, O., 1977, “An Equilibrium Characterization of the Term Structure,” Journal of Financial<br />
Economics, 5, 177–188.<br />
White, H., 1982, “Maximum Likelihood Estimation of Misspeci…ed <strong>Models</strong>,”Econometrica, 50, 1–26.<br />
27
Appendix: Assumptions, Proofs, Option <strong>Pricing</strong> Formulas<br />
A. Assumptions<br />
First, we collect the regularity conditions assumed in this paper. Recall that we want to estimate the pricing<br />
formula P (X) where X 2 R d is the state variable. We assume an investor has an economic model which<br />
implies a possibly misspeci…ed pricing formula f (X; ) for P (X). 2 R p .<br />
Assumption 1 There exists a unique function (X) such that f (X; (X)) = P (X). The range of (X) is<br />
in a compact set .<br />
When the model is correctly speci…ed, (X) is a constant. For all practical purposes, knowing f (X; (X))<br />
amounts to knowing the true pricing formula P (X).<br />
Assumption 2 P (X) and f (X; ) are bounded and thrice-continuously di¤erentiable <strong>with</strong> respect to X and<br />
<strong>with</strong> bounded derivatives.<br />
Assumption 3 (Sample) The sample consists of independent and identically distributed observations fx i ; y i g n i=1<br />
where<br />
y i = P (x i ) + " i :<br />
" i is an independent random error satisfying E [" i j X = x i ] = 0, Var[" i j X = x i ] = v (x i ) > 0. v () is continuous<br />
and bounded.<br />
Assumption 4 inf X; kf (X; ) f T (X; )k > 0. For any x 2 R d and any 2 , there exists an H > 0, a<br />
neighborhood N of , a non-random function G (; h) continuously di¤erentiable at 2 N and 0 h H,<br />
and random variables Z h N (0; (h)) indexed by h where (h) is continuous at h = 0 such that<br />
sup<br />
2N;0hH<br />
n 1<br />
sup<br />
0hH<br />
n X<br />
i=1<br />
n 1=2<br />
f x h i ; f T x h i ; G (; h)<br />
= O p<br />
n X<br />
i=1<br />
f x h i ; " i x h <br />
i<br />
Z h <br />
= o p (1)<br />
n 1=2<br />
where the sequence of independent random variables x h i<br />
n<br />
i=1 satis…es x<br />
h<br />
i x h for all i.<br />
Other than requiring uniformity over h H, Assumption 4 is standard in large sample asymptotics (see<br />
Newey and McFadden (1994)). Restricting to x<br />
h<br />
i<br />
x h is because the model will be estimated locally<br />
using observations less than h away from x when the model is misspeci…ed.<br />
28
Let p (X) denote the probability density function of X.<br />
Assumption 5 p (x) > 0 for all x 2 R d , p () is twice-continuously di¤erentiable.<br />
We next prove the propositions. Recall that n x; b h<br />
denotes the number of observations less than b h away<br />
<br />
from x. When X is d-dimensional, n x; b h<br />
= O p n b <br />
h d when n ! 1, b h ! 0, and n b h d ! 1.<br />
B. Proof of Proposition 1<br />
See Theorem 1 in Gozalo and Linton (2000).<br />
C. Proof of Proposition 2<br />
Using the standard large sample asymptotics argument (see for example Newey and McFadden (1994)),<br />
=<br />
p<br />
nx; b h<br />
<br />
b (x)<br />
0<br />
@ 1<br />
n x; b h<br />
kx i<br />
X<br />
xk b h<br />
(x)<br />
<br />
F i F T i<br />
1<br />
A<br />
1<br />
n 1=2<br />
x; b h<br />
kx i<br />
X<br />
xk b h<br />
F i (y i<br />
<br />
f (x i ; (x))) + O p n 1=2<br />
x; b + b h 2k+2<br />
h<br />
(C.1)<br />
To simplify notation, F i f (x i ; (x)). Without model misspeci…cation (i.e., if k ! 1), this is the standard<br />
asymptotics result. With misspeci…cation, (C.1) is modi…ed slightly because y i<br />
f (x i ; (x)) does not have<br />
mean zero for x i 6= x. The magnitude of the bias<br />
E 0
D. Proof of Proposition 3<br />
The crossvalidation criterion function<br />
By (2.10) and (2.9),<br />
CV (h) = 1 n<br />
= 1 n<br />
+ 2 n<br />
nX<br />
i=1<br />
h<br />
y i<br />
f<br />
nX<br />
" 2 i + 1 n<br />
i=1<br />
<br />
x i ; b i 2<br />
i;h (x i )<br />
nX<br />
i=1<br />
h<br />
P (x i )<br />
nX<br />
" i<br />
hP (x i ) f<br />
i=1<br />
f<br />
<br />
x i ; b i 2<br />
i;h (x i )<br />
<br />
x i ; b i<br />
i;h (x i ) :<br />
1<br />
n<br />
nX<br />
i=1<br />
h<br />
P (x i )<br />
f<br />
<br />
x i ; b i 2 <br />
i;h (x i ) = Op h 4k+4 + nh d 1<br />
:<br />
(D.1)<br />
We will later prove the following lemma.<br />
Lemma 1 Under the conditions of Proposition 3,<br />
1<br />
n<br />
nX<br />
" i<br />
hP (x i ) f<br />
i=1<br />
= o p<br />
1<br />
n<br />
nX<br />
i=1<br />
h<br />
P (x i )<br />
<br />
x i ; b i<br />
i;h (x i )<br />
f<br />
!<br />
<br />
x i ; b i 2<br />
i;h (x i ) :<br />
(D.2)<br />
Lemma 1 and (D.1) imply<br />
CV (h) = 1 nX<br />
" 2 i + O p<br />
h 4k+4 + nh d 1<br />
: (D.3)<br />
n<br />
i=1<br />
b h, which minimizes CV (h), satis…es<br />
b h = n 1=(4+4k+d) :<br />
It can then be calculated using (2.10) that the resulting estimation error is<br />
P (x) = f<br />
<br />
x; b <br />
(x) + O n (2+2k)=(4+4k+d) :<br />
30
E. Proof of Lemma 1<br />
Recalling P (x i ) = f (x i ; (x i )), Taylor expansion around gives<br />
<br />
P (x i ) f x i ; b <br />
i;h (x i )<br />
<br />
= f T (x i ; (x i )) (x i ) b i;h (x i ) + O p (xi ) b i;h (x i ) 2<br />
= F T i<br />
=<br />
2<br />
4<br />
0
nX<br />
The …rst term 1 n<br />
" i T h (x i ) has a zero mean and its variance is<br />
i=1<br />
V ar<br />
"<br />
1<br />
n<br />
#<br />
nX<br />
" i T h (x i )<br />
i=1<br />
= 1 X<br />
n<br />
n 2 V ar [" i T h (x i )] + 1 n 2<br />
i=1<br />
n X<br />
X<br />
Cov [" i T h (x i ) ; " u T h (x u )]<br />
i=1 u6=i<br />
(E.6)<br />
First, we bound the variance terms in (E.6).<br />
V ar [" i T h (x i )] = E " 2 h<br />
i E<br />
= E " 2 <br />
i<br />
= O p<br />
<br />
0
Next, we bound the second term in (E.5) involving U h . Expand w (x i ; x j ) around x j = x i ,<br />
w (x i ; x j ) = w (x i ; x i ) + w 2 (x i ; ex j ) (x j x i ) :<br />
where w 2 (; ) denotes derivative of w (; ) <strong>with</strong> respect to the second argument. ex j is in between x i and x j .<br />
From (E.2),<br />
w (x i ; x i ) = O p<br />
<br />
w 2 (x i ; ex j ) = O p<br />
<br />
nh d 1 <br />
nh d 1 :<br />
(E.12)<br />
Recall the model matches P (X) up to its 2k-th derivative (see (2.8)),<br />
U h (x) =<br />
=<br />
0
variance bounded by (recall that each remainder term depends only on observations less than h away from x i<br />
so the remainder terms are independent of each other if x i and x u are more than 2h apart)<br />
O p<br />
1<br />
n 2 n <br />
h 4k+4 + nh d 1 2<br />
+<br />
1<br />
n 2 n nhd h 4k+4 + nh d 1 2 <br />
= O p<br />
<br />
h 4k+4 + nh d 1 2<br />
n 1 + h d<br />
which implies that the sum of the remainder terms in (E.5) is of order<br />
O p<br />
h 4k+4 + nh d 1 h d 1=2 : (E.16)<br />
That n 1 = o p h d is because the crossvalidation criterion is minimized over h that satis…es nh d ! 1.<br />
Combining (E.11), (E.15) and (E.16), (E.5) becomes (notice the fact that 2 jabj a 2 + b 2 )<br />
1<br />
n<br />
<br />
= O p<br />
nX<br />
" i<br />
hP (x i ) f<br />
i=1<br />
<br />
x i ; b i<br />
i;h (x i )<br />
nh d <br />
1=2<br />
n 1=2 + h 2k+2 + nh d 1=2<br />
n 1=2 +<br />
h 4k+4 + nh d 1<br />
h d 1=2<br />
h 4k+4 + nh d 1 h d 1=2 <br />
= O p<br />
h 2k+2 n 1=2 +<br />
<br />
= O p h 2k+2 h d 1=4<br />
n h d 1=2 1=2 <br />
+ h 4k+4 + nh d 1<br />
h d <br />
1=2<br />
= O p<br />
<br />
h 4k+4<br />
h d <br />
1=2<br />
+ n h d 1=2 1 <br />
+ h 4k+4 + nh d 1<br />
h d <br />
1=2<br />
= O p<br />
<br />
h 4k+4 + nh d 1 h d 1=2 <br />
which, together <strong>with</strong> (D.1), proves Lemma 1.<br />
F. Option pricing formula<br />
F.1.<br />
CIR model<br />
When the short rate follows the CIR model in (3.1), the price of a call option <strong>with</strong> maturity and strike price<br />
K on a T -year Treasury zero-coupon bond <strong>with</strong> par $1 is given by Cox, Ingersoll, and Ross (1985),<br />
C (; T; r 0 ; K) = B (r 0 ; T ) 2 2r [ () + B(T )] ; 4<br />
2 ;<br />
KB (r 0 ; ) 2<br />
2r [ () + ] ; 4<br />
2 ; 2 ()2 r 0 e <br />
() +<br />
!<br />
2 () 2 r 0 e <br />
() + B(T )<br />
!<br />
34
where r 0 is the short rate at the time of option pricing and 2 (; n; c) denotes the cumulative probability<br />
distribution function of a non-central Chi-square distribution <strong>with</strong> degree of freedom n and non-centrality<br />
parameter c. The other terms used in the option pricing formula are<br />
B (r 0 ; T ) = A (T ) exp (B (T ) r 0 )<br />
A (T ) =<br />
<br />
2 exp 1 2<br />
(k + ) T<br />
(k + ) (exp (T ) 1) + 2<br />
B (T ) =<br />
2 (exp (T ) 1)<br />
(k + ) (exp (T ) 1) + 2<br />
! 2k<br />
2<br />
p k 2 + 2 2<br />
<br />
r 1 A (T )<br />
=<br />
B (T ) log K<br />
2<br />
() =<br />
2 (e 1)<br />
= + <br />
2 :<br />
F.2.<br />
Vasicek model<br />
When the short rate follows the Vasicek model in (3.2), the price of a call option <strong>with</strong> maturity and strike<br />
price K on a T -year Treasury zero-coupon bond <strong>with</strong> par $1 is given by Jamshidian (1989),<br />
C (; T; r 0 ; K) = B (r 0 ; T ) (z 1 ) KB (r 0 ; ) (z 2 )<br />
where r 0 is the short rate at the time of option pricing and () denotes the cumulative probability distribution<br />
function of a standard normal random variable. The other terms used in the option pricing formula are<br />
B (r 0 ; T ) = exp [A (T ) + B (T ) r 0 ]<br />
A (T ) =<br />
B (T ) =<br />
2<br />
4k B (T )2 (T + B (T ))<br />
1<br />
k 1 e kT <br />
<br />
<br />
2 <br />
2k 2<br />
35
z 1 = 1 B (r0 ; T )<br />
log<br />
p B (r 0 ; ) K<br />
z 2 = 1 p<br />
log<br />
B (r0 ; T )<br />
B (r 0 ; ) K<br />
<br />
+ p<br />
2<br />
<br />
p<br />
2<br />
s<br />
(1 e 2 ) 1 e (T ) 2<br />
p = <br />
2 3 :<br />
Observing r 0 , the price of a treasury future that delivers a T -year zero coupon bond in years can be<br />
calculated according to Chen (1992),<br />
F (; T; r 0 ) = exp [C (; T ) + D (; T ) r 0 ]<br />
where<br />
C (; T ) = A (T ) + 1<br />
4k B (T ) e 2k e k 1 B (T ) 2 + e k B (T ) 2 + 4k <br />
D (; T ) = e k B (T ) :<br />
36
Table 1. Simulation<br />
This table reports the Treasury option pricing simulation result comparing four estimation methods: the<br />
parametric estimator using the correct model (Cox, Ingersoll, and Ross (1985) process), the parametric estimator<br />
using a misspeci…ed model (Vasicek (1977) model), the robust parametric estimator proposed in this<br />
paper which uses the misspeci…ed Vasicek (1977) model but explicitly adjusts for misspeci…cation, and the<br />
nonparametric estimator. The simulation is iterated 100 times and each simulation sample path corresponds<br />
to …ve years of weekly observations.<br />
r<br />
Panel A shows the average root integrated mean squared error (RIMSE)<br />
P 2<br />
1 n<br />
de…ned as RIMSE <br />
bCi<br />
n i=1<br />
C i where C b and C are, respectively, the estimated and the true<br />
Treasury option prices in the simulation. Panel B shows the average regions of …t (h in (2.12)) in the proposed<br />
method over which parametric …t is used. Panel C shows the estimation RIMSE for parametric estimation<br />
using the correct CIR model where the closed-form option price C is perturbed to C (1 + ") where " is<br />
uniformly distributed over [ !; !]. This models potential numerical error if a numerical method instead of<br />
the closed-form formula is used to compute the option prices in the estimation. Panel D shows the RIMSE<br />
for a simulation using parametric estimation based on the correct CIR model and numerical integration to<br />
obtain option prices.<br />
A. Performance of the option price estimators<br />
RIMSE<br />
Parametric $0:00022<br />
Parametric (using misspeci…ed) $0:041<br />
Nonparametric $0:013<br />
Proposed (using misspeci…ed) $0:0015<br />
B. Robust parametric estimator: region of …t (h) along various dimensions<br />
C. Simulate numerical error<br />
h<br />
Interest rate 0:026<br />
Option maturity 0:01<br />
Bond maturity 0<br />
! RIMSE<br />
0:01% $0:00023<br />
0:1% $0:00061<br />
0:2% $0:0012<br />
0:3% $0:0017<br />
0:5% $0:0028<br />
1% $0:0056<br />
D. Performance of parametric estimation using correct model and numerical integration<br />
RIMSE<br />
Parametric (Numerical) $0:0063<br />
37
Table 2. CBOT Treasury option pricing<br />
This table reports the Treasury option pricing result using data of Treasury options traded on CBOT in<br />
the sample period from May 1990 to December 2006. Three pricing methods are compared: the parametric<br />
estimator, the robust parametric estimator, and the nonparametric estimator. Both the parametric and the<br />
robust parametric estimators use the possibly misspeci…ed option pricing formula (4.1) which assumes that the<br />
short rate follows the Vasicek (1977)<br />
r<br />
process. Panel A shows the average root integrated mean squared error<br />
P 2<br />
1 n<br />
(RIMSE) de…ned as RIMSE <br />
bCi<br />
n i=1<br />
C i where C b and C are, respectively, the estimated and<br />
the observed Treasury option prices. Also shown in panel A is the R-square in the regression of observed call<br />
option price on predicted option price. The estimation in panel A uses observations in the entire sample period.<br />
Panel B shows the out-of-sample RIMSE and R-square comparisons of the three estimation methods. The<br />
out-of-sample estimation uses …ve years’observations to obtain parameter estimates and then measures the<br />
RIMSE and R-square in the subsequent year using the estimated parameters. The …rst year of out-of-sample<br />
comparison is 1996.<br />
A. In-sample pricing performance<br />
B. Out-of-sample pricing performance<br />
Parametric Nonparametric Proposed<br />
RIMSE 0.476 0.383 0.212<br />
R 2 0.498 0.744 0.902<br />
RIMSE R 2<br />
Parametric Nonparametric Proposed Parametric Nonparametric Proposed<br />
1996 0.468 0.379 0.164 0.523 0.788 0.941<br />
1997 0.485 0.375 0.221 0.476 0.727 0.947<br />
1998 0.543 0.495 0.324 0.487 0.623 0.834<br />
1999 0.430 0.347 0.149 0.578 0.794 0.955<br />
2000 0.395 0.292 0.148 0.482 0.798 0.930<br />
2001 0.418 0.325 0.199 0.559 0.804 0.933<br />
2002 0.540 0.444 0.247 0.582 0.793 0.912<br />
2003 0.632 0.481 0.296 0.485 0.761 0.922<br />
2004 0.549 0.385 0.322 0.596 0.817 0.932<br />
2005 0.538 0.432 0.414 0.532 0.804 0.971<br />
2006 0.401 0.405 0.399 0.527 0.654 0.907<br />
Average 0.491 0.396 0.262 0.530 0.760 0.926<br />
38
ond option price<br />
Figure 1. Compare option prices along the dimension of option maturity. This …gure<br />
compares the option prices of CIR model (true model in simulation) and Vasicek model along the option<br />
maturity dimension. Prices from Vasicek model are shown in neighborhoods around option maturity of 1<br />
month, 3 months, 6 months, and 1 year. The parameter for CIR process is set to that in (3.4). The parameters<br />
for Vasicek process are set to those estimated in section 3, which di¤er across the four Vasicek price curves<br />
shown. The underlying bond maturity is set to 10 years and the short rate is set to 7% (approximately the<br />
mean interest rate) in the simulation.<br />
1.8<br />
Vasicek vs CIR bond option prices<br />
1.6<br />
1.4<br />
1.2<br />
τ = 0.5<br />
↓<br />
↑<br />
τ = 1<br />
1<br />
0.8<br />
τ = 0.25<br />
↓<br />
0.6<br />
τ = 1 month<br />
0.4<br />
↑<br />
τ = 0.083<br />
τ = 3 months<br />
τ = 6 months<br />
0.2<br />
τ = 1 year<br />
CIR (true)<br />
0<br />
0 0.2 0.4 0.6 0.8 1 1.2<br />
τ: option maturity<br />
39
ond option price<br />
Figure 2. Compare option prices along the dimension of bond maturity. This …gure<br />
compares the option prices of CIR model (true model in simulation) and Vasicek model along the bond<br />
maturity dimension. Prices from Vasicek model are shown in neighborhoods around bond maturity of 2, 5,<br />
10, and 30 years. The parameter for CIR process is set to that in (3.4). The parameters for Vasicek process<br />
are set to those estimated in section 3, which di¤er across the four Vasicek price curves shown. The option<br />
maturity is set to 3 months and the short rate is set to 7% (approximately the mean interest rate) in the<br />
simulation.<br />
1<br />
0.9<br />
0.8<br />
Vasicek vs CIR bond option prices<br />
← T = 5<br />
← T = 10<br />
T = 2<br />
T = 5<br />
T = 10<br />
T = 30<br />
CIR (true)<br />
0.7<br />
0.6<br />
0.5<br />
← T = 2<br />
0.4<br />
0.3<br />
T = 30 →<br />
0.2<br />
0 5 10 15 20 25 30<br />
T: bond maturity<br />
40
ond option price<br />
Figure 3. Compare option prices along the dimension of short rate. This …gure compares<br />
the option prices of CIR model (true model in simulation) and Vasicek model along the short rate dimension.<br />
Prices from Vasicek model are shown in neighborhoods around short rate of 0.04, 0.07, and 0.1, which are<br />
approximately the mean and mean plus/minus one standard deviation of the short rate. The parameter for<br />
CIR process is set to that in (3.4). The parameters for Vasicek process are set to those estimated in section<br />
3, which di¤er across the three Vasicek price curves shown. The option maturity is set to 3 months and the<br />
bond maturity is set to 10 years.<br />
1.1<br />
1<br />
0.9<br />
0.8<br />
Vasicek vs CIR bond option prices<br />
← r = .07<br />
← r = .10<br />
0.7<br />
0.6<br />
← r = .04<br />
0.5<br />
0.4<br />
r = .04<br />
r = .07<br />
r = .10<br />
CIR (true)<br />
0.02 0.04 0.06 0.08 0.1 0.12<br />
r: short rate<br />
41
obust prices<br />
nonparametric prices<br />
parametric prices<br />
Figure 4. Scatter plots of observed and estimated Treasury option prices. This …gure<br />
shows the scatter plots of observed Treasury option prices against option prices estimated, respectively, using<br />
parametric methods, nonparametrics, and the robust parametric pricing method which addresses possible<br />
misspeci…cation (labeled “robust prices”in the plot). The estimation covers the sample period from May 1990<br />
to December 2006.<br />
5<br />
observed vs parametric bond option prices<br />
4<br />
3<br />
2<br />
1<br />
0<br />
0 1 2 3 4 5<br />
observed prices<br />
observed vs nonparametric bond option prices<br />
4.5<br />
4<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
0<br />
0 1 2 3 4 5<br />
observed prices<br />
4<br />
observed vs robust bond option prices<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
0<br />
0 1 2 3 4 5<br />
observed prices<br />
42
RIMSE<br />
RIMSE<br />
RIMSE<br />
Figure 5. RIMSE for various regions of …t. This …gure shows root integrated mean squared error<br />
(RIMSE) in CBOT Treasury option pricing for various regions of …t along the dimensions of option maturity,<br />
bond maturity, and short rate. The robust parametric pricing method selects a region of …t separately for each<br />
dimension to minimize the RIMSE. In the plot for bond maturity, the horizontal axis refers to the number of<br />
nearest bond maturities. For example, 1 means using the nearest 1 bond maturity –bond maturities of 5, 10,<br />
and 30 years are included in 10-year Treasury option pricing. The sample period is May 1990 to Dec 2006.<br />
0.26<br />
RIMSE and region of fit for option maturity<br />
0.25<br />
0.24<br />
0.23<br />
0.22<br />
0.21<br />
0.2<br />
0.19<br />
0 5 10 15 20<br />
region of fit for option maturity (weeks)<br />
0.4<br />
RIMSE and region of fit for bond maturity<br />
0.35<br />
0.3<br />
0.25<br />
0.2<br />
0 1 2<br />
region of fit for bond maturity<br />
0.23<br />
RIMSE and region of fit for short rate<br />
0.225<br />
0.22<br />
0.215<br />
0.21<br />
0.205<br />
0.2<br />
0.195<br />
1 2 3 4 5 6 7 8<br />
region of fit for short rate (%)<br />
43