## Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

**EMM, GMM, QMLE, and MLE for a
**

Square-Root Interest Rate Diusion Model

1

Hao Zhou

Mail Stop 91

Federal Reserve Board

Washington, DC 20551

**First Draft: April 1997
**

Last Revised: August 2000

1 I would like to thank George Tau
hen for his valuable advi
e and Ronald Gallant for his

insightful
omments. The views expressed in this paper re
e
t those of the author and do not

represent those of the Board of Governors of the Federal Reserve System or other members of its

sta. I am grateful to an anonymous referee for his or her
onstru
tive suggestions for making the

study
omplete. This paper has beneted from dis
ussions with Matthew Pritsker, Christopher

Downing, Mark Coppejans, Chien-Te Hsu, and the parti
ipants of the Duke nan
ial e
onomi
s

seminar. Please
onta
t Hao Zhou with questions and
omments: Trading Risk Analysis Se
tion,

Division of Resear
h and Statisti
s, Federal Reserve Board, Washington DC 20551 USA; Phone

1-202-452-3360; Fax 1-202-452-3819; e-mail hao.zhoufrb.gov. An earlier draft of the paper was

distributed under the title \Finite Sample Properties of EÆ
ient Method of Moments and Maximum

Likelihood Estimations for a Square-Root Diusion Pro
ess".

Abstra
t

This paper performs a Monte Carlo study on EÆ
ient Method of Moments (EMM), Generalized Method of Moments (GMM), Quasi-Maximum Likelihood Estimation (QMLE), and

Maximum Likelihood Estimation (MLE) for a
ontinuous-time square-root model under two

hallenging s
enarios|high persisten
e in mean and strong
onditional volatility|that are

ommonly found in estimating the interest rate pro
ess. MLE turns out to be the most eÆ
ient of the four methods, but its nite sample inferen
e and
onvergen
e rate suer severely

from approximating the likelihood fun
tion, espe
ially in the s
enario of highly persistent

mean. QMLE
omes se
ond in terms of estimation eÆ
ien
y, but it is the most reliable in

generating inferen
es. GMM with lag-augmented moments has overall the lowest estimation

eÆ
ien
y, possibly due to the ad ho
hoi
e of moment
onditions. EMM shows an a
elerated
onvergen
e rate in the high volatility s
enario, while its overreje
tion bias in the

mean persisten
e s
enario is una
eptably large. Finally, under a stylized alternative model

of the US interest rates, the overidenti
ation test of EMM obtains the ultimate power for

dete
ting misspe
i
ation, while the GMM J-test is in
reasingly biased downward in nite

samples.

**Keywords: Monte Carlo Study, EÆ
ient Method of Moments, Maximum Likelihood
**

Estimation, Square-Root Diusion, Quasi-Maximum Likelihood, Generalized Method of Moments.

JEL
lassi
ation: C15; C22; C52

1 Introdu
tion

When estimating a
ontinuous time model in nan
e, one often fa
es the diÆ
ulty of partial

observability. Usually the
ontinuous time re
ord is not available, sin
e the data is dis
retely sampled. A further
ompli
ation is that the transitional density of the sto
hasti

pro
ess does not always have a
losed-form solution. Due to the la
k of a tra
table likelihood fun
tion, mu
h of the interest among resear
hers has turned to nonlikelihood-based

approa
hes. For instan
e, by exploiting the analyti
al solutions of the rst two moments, the

Quasi-Maximum Likelihood Estimation (QMLE), as dis
ussed in Bollerslev and Wooldridge

(1992)
onveniently
ir
umvents the need to evaluate density fun
tion, although the asymptoti
validity of QMLE imposes
ertain restri
tions on the innovation density (Newey and

Steigerwald 1997). Moreover, the Generalized Method of Moments (GMM) by Hansen (1982)

and Hansen and S
heinkman (1995) further redu
es the relian
e on distribution assumptions

by mat
hing the empiri
al moments with the theoreti
al ones. Meanwhile, the Simulated

Method of Moments (SMM) in time-series appli
ation (Ingram and Lee 1991, DuÆe and

Singleton 1993) minimizes the relian
e on distribution assumptions by mat
hing the empiri
al moments with the simulated ones. Both GMM and SMM are robust to the misspe
i
ation of likelihood fun
tions, while retaining a parametri
model to
ondu
t simulation or

proje
tion. However, these methods of moments suer from the ad ho
hoi
e of moment

onditions and must presume the existen
e of arbitrary population moments; and the
hisquare spe
i
ation test of the overidentifying restri
tions is subje
t to severe overreje
tion

bias (Hansen, Heaton, and Yaron 1996, Andersen and Srenson 1996). Furthermore, the

eÆ
ien
y loss of parameter estimates is
losely related to the high
ost in estimating the

weighting matrix, as the varian
e-
ovarian
e matrix is typi
ally heteroskedasti
and serially

orrelated (Andersen and Srenson 1996). The Wald test is also found to ex
eed its asymptoti
size due to the diÆ
ulty in estimating the residue spe
tral-density matrix (Burnside

and Ei
henbaum 1996).

The EÆ
ient Method of Moments (EMM), introdu
ed by Bansal, Gallant, Hussey, and

Tau
hen (1995) and Gallant and Tau
hen (1996
), endogenously sele
ts the moment
onditions during the pro
edure's rst step. A seminonparametri
s
ore generator (SNP) uses

the Fourier-Hermite polynomial to approximate the underlying transitional density. As an

1

**orthogonal series estimator, the SNP density has a fast uniform
onvergen
e, given the
**

smoothness of the underlying distribution fun
tion. A suitable model sele
tion
riterion,

e.g., the S
hwarz's Bayesian Information Criterion (BIC), is used to
hoose the dire
tion

and
omplexity of the auxiliary model expansion. The se
ond stage of EMM is simply an

SMM-type estimator, minimizing the quasi-maximum likelihood s
ore fun
tions that are
hosen appropriately in the rst stage. Sin
e the s
ore fun
tions are orthogonal, the weighting

matrix (i.e., the information matrix from the quasi-maximum likelihood estimator) should

be nearly serially un
orrelated. Hen
e the asymptoti
varian
e estimator approa
hes the

minimum bound, and the parameter estimates are asymptoti
ally as eÆ
ient as MLE. It

is proven that for an ergodi
sto
hasti
system with partially observed data, the eÆ
ien
y

of EMM approa
hes that of MLE, as the number of moment
onditions and the number of

lags entering ea
h moment in
rease with the sample size (Gallant and Long 1997). Another

salient feature of EMM is the
apability to dete
t a misspe
ied stru
tural model, if the

auxiliary model is ri
h enough su
h that Hermite polynomial s
ores approximate the true

s
ores fairly well (Tau
hen 1997). Under
orre
t spe
i
ation of the maintained model, the

normalized obje
tive fun
tion value
onverges in distribution to a
hi-square distribution.

Under misspe
i
ation, the unnormalized obje
tive fun
tion
onverges almost surely to a

onstant. For the parti
ular
hoi
e of a s
ore generator, this
onstant may be zero and the

hi-square test has little power against the alternatives. If the data generating pro
ess is adequately
aptured by a more
exible nonparametri
s
ore generator, the
onstant is positive

and reje
tion of misspe
i
ation is almost
ertain.

Re
ent Monte Carlo studies do
umented signi
ant eÆ
ien
y gains of EMM over GMM

(Andersen, Chung, and Srenson 1999), but with similar overreje
tion problems in spe
i
ation tests (Chuma
ero 1997). In more analyti
al fashion, Gallant and Tau
hen (1998a) show

that EMM outperforms the
onventional method of moments (CMM) for a representative

lass of e
onometri
models, given the same number of moments being sele
ted. However,

there is no universal theory regarding the eÆ
ien
y of EMM versus that of CMM, and the

omparison must be made
ase-by-
ase (Gallant and Tau
hen 1998a). Therefore
hoosing

EMM over QMLE, as in Dai and Singleton (2000), should be a
ompanied by solid argument

or Monte Carlo eviden
e. This paper
omplements these
omparative studies in several areas. First, the Monte Carlo setup is a
ontinuous time model, like many re
ent appli
ations

2

**of EMM, whi
h have fo
used on the sto
hasti
dierential equations. Se
ond, I
onsider the
**

relative eÆ
ien
y of EMM with respe
t to asymptoti
ally eÆ
ient MLE,
omputationally ef
ient QMLE, and empiri
ally attra
tive GMM. Third, both blindfold and edu
ated
hoi
e

of moment
onditions are examined, so as to best imitate the realisti
approa
hes taken

by resear
hers using EMM. Thus, the main
ontribution of this paper is to provide Monte

Carlo eviden
e whi
h shows that when the analyti
al density or moments are unavailable,

EMM performs reasonably well
ompared to the infeasible MLE, QMLE, or GMM. Further,

under two
hallenging s
enarios|mean persisten
e and volatility
luster, ea
h method has

its own strength and weakness, in terms of estimation eÆ
ien
y, parameter inferen
e, and

spe
i
ation test.

A square-root diusion pro
ess (Cox, Ingersoll, and Ross 1985) is
hosen as the vehi
le

for
ondu
ting the Monte Carlo study. On the one hand, the CIR model is simple enough

to give
losed-form solutions for both the transitional density and the asset pri
ing formula.

On the other hand, it is ri
h enough to generate a highly persistent volatility and nonGaussian error distribution. The square-root pro
ess seems to be a good starting point

to model more
ompli
ated nan
ial time series data. EMM estimation of the interest

rate diusions is reported by Gallant and Tau
hen (1998b), and the square-root model is

rmly reje
ted. With a
losed-form transitional density, the dynami
maximum likelihood

estimation was implemented for the two-fa
tor CIR model (Pearson and Sun 1994, DuÆe and

Singleton 1997). Gibbons and Ramaswamy (1993) employed a GMM estimator, using the

sto
hasti
Euler equations to generate the moment
onditions. Their results favor the squareroot model. The most re
ent interest in aÆne term stru
ture (Dai and Singleton 2000)
an

be viewed as an immediate extension of the multifa
tor square-root model. The CIR model

also has an expli
it marginal density in terms of the drift and volatility fun
tions, whi
h

motivated a nonparametri
spe
i
ation test (At-Sahalia 1996b). Conley, Hansen, Luttmer,

and S
heinkman (1997) implemented a GMM estimator for a subset of the parameters by

exploiting the reversibility of a stationary Markov
hain. Fisher and Gilles (1996) proposed

a general QMLE estimator for the CIR type diusion pro
esses.

The remaining se
tions are organized as following: Se
tion 2 dis
usses some properties of

the square-root model and
hara
terizes the implementations of MLE, QMLE, and GMM;

Se
tion 3 introdu
es the relatively new EÆ
ient Method of Moments estimator; Se
tion 4

3

**designs the Monte Carlo experiment and suggests the ben
hmark model
hoi
e; Se
tion 5
**

reports the major ndings; and Se
tion 6
on
ludes.

**2 Square-Root Model and Parametri
Estimator
**

This se
tion denes a maximum likelihood estimator for the square-root model, based on

a Poisson-mixing-Gamma
hara
terization of the likelihood fun
tion. A quasi-maximum

likelihood estimator is also available with analyti
al solutions to the rst two
onditional

moments. Augmenting these two moments with instrumental variables gives a generalized

method of moments estimator found in the literature.

**2.1 Probabilisti
Solution to Square-Root Model
**

It is a well-known result that the square-root model,

drt = (a0 + a1 rt )dt + b0 rt1=2 dWt ;

(1)

**satises the regularity
onditions for both a strong solution (pathwise
onvergent) and a weak
**

solution (
onvergent in probability) (Karatzas and Shreve 1997). A strong solution obviously

implies a weak one, but not vi
e versa. If (1) a0 > 0, (2) b0 > 0, (3) a1 < 0, and (4) b20 2a0 ,

then the square-root model has a unique fundamental solution (Feller 1951). The marginal

density is a Gamma distribution, and the transitional density is a type I Bessel fun
tion

distribution or a non
entral
hi-square distribution with a fra
tional order (Cox et al. 1985).

Intuitively,
ondition (3) gives mean reversion, and
ondition (4) ensures the stationarity.

The marginal Gamma distribution is

! 1 !r0

f (r0 j; ! ) =

r e ;

(2)

( ) 0

where = 2a0 =b20 , ! = 2a1 =b20 , and () is the Gamma fun
tion.1 The un
onditional mean

and varian
e are

a0

E (r0 ) =

= ;

(3)

a1

!

b20 a0

= :

(4)

V (r0 ) =

2a21 ! 2

1 In

**some textbooks and software, the se
ond parameter of the Gamma density is written in terms of 1=!
**

instead of ! (Johnson and Kotz 1970).

4

**Noti
e that the rst two moments merely identify the marginal distribution. Higher order
**

moments are simply nonlinear fun
tions of the rst two moments. The marginal density alone

an not identify all three parameters in the diusion pro
ess. Any GMM-type estimator must

add at least one lagged instrumental variable (Gibbons and Ramaswamy 1993). A reje
tion of

the marginal distribution
an reje
t the square-root model; however, a non-reje
tion does not

provide enough information for judging a parti
ular parameter setting (At-Sahalia 1996b).

Transitional information must be exploited to fully identify the dynami
stru
ture.

2 The
onditional density is

1

v q

f (r1 jr0 ; a0 ; a1 ; b0 ) =
e u v ( ) 2 Iq (2(uv ) 2 );

u

2a0

q = 2 1;

b0

(5)

(6)

**where
= 2a1 =(b20 (1 ea1 )), u =
r0 ea1 , and v =
r1 . Iq () is a modied Bessel fun
tion
**

of the rst kind with a fra
tional order q (Oliver 1972). The
onditional mean and varian
e

are

E (r1 jr0 ) = r0 ea1

a0

(1 ea1 );

a1

(7)

b2

b2 a

V (r1 jr0 ) = r0 ( 0 )ea1 (1 ea1 ) + 0 20 (1 ea1 )2 :

(8)

a1

2a1

A
ording to the stationary property, the limit of the transitional density as the time interval goes to innity, is exa
tly the marginal density. Therefore any estimation strategy or

spe
i
ation test that exploits the transitional density will naturally nest the ones that rely

on the marginal density.

It is a
ommon pra
ti
e in the literature to
all this distribution a \non
entral
hisquare distribution." However, the \integer order non
entral
hi-square distribution" does

not naturally extend to the \fra
tional order non
entral
hi-square distribution." The latter

arises
ommonly from the solution to a diusion pro
ess (Feller 1971), while the former

arises from the sample standard deviation of independent, nonidenti
al, non
entered, normal

random variables (Johnson and Kotz 1970).

2 The

one.

pro ess is sampled with weekly observations and the time interval of a week is normalized to be

5

**2.2 Analyti
al Foundations for MLE, QMLE, and GMM
**

In industry and a
ademi
s alike, one popular method in estimating the square-root model for

interest rates is the Dis
retized Maximum Likelihood Estimation (DMLE), i.e., a misspe
ied

QMLE based on the time dis
retization of the
onditional mean E (r1 jr0 ) a0 +(1+ a1 )r0 and

varian
e V (r1 jr0 ) b20 r0 . As pointed out by Lo (1988), DMLE is generally not
onsistent.

The parameter estimates are asymptoti
ally biased, sin
e both moments are misspe
ied.

When implementing MLE for the square-root model, the Bessel fun
tion representation

of the likelihood fun
tion is not at all a
onvenient form (Pearson and Sun 1994, DuÆe and

Singleton 1997). An alternative Poisson-mixing-Gamma
hara
terization
an be inferred

from the simulation strategy suggested by Devroye (1986). Within the admissible parameter

region, one
an substitute the Bessel fun
tion with an innite series expansion (Oliver 1972).

p

With appropriate transformations (y = v , = q + 1, and = 2u), the alternative mixing

formula
omes out ni
ely,

f (y ) =

=

1 y j + 1e

X

j =0

1

X

y

(j + )

2 j

2 e

j!

2

2

(9)

2

Gamma(y jj + ; 1) Poisson(j j ):

2

j =0

**One needs to be
autious that the Poisson weights are not
onstant, but rather
ondition on
**

the previous realization r0 . This formula
orresponds to the \Poisson driven Gamma pro
ess"

in Feller (1971). The only dieren
e is that = q + 1 remains a fra
tional number, not an

integer. The evaluation of the log-likelihood fun
tion in MLE is greatly simplied when using

the Poisson-mixing-Gamma formula. It is fairly easy to a
hieve the single pre
ision 10 8 by

trun
ating the Poisson distribution around 100. One
an also avoid any
ompli
ation of

omplex value or non-
onvergen
e in evaluating the Bessel fun
tion.3

The exa
t expressions for the
onditional mean and varian
e (equations 7 and 8) suggest

3 This

mixing approa
h to likelihood fun
tion works well for the high volatility
ase, but the approa
h

fails for the mean persisten
e
ase, be
ause the latter is
lose to the unit root or nonstationary region (see

Se
tion 4.1 for details on
hoosing the ben
hmark s
enario) and be
ause the evaluation of the likelihood

fun
tion easily diverges. To implement MLE in the mean persisten
e
ase, I adopt the asymptoti
expansion

formula (Oliver 1972) in the same fashion as Pearson and Sun (1994).

6

**a quasi-maximum likelihood estimator (QMLE) for the square-root model,
**

TX1

1

exp

log q

max

fa0 ;a1 ;b0 g t=1

2V (rt+1 jrt )

(

)

[rt+1 E (rt+1 jrt )℄2

:

2V (rt+1 jrt )

(10)

**QMLE is shown to be root-n
onsistent (Bollerslev and Wooldridge 1992) and asymptoti
ally
**

normal (Newey and Steigerwald 1997) under some mild regularity
onditions.

For
omparison purposes, a generalized method of moments (GMM) estimator
an be

onstru
ted from the following moment ve
tor

2

6

ft (a0 ; a1 ; b0 ) = 664

rt+1 E (rt+1 jrt )

rt [rt+1 E (rt+1 jrt )℄

V (rt+1 jrt ) [rt+1 E (rt+1 jrt )℄2

rt fV (rt+1 jrt ) [rt+1 E (rt+1 jrt )℄2 g

3

7

7

7:

5

(11)

P

**The parameter is estimated by min gT0 W gT , where gT = 1=T Tt=11 ft and W is the asymptoti
**

varian
e-
ovarian
e matrix of gT (Hansen 1982). A two step estimator is adopted here.

2.3 Inferen
e

Sin
e we know the true parameter value for MLE, the likelihood ratio tests
an determine how

often the
onden
e region
entered at the estimated parameter value
ontains the truth.

The same prin
iple may be extended to QMLE if the expe
ted quasi-likelihood fun
tion

is uniquely maximized at the truth (satisfying identi
ation requirement). In addition to

the
orre
t spe
i
ation of the mean and varian
e, the remaining innovation error must be

entered at zero or have a symmetri
distribution (Newey and Steigerwald 1997). QMLE

for the square-root model meets these
onditions. Also, the sample-size normalized
riterion

fun
tion value in GMM provides a spe
i
ation test distributed as a Chi-square (1).

The standard error for an individual parameter is estimated by the inverted Hessian

formula in MLE, the OPG-Hessian-OPG formula in QMLE, and the Gradient-Weighting

Matrix-Gradient formula in GMM.

2.4 Simulation

This
hara
terization (equation 9) also denes a
omposite method to simulate the squareroot pro
ess (Devroye 1986). First, one draws a random number j from the Poisson(j j 2=2).

7

**Then, one draws another random number y from the Gamma(y jj + ; 1). Finally, one
**

al
ulates the desired state variable r1 by r1 = y=
. Noti
e that the realized r1 is the

onditioning value r0 in the next draw. The initial value r0 , when starting a simulation run,

an be set to the theoreti
al un
onditional mean. To pass on the transient ee
t, the rst

1000 realizations
an be dis
arded.

**3 EÆ
ient Method of Moments Estimation
**

This se
tion des
ribes the EMM estimator in a univariate
ase. For formal dis
ussions, see

Gallant and Tau
hen (1996
), Tau
hen (1997), and Gallant and Long (1997).

**3.1 Approximating True Density with Auxiliary Model
**

Denote the invariant probability measure implied by the underlying data generating pro
ess

as the p-model. It is assumed that the dire
t maximum likelihood estimation of the p-model

is not available. However, any smooth density fun
tion
an be approximated arbitrarily
lose

by a Hermite polynomial expansion.

Consider a s
alar
ase. Let y be the random variable, x be the lagged y , and be the

parameter. The auxiliary f -model has a density fun
tion dened by a modied Hermite

polynomial,

f (y jx; ) = C f[P (z; x)℄2 (y jx; x2 )g;

(12)

where P is a polynomial with degree Kz in z and thus the square of P makes the density

positive. The argument of the polynomial is z , whi
h is the transformation z = (y x )=x .

The
oeÆ
ient of the polynomial is another polynomial of degree Kx in x. The
onstant in

the polynomial is set to 1 for identi
ation. C is a normalizing fa
tor to make the density

proper.4 () is a normal density of y with
onditional mean x and
onditional varian
e x2 .

The length of the auxiliary model parameter is determined by the lag in mean L, lag in

varian
e Lr , lag in polynomial
oeÆ
ient Lp , polynomial degree Kz , and polynomial degree

Kx . Let fy~t gnt=1 be the observed data and x~t 1 be the lagged observations. The sample mean

4 In

the
ase of multivariate density, the intera
tion terms in both the Hermite polynomial and the

oeÆ
ient polynomial
an be set to zero, in order for a gradual expansion of the auxiliary model. See above

referen
es for details.

8

log-likelihood fun tion is dened by

Ln(; f g

y~t nt=1 )

n

1X

log[f (~yt jx~t 1 ; )℄:

=

n t=1

(13)

A quasi-maximum-likelihood estimator is obtained by

~ = arg max

Ln(; fy~tgnt=1 ):

(14)

**The dimension of the auxiliary f -model, the length of , is sele
ted by S
hwarz's Bayesian
**

Information Criterion (BIC). There are dierent
hoi
es of information
riteria for optimal

model sele
tion. For a nite dimensional stationary pro
ess, BIC with a larger penalty for

model
omplexity proves to be
onsistent, while the Akaike's Information Criterion (AIC)

will overt the model. On the other hand, if the true dimension is innity or in
reases to

innity with the sample size, AIC with a smaller penalty for model
omplexity is optimal

(Zheng and Loh 1995). The dimensions of the f-model needs to be as large as the p-model

to meet the identi
ation
ondition.5

**3.2 Mat
hing Auxiliary S
ores with Minimum Chi-Square
**

From the rst-stage seminonparametri
estimates, one obtains the tted s
ores as the moment
onditions,

n

1X

mn (~) =

log f (~yt jx~t 1 ; ~):

(15)

n t=1

In the se
ond stage, a SMM-type estimator is implemented in the following way. Although

the dire
t MLE for p-model is assumed to be impossible, the simulation from the stru
tural

model (e.g., the sto
hasti
dierential equation) is readily available. Let fy^t gNt=1 be a long

simulation from a
andidate value of , the parameter of the maintained stru
tural model.

The auxiliary s
ore fun
tions
an be reevaluated by numeri
al integration of of the s
ore

fun
tions with the simulated data,

N

1X

log f (^yt jx^t 1 ; ~);

m

^ N (; ~) =

N t=1

5 Even

(16)

**when the underlying p-model has a xed dimension (i.e., a xed number of parameters), the
**

auxiliary f -model
an still have an in
reasing order in nite sample sizes, be
ause of the approximation from

auxiliary model to true model.

9

and the minimum hi-square estimator is simply,

^ = arg min

fm^ N (; ~)0I~ 1 m^ N (; ~)g;

(17)

**where the weighting matrix I~ 1 is estimated by the mean-outer-produ
t of s
ores from the
**

auxiliary model

n

~I = 1 X[ log f (~yt jx~t 1 ; ~)℄[ log f (~yt jx~t 1 ; ~)℄0 :

(18)

n t=1

**Remember that the nite sample eÆ
ien
y loss of the GMM-type estimators is largely
**

attributed to the high
ost and low a
ura
y in estimating the serially
orrelated weighting

matrix. In EMM the moment
onditions (s
ore fun
tions) of the rst step QMLE are orthogonal by
onstru
tion; hen
e the information matrix is diagonal or nearly diagonal (i.e.,

serially un
orrelated). EMM will be asymptoti
ally as eÆ
ient as MLE if the following
onditions are met: the dimension of the auxiliary model is suÆ
iently large (K ! 1), the lag

in the the auxiliary model is suÆ
iently long (L ! 1), and the simulation from the maintained stru
tural model is suÆ
iently long (N ! 1). This en
ompasses both Markovian

and non-Markovian
ases (Gallant and Long 1997).

**3.3 Overreje
tion and Misspe
i
ation
**

The normalized
riterion fun
tion value in the EMM estimation,

2n = nm

^ N (^; ~)0 I~ 1 m

^ N (^; ~);

(19)

forms a spe
i
ation test for the overidentifying restri
tions. Under the
orre
t spe
i
ation

of the maintained model (Tau
hen 1997), we have
2n D! X 2 (l lp ), where the degree of

freedom equals the parameter length of the auxiliary model minus that of the stru
tural

model. However, if the maintained model is misspe
ied (Tau
hen 1997), we have

2n

n

! 2 = m^ N (; )0I 1 m^ N (; ) > 0;

a:s:

where , , and I are the asymptoti
pseudo-true values under the maintained misspe
i
ation. As long as the sample size n is large enough, the SNP s
ore generator will be ri
h

enough su
h that the false passage
2 = 0 will not o
ur under misspe
i
ation.

10

**4 Monte Carlo Design and Ben
hmark Choi
e
**

Te
hni
al aspe
ts of the Monte Carlo experiment are summarized here. Most importantly,

two demanding s
enarios|mean persisten
e and high volatility|are
hosen as the ben
hmarks. Also, a maintained model of interest rate with sto
hasti
volatility and level feedba
k

is designated as the data generating pro
ess for
omparing the power of dete
ting misspe
i
ation.

4.1 Experimental Design

All
omputations are performed on Sun Unix Workstations|Spar
10, Spar
20, and Pentium

II-300. Programs for generating random samples and GMM, QMLE, and MLE estimations

of the square-root model are written in FORTRAN language. The FORTRAN
odes for

SNP and EMM are modied from SNP Version 8.5 (Gallant and Tau
hen 1996b) and EMM

Version 1.3 (Gallant and Tau
hen 1996a), in
orporating automati
SNP sear
h by BIC in

some of the EMM estimations. NPSOL (Gill, Murray, Saunders, and Wright 1991) is the

optimization routine used for all of the programs.

The number of Monte Carlo repli
ations is
hosen as 1000 for ea
h s
enario and ea
h estimator. Two nite sample sizes|500 and 1500 weekly observations|are used for
ontrasting

the asymptoti
behavior of ea
h estimator. To generate the pseudo-random samples, the

Poisson-mixing-Gamma formula in Se
tion 2 is implemented, and the 1000 initial stret
h

is dis
arded ea
h time in order to pass on the transient ee
t. The se
ond stage of EMM

estimation is similar to SMM. The simulation size should be at least 30,000. EMM does

stabilize from 50,000 to 75,000 and
ould have more Monte Carlo errors at 100,000. Thus

50,000 turns out to be a
onservative but e
onomi
al
hoi
e (Gallant and Tau
hen 1998b).

To
hoose a SNP s
ore generator in the EMM estimation, one
an either let BIC automati
ally de
ide the SNP dimension in ea
h repli
ation, or, one
an use a \posterior" xed

SNP spe
i
ation. For the
omparisons with MLE, QMLE, and GMM estimators, I use

the xed SNP s
ore in EMM estimation, su
h that the overreje
tion test statisti
s has the

same degree of freedom for ea
h sample size. However, in order to
hoose an appropriate

xed SNP s
ore, I rst run EMM 1000 times with the automati
SNP s
ore generator. I

then adopt one parti
ular s
ore for the 500 sample size and another for the 1500 sample

11

**size. These spe
i
ations are at relatively higher dimensions, with abundant o
urren
e but
**

without severe reje
tion. A by-produ
t of the EMM estimation with automati
SNP s
ore is

that some light
an be shed on how the overreje
tion bias varies with the number of moments

hosen by the SNP estimation.6

In terms of the
omputing time, ea
h QMLE run takes about 3-5 se
onds; ea
h GMM

run takes about 5-10 se
onds; ea
h MLE run takes about 5-10 minutes; and ea
h EMM run

takes about 1-2 hours. Note that MLE requires numeri
al approximation to the likelihood

fun
tion and also that EMM uses numeri
al integration to evaluate the moment
onditions.

EMM estimator is programmed up and ready to be implemented (Gallant and Tau
hen

1996b, Gallant and Tau
hen 1996a). One only needs to in
orporate the square-root model

in the simulation
ode. The developing time for QMLE and EMM is minimal, while GMM

espe
ially MLE requires a lot of ne-tuning. Overall, EMM is still the most
omputationally

intensive method, although there are numeri
al te
hniques to speed up the estimation (e.g.,

parallel pro
essing or antitheti
simulation).

**4.2 Ben
hmark Model
**

To sele
t a suitable parameter setting, we start with the empiri
al result from Gallant and

Tau
hen (1998b), drt = (0:02491 0:00285rt)dt + 0:0275rt1=2 dWt . Using equations 3, 4, and

6-8, one
an
al
ulate the un
onditional mean and varian
e, the Bessel fun
tion order, and

the
onditional mean and varian
e. It is not diÆ
ult to see that this original spe
i
ation,

S
enario 1 in Table 1, features low mean-reversion (E (rt+1 jrt ) is nearly the unit-root) and

low
onditional volatility (V (rt+1 jrt ) is
lose to zero). Also the un
onditional varian
e is

unusually small, representing an abnormally quiet pro
ess. The order of the Bessel fun
tion,

twi
e of whi
h
orresponds to the degree of freedom for an integer order non
entral
hi-square

distribution, is so large that the
onditional density looks almost Gaussian. Not withstanding

all these short
omings, s
enario 1 is still the typi
al empiri
al result. It poses an important

hallenge to resear
hers in tting the highly persistent, nearly unit-root
onditional mean

pro
ess, although its volatility stru
ture and innovation density are not ri
h enough. Hen
e

6 This

automati
-plus-xed SNP-EMM pro
edure mimi
s the realisti
situation, when an empiri
al resear
her not only relies on BIC as an obje
tive
riterion in model sele
tion but also in
orporates prior

subje
tive information to expand the auxiliary model|e.g, the E-GARCH s
ore generator in Andersen et al.

(1999).

12

it will only rarely generate the high order ARCH or the high degree polynomial in the SNP

s
ore. Based on the empiri
al results in Gallant and Tau
hen (1998b), I use the established

SNP s
ore for S
enario 1|s14140 for the 500 sample size and s14141 for the 1500 sample

size7 |and I report only the xed s
ore EMM estimations. The S
enario 1 in Table 1 is

termed an LMR-LCV spe
i
ation (low-mean-reversion and low-
onditional-volatility).

If one in
reases only the varian
e parameter b0 from S
enario 1 to 2-4 in Table 1, the Bessel

fun
tion order q de
reases gradually from the Gaussian-like spe
i
ation, but the
onditional

volatility is still negligible. Alternatively, one
an in
rease both the mean parameters a0 and

a1 by a fa
tor of 100 and the varian
e parameter b0 by a fa
tor of 10 from S
enario 1 to

5 as in Table 1, while holding the un
onditional mean and varian
e
onstant. This
hange

will in
rease the
onditional volatility slightly, but the Bessel fun
tion order q is still quite

large (resembling a Gaussian-like distribution). If one in
reases the varian
e parameter b0

from S
enario 5 to 6-8 as shown in Table 1, both high
onditional volatility and small Bessel

fun
tion order are a
hieved. S
enario 8 is ri
h enough in both the
onditional volatility

and non-Gaussian innovation, hen
e S
enario 8 is suitable for examining the automati
SNP

s
ore EMM estimations. Strong volatility
luster is another
hallenge in tting the short

interest rate, although the high persisten
e in mean is sa
ri
ed somewhat. S
enario 8 in

Table 1, drt = (2:491 0:285rt )dt + 1:1rt1=2 dWt , is termed an HMR-HCV spe
i
ation

(high-mean-reversion and high-
onditional-volatility).

One should not arbitrarily relate either S
enario 1 or S
enario 8 alone to the real interest

rate, but put together they represent the most important features in short rate pro
ess.

There is a fundamental
on
ern with how
exible the square-root model
an be. In order

to t the real interest rate data, one would like to hold the un
onditional mean ( a0 =a1 )

onstant without explosion (b0 4a20 ), resembling a stationary interest rate pro
ess. In

addition one would like to a
hieve high persisten
e in both
onditional mean and varian
e.

With only three parameters to manipulate, it seems impossible to satisfy all four
onstraints

simultaneously. This indi
ates that the square-root model may not be
exible enough to

model the interest rate dynami
s. To nd a more suitable model requires a fourth degree of

7 s14140 means 1 lag in mean, 4 lag in varian
e, 1 lag in
oeÆ
ient polynomial, 4 degree Hermite polynomial, and 0 degree
oeÆ
ient polynomial. s14141 only diers in 1 lag
oeÆ
ient polynomial. There is more

detailed dis
ussion of the SNP stru
ture in Se
tion 3.1.

13

freedom, for example, a sto hasti volatility omponent (Gallant and Tau hen 1998b).

**4.3 Testing Misspe
i
ation
**

The goal is to nd a true data generating pro
ess, of whi
h an adequate auxiliary s
ore will

not a
ommodate a misspe
ied model. As dis
ussed in Se
tion 1, the square-root model

is widely used in tting the short rate pro
ess, but most serious studies have reje
ted this

spe
i
ation. So it is natural to adopt the square-root pro
ess as a misspe
ied model and

use a non-reje
ted model as the true data generating pro
ess (for the short interest rate).

A re
ent study by Gallant and Tau
hen (1998b) gave the most favorable eviden
e for the

following spe
i
ation,

drt = (0:014 0:002rt)dt + (0:043 0:018rt )eut dW1t ;

(20)

dut = ( 0:006rt

(21)

0:157ut)dt + (0:593 0:052ut)dW2t :

The short rate pro
ess rt has a linear drift and a linear diusion, with the diusion multiplied

by an unobserved (exponential) sto
hasti
volatility term eut . The short rate rt is only

partially observable in dis
rete time. The latent sto
hasti
volatility pro
ess ut also has

a linear drift and a linear diusion, with the drift in
luding a short rate level feedba
k.

This model adequately passed the spe
i
ation test for various simulation sizes and greatly

outperformed the
ompeting models in terms of reproje
ting the
onditional density and the

onditional volatility. It would be a very suitable
hoi
e for the true data generating pro
ess.

When we t the misspe
ied square-root model drt = (a0 + a1 rt )dt + b0 rt1=2 dWt to the

interest rate data simulated from the maintained true spe
i
ation (equations 20 and 21),

the drift is
orre
tly spe
ied as a linear fun
tion, and the misspe
i
ation
omes only into

the diusion. From It^o's formula we know that the
onditional mean is also
orre
tly spe
ied as linear; therefore the drift parameters are
onsistent estimates (At-Sahalia 1996a). It

is equivalent to the
ase where Ordinary Least Square is
onsistent, but una
ounted heteroskedasti
and/or
orrelated error stru
ture may
ause very noisy and ineÆ
ient estimates.

The misspe
ied diusion generates in
onsistent estimates of the
onditional varian
e, whi
h

may
ause serious distortions in pri
ing dis
ount bond yields or other interest rate sensitive

derivatives. Therefore dete
ting the misspe
i
ation in diusion or volatility is a
riti
al

hallenge to resear
hers and pra
titioners.

14

**5 Monte Carlo Results
**

Tables 2 through 6 and Figures 1 through 9 summarize the major ndings of this paper. The

dis
ussions are organized along topi
s, and extensive
omparisons are made a
ross QMLE,

EMM, GMM, and MLE. The main fo
uses are nite sample eÆ
ien
y, overreje
tion bias of

EMM under the null, and the dete
tion of maintained misspe
i
ation in EMM and GMM.

Both the automati
s
ore generator and the xed s
ore generator are used in EMM. The

likelihood ratio tests in MLE and QMLE provide joint inferen
es under the null.

5.1 Simulation S
hemes

The Poisson-mixing-Gamma formula is a useful
hara
terization of the transitional density

of the square-root model. The simulation a
ura
y based on the distribution fun
tion
an

provide an independent
he
k for the derivations in Se
tion 2. The Monte Carlo study is

not alone in needing a reliable simulator; further appli
ations of MLE with this formula

also require some justi
ation. A
ornerstone of the EMM estimator|a simulation-based

estimator|is the dis
retized approximation to sto
hasti
dierential equations. EMM uses

a weak-order 2 s
heme (Kloeden and Platen 1992). To assess these simulation approa
hes,

some empiri
al statisti
s from long realizations (100,000) are
ompared to their theoreti
al

ounterparts.

Table 2 lists the
al
ulation of two moments and three quantiles. Clearly both s
hemes

from exa
t distribution and time dis
retization work reasonably well. In both persistent

mean and strong volatility
ases, the simulated moments and quantiles are very
lose to the

model implied ones. It is not surprising that the probabilisti
method is slightly better than

the dis
retized method, although the dieren
e is negligible.

5.2 S
ore Generator

An important feature of EMM is the endogenous moment sele
tion by a seminonparametri

s
ore generator (SNP), whi
h
ontrasts with the ad ho
hoi
e of moment
onditions in some

less sophisti
ated GMM or SMM estimators. The optimal SNP sear
h and the inexpensive

weighting matrix estimate are key to the eÆ
ien
y argument, and hopefully they also improve the overreje
tion test. It is worthwhile to
he
k whether the SNP s
ore
aptures the

15

**distribution features of dierent dependent stru
tures before laun
hing the full-s
ale Monte
**

Carlo experiment. Tables 3 (500 sample size) and Table 4 (1500 sample size) report the SNP

sear
hes for the 8 s
enarios in Table 1. For ea
h setting, the frequen
ies of all kinds of model

hoi
es among 100 repli
ations are listed. Model dimension is represented by a ve-digit

number, whi
h stands for,
onse
utively, lag in mean, lag in varian
e, lag in polynomial,

degree of Hermite polynomial, and degree of Hermite
oeÆ
ient polynomial.

S
enario 1 in Tables 3 and 4 is the LMR-LCV
ase (low-mean-reversion, low-
onditionalvolatility). Not surprisingly the Gaussian auto-regression spe
i
ation of 10100 dominates

other
hoi
es. This nding is
onsistent with the fa
t that the true density is nearly Gaussian

under this parameter setting (see Table 1). Moving from S
enario 1 to 2, 3, and 4, the

onditional volatility in
reases gradually, sin
e the varian
e parameter b0 is altered (see Table

1). The dominating
hoi
e is still Gaussian, and the
han
es of ARCH and/or non-Gaussian

spe
i
ations in
rease slightly. Moving toward S
enarios 5-8, both mean parameters a0 and

a1 as well as varian
e parameter b0 are altered (see Table 1), and ultimately one rea
hes

the HMR-HCV
ase (high-mean-reversion, high-
onditional-volatility). It is
lear that the

SNP sear
h favors the nonlinear, nonparametri
AR-ARCH spe
i
ation. This is
onsistent

with the low Bessel fun
tion order and high
onditional varian
e (Table 1). Largely due to

this \distribution-dependent" or \data-dependent" s
ore generator, the EMM estimator is

laimed to be asymptoti
ally eÆ
ient and hopefully more reliable in the spe
i
ation test.

Also evident from Tables 3 and 4 is that larger sample sizes enable the SNP to pi
k

up higher model dimensions. In fa
t, the asymptoti
eÆ
ien
y argument requires that the

number of moment
onditions and the lags entering ea
h moment in
rease with the sample

size (Gallant and Long 1997).

A salient question is whether the stru
tural model
an be identied when the SNP sear
h

does pi
k the Gaussian-AR(1) s
ore. This
orresponds to a quasi-maximum likelihood estimator based on the innovation assumption

zt = (rt

0

1 rt 1 )=2 N (0; 1):

(22)

Sin
e the
onditional mean is
orre
tly spe
ied, the QMLE of 0 and 1 is a
onsistent estimator of ea1 and a0 =a1 (1 ea1 ). It is just an Ordinary Least Square with a heteroskedasti

and serially
orrelated error term (At-Sahalia 1996a). The
onditional varian
e is mis16

**spe
ied as the
onstant 2 . However, a
ording to the theory of misspe
ied maximum
**

likelihood estimation (White 1994), the estimator ^ 2 may
onverge to a pseudo-true value

2 . The key argument is that the misspe
ied asymptoti
varian
e 2 must be a fun
tion

of the true varian
e parameter b0 , sin
e both
onditional varian
e and un
onditional varian
e are determined by b0 . These asymptoti
relations, two expli
it and one impli
it, are

indeed the binding fun
tions in the language of Indire
t Inferen
e (Gourierous, Monfort, and

Renault 1993). Obviously the stru
tural parameters a0 , a1 , and b0 are exa
tly identied by

the auxiliary parameters 0 , 1 , and 2 . EMM is thus a feasible rst-order approximation

toward the Indire
t Inferen
e (Gallant and Long 1997).

**5.3 Estimation Bias and Finite Sample EÆ
ien
y
**

Table 5 (persistent mean
ase ) and Table 6 (strong volatility
ase) report the mean bias,

medium bias, and root-mean-squared error (RMSE) a
ross MLE, QMLE, GMM, and EMM

between 500 and 1500 sample sizes.

First look at the
ase of persistent mean (Table 5). The biases are very large for the drift

estimates (a0 and a1 ) but quite small for the diusion parameter (b0 ). The biases redu
e

with the sample size, ex
ept for MLE and GMM. MLE is the most eÆ
ient in a
hieving the

smallest RMSE; however, the drift parameter estimates diverge|RMSE does not shrink with

the sample size.8 EMM seems to be more eÆ
ient than GMM but less eÆ
ient than QMLE.

The diusion parameter estimation in GMM does not
onverge, similar to the drift parameter

in MLE. The drift parameter estimates in QMLE and the inter
ept of drift in GMM seem

to
onverge faster than root-n. The
onvergen
e rate of EMM in these nite samples is very

p

lose to 3. Overall the drift estimates are more biased and noisy than the diusion estimate,

and the
lose-to-unit-root mean persisten
e
auses some unusual
onvergen
e problems in

nite samples. The rank in order of eÆ
ien
y from highest to lowest would be MLE, QMLE,

EMM, GMM, as expe
ted.

In the
ase of strong volatility (Table 6), all parameter estimates have very small biases

(relative to parameter value), and all biases shrink appropriately as the sample size in
reases.

8 When

the square-root is
lose to the non-stationary region, the mixing formula for approximating the

likelihood easily diverges in MLE estimations. One has to rely on the asymptoti
expansion whi
h introdu
es

the asymptoti
biases. It should be pointed out that the simulation s
heme based on the mixing formula is

still sound, sin
e the parameter is xed in simulations.

17

**In terms of the eÆ
ien
y, the RMSE's of MLE, QMLE, and GMM are shrinking approxip
**

mately at the rate 3, but the RMSE of EMM de
reases faster than root-n. Re
all that

EMM should be asymptoti
ally as eÆ
ient as MLE, as the auxiliary SNP s
ore generator

adopts an in
reasing dimension with the sample size. Overall MLE a
hieves the highest

eÆ
ien
y, QMLE
omes se
ond at T = 500 and third at T = 1500, EMM is third at T = 500

and se
ond at T = 1500, GMM is best for the inter
ept parameter of drift but is worst for

the slope parameter of drift and the diusion parameter. The nding that some parameter

estimates in GMM and QMLE are slightly more eÆ
ient than MLE,
an be attributed to the

fa
t that \exa
t" likelihood fun
tion in MLE needs to be numeri
ally approximated while

the moment
onditions in GMM or QMLE are in
losed forms.

5.4 Parameter Inferen
e

The nite sample distributions of the standardized t-test statisti
s for individual parameters

are summarized in Figure 1 (persistent mean) and Figure 2 (strong volatility). Standard

Gaussian kernel smoothing is adopted here. QMLE works well in both
ases; EMM is

reliable in the latter
ase, while MLE and GMM seem not to perform in either
ase.

In the
ase of mean persisten
e (Figure 1), MLE seems to have asymptoti
biases for the

drift parameter (underestimating a0 and overestimating a1 ), but the diusion parameter is

perfe
tly approximated by its asymptoti
distribution. QMLE works equally well for the

diusion parameter, while its estimates for the drift parameters, though having some nite

sample bias (a0 has upward bias and a1 has downward bias), are dissipating asymptoti
ally.

GMM is not biased in estimating the drift, but the standard error is too small (high peak in

the middle) and the estimation variation is too large (fat tails on both sides). Its inferen
e

for the diusion is s
attering everywhere. EMM suers from high variations for all the

parameters.

The
ase of strong volatility (Figure 2) looks mu
h dierent. MLE has almost negligible

biases, but it still understates the standard error and produ
es fat tails. QMLE works

perfe
tly well, espe
ially in the tails where the 1%, 5%, and 10% t-tests are usually
ondu
ted.

GMM suers heavily from both understating and overstating the standard errors. EMM

omes
lose to QMLE, and its bias and variation are shrinking with the in
reasing sample

sizes.

18

**Several fa
tors may
ontribute to the unusual t-test distributions: (1) high persisten
e
**

in the mean makes the simulated data look like a unit-root and the MLE, GMM, and EMM

estimators
an not distinguish it from a persistent yet stationary situation in nite samples;

(2) the so-
alled \exa
t" likelihood in MLE is approximated by its series expansion (in strong

volatility
ase) or its asymptoti
expansion (in persistent mean
ase); (3) the t-test is a Wald

test and la
ks invarian
e to nonlinear transformations.9

**5.5 Overreje
tion Bias and Number of Moments
**

The Monte Carlo results on EMM with automati
SNP s
ore generator
an reveal some

onne
tions between the number of overidentifying moments and the overreje
tion rate.

Figures 3 and 4 report this experiment for S
enario 8|the
ase of strong volatility|whi
h

an generate ri
her SNP s
ores. \Number of Overidentied Moments" refers to the dieren
e

between the number of moments
hosen automati
ally by BIC (whi
h varies ea
h trial) and

the number of stru
tural parameters in the square-root model (whi
h is xed at 3). The

o
urren
e
urve is the per
entage of how many times BIC
hooses a parti
ular number of

moments over the 1000 Monte Carlo repli
ations. Further, the reje
tion
urves (xed at

the 5% level) tell you the per
entage of how many times EMM reje
ts the null square-root

spe
i
ation at this parti
ular \Number of Overidentied Moments" over the number of

times that BIC pi
ks this parti
ular SNP s
ore.10

The asymptoti
size of the spe
i
ation test is xed at 5%. The o
urren
e rates show the

frequen
ies of dierent numbers of moment
onditions in 1000 repli
ations. On average, the

5% gross overreje
tion rate in automati
s
ore EMM is about 20% for T = 500 and about

25% for T = 1500. Some important features need to be mentioned. First, the reje
tion

urve does not uniformly shoot up when more moment
onditions are in
luded, sin
e these

moments are optimally sele
ted by the SNP s
ore generator. Se
ond, the reje
tion rates are

more stable at T = 1500 than T = 500, as more moments and lags are in
luded. Third,

the reje
tion rate
ould be remarkably small for
ertain low dimensions as well as for some

high dimensions. Sin
e BIC tends to undert the auxiliary model in small samples, the

9 For

**example, the drift fun
tion in the square-root model a0 + a1 rt
an be reparameterized as ( rt ),
**

then be
omes a nonlinear transformation of a0 and a1 .

10 Figures 7 and 8 are the same, ex
ept that the data is simulated from the alternative sto
hasti
volatility

pro
ess while the EMM is
arried out for a misspe
ied square-root pro
ess.

19

**higher level unreje
ted s
ore is more likely
apturing the true distribution. If the lower
**

level unreje
ted s
ore did sele
t the true spe
i
ation, the reje
tion rate is likely to shoot

up beyond that level. The impli
ation for empiri
al work is that an SNP sear
h should go

beyond the rst optimal
hoi
e by BIC.

**5.6 Spe
i
ation Test
**

Based on the insights gathered from the EMM experiments with automati
SNP s
ore generator, I
hoose some \edu
ated" xed SNP s
ores for the strong volatility
ase. For the

500 sample size I use s10111 (1 lag in mean, 0 lag in varian
e, 1 lag in polynomial
oeÆ
ient, 1 degree in Hermite polynomial, and 1 degree in
oeÆ
ient polynomial), and for the

1500 sample I use s10121 (1 lag in mean, 0 lag in varian
e, 1 lag in polynomial
oeÆ
ient,

2 degrees in Hermite polynomial, and 1 degree in
oeÆ
ient polynomial). In the
ase of

mean persisten
e, whi
h is widely estimated in empiri
al studies, I use the established SNP

s
ores|s14140 for T = 500 and s14141 for T = 1500. In Figures 5 and 6, these EMM

J-tests results are
ontrasted with GMM. With a knowledge of the true parameters, one

an also perform a likelihood ratio test in MLE or a quasi-likelihood ratio test in QMLE to

see whether the
onden
e ball
on
entrated at the estimated parameter
ontains the true

parameter as often as suggested by the
hi-square (3) random variable. This approa
h is

not available to the empiri
al resear
her, sin
e no true parameter is known. However, in the

Monte Carlo setting, one
an use these \infeasible" tests to judge the reliability of MLE or

QMLE.

When testing the mean persisten
e s
enario (Figure 5), QMLE gives the best inferen
e

as the overreje
tion bias is small and is redu
ing with in
reases in the sample size. GMM

and EMM have severe overreje
tion biases, although the biases shrink rapidly in a large

sample size. EMM seems to be worse than GMM, even with improved estimates of the

weighting matrix. The MLE LR-test diverges as the sample size in
reases, whi
h
an be

attributed to the asymptoti
bias of parameter estimates introdu
ed by approximating the

likelihood fun
tion. Turning to the
ase of strong volatility (Figure 6), one
an see that the

QMLE LR-test and the EMM J-test have almost perfe
t size, and the overreje
tion bias is

negligibly small. On the
ontrary, the MLE LR-test has severe underreje
tion, whi
h does

not shrink asymptoti
ally. Even worse, the GMM J-test has a large underreje
tion bias,

20

**whi
h is diverging with in
reases in sample sizes.
**

There are at least four sour
es of overreje
tion or underreje
tion bias in GMM-type

estimators: ina
urate and
ostly estimates of the weighting matrix; unseasoned sele
tion

of the moment
onditions; an inadequate number of moments to
apture the distribution

feature; and simply a small sample bias. A generi
EMM approa
h over
omes the rst two

problems by adopting a serially un
orrelated information matrix and an optimal SNP s
ore

generator. The
onservative BIC pro
edure in EMM may
hoose too few moments, but

one
an re
tify this problem by using additional information and extending the SNP sear
h

beyond the BIC
hoi
e. The remaining small sample bias
an be remedied by enlarging the

sample size. The above arguments
arry through here, ex
ept in the
ase of persistent mean,

where the EMM estimator may mistake the data as
oming from a unit-root pro
ess.

**5.7 Dete
ting Misspe
i
ation
**

The power for dete
ting misspe
i
ation is examined in two aspe
ts: (1) rst, using automati
SNP s
ore EMM pro
edure to study the power in relation to the number of moments,

whi
h is optimally
hosen by BIC; (2) se
ond, using a xed SNP-s
ore EMM pro
edure to

ompare the power with GMM, where EMM adopts an \edu
ated"
hoi
e of the moment

onditions while GMM adopts the simple lag-augmented moment
onditions.11

In the rst stage, the ben
hmark sto
hasti
volatility model (equations 20 and 21 in

Se
tion 4.3) is used to simulate 1000 repli
ations for the 500 and 1500 sample sizes, and

then a square-root diusion pro
ess is tted to the data. This time I let BIC automati
ally

hoose the best SNP s
ore generator. Sin
e the drift is linear, the
onditional mean with

lag one is
orre
tly spe
ied. For the 500 sample size, 91% of the trials sele
t lag 1; and

for the 1500 sample size, 93% of the trials sele
t lag 1. The
hoi
e of
onditional standard

deviation is all over the pla
e, due to the nature of nonlinear sto
hasti
volatility. For T =

500, the sele
tion is s
attered mainly from lag 1 to lag 4, and for T = 1500, it is s
attered

mainly from lag 3 to lag 6. The
hoi
es of Kz and Kx are predominantly zero. Figures 7 and

11 One

should be
autioned that the alternative model of interest rate pro
ess adopted here is only one

spe
ial
ase, and a full s
ale study of misspe
i
ation issue is
learly outside the s
ope of this paper. Even

in the following limited example, the performan
e of EMM is not unrelated with the established ARCH

(Engle 1982) and GARCH (Bollerslev 1986) ltering of the sto
hasti
volatility pro
ess. In fa
t, the power

of EMM will be optimal if the rst stage SNP auxiliary model|with an ARCH (earlier version) or GARCH

(re
ent version) leading term|adequately
aptures the interest rate dynami
s.

21

**8 plot the 5% reje
tion rates against the number of overidentied moments along with the
**

o
urren
e rates of these moment
hoi
es. The highlight is that the probability of reje
ting

a misspe
ied model does
onverge to one very qui
kly. At T = 500, the 5% level reje
tion

rate is around 80-90% for a range of overidentied moments between 1 to 6, and beyond

that the reje
tion rate is almost 100% (Figure 7). At T = 1500, the reje
tion rate is always

lose to 100%, ex
ept in an exa
tly identied
ase (Figure 8).

In the se
ond stage, similar to the study of the overreje
tion issue, I x the SNP s
ore

generator and look at the reje
tion rate uniformly along the 1%-100% test level. The xed

SNP s
ore generator for the 500 sample size is s13100 (1 lag in mean and 3 lags in varian
e),

and the
hi-square test has a degree of freedom that is 3. When T = 1500, the s
ore is

xed at s15100 (1 lag in mean and 5 lags in varian
e), with 5 degrees of freedom. Figure 9

gives the reje
tion plot for the EMM J-test statisti
s, in
omparison with GMM whi
h has 1

overidentied moment. The upshot is that the EMM has the power to dete
t a misspe
ied

model, and the power qui
kly
onverges to one as the sample size in
reases from 500 to 1500.

In
ontrast, the GMM has a serious underreje
tion problem for the maintained misspe
i
ation, and the underreje
tion bias be
omes larger as sample size in
reases|ultimately loosing

the power to dete
t misspe
i
ation. The explanation is quite simple: EMM
hooses
arefully a SNP moment generator by BIC, while standard GMM simply uses the lag-augmented

instruments.

6 Con
lusions

This paper performs a Monte Carlo study on EÆ
ient Method of Moments (EMM), Generalized Method of Moments (GMM), Quasi-Maximum Likelihood Estimation (QMLE), and

Maximum Likelihood Estimation (MLE) for a
ontinuous-time square-root model under two

hallenging s
enarios|high persisten
e in mean and strong
onditional volatility|that are

ommonly en
ountered when estimating the empiri
al interest rate pro
ess.

MLE a
hieves the highest eÆ
ien
y, while its inferen
es on individual parameters and

overall spe
i
ation are not very reliable and are even misleading on some o
asions. QMLE

is less eÆ
ient in
omparison to MLE, but QMLE stands out as the best inferen
e tool in

both the individual t-test and overall LR-test. EMM shows a
onvergen
e rate faster than

22

**root-n, due to the expanding SNP s
ore
hoi
e by BIC as sample size in
reases. EMM also
**

provides better inferen
e than GMM or MLE in a high volatility s
enario. In the
ase of

persistent mean|
lose to unit root in small samples|some asymptoti
s of MLE and GMM

break down, as parameter estimates and test statisti
s diverge.

A number of lessons
an be learned from this study: (1) MLE is not ne
essarily the best

hoi
e if the numeri
al approximation to the density is
omplex and/or the approximation

tends to diverges near the non-stationary region; (2) QMLE is simple to implement and
an

be very reliable when the spe
i
ation information is easily in
orporated in the
losed-form

onditional mean and varian
e; (3) if there is no new information to be in
orporated into

the moment
onditions, GMM
an not be superior to QMLE; (4) when the true density or

moment fun
tions are not known, EMM is the only
hoi
e; its small sample performan
e is

not ne
essarily inferior to the infeasible MLE or QMLE and is most likely superior to the

infeasible GMM.

23

Referen
es

At-Sahalia, Ya
ine (1996a), \Nonparametri
Pri
ing of Interest Rate Derivatives," E
onometri
a , vol. 64, 527{560.

At-Sahalia, Ya
ine (1996b), \Testing Continuous-Time Models of the Spot Interest Rate,"

The Review of Finan
ial Studies , vol. 9, 385{426.

Andersen, Torben G., Hyung-Jin Chung, and Bent E. Srenson (1999), \EÆ
ient Method of

Moments Estimation of a Sto
hasti
Volatility Model: A Monte Carlo Study," Journal

of E
onometri
s , vol. 91, 61{87.

Andersen, Torben G. and Bent E. Srenson (1996), \GMM Estimation of a Sto
hasti
Volatility Model: A Monte Carlo Study," Journal of Business and E
onomi
Statisti
s , vol. 14,

328{352.

Bansal, Ravi, A. Ronald Gallant, Robert Hussey, and George Tau
hen (1995), \Nonparametri
Estimation of Stru
tural Models for High-Frequen
y Curren
y Market Data,"

Journal of E
onometri
s , vol. 66, 251{287.

Bollerslev, Tim (1986), \Generalized Autoregressive Conditional Heteros
edasti
ity," Journal of E
onometri
s , vol. 31, 307{327.

Bollerslev, Tim and Jeery Wooldridge (1992), \Quasi-Maximum Likelihood Estimators and

Inferen
e in Dynami
Models with Time-Varying Covarian
es," E
onometri
Review ,

vol. 11, 143{172.

Burnside, Craig and Martin Ei
henbaum (1996), \Small-Sample Properties of GMM-Based

Wald Tests," Journal of Business and E
onomi
Statisti
s , vol. 14, 294{308.

Chuma
ero, Romulo A. (1997), \Finite Sample Properties of the EÆ
ient Method of Moments," Studies in Nonlinear Dynami
s and E
onometri
s , vol. 2, 35{51.

Conley, Tim, Lars Peter Hansen, Erzo Luttmer, and Jose S
heinkman (1997), \Short Term

Interest Rates as Subordinated Diusions," Review of Finan
ial Studies , vol. 10, 525{

578.

24

**Cox, John C., Jonathan E. Ingersoll, and Stephen A. Ross (1985), \A Theory of the Term
**

Stru
ture of Interest Rates," E
onometri
a , vol. 53, 385{407.

Dai, Qiang and Kenneth J. Singleton (2000), \Spe
i
ation Analysis of AÆne Term Stru
ture

Models," Journal of Finan
e , forth
oming.

Devroye, Lu
(1986), Non-Uniform Random Variate Generation , Spinger-Verlag.

DuÆe, Darrell and Kenneth Singleton (1993), \Simulated Moments Estimation of Markov

Models of Asset Pri
es," E
onometri
a , vol. 61, 929{952.

DuÆe, Darrell and Kenneth Singleton (1997), \An E
onometri
Model of the Term Stru
ture

of Interest-Rate Swap Yields," Journal of Finan
e , vol. 52, 1287{1321.

Engle, Robert F. (1982), \Autoregressive Conditional Heteros
edasti
ity with Estimates of

the Varian
e of U.K. In
ation," E
onometri
a , vol. 50, 987{1008.

Feller, Wiliam (1951), \Two Singular Diusion Problems," Annals of Mathemati
s , vol. 54,

173{182.

Feller, Wiliam (1971), An Introdu
tion to Probability Theory and Its Appli
ations , vol. 2,

John Wiley & Sons, In
., Prin
eton University, 2nd ed.

Fisher, Mark and Christian Gilles (1996), \Estimating Exponential AÆne Models of the

Term Stru
ture," Working Paper .

Gallant, A. Ronald and Jonathan R. Long (1997), \Estimating Sto
hasti
Dierential Equations EÆ
iently by Minimum Chi-Square," Biometrika , vol. 84.

Gallant, A. Ronald and George Tau
hen (1996a), User's Guide for EMM: A Program for

EÆ
ient Method of Moments Estimation , 1st ed.

Gallant, A. Ronald and George Tau
hen (1996b), User's Guide for SNP: A Program for

Nonparametri
Time Series Analysis , 8th ed.

Gallant, A. Ronald and George Tau
hen (1996
), \Whi
h Moment to Mat
h?" E
onometri

Theory , vol. 12, 657{681.

25

**Gallant, A. Ronald and George Tau
hen (1998a), \The Relative EÆ
ien
y of Method of
**

Moments Estimators," Working Paper .

Gallant, A. Ronald and George Tau
hen (1998b), \Reproje
ting Partially Observed Systems with Appli
ation to Interest Rate Diusions," Journal of the Ameri
an Statisti
al

Asso
iation , vol. 93, 10{24.

Gibbons, Mi
hael R. and Krishna Ramaswamy (1993), \A Test of the Cox, Ingersoll, and

Ross Model of the Term Stru
ture," Review of Finan
ial Studies , vol. 6, 619{658.

Gill, Philip E., Walter Murray, Mi
hael A. Saunders, and Margaret H. Wright (1991), \User's

Guide for NPSOL (Version 4.06): A Fortran Pa
kage for Nonlinear Programming,"

Te
h. rep., Stanford University.

Gourierous, C., A. Monfort, and E. Renault (1993), \Indire
t Inferen
e," Journal of Applied

E
onometri
s , vol. 8, s85{s118.

Hansen, Lars Peter (1982), \Large Sample Properties of Generalized Method of Moments

Estimators," E
onometri
a , vol. 50, 1029{1054.

Hansen, Lars Peter, John Heaton, and Amir Yaron (1996), \Finite-Sample Properties of

Some Alternative GMM Estimators," Journal of Business and E
onomi
Statisti
s ,

vol. 14, 262{280.

Hansen, Lars Peter and Jose Alexandre S
heinkman (1995), \Ba
k to the Future: Generalized Moment Impli
ations for Continuous Time Markov Pro
ess," E
onometri
a ,

vol. 63, 767{804.

Ingram, Beth F. and B. S. Lee (1991), \Simulation Estimation of Time Series Models,"

Journal of E
onometri
s , vol. 47, 197{205.

Johnson, Norman L. and Samuel Kotz (1970), Distributions in Statisti
s: Continuous Univariate Distributions , vol. 2, John Wiley & Sons.

Karatzas, Ioannis and Steven E. Shreve (1997), Brownian Motion and Sto
hasti
Cal
ulus ,

Springer.

26

**Kloeden, Peter E. and E
khard Platen (1992), Numeri
al Solution of Sto
hasti
Dierential
**

Equations , Appli
ations of Mathemati
s, Springer-Verlag.

Lo, Andrew W. (1988), \Maximum Likelihood Estimation of Generalized It'^o Pro
ess with

Dis
retely Sampled Data," E
onometri
Theory , vol. 4, 231{247.

Newey, Whitney K. and Douglas G. Steigerwald (1997), \Asymptoti
Bias for QuasiMaximum-Likelihood Estimators in Conditional Heteros
edasti
ity Models," E
onometri
a , vol. 65, 587{599.

Oliver, F. W. J. (1972), Handbook of Mathemati
al Fun
tions with Formulas, Graphs, and

Mathemati
al Tables , John Wiley & Sons.

Pearson, Neil D. and Tong-Sheng Sun (1994), \Exploiting the Conditional Density in Estimating the Term Stru
ture: An Appli
ation to the Cox, Ingersoll, and Ross Model,"

Journal of Finan
e , vol. 49, 1279{1304.

Tau
hen, George (1997), \New Minimum Chi-Square Methods in Empiri
al Finan
e," in

\Advan
es in E
onometri
s, Seventh World Congress," (edited by Kreps, D. and K. Wallis), Cambridge University Press, Cambridge UK.

White, Halbert (1994), Estimation, Inferen
e, and Spe
i
ation Analysis , Cambridge University Press, University of California, San Diego.

Zheng, Xiaodong and Wei-Yin Loh (1995), \Consistent Variable Sele
tion in Linear Models,"

Journal of the Ameri
an Statisti
al Asso
iation , vol. 90, 1029{1054.

27

**Table 1: Ben
hmark Model Choi
e
**

The square-root model is drt = (a0 + a1 rt )dt + b0 rt1=2 dWt . S
enario 1 is taken from Gallant

and Tau
hen (1998b). In S
enarios 2-4, the varian
e parameter b0 is in
reased by a fa
tor

of 2, 3, and 4 respe
tively. From S
enario 1 to S
enario 5, the mean parameters a0 and a1

are multiplied by 100 and the varian
e parameter b0 is multiplied by 10. From S
enario 5 to

S
enarios 6-8, the varian
e parameter b0 is in
reased by a fa
tor of 2, 3, and 4 respe
tively.

E (rt ), V (rt ), q , E (rt+1 jrt ), and V (rt+1 jrt ) are
al
ulated using equations 3, 4, and 6-8.

E (rt )

V (rt )

Bessel q

E (rt+1 jrt )

V (rt+1 jrt )

E (rt )

V (rt )

Bessel q

E (rt+1 jrt )

V (rt+1 jrt )

S
enario 1

S
enario 2

S
enario 3

S
enario 4

a0 = 0:02491 a0 = 0:02491 a0 = 0:02491 a0 = 0:02491

a1 = 0:00285 a1 = 0:00285 a1 = 0:00285 a1 = 0:00285

b0 = 0:0275

b0 = 0:055

b0 = 0:0825

b0 = 0:11

8.74

8.74

8.74

8.74

1.16

4.64

10.44

18.55

64.88

15.47

6.32

3.12

0.997rt +0.025 0.997rt +0.025 0.997rt+0.025 0.997rt +0.025

0.001rt +0.000 0.003rt +0.000 0.007rt+0.000 0.012rt +0.000

S
enario 5

S
enario 6

S
enario 7

S
enario 8

a0 = 2:491

a0 = 2:491

a0 = 2:491

a0 = 2:491

a1 = 0:285

a1 = 0:285

a1 = 0:285

a1 = 0:285

b0 = 0:275

b0 = 0:55

b0 = 0:825

b0 = 1:1

8.74

8.74

8.74

8.74

1.16

4.64

10.44

18.55

64.88

15.47

6.32

3.12

0.75rt +2.17

0.75rt +2.17

0.75rt +2.17

0.75rt +2.17

0.05rt +0.07

0.20rt +0.29

0.45rt +0.64

0.79rt +1.14

28

**Table 2: Comparing Simulation S
hemes (100,000 Length)
**

For the square-root model drt = (a0 + a1 rt )dt + b0 rt1=2 dWt , the marginal distribution is

a Gamma (equation 2), and the theoreti
al values are
al
ulated a
ordingly. Simulation

by distribution is based on the Poisson-mixing-Gamma formula (equation 11) and is subsequently implemented in both the Monte Carlo data generation and the Maximum Likelihood

Estimation. Simulation by dis
retization is based on the weak-order 2 s
heme (Gallant and

Long 1997) and unders
ores the EÆ
ient Method of Moments. S
enario 1 (LMR-LCV)

is the Low-Mean-Reversion Low-Conditional-Varian
e ben
hmark, and S
enario 8 (HMRHCV) is the High-Mean-Reversion High-Conditional-Varian
e alternative. Both s
enarios

are dened in Table 1.

S
enario 1

Simulated by Simulated by Theoreti
al

LMR-LCV

Distribution Dis
retization

Value

Mean

8.72

8.79

8.74

Varian
e

1.18

1.09

1.16

5% Quantile

6.91

7.19

7.05

Median

8.72

8.72

8.70

95% Quantile

10.50

10.57

10.58

S
enario 8

Simulated by Simulated by Theoreti
al

HMR-HCV Distribution Dis
retization

Value

Mean

8.75

8.70

8.74

Varian
e

18.34

18.22

18.55

5% Quantile

3.04

3.03

3.05

Median

8.08

8.02

8.05

95% Quantile

16.70

16.67

16.81

29

**Table 3: SNP Sear
h for the Sample Size of 500
**

(This note applies to both Table 3 and Table 4.) These results are from 100 repli
ations

of ea
h s
enario with the sample sizes of 500 and 1500. The information
riterion used in

moment sele
tion is S
hwarz's BIC. S
enarios 1-8 are the same as those in Table 1. Ea
h

model spe
i
ation is
hara
terized by a 5-digit number. Ea
h digit
onse
utively stands for

lag in mean, lag in varian
e, lag in polynomial, degree of Hermite polynomial, and degree of

Hermite
oeÆ
ient polynomial.

S
enario 1 S
enario 2 S
enario 3 S
enario 4

Model %

Model %

Model %

Model %

10100 95

10100 96

10100 92

10100 88

11100 3

20100 3

10110 4

11100 7

20100 1

10110 1

11100 2

10110 1

21100 1

20100 2

10120 1

20100 1

21100 1

11111 1

S
enario 5 S
enario 6 S
enario 7 S
enario 8

Model %

Model %

Model %

Model %

10100 90

10100 77

10100 46

10111 25

10110 5

10110 12

10110 25

10100 18

20100 3

11110 4

10111 8

10110 15

11100 1

11100 3

10121 7

10121 11

10120 1

10111 1

10120 6

11110 10

10120 1

11110 3

11120 6

12110 1

10130 1

10120 4

20100 1

11100 1

12120 3

11120 1

10131 2

11130 1

11130 2

12110 1

10112 1

11111 1

15110 1

21120 1

30

**Table 4: SNP Sear
h for the Sample Size of 1500
**

S
enario 1 S
enario 2 S
enario 3 S
enario 4

Model %

Model %

Model %

Model %

10100 96

10100 96

10100 85

10100 57

10110 2

11100 4

11100 11

11100 15

20100 1

20100 2

12100 9

11100 1

12100 1

13100 5

13100 1

11111 2

15100 2

16100 2

11121 1

14100 1

14110 1

15110 1

16110 1

16111 1

18100 1

25100 1

S
enario 5 S
enario 6 S
enario 7 S
enario 8

Model %

Model %

Model %

Model %

10100 89

10111 43

10111 41

12120 12

10110 9

10100 30

10121 13

10131 11

10120 1

10110 10

11110 11

10121 10

11110 1

10120 6

11111 8

11110 8

10121 4

11121 8

10121 7

11110 2

10131 7

11140 7

12120 2

10111 5

10111 5

11111 1

12120 3

11130 5

20111 1

10120 2

13120 5

20120 1

10141 1

10122 4

12110 1

12130 4

11111 1

11141 1

11150 1

11160 1

12110 1

14120 1

21120 1

21140 1

22120 1

31

**Table 5: Finite Sample Bias and EÆ
ien
y Comparison for S
enario 1
**

For ea
h sample size, 1000 Monte Carlo repli
ations are generated from the square-root pro
ess, and the model is estimated by MLE, QMLE, GMM, and EMM respe
tively. S
enario 1

is the Low-Mean-Reversion Low-Conditional-Varian
e (LMR-LCV)
ase. EMM estimation

in this table uses a xed SNP s
ore generator.

Mean Bias

Median Bias

RMSE

True Value T = 500 T = 1500 T = 500 T = 1500 T = 500 T = 1500

Maximum Likelihood Estimation

a0 = 0.02491 -0.0123

-0.0130 -0.0119

-0.0126

0.0125

0.0131

a1 =-0.00285 0.0014

0.0015

0.0014

0.0014

0.0014

0.0015

b0 = 0.02750 -4.4e-5

2.5e-6 -4.6e-5

2.1e-5

0.0009

0.0005

Quasi-Maximum Likelihood Estimation

a0 = 0.02491 0.0994

0.0285

0.0803

0.0209

0.1343

0.0437

a1 =-0.00285 -0.0113

-0.0033 -0.0091

-0.0025

0.0153

0.0050

b0 = 0.02750

3.0e-5

1.2e-5

4.1e-5

1.9e-5

0.0009

0.0005

Generalized Method of Moments

a0 = 0.02491 0.0019

0.0023

4.7e-5

-5.1e-7

0.2418

0.0960

a1 =-0.00285 -0.0012

0.0022 -7.9e-5

-5.8e-5

0.0539

0.0481

b0 = 0.02750 0.0040

0.0075 -7.3e-6

6.1e-6

0.1256

0.1264

EÆ
ient Method of Moments

a0 = 0.02491 0.0451

0.0407

2.6e-4

0.0085

0.1252

0.0944

a1 =-0.00285 -0.0054

-0.0048 -8.1e-5

-0.0012

0.0149

0.0112

b0 = 0.02750 -0.0015

-0.0003 -4.8e-6

-4.3e-7

0.0076

0.0041

32

**Table 6: Finite Sample Bias and EÆ
ien
y Comparison for S
enario 8
**

For ea
h sample size, 1000 Monte Carlo repli
ations are generated from the square-root

pro
ess, and the model is estimated by MLE, QMLE, GMM, and EMM respe
tively. S
enario 8 is the How-Mean-Reversion How-Conditional-Varian
e (HMR-HCV)
ase. EMM

estimation in this table uses a xed SNP s
ore generator.

Mean Bias

Median Bias

RMSE

True Value T = 500 T = 1500 T = 500 T = 1500 T = 500 T = 1500

Maximum Likelihood Estimation

a0 = 2.491 -0.0832

-0.0663 -0.0679

-0.0524

0.1337

0.0923

a1 =-0.285 0.0085

0.0058

0.0029

0.0010

0.0251

0.0161

b0 = 1.100

0.0024

-0.0016

0.0060

0.0000

0.0432

0.0263

Quasi-Maximum Likelihood Estimation

a0 = 2.491

0.0742

0.0224

0.0022

0.0006

0.3613

0.2111

a1 =-0.285 -0.0100

-0.0020 -0.0071

-0.0011

0.0448

0.0258

b0 = 1.100

0.0023

0.0003

0.0015

-0.0001

0.0430

0.0246

Generalized Method of Moments

a0 = 2.491

0.0003

0.0039

0.0001

0.0001

0.1023

0.0541

a1 =-0.285 0.0186

0.0110

0.0001

0.0005

0.0784

0.0561

b0 = 1.100

0.0097

0.0041 -0.0023

-0.0018

0.0838

0.0581

EÆ
ient Method of Moments

a0 = 2.491

0.1323

-0.0067

0.0433

-0.0173

0.4891

0.2000

a1 =-0.285 -0.0310

-0.0022 -0.0199

-0.0000

0.0694

0.0257

b0 = 1.100 -0.0218

-0.0137 -0.0091

-0.0122

0.0618

0.0296

33

MLE: a0

MLE: a1

MLE: b0

0.5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

0

−5

0

QMLE: a0

5

0

−5

0

QMLE: a1

5

0.5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

0

−5

0

GMM: a0

5

0

GMM: a1

5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

−5

0

EMM: a

5

0

EMM: a

5

−5

0

GMM: b0

5

−5

0

EMM: b

5

0

5

1

0

0.5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

0

0

5

0

−5

0

−5

0

QMLE: b0

0

−5

0.5

0

−5

5

0

−5

0

5

−5

**Figure 1: Sampling Distributions of t-Statisti
s for S
enario 1.
**

The notations are respe
tively: \- - -" t-test statisti
s for 500 sample size; \|{" t-test

statisti
s for 1500 sample size; \-.-.-" Normal (0,1) density as the referen
e.

34

MLE: a0

MLE: a1

MLE: b0

0.5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

0

−5

0

QMLE: a0

5

0

−5

0

QMLE: a1

5

0.5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

0

−5

0

GMM: a0

5

0

GMM: a1

5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

−5

0

EMM: a

5

0

EMM: a

5

−5

0

GMM: b0

5

−5

0

EMM: b

5

0

5

1

0

0.5

0.5

0.5

0.4

0.4

0.4

0.3

0.3

0.3

0.2

0.2

0.2

0.1

0.1

0.1

0

0

0

5

0

−5

0

−5

0

QMLE: b0

0

−5

0.5

0

−5

5

0

−5

0

5

−5

**Figure 2: Sampling Distributions of t-Statisti
s for S
enario 8.
**

The notations are respe
tively: \- - -" t-test statisti
s for 500 sample size; \|{" t-test

statisti
s for 1500 sample size; \-.-.-" Normal (0,1) density as the referen
e.

35

100

Percentage of Rejection and of Occurrence

90

80

70

60

50

40

Rejection Curve

30

20

Occurrence Curve

10

0

0

2

4

6

8

Number of Overidendified Moments

10

12

**Figure 3: 5% Overreje
tion Rate of EMM T = 500 with Automati
S
ore Generator.
**

The o
urren
e rate is the frequen
y of the same moment
hoi
e divided by 1000. The

reje
tion rate is the frequen
y of reje
tions divided by the number of o
urren
es.

36

100

Percentage of Rejection and of Occurrence

90

80

70

60

Rejection Curve

50

40

30

20

10

Occurrence Curve

0

0

2

4

6

8

Number of Overidendified Moments

10

12

**Figure 4: 5% Overreje
tion Rate of EMM T = 1500 with Automati
S
ore Generator.
**

The o
urren
e rate is the frequen
y of the same moment
hoi
e divided by 1000. The

reje
tion rate is the frequen
y of reje
tions divided by the number of o
urren
es.

37

MLE LR−Test T = 500

MLE LR−Test T = 1500

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

QMLE LR−Test T = 500

0

100

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

GMM J−Test T = 500

0

100

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

EMM J−Test T = 500

0

100

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

0

100

0

20

40

60

80

QMLE LR−Test T = 1500

100

0

20

40

60

80

GMM J−Test T = 1500

100

0

20

40

60

80

EMM J−Test T = 1500

100

0

20

100

40

60

80

**Figure 5: Spe
i
ation Test for S
enario 1.
**

The likelihood ratio tests for MLE and QMLE are against the true parameter values, and

the J-tests for GMM and EMM are against the overidentifying restri
tions.

38

MLE LR−Test T = 500

MLE LR−Test T = 1500

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

QMLE LR−Test T = 500

0

100

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

GMM J−Test T = 500

0

100

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

EMM J−Test T = 500

0

100

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

0

100

0

20

40

60

80

QMLE LR−Test T = 1500

100

0

20

40

60

80

GMM J−Test T = 1500

100

0

20

40

60

80

EMM J−Test T = 1500

100

0

20

100

40

60

80

**Figure 6: Spe
i
ation Test for S
enario 8.
**

The likelihood ratio tests for MLE and QMLE are against the true parameter values, and

the J-tests for GMM and EMM are against the overidentifying restri
tions.

39

100

Percentage of Rejection and of Occurrence

90

80

Rejection Curve

70

60

50

40

Occurrence Curve

30

20

10

0

0

1

2

3

4

5

6

7

Number of Overidendified Moments

8

9

10

11

**Figure 7: 5% Reje
tion of Misspe
ied Model T = 500
**

The o
urren
e rate is the frequen
y of the same moment
hoi
e divided by 1000. The

reje
tion rate is the frequen
y of reje
tions divided by the number of o
urren
es.

40

100

Rejection Curve

Percentage of Rejection and of Occurrence

90

80

70

60

50

40

Occurrence Curve

30

20

10

0

0

2

4

6

8

10

Number of Overidendified Moments

12

14

16

**Figure 8: 5% Reje
tion of Misspe
ied Model T = 1500
**

The o
urren
e rate is the frequen
y of the same moment
hoi
e divided by 1000. The

reje
tion rate is the frequen
y of reje
tions divided by the number of o
urren
es.

41

GMM J−Test T = 500

GMM J−Test T = 1500

100

100

80

80

60

60

40

40

20

20

0

0

20

40

60

80

0

100

0

20

**EMM J−Test T = 500
**

100

80

80

60

60

40

40

20

20

0

20

40

60

60

80

100

80

100

EMM J−Test T = 1500

100

0

40

80

0

100

0

20

40

60

**Figure 9: Power to Dete
t Misspe
ied Model.
**

The GMM J-test is a
hi-square (1), and the EMM J-test is a
hi-square (3) for T = 500

and
hi-square (5) for T = 1500.

42

- Diseño experimentaluploaded bycarolina
- Growth performance of station tested rams in Sloveniauploaded byGregor Gorjanc
- The Exponentiated Complementary Exponential Geometric Distributionuploaded bySEP-Publisher
- 342-2uploaded bycorbinlm
- Max Likelihooduploaded byravibabukancharla
- shenuploaded byapi-284511189
- Lion 5 Paperuploaded bymilp2
- Actuarial Science Major(SFU)uploaded byAnonymous YjLLXDZuD8
- 22 Final Paper Oyugiuploaded byKostasSoufras
- Misra GamesWorkshopSlides 2013uploaded byLucas Piroulus
- jop43liangwebuploaded byOnkar Bagaria
- The Conquest of South American Inflationuploaded bygerom
- Rodriguez 2008-Bergmann’s Rule and the Geography Ofuploaded byTroocks
- Estimation Theoryuploaded byMurash javaid
- AMJ-2005- Summary and Respuploaded byKen Simons
- A NewmodifiedWeibulldistributionuploaded byFeñita Isabêl Arce Nuñez
- SA AHMAD FPE 220 (2004) 189-198uploaded byArun Ebenezer
- technical problem solving syllabusuploaded byapi-269153238
- Chang2014 Kappauploaded byIrena
- Determining Returns Exercisesuploaded byRicarr Wee Chiong
- Dynamic Analysis of Consumer Response to Marketing Strategiesuploaded bygeorgeserediuc7898
- Airline Passengers time series with Exceluploaded bySpider Financial
- Predicting Type II Censored Data From Factorial Experiments Using Modified Maximum Likelihood Predictoruploaded byapi-3730914
- Duasa (2007)uploaded byirushad
- Chapter 14 Solutionsuploaded byDonna Stamps
- PanelDataNotes-20uploaded byToan
- Biometrics Information Handbook Seriesuploaded byGrazyella Yoshida
- text_classification.pdfuploaded byAsma Mushtaq
- 9709_w06_qp_6uploaded bySarah Phillips
- Workaholism, Scale Development and Preliminary Empirical Testinguploaded bysildia009

- US Federal Reserve: 08-06-07amex-charge-agremntuploaded byThe Fed
- US Federal Reserve: mfruploaded byThe Fed
- US Federal Reserve: pre 20071212uploaded byThe Fed
- US Federal Reserve: agenda financing-comm-devuploaded byThe Fed
- US Federal Reserve: eighth district countiesuploaded byThe Fed
- US Federal Reserve: js3077 01112005 MLTAuploaded byThe Fed
- US Federal Reserve: er680501uploaded byThe Fed
- US Federal Reserve: nib researchuploaded byThe Fed
- US Federal Reserve: 2002-37uploaded byThe Fed
- US Federal Reserve: Check21ConfRptuploaded byThe Fed
- US Federal Reserve: gordenuploaded byThe Fed
- US Federal Reserve: rpts1334uploaded byThe Fed
- US Federal Reserve: rs283totuploaded byThe Fed
- US Federal Reserve: Paper-Wrightuploaded byThe Fed
- US Federal Reserve: callforpapers-ca-research2007uploaded byThe Fed
- US Federal Reserve: 0205estruploaded byThe Fed
- US Federal Reserve: s97mishkuploaded byThe Fed
- US Federal Reserve: 0205mccauploaded byThe Fed
- US Federal Reserve: 0306apgauploaded byThe Fed
- US Federal Reserve: brq302scuploaded byThe Fed
- US Federal Reserve: iv jmcb finaluploaded byThe Fed
- US Federal Reserve: 0205vanduploaded byThe Fed
- US Federal Reserve: wp05-13uploaded byThe Fed
- US Federal Reserve: Paper-Wrightuploaded byThe Fed
- US Federal Reserve: 78199uploaded byThe Fed
- US Federal Reserve: 0205aokiuploaded byThe Fed
- US Federal Reserve: domsuploaded byThe Fed
- US Federal Reserve: PBGCAMRuploaded byThe Fed
- US Federal Reserve: lessonslearneduploaded byThe Fed
- US Federal Reserve: ENTREPRENEURuploaded byThe Fed

Close Dialog## Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

Loading