Cumby Huizinga - Testing Autocorrelation OLS
Cumby Huizinga - Testing Autocorrelation OLS
Cumby Huizinga - Testing Autocorrelation OLS
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to Econometrica.
http://www.jstor.org
1. INTRODUCTION
We would like to thank seminar participants at the University of Chicago and Lars Hansen, in
particular, as well as two anonymous referees for helpful comments. Huizinga gratefully acknowledges funding from the University of Chicago Graduate School of Business and the Sloan
Foundation.
2 See Hansen (1982) for a description of Generalized Method of Moments estimators. Cumby,
Huizinga, and Obstfeld (1983) describe the two-step two-stage least squares estimator.
3 Godfrey (1978a) describes a test that is valid with some instrumental variables estimators, but
the test is not valid in the presence of conditionally heteroscedastic errors or with instrumental
variables estimators such as two-step two-stage least squares. The test, like Durbin's h test, is also
restricted to testing the significance of the first autocorrelation of the regression error.
4See, for example, work on returns in the foreign exchange market by Hansen and Hodrick
(1980), the study of real interest rates by Huizinga and Mishkin (1984), the investigation of stock
returns by Fama and French (1988), and work on the term structure of interest rates by Mishkin
(1990). Hall (1988), Hansen and Singleton (1990), and Christiano, Eichenbaum, and Marshall (1991)
address the issue of time aggregated data.
185
186
ROBERT
E. CUMBY
therefore necessitate nonlinear estimation.5 In contrast, the 1 test described in this paper
avoids the use of nonlinear- estimation because it is based solely on the sample
autocorrelations of regression residuals and a consistent measure of their asymptotic
covariance matrix. The 1 test thus reflects a desire for simplicity, and for ensuring that
regression diagnostics do not become more costly or more difficult to compute than the
original regression.
The third situation is conditional heteroscedasticity of the error term, a situation that
is frequently detected in empirical studies.6 As discussed below, the presence of conditional heteroscedasticity can seriously undermine tests for serial correlation of regression
errors that ignore its presence. The 1 test can be used with either conditionally
heteroscedastic or homoscedastic errors.
The outline of the paper is as follows. In Section 2, we derive the asymptotic
distribution of the sample autocorrelations at lags q + 1 to q + s of regression residuals
from a model where the regression errors are a qth order moving average with possibly
conditionally heteroscedastic innovations. The regression is assumed to be estimated by
instrumental variables, with instruments that are predetermined, but not necessarily
strictly exogenous. We note how the distribution simplifies when the regression errors
are conditionally homoscedastic and when all regressors are predetermined or strictly
exogenous variables so that ordinary least squares is appropriate. Based on this asymptotic distribution of the sample autocorrelations of regression residuals, a test of the
hypothesis that the true regression errors are a qth order moving average process is
presented in Section 3. Section 4 contains summary remarks.
Y=X'
+E,
(t
T),
where yt and E, are scalar random variables, X, is a 1 x k vector of the k scalar random
variables X1 t, X2, t.... Xk, t, and 8 is a k x 1 vector of unknown parameters. The vector
of regressors, Xt, may include jointly endogenous variables (those contemporaneously
correlated with Et), predetermined variables (those uncorrelated with Et+j for j > 0 but
correlated with Et_ for some j > 0), or strictly exogenous variables (those uncorrelated
with ?t+j for all j).
The regression errors Et are assumed to have mean zero and satisfy two other
conditions. First, though they are allowed to be conditionally heteroscedastic, they are
assumed to be unconditionally homoscedastic. Second, for a known q > 0, their autocorrelations at all lags greater than q are required to be zero.
It is also assumed that there exists a 1 x h vector of instrumental variables Zt,
comprised of h > k scalar random variables Z1,t, Z2,t I.. . Zh, t, each of which is uncorrelated with Et. Zt is required to be predetermined, but not necessarily strictly exogenous,
5This is true of the Box-Pierce test, the likelihood ratio test, and, as discussed in Godfrey
(1978c), the Lagrange multiplier test. It is also true of a GMM approach that jointly estimates the
parameters of primary interest and the residual autocorrelations.
6 Diebold (1986) proposes a test for serially correlated errors in the presence of the ARCH
distribution described by Engle (1982). The I test discussed here is more generally applicable than
Diebold's test because it does not assume any particular functional form for the conditional
heteroscedasticity.
THE AUTOCORRELATION
TESTING
STRUCTURE
187
E(Et) = ?'
(3)
E(EtEt -n)l/(.-
'
2E(
= Pn
and
(4)
E(Et
|Zt
Zt-
1 *
Et-q-D1
Ft-q-21
.. *
)?
Q=
lim (1/T)E(Z'E?'Z)
d = (X'ZA
'Z'X) -XZA
- lz'y
for some observable matrix AT. This formulation is general enough to include ordinary
least squares (AT = (X'X/T) and Z = X), two-stage least squares (AT = (Z'Z/T)), and
two-step two-stage least squares (AT is proportional to a consistent estimate of f2). The
asymptotic covariance matrix for the estimator d is denoted Vd,with
(7)
Vd = DQD'
(8)
1Z'X) -XZA
The objective of this section of the paper is to derive, within the framework of the
model described by equations (1)-(8), the asymptotic covariance matrix of the sample
autocorrelations of the regression residuals, t= Yt- Xt d. In the following section we
show how a consistent estimate of this covariance matrix can be used to test the
hypothesis that the s x 1 vector p = (Pq+1i'... Pq+sY= 0
Let r= [rq+, rq+2,..., q+sI' and
T
A
At
?tt-n
(9)
At2
t= 1
(10)
vr
=vTr + -aFT(d-
),
(ll)
n
~t=n+l1
=T
t=1
188
and the jth row of the s x k matrix dr/d8 is evaluated at d*, which lies between d and
8. Equation (10) shows that the asymptotic covariance matrix of r (the vector of sample
autocorrelations of the regression residuals) can be derived as the asymptotic covariance
matrix for the sum of r (the vector of sample autocorrelations of the true disturbances)
and dr/d8(d - 8). Only when ar/la can safely be ignored will the sampling variation in
the estimation of 8 not affect the sampling variation in the estimation of p.
Let the s X k matrix B have i, jth element,
(12)
B(i,j)
=-[E(ct_qjiXj,t)
+E(
tXj,t-q-i)] /E(Kt)
We show in the Appendix that B = plim dr/la and thus that BVdB' is the asymptotic
covariance matrix of ar/d8(d - 8). In most models, the implication of equation (4) that
... ) = 0 will be sufficient to ensure Xi, t-q-i is predetermined with
t-q-2,
E(?tt_q_l
respect to Et and thus that the second term of the sum in equation (12) is zero.
To complete the notation, let jit = tE q i for i =1,...,s,
jt=EtZj,t for i=
1,.. ., h, the ijth element of the s X s matrix VJ be given by
q
(13)
Vr(i, j)
E(~i,t(j,t-n)~
n= -q
(14)
C(i, j)
-2w
E(git(i t',t-n)-
n=-q
In the Appendix we show that Vr is the asymptotic covariance matrix of r and that the
asymptotic covariance matrix of r with dr/d8(d - 8) is BDC'. Proposition 1 combines
these findings in giving the key result of this section.7
PROPOSITION1: Given equations (1) through (14) and the regularityconditions stated in
A
the Appendix, FTYrPN(O, VJ^),where V^= Vr+ BVdB' + CD'B' + BDC'.
Proposition 1 states that, in general, having to estimate the residuals will affect the
asymptotic distribution of their sample autocorrelations. The following special cases of
the general model provide further insight into Proposition 1 and help clarify the
relationship between tests based on the asymptotic distribution of r and tests of residual
autocorrelation proposed elsewhere in the literature.
Case (i): Strictly Exogenous Regressors. Since B = 0 when the regressors are strictly
exogenous, VJ= Vr and one can safely ignore the fact that the true residuals are
unavailable.
Case (ii): Conditionally Homoscedastic Residuals. We show in the Appendix that
when the residuals are conditionally homoscedastic, Vr and C can be rewritten as
q
(15)
Vr(i,i) =
Pn-i+jPn
n= -q
and
(16)
C(i,j)=
q
E
PnE(et-q-iZjZt-n).
n= -q
TESTING
THE AUTOCORRELATION
STRUCTURE
189
The well known result that the sample autocorrelations of a serially uncorrelated series
are independent and asymptotically normal with variance 1/T follows from (15) with
q = 0. When q > 0 the sample autocorrelations are not independent and, though asymptotically normal, do not have variance 1/T.
Case (iii): Conditionally Homoscedastic Residuals, Predetermined Regressors, and
q = 0. When the regressors are predetermined, ordinary least squares yields consistent
estimates of 8; we can set Z = X, AT= X'X/T, and the second term of B will be zero.
Combining this with the assumption of conditional homoscedasticity (so that equation
(16) is valid) and q = 0 (so that p,n= 0 for n # 0) yields C = -o,2B. Furthermore,
Vd= oQ2plim(X'X/T)-1 = o2D so that BDC'= -BVdB'. Finally, it follows from equation (15) that in this case VJ= I, and thus VJ-= I - BVdB'.
Unlike the case of strictly exogenous regressors, when regressors are merely predetermined one cannot safely ignore the use of regression residuals rather than the true
= I - BVdB' can be used
disturbances in estimating autocorrelations. The expression VP/
to derive the well-known Durbin (1970) h test. Durbin (1970) considers testing whether
the autocorrelation of the error term at lag one is zero in a model with lagged dependent
variables and strictly exogenous variables as regressors. In this case B will contain all
zeros except a single value of minus one in the position corresponding to the dependent
variable lagged once. Using Vdl to denote the estimated variance of the coefficient on
this variable, the asymptotic variance of the first autocorrelation of the regression
residuals is seen to be 1/T - Vdl, which matches the formula given by Durbin (1970).8
Case (iv): Conditionally Homoscedastic Residuals, Only Lagged Dependent Variables, and q = 0. When the regression error is conditionally homoscedastic, X, contains
only k lagged values of yt, and q = 0, we have a special case of the model considered by
Box and Pierce (1970), who propose testing the hypothesis of zero correlation in the
regression error by comparing Q, = TP'r to the critical value of a chi-squared random
variable with s - k degrees of freedom. Understanding the logic behind the Box-Pierce
test and why the test in general fails when regressors other than lagged dependent variables are present becomes quite simple using the result from case (iii) that VJ=
I-BVdB'.
Specifically, it can be shown that when X, contains only lagged values of yt, Vd
approaches (B'B)- 1 as s increases. It follows that as s increases, V^ approaches
I - B(B'B)- B', an idempotent matrix of rank s - k. Hence, for both large s and large
T, Q, will be approximately distributed as a chi-square with s - k degrees of freedom.9 10
If, however, X, contains any variables other than lagged dependent variables, Vd will not
in general approach (B'B)1 and it is unlikely, though not impossible, that I - BVdB'
will be an idempotent matrix.
3. TESTING RESIDUAL AUTOCORRELATIONS
EQUAL TO ZERO
The results presented in Section 2 can be used to develop a Wald test of the null
hypothesis that the regression error in equation (1) is uncorrelated with itself at lags
8
Godfrey (1978b) also considers the case of lagged endogenous and/or strictly exogenous
regressors, conditionally homoscedastic errors, and q = 0. Among other things, he extends Durbin
(1970) by showing that the asymptotic covariance matrix for a vector of sample autocorrelations of
regression residuals is I - BVdB', the formula derived above.
If W is an n x 1 random normal vector with mean 0 and n x n covariance matrix V whose
trace is nonzero, then W'W is distributed as a chi-square random variable with n - m degrees of
freedom if and only if V is idempotent and has rank n - m. See Johnson and Kotz (1970, pages
177-178).
10 Ljung (1986) investigates how large s must be before the
Qs statistic approaches the chi-square
distribution. She finds that in samples of 50 or 100 observations, s > 10 is sufficient for all AR(1)
models examined and that s > 2 is sufficient for AR(1) models with the autoregressive parameter
below .9.
190
+CD'B+BDC']
Vd.
r X2(s).
Proposition 2 states that if Vr, B, C, D, and Vd can be estimated consistently, then the
Iq statistic will be asymptotically distributed as a chi-square random variable with s
degrees of freedom.12 In the remaining part of this section we discuss how consistent
estimates of Vr, B, C, D, and Vd can be formed.
Define the (h + s) x 1 vector mtby
(17)
'7 = (1)t'
t)-
11
mtis proportional
tE)
( ?
-E2i)
so that
BVdB' BDC'
CD'B'
Vr .
It follows from equation (20) that a consistent estimate of the asymptotic covariance
matrix of r can be obtained from consistent estimates of P and 'I. It also follows that if
the consistent estimate of 'I is positive definite in the sample, the resulting lq will be
positive.
11
In many instances, instrumental variables are chosen as lagged endogenous variables so that
rejecting the null hypothesis may call into question the validity of equation (4). In such cases it may
be preferable to think of the null hypothesis being tested as a joint hypothesis concerning the serial
correlation of the residuals and the validity of the instruments. Viewed in this way, the test
described in this paper becomes an alternative to the J statistic proposed in Hansen (1982).
12 Godfrey (1978b) considers a model with lagged endogenous and strictly exogenous regressors,
conditionally homoscedastic errors, and q = 0. He shows that using r to test p = 0 is equivalent to
the Lagrange multiplier test of the null hypothesis that the error term is serially uncorrelated
against the alternatives that the error is MA(s) or AR(s) for s > 0. Hence, in some models, the test
described in Proposition 2 is asymptotically equivalent to a likelihood ratio test. However, Godfrey
(1978c) shows that in the same model but with q > 0, computation of the Lagrange multiplier test of
the null hypothesis that the error term is MA(q) against the alternatives that the error is MA(q + s)
or AR(q + s) requires that the moving average parameters be estimated. In this model the test
described in Proposition 2 may not possess all the desirable properties of a Lagrange multiplier or
likelihood ratio test, but will be less computationally burdensome than those tests. A procedure that
would not require full maximum likelihood estimation of the moving average parameters is to
implement a C(a) test. Such a test would be asymptotically equivalent to the likelihood ratio and
Lagrange multiplier tests and would only require that the derivatives of the likelihood function be
evaluated at initial consistent estimates. See Godfrey (1989, pp. 27-28).
TESTING
THE AUTOCORRELATION
191
STRUCTURE
(21)
U(i, j)
for
i-j-q
> 0,
=0
otherwise,
so that the jth column of U is the vector of regression residuals lagged q +j times.
Then,
A
(22)
62 = (1/T)
t
t= 1
(23)
B=
U'X/T)/dE2,
and
(24)
T(X'ZAT1ZTX)'X'ZAiI
(25)
7t
t?Ft-q-1,
At
be
.. I**?tt-q-s)
(26)
Rn= (1/T)
E,
19t/-n
t=n + 1
(27)
WN R
P=
n_ -N
C= U V'Z/Tdr2,
Aj
where
192
ROBERT
E. CUMBY
While the analysis of this paper centers on the asymptotic distribution of simple
autocorrelations, the results are also relevant for the asymptotic distribution of partial
autocorrelations of regression residuals. Regressing E' on Et'q- 1,.
t_q_s
yields the
estimated coefficient vector b = Fr, where F = (U'U) converges in probability to a
s x s matrix F. As a result, b converges in distribution to a normal random variable with
mean zero and covariance matrix Vb = FVjF' when the null hypothesis is true, and the
standard Wald statistic for testing b = 0 will be numerically identical to the 1 statistic
described in Proposition 2.14 In the special case of q = 0, F is an identity matrix so that
even though b will not equal r in finite samples, one can replace r with b in Proposition
2 and obtain a valid test.
A
4. CONCLUDING REMARKS
The distribution of the 1 test described in this paper is based on asymptotic distribution theory and thus leaves open the issue of how well the test will work in finite samples.
In Cumby and Huizinga (1990) we report the results of a series of Monte Carlo
experiments that investigate the performance of the 1 test in a number of models. In
each model, the performance of the 1 test is evaluated both in sample sizes of 50 and 100
observations, both with errors that are conditionally homoscedastic and errors that are
conditionally heteroscedastic, and both with q = 0 and q = 2.
Several conclusions emerge from these Monte Carlo experiments. First, the small
sample distribution of the 1 statistic is close to the asymptotic distribution in the tails, so
that the test is reliable in terms of its size. Second, performance of the 1 test in the case
of q = 2 is sensitive to how the matrix 'I in equation (18) is estimated. The best results
were obtained using equation (27) and the "Gaussian" weights wN = exp ( - i2/2N2).15
Third, while the presence of conditionally heteroscedastic errors does not present any
problems for the 1 test, it often makes the popular portmanteau tests suggested by Box
and Pierce (1970) and Ljung and Box (1978) quite unreliable.
Stern School of Business, New York University, 100 TrinityPlace, New York, NY 10006,
U.S.A.
and
Graduate School of Business, University of Chicago, 1101 E. 58th Street, Chicago, IL
60637, U.S.A.
Manuscript
receivedJanuary,1989;final revisionreceivedMarch,1991.
APPENDIX
This appendixprovidesa proof of the main propositionin the text. Let y,, E,, X1,,. . .,
T. Define X,
Zl, , . . .,SZh,t be scalarrandomvariableson whichwe have observationsfor t = 1.
and Z, to be the 1 x k and 1 x h vectors(X1,, ..., Xk,I) and (Z1,
Zht,), and define y, X, and
Z to be the T x 1, T x k, and T x h matrices(y1.
YT), (X,.,
XTi)',and (Z ..., ZT)'. Define
,)' for j, = ElZ,1 and i,I = El Elq_j, and let AT be an observable
mt= (w1,1' * Oh, t E1,
14
Since the estimated covariance matrix for b reported by standard regression packages will not
in general be a consistent estimate of Vb, testing b = 0 with the typical F test reported by these
packages is not an asymptotically valid procedure.
15
See Brillinger (1975, p. 55) for a discussion of the Gaussian weighting scheme. Details on the
optimal choice of N are discussed in Cumby and Huizinga (1990). The performance of the I test was
substantially poorer when IF was estimated with equation (27) and the modified Bartlett weights
discussed by Anderson (1971) and Newey and West (1987).
TESTING
THE AUTOCORRELATION
193
STRUCTURE
h x h matrix. We assume that for a known constant q and unknown k x 1 vector of constants 8,
(Al)
(A2)
Yt=Xts+Et,
(A3)
E(Et) = 0,
(A4)
(A5)
(A6)
IFE(7
(A7)
(I/T)
Et-q-21 .. )
IESt-q-1,
?0
q,
t 1t-q)
+
+E(-qT-qj-q+1)
+E(ijtij'+q-i)
plim X'Z exists and has rank k,
+E(qt-q'j+q)
is positive definite,
and
(A8)
PROOF: The proof can be found in Cumby, Huizinga, and Obstfeld (1983).
LEMMA 2: Given
plim dr/d8 = B.
the assumptions
and
(A1)-(A8)
above,
(drjdaj)1 =d*=d,
where
et =y-X
id
[xj,
tet -q-i
+ Xj,t_q-iet ]
Le 2
+2
L e,et-q-i E Xi, e,
[Ee7]2 2
2]e
Therefore, using (Al), the fact that d is a consistent estimate of 8 (Lemma 1), and the fact that d:*
lies between d and 8, we get
plim
dri
E(Xj,t?t-q-i)
E(Xj,t-q-iEt)
0|.2
0.q2
E_Xj, t?_t)
qi
Since pj = 0 for j > q, the third term in this sum is zero and the lemma is proved.
The proof of Proposition 1 is now straightforward.
Q.E.D.
194
ROBERT
E. CUMBY
PROOF:
r^
N(O,VI^)with
dr
Tr+ VT_ (d -
),
where the ith row of dr/d8 is evaluated at d*, which lies between d and S. Stacking the terms on
the right-hand side of this expression and substituting the definitions of d and U gives
-
drIdS T( X'ZA
[ar/d6(d-b)|
[ I
L/
T(IAT)
X'ZAT
Z'X)
Ii1xLU)?XZT
pZtg/
plim[dr/d T( X'ZA-T1Z'X)
0
(EIET)
I]
X ZAT
[BD
[0
I0
Z[E FT
N(O, W')
for
'aCIOE
,l 2
Thus,
A
,FT drldS(d - )
N(?' 0't4")
where
?t?t=
BDC'
BVdB'
CD'B'
VI
r' is the sum of the two random vectors that are asymptotically normally distributed with
Since FT
covariance matrix 01IP', it follows that fTr is asymptotically normally distributed with covariance
Q.E.D.
matrix given by BVdB' + BDC' + CD'B' + V, and the proof is complete.
In the text, we discuss how the asymptotic distribution of Vr^ is affected when the assumption
that Et is conditionally homoscedastic,
(A9)
Zt-11 ..?t-q-1,1t-q-2,1
E(tt-nIlZt,
=E(1tEt1n),
0 6 n <q,
is added to assumptions (A1)-(A8) above. In particular, equations (15) and (16) give forms of
equations (13) and (14) which are claimed to be valid when this assumption is added. To verify that
equation (15) is in fact correct, note that when (A9) holds and - q < n 6 q,
(J-4E(itjtn)
=
=
(E
(E4E(etEt-qi_t-nqj
OE
t-n-q-j)
4E{Et-q-ijt-n-q-jE(-t-t-n
=~4E(,t-qjiet-nqj)E(e-te-t)
in-n)
Pn+s-iPn-
I-t-q-iI
?t-n-q-j))
TESTING
THE AUTOCORRELATION
STRUCTURE
195
REFERENCES
T. W. (1971): The Statistical Analysis of Time Series. New York: John Wiley and Sons.
Box, G. E. P., AND D. A. PIERCE (1970): "Distribution of Residual Autocorrelations in Autoregressive-Integrated Moving Average Time Series Models," Journal of the American Statistical Association, 65, 1509-1526.
BREUSCH, T. S., AND L. GODFREY (1981): "A Review of Recent Work on Testing for Autocorrelation in Dynamic Simultaneous Models," in Macroeconomic Analysis, ed. by D. Currie, R. Nobay,
and D. Peel. London: Croon-Helm.
BRILLINGER, D. R. (1975): Time Series: Data Analysis and Theory. New York: Holt, Rinehart, and
Winston.
CHRISTIANO, L. J., M. EICHENBAUM, AND D. MARSHALL (1991): "The Permanent Income Hypothesis
Revisited," Econometrica, 59, 397-423.
CUMBY, R. E., AND J. HUIZINGA (1990): "Testing the Autocorrelation Structure of Disturbances in
Ordinary Least Squares and Instrumental Variables Regressions," NBER Technical Working
Paper No. 90.
CUMBY, R. E., J. HUIZINGA, AND M. OBSTFELD (1983): "Two-Step Two-Stage Least Squares
Estimation in Models with Rational Expectations," Journal of Econometrics, 21, 333-355.
DIEBOLD, F. X. (1986): "Testing for Serial Correlation in the Presence of ARCH," Proceedings of
the American Statistical Association, Business and Economics Statistics Section, 323-328.
DURBIN, J. (1970): "Testing for Serial Correlation in Least Squares Regression Models when some
of the Regressors Are Lagged Dependent Variables," Econometrica, 38, 410-421.
ENGLE, R. F. (1982): "Autoregressive Conditional Heteroscedasticity With Estimates of the Variance of United Kingdom Inflation," Econometrica, 50, 987-1008.
FAMA, E. F., AND K. R. FRENCH (1988): "Permanent and Transitory Components of Stock Prices,"
Journal of Political Economy, 96, 246-273.
GODFREY, L. G. (1978a): "A Note on the Use of Durbin's h Test when the Equation Is Estimated
by Instrumental Variables," Econometrica, 46, 225-228.
(1978b): "Testing Against General Autoregressive and Moving Average Error Models when
the Regressors Include Lagged Dependent Variables," Econometrica, 46, 1293-1301.
(1978c): "Testing for Higher Order Serial Correlation in Regression Equations when the
Regressors Include Lagged Dependent Variables," Econometrica, 46, 1303-1310.
(1989): Misspecification Tests in Econometrics. New York: Cambridge University Press.
HALL, R. E. (1988): "Intertemporal Substitution in Consumption," Journal of Political Economy, 96,
339-357.
HANNAN, E. J. (1973): "Central Limit Theorems for Time Series Regressions," Zeitschrift fur
Wahrscheinlichkeitstheorieund VerwandteGebiete, 26, 157-170.
HANSEN, L. P. (1982): "Large Sample Properties of Generalized Method of Moments Estimators,"
Econometrica, 50, 1029-1054.
HANSEN, L. P., AND R. J. HODRICK (1980): "Forward Exchange Rates as Optimal Predictors of
Future Spot Rates: An Econometric Analysis," Journal of Political Economy, 88, 829-853.
HANSEN, L. P., AND K. J. SINGLETON (1990): "Efficient Estimation of Linear Asset Pricing Models
with Moving Average Errors," NBER Technical Working Paper No. 86.
HUIZINGA, J., AND F. S. MISHKIN (1984): "Inflation and Real Interest Rates on Assets with Different
Risk Characteristics," Journal of Finance, 34, 699-714.
JOHNSON, N. L., AND S. KOTZ (1970): Continuous UnivariateDistributions-2. New York: John Wiley
and Sons.
IJUNG, G. M. (1986): "Diagnostic Testing of Univariate Time Series Models," Biometrika, 73,
725-730.
IJUNG, G. M., AND G. E. P. Box (1978): "On a Measure of Lack of Fit in Time Series Models,"
Biometrika, 65, 297-303.
McLEOD,A. I. (1978): "On the Distribution of Residual Autocorrelations in Box-Jenkins Models,"
Journal of the Royal Statistical Society, 40, 296-302.
MISHKIN, F. S. (1990): "What Does the Term Structure Tell Us About Future Inflation?" Journal of
Monetary Economics, 25, 77-95.
NEWEY, W. K., AND K. D. WEST (1987): "A Simple, Positive Definite, Heteroscedasticity and
Autocorrelation Consistent Covariance Matrix," Econometrica, 55, 703-708.
SARGAN, J. D. (1976): "Testing for Misspecification After Estimating Using Instrumental Variables,"
London School of Economics Working Paper.
ANDERSON,