406
The Nobel Prizes
Under an asset pricing interpretation, (Y
t+l
)′Z
t
is a synthetic payoff vector
with a corresponding price vector (Q
t
)′Z
t
. Finally, we may form the uncondi-
tional expectation by averaging over the coarser conditioning information set
F
t
:
E
S
t
+
S
t
⎛
⎝⎜
⎞
⎠⎟
Y
t
+
( )
′ −
Q
t
( )
′ |F
t
⎡
⎣
⎢
⎤
⎦
⎥ = 0. (3)
This becomes an estimation problem once we parameterize the SDF in terms
of observables and unknown parameters to be estimated.
Hansen and Singleton (1982) is an initial example of this approach.
11
In
that work we consider the case in which the SDF process can be constructed
from observables along with some unknown parameters. Economics comes
into play in justifying the construction of the SDF process and sometimes in
the construction of returns to investment. From an econometric perspective,
time series versions of Laws of Large Numbers and Central Limit Theorems give
us approximate ways to estimate parameters and test restrictions as in Hansen
(1982b).
In Hansen (1982b), I also studied statistical efficiency for a class of GMM
estimators given a particular choice of Z in a manner that extends an approach
due to Sargan (1958, 1959).
12
When (3) has more equations than unknown pa-
rameters, multiple GMM estimators are the outcome of using (at least implic-
itly) alternative linear combinations of these equations equal to the number of
parameters. Since there are many possible ways to embark on this construction,
there is a family of GMM estimators. This family of estimators has an attainable
efficiency bound derived and reported in Hansen (1982b).
13
When the number
11
An earlier application of GMM inference is found in my work Hansen and Hodrick
(1980). In that paper we studied the empirical relationship between the logarithm of a
future spot exchange and the logarithm of the current forward rate and other possible
predictors. We applied ordinary least squares in our work, but with corrected standard
errors. Others were tempted to (and in fact did) apply generalized least squares (GLS) to
“correct for” serial correlation, but applied in this setting GLS is statistically inconsistent.
The counterpart to the moment conditions studied here are the least squares orthogo-
nality conditions. The contract interval played the role of l in this least squares analysis
and was typically larger than one. In subsequent work, Hansen and Hodrick (1983), we
used a SDF formulation to motivate further empirical characterizations, which led us
to confront over-identification. See also Bilson (1981) and Fama (1984) who featured a
cross-currency analysis.
12
See Arellano (2002) for a nice discussion relating GMM estimation to the earlier work
of Sargan.
13
See Hansen (2007b) for a pedagogical discussion of GMM estimation including discus-
sions of large sample statistical efficiency and tests.
6490_Book.indb 406
11/4/14 2:30 PM
Uncertainty Outside and Inside Economic Models 407
of equations exceeds the number of free parameters, there is also a direct way to
test equations not used formally in estimation. While nesting estimators into a
general GMM framework has great pedagogical value, I was particularly inter-
ested in applying a GMM approach to problems requiring new estimators as in
many of the applications to financial economics and elsewhere.
14
Notice that the model, as written down in equation (3), is only partially
specified. Typically we cannot invert this relation, or even its conditional coun-
terpart, to deduce a full time series evolution for economic aggregates and fi-
nancial variables.
15
Other relations would have to be included in order to obtain
a full solution to the problem.
3.2 Further econometric Challenges
I now digress temporarily and discuss some econometric extensions that I and
others contributed to.
3.2.1 S
emiPaRameTRic
e
fficiency
Since the model is only partially specified, the estimation challenge leads di-
rectly to what is formally called a semiparametric problem. Implicitly the re-
mainder of the model can be posed in a nonparametric manner. This gives
rise to a problem with a finite-dimensional parameter vector of interest and an
infinite-dimensional “nuisance” parameter vector representing the remainder
of the model. This opens the door to the study of semiparametric efficiency of
a large class of estimators as will be evident from the discussion that follows. In
typical GMM problems, the actual introduction of the nuisance parameters can
be sidestepped.
Relation (2) conditions on the information set of economic agents. We have
great flexibility in choosing the matrix process Z. The entries of Z
t
should be
in the F
t
information set, but this still leaves many options when building a
Z process. This flexibility gives rise to an infinite class of estimators. In Han-
sen (1982b), I studied statistical efficiency given a particular choice of Z. This
14
Other econometricians have subsequently found value in unifying the treatment of
GMM estimators into a broader type of extremum estimators. This, however, misses
some of the special features of statistical efficiency within a GMM framework and does
not address the issue of how to construct meaningful estimators from economic models.
15
For those reluctant to work with partially specified models, Lucas (1978) showed how
to close a special case of this model by considering an endowment economy. But from
an empirical standpoint, it is often not necessary to take the endowment nature of the
economy literally. The consumption from the endowment economy may be conceived of
as the equilibrium outcome of a model with production and preserves the same pricing
relations.
6490_Book.indb 407
11/4/14 2:30 PM