420
The Nobel Prizes
beliefs inside the model for two reasons. First, investor beliefs may differ from
those implied by the model even if other components of the model are correctly
specified. For instance, when historical evidence is weak, there is scope for be-
liefs that are different from those revealed by infinite histories of data. Second,
if some of the model ingredients are not correct but only approximations, then
the use of model-based beliefs based on an appeal to rational expectations is less
compelling. Instead there is a rationale for the actors inside the model to adjust
their beliefs in face of potential misspecification.
For reasons of tractability and pedagogical simplicity, throughout this and
the next section I use a baseline probability model to represent conditional ex-
pectations, but not necessarily the beliefs of the people inside the model. Pre-
suming that economic actors use the baseline model with full confidence would
give rise to a rational expectations formulation, but I will explore departures
from this approach. I present a tractable way to analyze how varying beliefs
will alter this baseline probability model. Also, I will continue my focus on the
channel by which SDFs affect asset values. A SDF and the associated risk prices,
however, are only well-defined relative to a baseline model. Alterations in beliefs
affect SDFs in ways that can imitate risk aversion. They also can provide an ad-
ditional source of fluctuations in asset values.
My aim in this section is to study whether statistically small changes in be-
liefs can imitate what appears to be a large amount of risk aversion. While I fea-
ture the role of statistical discipline, explicit considerations of both learning and
market discipline also come into play when there are heterogeneous consumers.
For many environments there may well be an intriguing interplay between these
model ingredients, but I find it revealing to narrow my focus. As is evident from
recent work by Blume and Easley (2006), Kogan et al. (2011) and Borovička
(2013), distorted beliefs can sometimes survive in the long run. Presumably
when statistical evidence for discriminating among models is weak, the impact
of market selection, whereby there is a competitive advantage of confidently
model specification a model. A model builder may impose these restrictions prior to
looking at the data. The expectations become “rational” once the model is fit to data,
assuming that the model is correctly specified. I used GMM and related methods to ex-
amine only a portion of the implications of a fully specified, fully solved model. In such
applications, an empirical economist is not able to use a model solution to deduce the be-
liefs of economic actors. Instead these methods presume that the beliefs of the economic
actors are consistent with historical data as revealed by the Law of Large Numbers. This
approach presumes that part of the model is correctly specified, and the data are used as
part of the implementation of the rational expectations restrictions.
6490_Book.indb 420
11/4/14 2:30 PM
Uncertainty Outside and Inside Economic Models 421
knowing the correct model, will at the very least be sluggish. In both this and the
next section, I am revisiting a theme considered by Hansen (2007a).
5.1 Martingale Models of belief Perturbations
Consider again the asset pricing formula but now under an altered or perturbed
belief relative to a baseline probability model:
E
S
t
+
S
t
⎛
⎝⎜
⎞
⎠⎟
Y
t
+
⎪F
t
⎡
⎣
⎢
⎤
⎦
⎥ =
Q
t
(12)
where the
Ẽ is used to denote the perturbed expectation operator and
S̃ is the
SDF derived under the altered expectations. Mathematically, it is most conve-
nient to represent beliefs in an intertemporal environment using a strictly posi-
tive (with probability one) stochastic process M with a unit expectation for all t
≥ 0. Specifically, construct the altered conditional expectations via the formula:
E B
τ
⎪F
t
⎡⎣
⎤⎦ = E
M
τ
M
t
⎛
⎝⎜
⎞
⎠⎟
B
τ
⎪F
t
⎡
⎣
⎢
⎤
⎦
⎥
for any bounded random variable B
τ
in the date τ ≥ t information set F
τ
. The
martingale restriction imposed on M is necessary for the conditional expecta-
tions for different calendar dates to be consistent.
29
Using a positive martingale M to represent perturbed expectations we re-
write (12) as:
E
M
t
+
S
t
+
M
t
S
t
⎛
⎝⎜
⎞
⎠⎟
Y
t
+
⎪F
t
⎡
⎣
⎢
⎤
⎦
⎥ =
Q
t
which matches our original pricing formula (1) provided that
S = MS̃. (13)
29
The date zero expectation of random variable B
t
that is in the F
t
information set may
be computed in multiple ways
E B
t
⎪F
0
⎡⎣
⎤⎦ = E
M
τ
M
0
⎛
⎝⎜
⎞
⎠⎟
B
t
⎪F
0
⎡
⎣
⎢
⎤
⎦
⎥ = E
M
t
M
0
⎛
⎝⎜
⎞
⎠⎟
B
t
⎪F
0
⎡
⎣
⎢
⎤
⎦
⎥
for any τ ≥ t. For this equality to hold for all bounded random variables B
t
in the date t
information set, E(M
τ
⎪F
t
) = M
t
. This verifies that M is a martingale relative to {F
t
: t ≥ 0}.
6490_Book.indb 421
11/4/14 2:30 PM