430
The Nobel Prizes
deduce bounds on the expected utility consequences of alternative decisions,
and more generally a mapping from alternative priors into alternative expected
outcomes. Building on discussions in Walley (1991) and Berger (1994), there
are multiple reasons to consider a family of priors. This family could represent
the views of alternative members of an audience, but they could also capture
the ambiguity to a single decision maker struggling with which prior should be
used. Ambiguity aversion as conceived by Gilboa and Schmeidler (1989) and
others confronts this latter situation by minimizing the expected utility for each
alternative decision rule. Max-min utility gives a higher rank to a decision rule
with the larger expected utiltiy outcome of this minimization.
38
Max-min utility has an extension whereby the minimization over a set of
priors is replaced by a minimization over priors subject to penalization. The
penalization limits the scope of the prior sensitivity analysis. The penalty is mea-
sured relative to a benchmark prior used as a point of reference. A discrepancy
measure for probability distributions, for instance some of the ones I discussed
previously, enforce the penalization. See Maccheroni et al. (2006) for a general
analysis and Hansen and Sargent (2007) for implications using the relative en-
tropy measure that I already mentioned. Their approach leads to what is called
variational preferences.
For either form of ambiguity aversion, with some additional regularity con-
ditions, a version of the Min-Max Theorem rationalizes a worst-case prior. The
chosen decision rule under ambiguity aversion is also the optimal decision rule
if this worst-case prior were instead the single prior of the decision maker. Dy-
namic counterparts to this approach do indeed imply a martingale distortion
when compared to a benchmark prior that is among the set of priors that are en-
tertained by a decision maker. Given a benchmark prior and a dynamic formu-
lation, this worst-case outcome implies a positive martingale distortion of the
type that I featured in Section 5. In equilibrium valuation, this positive martin-
gale represents the consequences of ambiguity aversion on the part of investors
inside the model. This martingale distortion emerges endogenously as a way to
confront multiple priors that is ambiguity averse or robust. In sufficiently simple
environments, the decision maker may in effect learn the model that generates
the data in which case the martingale may converge to unity.
There is an alternative promising approach to ambiguity aversion. A deci-
sion theoretic model that captures this aversion can be embedded in the analy-
sis of Segal (1990) and Davis and Pate-Cornell (1994), but the application to
38
See Epstein and Schneider (2003) for a dynamic extension that preserves a recursive
structure to decision making.
6490_Book.indb 430
11/4/14 2:30 PM
Uncertainty Outside and Inside Economic Models 431
ambiguity aversion has been developed more fully in Klibanoff et al. (2005)
and elsewhere. It is known as a smooth ambiguity model of decision making.
Roughly speaking, distinct preference parameters dictate behavior responses
to two different sources of uncertainty. In addition to aversion to risk given a
model captured by one concave function, there is a distinct utility adjustment
for ambiguity aversion that emerges when weighting alternative models using
a Bayesian prior. While this approach does not in general imply a martingale
distortion for valuation, as we note in Hansen and Sargent (2007), such a distor-
tion will emerge with an exponential ambiguity adjustment. This exponential
adjustment can be motivated in two ways, either as a penalization over a fam-
ily of priors as in variational preferences or as a smooth ambiguity behavioral
response to a single prior.
6.2 unknown Models and ambiguity aversion
I now consider an approach with an even more direct link to the analysis in
Section 5. An important initiator of statistical decision theory, Wald (1939), ex-
plored methods that did not presume a priori weights could be assigned across
models. Wald (1939)’s initial work generated rather substantial literatures in sta-
tistics, control theory and economics. I am interested in such an approach as a
structured way to perform an analysis of robustness. The alternative models rep-
resented as martingales may be viewed as ways in which the benchmark prob-
ability model can be misspecified. To explore robustness, I start with a family of
probability models represented as martingales against a benchmark model. Dis-
crepancy measures are most conveniently expressed in terms of convex func-
tions of the martingales as in Section 5. Formally the ambiguity is over models,
or potential misspecifications of a benchmark model.
What about learning? Suppose that the family of positive martingales with
unit expectations is a convex set. For any such martingale M in this set and some
0 < ω < 1, construct the mixture ωM + (1 – ω) is a positive martingale with unit
expedations. Notice that
ωM
t
+
τ
+(1−
ω)1
ω
M
t
+(1−
ω)1
=
ωM
t
M
t
+
τ
M
t
⎛
⎝⎜
⎞
⎠⎟
+(1−
ω)1
ωM
t
+(1−
ω)1
.
The left-hand side is used to represent the conditional expectations operator
between dates t + τ and t. If we interpret ω as the prior assigned to model M and
(1 – ω) as the prior assigned to a benchmark model, then the right-hand side
6490_Book.indb 431
11/4/14 2:30 PM