424
The Nobel Prizes
bound. If the minimized value is zero, the probability distortion vanishes and
investors eventually settle on the benchmark model as being correct.
A straightforward derivation shows that even when we change the roles of
the benchmark model and the alternative model, the counterpart to κ(M) re-
mains the same.
31
Why is Chernoff entropy interesting? When this common
decay rate is small, even long histories of data are not very informative about
model differences.
32
Elsewhere I have explored the connection between this
Chernoff measure and Sharpe ratios commonly used in empirical finance, see
Anderson et al. (2003) and Hansen (2007a).
33
The Chernoff calculations are of-
ten straightforward when both models (the benchmark and perturbed models)
are Markovian. In general, however, it can be a challenge to use this measure
in practice without imposing considerable a priori structure on the alternative
models.
In what follows, I will explore discrepancy
measures that are similar to
this Chernoff measure but are arguably more tractable to implement. What I
describe builds directly on my discussion of GMM methods and extensions.
Armed with factorization (13), approaches that I suggested for the study of SDFs
can be adapted to the study of belief distortions. I elaborate in the discussion
that follows.
5.3 ignored belief distortions
Let me return to GMM estimation and model misspecification. Recall that the
justification for GMM estimation is typically deduced under the premise that
the underlying model is correctly specified. The possibility of permanent belief
distortions, say distortions for which κ(M) > 0, add structure to the model mis-
specification. But this is not enough structure to identify fully the belief distor-
tion unless an econometrician uses sufficient asset payoffs and prices to reveal
the SDF. Producing bounds with this extra structure can still proceed along the
lines of those discussed in Section 3.2.3 with some modifications. I sketch below
one such approach.
31
With this symmetry and other convenient properties of κ(M), we can interpret the
measure as a metric over (equivalence classes of ) martingales.
32
Bayesian and max-min decision theory for model selection both equate decay rates in
type I and type II error rates.
33
The link is most evident when a one-period (in discrete time) or local (in continuous
time) measure of statistical discrimination is used in conjunction with a conditional nor-
mal distribution, instead of the large t measure described here.
6490_Book.indb 424
11/4/14 2:30 PM
Uncertainty Outside and Inside Economic Models 425
Suppose the investors in the model are allowed to have distorted beliefs, and
part of the estimation is to deduce the magnitude of the distortions. How big
would these distortions need to be in a statistical sense in order to satisfy the
pricing restrictions? What follows makes some progress in addressing this ques-
tion. To elaborate, consider again the basic pricing relation with distorted beliefs
written as unconditional expectation:
E
M
t
+
S
t
+
M
t
S
t
⎛
⎝⎜
⎞
⎠⎟
(Y
t
+
′)Z
t
−(Q
t
′)Z
t
⎡
⎣
⎢
⎤
⎦
⎥ = 0.
As with our discussion of the study of SDFs without parametric restrictions,
we allow for a multiplicity of possible martingales
and impose bounds on expec-
tations of convex functions of the ratio
M
t
+
M
t
.
To deduce restrictions on M, for notational simplicity I drop the t subscripts
and write the pricing relation as:
E(m
s
′
y
− ′
q )
= 0
E(
m
−1) = 0.
(17)
To bound properties of m solve
inf
m
>0
E[
φ(m)] (18)
subject to (17) where ϕ is given by equation (5). This formulation nests many
of the so-called F-divergence measures for probability distributions including
the well known Kullback-Leibler divergence (θ = –1, 0). A Chernoff-type mea-
sure can be imputed by computing the bound for –1 < θ < 0 and optimizing
after an appropriate rescaling of the objective by θ(1 + θ). As in the previous
analysis of Section 3.2.3, there may be many solutions to the equations given in
(17). While the minimization problem selects one of these, I am interested in
this optimization problem to see how small the objective can be in a statistical
sense. If the infimum of the objective is small, then statistically small changes in
distributions suffice to satisfy the pricing restrictions. Such departures allow for
“behavior biases” that are close statistically to the benchmark probabilities used
in generating the data.
I have just sketched an unconditional approach to this calculation by al-
lowing conditioning information to be used through the “back door” with
the specification of Z but representing the objective and constraints in terms
6490_Book.indb 425
11/4/14 2:30 PM