410
The Nobel Prizes
for what alternative? Thus a complementary approach is to characterize proper-
ties of the family of s’s that do satisfy (4). These solutions might well violate the
parametric restriction.
The interesting challenge is how to characterize the family of SDFs that solve
(4) in useful ways. Here I follow a general approach that is essentially the same
as that in Almeida and Garcia (2013). I choose this approach both because of its
flexibilty and because it includes many interesting special cases used in empiri-
cal analysis. Consider a family of convex functions ϕ defined on the positive real
numbers:
18
φ(
r) =
1
θ(1+θ)
(
r)
1
+
θ
−1
⎡⎣
⎤⎦ (5)
for alternative choices of the parameter θ. The specification θ = 1 is commonly
used in empirical practice, in which case ϕ is quadratic. We shall look for lower
bounds on the
E
φ
s
Es
⎛
⎝⎜
⎞
⎠⎟
⎡
⎣⎢
⎤
⎦⎥
by solving the convex optimization problem:
19
λ = inf
s
>0
E
φ
s
Es
⎛
⎝⎜
⎞
⎠⎟
⎡
⎣⎢
⎤
⎦⎥
subject to E s
′
y
− ′
q
[
]
= 0. (6)
By design we know that
E
φ
s
Es
⎛
⎝⎜
⎞
⎠⎟
⎡
⎣⎢
⎤
⎦⎥
≥
λ.
Notice that
E φ
s
Es
⎛
⎝⎜
⎞
⎠⎟
⎡
⎣⎢
⎤
⎦⎥
hence λ are nonnegative by Jensen’s Inequality be-
cause ϕ is convex and ϕ (1) = 0. When θ = 1,
18
This functional form is familiar from economists’ use of power utility (in which case
we use –ϕ to obtain a concave function), from statisticians’ use of F-divergence measures
between two probability densities, the Box-Cox transformation, and the applications in
the work of Cressie and Read (1984).
19
Notice that the expectation is also an affine transformation of the moment generating
function for log s.
6490_Book.indb 410
11/4/14 2:30 PM
Uncertainty Outside and Inside Economic Models 411
2E
φ
s
Es
⎛
⎝⎜
⎞
⎠⎟
⎡
⎣⎢
⎤
⎦⎥
is the ratio of the standard deviation of s to its mean and 2λ is the greatest
lower bound on this ratio.
From the work of Ross (1978) and Harrison and Kreps (1979), arbitrage
considerations imply the economically interesting restriction s > 0 with prob-
ability one. To guarantee a solution to optimization problem (6), however, it is
sometimes convenient to include s’s that are zero with positive probability. Since
the aim is to produce bounds, this augmentation can be justified for mathemati-
cal and computational convenience. Although this problem optimizes over an
infinite-dimensional family of random variables s, the dual problem that opti-
mizes over the Lagrange multipliers associated with the pricing constraint (4) is
often quite tractable. See Hansen et al. (1995) for further discussion.
Inputs into this calculation are contained in the pair (y, q) and a hypothetical
mean Es. If we have time series data on the price of a unit payoff at date t + l,
Es can be inferred by averaging the date t prices over time. If not, by changing
Es we can trace out a frontier of solutions. An initial example of this is found in
Hansen and Jagannathan (1991) where we constructed mean-standard devia-
tion tradeoffs for SDFs by setting θ = 1.
20,21
While a quadratic specification of ϕ (θ = 1)
has been the most common one
used in empirical practice, other approaches have been suggested. For instance,
Snow (1991) considers larger moments by setting θ to integer values greater
than one. Alternatively, setting θ = 0 yields
E
φ
s
Es
⎛
⎝⎜
⎞
⎠⎟
⎡
⎣⎢
⎤
⎦⎥
=
E s log
s
− log Es
(
)
⎡⎣
⎤⎦
Es
,
20
This literature was initiated by a discussion in Shiller (1982) and my comment on that
discussion in Hansen (1982a). Shiller argued why a volatility bound on the SDF is of
interest, and he constructed an initial bound. In my comment, I showed how to sharpen
the volatility bound, but without exploiting that s > 0. Neither Shiller nor I explored
mean-standard deviation tradeoffs that are central in Hansen and Jagannathan (1991). In
effect, I constructed one point on the frontier characterized in Hansen and Jagannathan
(1991).
21
When θ is one, the function φ continues to be well defined and convex for negative
real numbers. As noted in Hansen and Jagannathan (1991), if the negative choices of s
are allowed in the optimization problem (which weakens the bound), there are quasi-
analytical formulas for the minimization problems with simple links to Sharpe ratios
commonly used in empirical finance.
6490_Book.indb 411
11/4/14 2:30 PM