Вы находитесь на странице: 1из 22

Estimating Long Run Risk: A Panel Data Approach

Peng Wang
+
Hong Kong University of Science and Technology
December, 2010
Abstract
The long-run risk (LLR) model of Bansal and Yaron (2004) provides a building
block for the resolution of the equity-premium puzzle, among many other puzzles.
However, in their calibration exercise, the magnitude of this long-run risk of consump-
tion is rather small and dicult to detect using only several macro data series. It is
hard for traditional small-scale models to estimate this component due to the very
small signal-noise ratio. For the LLR argument to work in practice, the representa-
tive consumer must be able to make inference about the long-run risk based on the
information set available. The original paper of Bansal and Yaron (2004) assumes that
the representative agent observes the long-run risk, and then the agents consumption
plan becomes a function of the underlying long-run risk. This paper assumes that the
agent is endowed with a rich information set of dividend-price ratios, from which the
consumer may learn the long-run risk. Then I propose a panel data approach such that
the small risk component of consumption can be accurately estimated, and thus lends
support for the presence of the long-run risk. Based on the factor representation for
cross-sectional asset price-dividend ratios implied by the theoretical model, a two-stage
factor augmented maximum likelihood estimator will provide more robust estimation
results than existing small-scale estimation methods. In the rst stage, the principal
components estimators oer precise estimates of the space spanned by the state vari-
ables, namely the long-run component of consumption and the stochastic volatility.
In the second stage, maximum likelihood estimation is carried out to recover the true
state variables and the structural parameters. The method used in this paper has
far-reaching implications for a more general class of the dynamic general equilibrium
models, where the estimation of the state variables is of direct relevance.

Department of Economics, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon,
Hong Kong. Tel: +852-23587630. Fax: +852-23582084. Email: pwang@ust.hk
1
1 Introduction
This paper concerns an application of large dimensional factor models to the analysis of
dynamic stochastic economic models. Specically I consider the analysis of structural dy-
namic factor models where both the cross-sectional dimension and the time dimension tend
to innity, which is called the large-(`. 1) framework. The economic model I consider is
the long-run risk (LLR) model of Bansal and Yaron (2004). The LLR model is of particular
interest to motivate the use of large dimensional factor models in estimating economically
important time series which is statistically dicult to measure using aggregate data.
In the LLR model, all economic variables of interest are functions of the long-run risk of
consumption growth.
1
This component is very small in magnitude but is crucial in resolving
the equity premium puzzle. It is hard for traditional small-scale models to estimate this
component due to the very small signal-noise ratio. However, modern factor models provide
a way to consistently estimate this long-run component. The asymptotic property of the
estimators is built upon double asymptotics of large (`. 1) . in which both the cross-section
dimension and the time dimension tend to innity. Such a methodology allows one to make
inference about the long-run component itself besides its parametric dynamic process. The
estimators in this paper will be structural in the sense that the factor structure is implied
by the model. Specically, the logarithm of cross-sectional asset price-dividend ratios is
approximately a linear function of the long run component, which is the state variable of
the economic model. If we treat the price-dividend ratios as measured with i.i.d. errors,
principal components estimators of the long run component oer consistent estimators for
the state variables, which further help the estimation of deep parameters.
There is a large body of work that has studied the implications of large factor models
to economic analysis. Bernanke, et al (2004) proposed a factor-augmented VAR model to
overcome the limited-information problem in the conventional small VAR models and was
able to resolve the price puzzle. Boivin and Giannoni (2006) studied the DSGE models in
a data rich environment, using the implied factors to augment a state-space representation of
the DSGE model to overcome the measurement errors problem. Other applications of factor
models include Ng and Ludvigson (2007, 2009), who studied the risk-return tradeo and the
bond risk premia, Ng and Moench (2010), who studied the housing market dynamics, Mum-
taz and Surico (2009), who studied the transmission of international shocks, etc. My work
diers from all the above in that I am considering a factor-augmented maximum likelihood
approach for a general likelihood function, not necessarily for a VAR or linear state-space
1
This is indeed a common feature of the DSGE models where linearized or log-linearized solutions are
available.
2
system. A two-step method will be shown to be parsimonious while eective.
2 Modeling the consumption growth
Centered around Bansal and Yaron (2004)s model setup are two key assumptions. Firstly,
the consumption growth contains a small stationary component, which is highly persistent.
They named this component the long-run risk of consumption growth. Secondly, the economy
is characterized by a representative agent, who is endowed with an Epstein-Zin recursive
preference. With such a preference, the agent cares about the consumption growth risk
far into the future in contrast to the case with additive utility functions. Although the
consumption growth process is statistically indistinguishable from an i.i.d. process, the role
of the predictable long run component is amplied by the Epstein-Zin preference, and the
feedback is shown in asset prices.
Let C
t
be the consumption at time t. and c
t
= log (C
t
) log (C
t1
) be a measure for
the consumption growth, which is modeled as follows:
c
t+1
= j
c
+r
t
+o
t
j
t+1
r
t+1
= jr
t
+,
c
o
t
c
t+1
o
2
t+1
o
2
= (o
2
t
o
2
) +o
&
n
t+1
where r
t
is a small but persistent component capturing the long-run risk in the consumption
growth, and o
2
t
is the conditional volatility of the consumption growth.
2.1 Representative agents problem
The representative agent has the Epstein-Zin recursive preference, characterized by
\
t
= [(1 o)C
1

t
+o(1
t
[\
1
t+1
])
1

]

1
.
where is the coecient of relative risk aversion, o =
1
11
. and is the intertemporal
elasticity of substitution (IES). Time t budget constraint is given by
\
t+1
= (\
t
C
t
)1
c,t+1
.
where 1
c,t+1
=
1
t+1
+C
t+1
1t
is the gross return on consumption claim.
3
The agents problem at time t is to maximize \
t
subject to the time t budget constraint,
which results in the following Euler equation
1
t
(`
t+1
1
),t+1
) = 1. or
1
t
(exp(:
t+1
+:
),t+1
)) = 1
with :
t
= log(`
t
) and :
),t
= log(1
),t
). The log version of intertemporal marginal rate of
substitution (IMRS) is given by
:
t+1
= o log o
o

c
t+1
+ (o 1):
c,t+1
2.2 Estimating the LLR: from the agents Prospective
The relevant state variables of this economy are r
t
and o
2
t
. For any dividend paying asset ,.
the log of price-dividend ratio is a linear function of state variables.
.
),t
=
0,)
+
1,)
r
t
+
2,)
o
2
t
.
where
i,)
is function of model parameters (Bansal and Yaron, 2004, and BKY, 2007).
In Bansal and Yaron (2004)s economy, the long-run component r
t
and the stochastic
volatility o
2
t
are both in the agents information set. The assumption that the agent observes
not only the consumption growth but also r
t
and o
2
t
seems a rather strong one because the
latter two series are statistically insignicant. The rational expectations assumption equates
the agents belief of the hidden process to that of the objective process. I will show that if
the agent can fully utilize the information from a large cross section of observed asset prices,
r
t
and o
2
t
can be treated as known. To x idea, I start from the agents prospective and use
a simple setup where the parametrized dynamic process of consumption growth is known to
the agent, however, the historical time series of r
t
and o
2
t
are unobserved. Thus the agent
must draw inference and make predictions of r
t
and o
2
t
based on observables.
I will compare two scenarios categorized by the agents information set. The small in-
formation scenario assumes that the agent uses only the information from aggregate series,
namely, the consumption growth and the asset price index, to estimate the hidden states.
The large information scenario assumes that the agent incorporate information from the
disaggregate asset prices when estimating the hidden states. I will show that the latter dis-
aggregate approach is superior to the former aggregate method in the sense that the hidden
states can be precisely estimated.
4
The consumption process is calibrated as in Bansal and Yaron (2004).
c
t+1
= j
c
+r
t
+o
t
j
t+1
r
t+1
= jr
t
+,
c
o
t
c
t+1
o
2
t+1
o
2
= (o
2
t
o
2
) +o
&
n
t+1
The error terms j
t
. c
t
. and n
t
are i.i.d. `(0. 1). The following table summarizes parameter
values.
Parameters j
c
,
c
o o
&
0.0015 0.979 0.044 0.0078 0.987 0.23 10
5
I simulate the time series c
t
. r
t
. o
2
t
according to the calibrated model. Then I generate
a panel of logarithm of dividend-price ratios according to
.
),t
=
0,)
+
1,)
r
t
+
2,)
o
2
t
+ on
),t
. , = 1. .... 100. t = 1. .... 100
where ` = 100. 1 = 100.
i,)
are i.i.d. draws from l(0. 1). and n
),t
is i.i.d. `(0. 1) serving
the role as the measurement error. The aggregate index of log dividend-price ratio .
0,t
is
formed as the sum of the cross-sectional average of .
),t
and an i.i.d. `(0. o
2
) measurement
error:
.
0,t
=
1
`
.

)=1
.
),t
=

0
+

1
r
t
+

2
o
2
t
+ n
t
+ on
0,t
j

0
+

1
r
t
+

2
o
2
t
+ on
0,t
.
where the last part comes from the fact that n
t
j
0 as ` .
2.2.1 Estimating the hidden state variables
Given the available information, an agent will look at the state space representation of the
model, and optimally update his prediction about the hidden state variables using some
ltering technique. A very unique feature of the above model is the low signal-noise ratio.
The conditional variances of c
t
and .
),t
are much larger than that of r
t
and o
2
t
. It is thus very
dicult to eectively detect the process of hidden states using only two observed aggregate
series. In contrast, in spite of the high noise, the disaggregate approach can eectively detect
the hidden state. To focus on the main idea, I rst use a quasi-maximum likelihood method,
treating o
t
as a constant although a time-varying o
t
is used to generate the data. In this way,
the agent is able to form a linear state-space system and use the Kalman lter to update his
5
belief about the hidden long-run component r
t
.
2
The following graph compares the estimates
from two information scenarios with the true long-run component.
0 20 40 60 80 100
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
x 10
-3
LLR component of consumption growth
x
x1(small info)
x2(large info)
The following graph shows the relative magnitude of hidden states with respect to the con-
sumption growth.
0 20 40 60 80 100
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
0.015
0.02
0.025
Consumpti on growth
c
t

c
+x
t

t
2
In general, the state-space system is of a nonlinear form due to the presence of
t
: Instead of using a
computationally intensive nonlinear ltering approach, I will consider a two-step factor-augmented maximum
likelihood approach later.
6
Given the small magnitude of the long-run component, the conditional variance of the ob-
served consumption growth is largely due to the error term. It comes at no surprise that
small information approach yields unsatisfactory results. However, the large information ap-
proach is able to accumulate the signal from the disaggregate data and provides consistent
estimation of the small long-run component. Although the agent is assumed to have the
false belief that o
t
is a constant, the disaggregate approach is able to precisely estimate the
entire time series of r
t
. This is due to the fact that o
2
t
is a slow moving series, the variance
of which is negligible compared to that of r
t
and c
t
.
2.2.2 Monte Carlo experiments
In this section, I conduct some Monte Carlo experiments to demonstrate how the size of
information aects the estimation of hidden state variables. For the same data generating
process, I will consider dierent combinations of ` and 1. I will also consider dierent
DGPs for the coecients
i,)
in the log dividend-price ratio equation. The following table
summarizes the Monte Carlo average of mean squared errors of ^ r for each model design
for ` = 200 repetitions (Monte Carlo standard errors are in parenthesis): `o1 (^ r) =
1
AT

A
n=1

T
t=1
_
^ r
(n)
t
r
(n)
t
_
2
. All values are reported as units of 10
5
.

i,)
~ `(0. 1)
i,)
~ l[0. 1]
1 ` Small Info. Large Info. Small Info. Large Info.
50 5 0.3249(0.4118) 0.0891(0.0516) 0.3765(0.4073) 0.1140(0.0694)
10 0.3071(0.3849) 0.0657(0.0353) 0.4372(0.5409) 0.0990(0.0629)
50 0.1993(0.1834) 0.0316(0.0124) 0.4106(0.4776) 0.0561(0.0267)
100 0.1805(0.1726) 0.0212(0.0065) 0.4495(0.5318) 0.0397(0.0160)
200 0.1810(0.1614) 0.0141(0.0036) 0.4221(0.4943) 0.0264(0.0085)
100 5 0.3813(0.5106) 0.0917(0.0411) 0.4857(0.5282) 0.1135(0.0532)
10 0.3202(0.3108) 0.0688(0.0306) 0.5127(0.6447) 0.1020(0.0558)
50 0.2466(0.2204) 0.0325(0.0091) 0.5343(0.5896) 0.0547(0.0159)
100 0.2196(0.1550) 0.0217(0.0058) 0.4900(0.6127) 0.0396(0.0114)
200 0.2140(0.1504) 0.0137(0.0029) 0.4347(0.4599) 0.0276(0.0068)
200 5 0.4605(0.5367) 0.0917(0.0357) 0.4979(0.5018) 0.1171(0.0448)
10 0.3183(0.3523) 0.0697(0.0228) 0.5520(0.6837) 0.1051(0.0394)
50 0.2681(0.1714) 0.0317(0.0075) 0.5392(0.5827) 0.0543(0.0131)
100 0.2308(0.1143) 0.0212(0.0039) 0.5279(0.5549) 0.0398(0.0090)
200 0.2486(0.1506) 0.0140(0.0022) 0.4775(0.4726) 0.0265(0.0045)
7
To summarize, it comes at no surprise that the large information approach uniformly
outperform the small information method. I will focus on the role of the size of information
on the estimation precision using the large information approach. For ` xed, increasing
the time dimension 1 seems not to be necessarily accompanied by a decline in the mean
squared errors. On the other hand, for a given 1. a larger cross-sectional dimension always
helps the estimation of the hidden long-run component. Such properties hold for both DGPs
for the loadings coecients
i,)
.
3 Estimation: from the econometricians prospective
The previous exercise focuses on the state estimation from the representative agents prospec-
tive, where all the model parameters are treated as known. Such an estimation practice can
be viewed as the learning process of the agent, who update his belief about the hidden state
variables using available information set.
The following part will look at the state estimation problem from an econometricians
prospective, where all the model parameters have to be estimated. And the parameter
estimates will be used to form the estimates for the hidden states.
3.1 GMM/SMM estimation
Exploring moment conditions implied by the Euler equation, a simulated method of moment
(SMM) procedure based on the GMM principle is proposed by Bansal, Kiku, and Yaron
(BKY, 2007). The Euler equation for an asset with return :
),t
is given by
1
t
exp [:
t+1
+:
),t+1
] = 1.
where the logarithm of IMRS is given by
:
t+1
= o log o
o

c
t+1
+ (o 1) :
c,t+1
.
Crucial to the construction of moment conditions is to construct estimates for the consump-
tion return process :
c,t+1
.
Using a log-linear approximation for log(
1
t+1
C
t+1
+1). BKY approximate :
c,t+1
= log(
1
t+1
C
t+1
+
1) + c
t+1
.
t
by
:
c,t+1
- /
0
+/
1
.
t+1
+ c
t+1
.
t
where .
t
= log(1
t
,C
t
) is the price-dividend ratio for consumption claim, /
1
=
exp( :)
1+exp( :)
.
8
/
0
= log(1 + exp( .)) /
1
.. with . = 1(.
t
).
The log of price-dividend ratio is a linear function of state variables,
.
t
=
0,c
+
1,c
r
t
+
2,c
o
2
t
.
where
i,c
is function of model parameters. Let = [
0,c

1,c

2,c
] = ( .). 1
t
= [1 r
t
o
2
t
]
0
.
then . is solved as the following xed-point problem
. = ( .)

1 . where

1 = 1(1
t
).
To close the construction of :
c,t+1
. one needs to recover the time series of state variables
r
t
and o
2
t
and thus to recover the time series of .
t
. BKY (2007) applied a two-step method.
In the rst step, r
t
is identied by regressing the consumption growth on the risk-free rate
and the market price-dividend ratio,
c
t+1
= /
0
a
1
t
+o
t
j
t+1
. with 1
t
= [1. .
n,t
. :
),t
]
0
.
In the second step, the conditional volatility o
2
t
is obtained by regressing the squared con-
sumption residual on the same set of observables,
o
2
t
= 1
t
_
(c
t+1
/
0
a
1
t
)
2
_
= /
0
o
1
t
Once the construction of the consumption return process is done, we are able to formulate
the pricing kernel :
t+1
for the GMM estimation. Due to the lack of a closed-form solution,
computationally intensive simulation method such as the simulated method of moments is
implemented in BKY (2007) to estimate the model.
The method proposed by this paper avoids the estimation of the consumption return
process, which requires numerically solving a xed-point problem. Instead of looking at
model restrictions on asset returns, I emphasize the importance of the linear approaximate
solution of log dividend-price ratios.
3.2 Estimate state variables using cross-section P/D data
Let .
),t
=
0,)
+
1,)
r
t
+
2,)
o
2
t
+ c
),t
. where c
),t
is the unobserved idiosyncratic error (or
the measurement error) of asset ,s log price-dividend ratio. Then the model implies the
following factor representation
2
t
= j
Z
+ 1
t
+1
t
. or 2 = j
Z
1
0
+ 1 +1
9
with 2
.T
being an (possibly unbalanced) panel of asset price-dividend ratios, j
Z
being the
time-invariant individual xed eect,
.2
being the factor loadings, 1
t
= [r
t
. o
2
t
]
0
being the
unobserved factors. Also we can include the risk-free rate in 2
t
.
The rst two leading eigenvectors
^
1 of the matrix
(Zj
Z
1
0
)
0
(Zj
Z
1
0
)
.T
oer a consistent
estimator for the column space of true 1.
3
Then we may denote the estimator for the state
vector as,
_
^ r
t
^ o
2
t
_
= 1
^
1
t
. or
_
^ r
t
= 1
11

^
1
1t
+1
12

^
1
2t
^ o
2
t
= 1
21

^
1
1t
+1
22

^
1
2t
.
for some rotation matrix 1. We still need an estimator for the rotation matrix. Here we
exploit the dynamic structure of consumption growth. Rewrite the demeaned consumption
growth process as
c
t+1
= 1
11
1
1t
+1
12
1
2t
+o
t
j
t+1
.
The above equation is a member of a larger system, namely,
_
c
t+1
2
t
_
=
_
1
11
. 1
12

__
1
1t
1
2t
_
+
_
o
t
j
t+1
1
t
_
.
To estimate the predictive component of c
t+1
. one may apply the principal components
estimator for the above system of equations to obtain a consistent estimator ^ r
t
=
^
1
11

^
1
1t
+
^
1
12

^
1
2t
.
4
where ^ r
t
is the estimated long-run component of consumption growth.
Alternatively, on may estimate
^
1
t
by using only 2
t
and regress c
t+1
on
^
1
1t
and
^
1
2t
to
obtain ^ r
t
=
^
1
11

^
1
1t
+
^
1
12

^
1
2t
. The alternative strategy works because 1
_
o
t
j
t+1
[1
_
=
1
_
1
_
o
t
j
t+1
[1. 2. o
_
[1
_
= 1
_
o
t
j
t+1
[2. o
_
= 0 and
1
T

T
t=1
_
_
_
^
1
t
H1
t
_
_
_
2
= o
j
(1). The
large sample properties of ^ r
t
from both approaches are the same.
To estimate the conditional volatility o
2
t
. BKY (2007) explored the conditional expecta-
tion representation of o
2
t
: o
2
t
= 1
t
[(c
t+1
r
t
)
2
]. The projection of (c
t+1
^ r
t
)
2
onto the
time-t information set, e.g. ^ r
c

t
c=1
provides a possible way to recover the process
_
^ o
2
t
_
. My
3
Bai and Ng (2002) showed that when (N; T) ; consistency of the factors can be established in the
sense that
1
T

T
t=1
_
_
_
^
F
t
HF
t
_
_
_
2
= o
p
(1) for some deterministic matrix H: Bai (2003) further showed that
_
N
_
^
F
t
HF
t
_

d
N(0; V
t
) under very general conditions on the model.
4
Although in the factor model, x
it
=
0
i
f
t
+e
it
;the factors f
t
and factor loadings
i
cannot be separately
identied without further restrictions, the common component c
it
=
0
i
f
t
is consistently estimated by ^ c
it
=
^

0
i
^
f
t
, see Bai and Ng (2002) and Bai (2003).
10
estimation strategy for o
2
t
diers from BKY in two ways. First, I will explore the conditional
likelihood function parametrized by 1
21
and 1
22
. Second, I will augment the likelihood func-
tion by the principal components estimators
^
1. Such a strategy is able to obtain a consistent
estimator for o
2
t
given the consistency of the MLE estimators
^
1
21
and
^
1
22
.
Explore the assumption that conditional on r
t
and o
2
. c
t+1
is independent normal. We
may formulate the conditional likelihood
,(c[r. o
2
) =
T

t=1
1
_
2:o
t
exp[0.5(
c
t+1
r
t
o
t
)
2
]. or
log ,(c[r. o
2
) = 0.5
T

t=1
(
c
t+1
r
t
o
t
)
2
0.5

log(o
2
t
) 0.51 log(2:).
Replacing r
t
and o
2
t
by their principal component estimators (parametrized by the rotation
matrix 1), one obtains an approximate likelihood function,
1(1
21
. 1
22
) = 0.5
T

t=1
(
c
t+1
^ r
t
^ o
t
)
2
0.5
T

t=1
log(^ o
2
t
) 0.51 log(2:)
= 0.5
T

t=1
(c
t+1
^ r
t
)
2
1
21

^
1
1t
+1
22

^
1
2t
0.5
T

t=1
log(1
21

^
1
1t
+1
22

^
1
2t
) 0.51 log(2:)
Maximize the above objective function to obtain consistent estimators for 1
21
. 1
22
. Then we
can construct the time series of state variables:
_
^ r
t
^ o
2
t
_
=
^
1
^
1
t
A natural extension of the above method is to add the dynamics of o
2
t
to the above
likelihood. Instead of studying the likelihood for c
t+1
conditioning on r
t
and o
2
t
. we may
11
formulate the joint likelihood of c
t+1
. r
t
. and o
2
t
:
,(c. o
2
. r)
= ,(c[o
2
. r) ,(r[o
2
) ,(o
2
)
=
T

t=1
,(c
t+1
[o
2
t
. r
t
) ,(r
t
[r
t1
. o
2
t1
) ,(o
2
t
[o
2
t1
)
=
T

t=1
1
_
2:o
t
exp[0.5(
c
t+1
r
t
o
t
)
2
]
1
_
2:,
c
o
t1
exp[0.5(
r
t
jr
t1
,
c
o
t1
)
2
]

1
_
2:o
&
exp[0.5(
o
2
t
o
2
t1
+ (1 ) o
2
o
&
)
2
]
-
T

t=1
1
_
2:^ o
t
exp[0.5(
c
t+1
^ r
t
^ o
t
)
2
]
1
_
2:,
c
^ o
t1
exp[0.5(
^ r
t
j^ r
t1
,
c
^ o
t1
)
2
]

1
_
2:o
&
exp[0.5(
^ o
2
t
^ o
2
t1
+ (1 ) o
2
o
&
)
2
].
In this way, we are able to jointly estimate 1
21
. 1
22
. o
&
. . o. ,
c
. j .
An alternative computationally simple approach is to explore the orthogonality condition
implied by the model. First notice that
1
_
o
2
t
j
2
t+1
[o
2
t
. r
t
_
= o
2
t
1
_
j
2
t+1
[o
2
t
. r
t
_
= o
2
t
.
More over, let 1 be the 1 2 factor matrix which represent the space spanned by r and o
2
.
then
o
2
t
= 1
_
o
2
t
j
2
t+1
[o
2
t
. r
t
_
= 1
_
o
2
t
j
2
t+1
[1
t
_
.
This implies that the projection of o
2
t
j
2
t+1
onto 1
t
yields o
2
t
. In practice, one may regress
^ o
2
t
^ j
2
t+1
on
^
1
t
and use the tted value as an estimator of o
2
t
.
3.2.1 Simulation studies
I will simulate the model using the same parameters as previous sections. From an econo-
metricians prospective, the consumption growth and the asset dividend-price ratios are the
observables. Both the state variables r
t
. o
2
t
and the model parameters are unknown and
will be jointly estimated using the optimality conditions implied by the model.
I will focus on the two step approach, namely, obtain the principal component estimators
for the space spanned by the hidden states and then use the likelihood method to estimate
the model parameters.
12
The following graph shows a typical estimation result from simulated data.
0 20 40 60 80 100
-2
-1
0
1
2
3
4
5
x 10
-3
LLR component of consumption growth
x
x1(small info)
x2(large info)
x3(pca)
The series r
1
and r
2
are estimated from the agents prospective, using small and large
information respectively. The series r
3
is the econometricians two-step estimates, treating
all model parameters as unknown. As demonstrated by this graph, both r
2
and r
3
are very
close to the true series r. while the small information approach fails to obtain reasonable
estimates. The results of two-step method is very close to that of the large information
method, in spite of the fact that the former estimates model parameters and the states
together. The simple regression approach is used to recover the time series of the conditional
variance. As shown in the following graph, the estimates are reasonably close to the true
series. A more detailed Monte Carlo study will be conducted to examine the properties of
13
the estimator.
0 20 40 60 80 100
5
6
7
8
9
10
11
x 10
-5
Conditional Variance

2
hat
To further compare the econometricians two-step method and the agents large informa-
tion method. I further conduct a Monte Carlo experiment for dierent model sizes. For the
same data generating process, I consider dierent combinations of ` and 1. I will also con-
sider dierent DGPs for the coecients
i,)
in the log dividend-price ratio equation. Again,
the following table summarizes the Monte Carlo average of mean squared errors of ^ r for
each model design for ` = 200 repetitions (Monte Carlo standard errors are in parenthesis):
14
`o1 (^ r) =
1
AT

A
n=1

T
t=1
_
^ r
(n)
t
r
(n)
t
_
2
. All values are reported as units of 10
5
.

i,)
~ `(0. 1)
i,)
~ l[0. 1]
1 ` Agent Econometrician Agent Econometrician
50 5 0.0930(0.0532) 0.3726(0.2553) 0.1119(0.0799) 0.3708(0.2689)
10 0.0750(0.0427) 0.3925(0.2912) 0.0959(0.0595) 0.3875(0.2826)
50 0.0328(0.0123) 0.3188(0.2617) 0.0533(0.0214) 0.3853(0.2787)
100 0.0210(0.0064) 0.2871(0.2737) 0.0384(0.0172) 0.3509(0.2505)
200 0.0145(0.0041) 0.3080(0.2278) 0.0269(0.0101) 0.3519(0.2393)
100 5 0.0908(0.0396) 0.2855(0.1479) 0.1264(0.0787) 0.3301(0.2269)
10 0.0706(0.0286) 0.2480(0.1222) 0.1008(0.0509) 0.3239(0.1910)
50 0.0321(0.0096) 0.2021(0.1199) 0.0577(0.0196) 0.2788(0.1437)
100 0.0212(0.0056) 0.1792(0.1414) 0.0389(0.0111) 0.2454(0.1237)
200 0.0138(0.0029) 0.1706(0.1578) 0.0277(0.0071) 0.1951(0.1105)
200 5 0.0968(0.0387) 0.2719(0.1368) 0.1202(0.0450) 0.2828(0.1519)
10 0.0731(0.0257) 0.2459(0.1065) 0.1061(0.0492) 0.2991(0.1574)
50 0.0312(0.0068) 0.1449(0.0629) 0.0557(0.0146) 0.2347(0.0878)
100 0.0217(0.0038) 0.1174(0.0707) 0.0383(0.0071) 0.1912(0.0678)
200 0.0139(0.0021) 0.0897(0.0708) 0.0266(0.0046) 0.1428(0.0626)
For 1 as small as 50. the two-step method not necessarily improves estimation as ` becomes
larger. This is consistent with the large sample theory of the second step, which requires
large 1 to obtain a consistent estimator of the regression coecient. Even if we observe the
factor 1
t
. the convergence rate of the OLS estimator
^
1 for the rotation matrix is
_
1. When
1 is as large as 100 or 200. combined with large `. the two-step method is able to provide
reasonably good estimation results. In general, a larger ` is accompanied by an improved
estimation precision. Such a pattern exists in both DGPs of
i,)
.
4 Issues related to the panel data
Data for the cross-sectional dividend price ratios can be obtained from Fama and Frenchs
website. I will use the 100 portfolio data for estimating the long-run component. There
are several possible issues that might complicate the empirical studies. First, individual
xed eects might present in the panel data model. In the conventional large-` small-1
framework, dierencing method is usually used to remove the individual xed eects and
15
then the model parameters are estimated based on the transformed model. In an large-
(`. 1) framework, the individual xed eects can be consistently estimated (see Hahn, J.,
Kuersteiner, 2002 for an example) without resorting to the dierencing method.
Second, there are some missing observations in the 100 portfolio data constructed by
Fama and French. Although the principal component methods do not directly apply for
such a data set, the expectation-maximization (EM) algorithm of Stock and Watson (1998)
can be applied to handle panel data with missing observations. Consider the panel data A
it
.
i = 1. .... `. t = 1. .... 1. Let 1
it
be a binary variable which takes value 0 if the observation
(i. t) is missing and 1 otherwise. The EM algorithm iterates between an expectation step
and a maximization step until some convergence criteria are met.
To initialize the EM algorithm, I replace the missing observations of a portfolio / by i.i.d.
random variables l
It
which are uniformly distributed with mean equal to the sample mean
of that portfolio. Such a replacement is done for all ` portfolios to form a balanced panel
A
(0)
of dimension 1 `:
A
(0)
it
=
_
A
it
. if 1
it
= 1
l
it
. if 1
it
= 0
.
Then the maximization step can be carried out using A
(0)
. In a typical (/) t/ max-
imization step, we start with a balanced panel A
(I)
. Principal components estimators
^
`
(I)
i
and
^
,
(I)
t
are obtained from A
(I)
by solving the restricted least squares problem
_
^
`
(I)
i
.
^
,
(I)
t
_
= arg min
fA
i
,)tg
.

i=1
T

t=1
_
A
(I)
it
`
0
i
,
t
_
2
:.t.
1
1
T

t=1
,
t
,
0
t
= 1
v
.
.

i=1
`
i
`
0
i
is diagonal.
Let
^

(I)
be the `: loading matrix and
^
1
(I)
be the 1 : factor matrix. Then
^
1
(I)
consists
the leading : eigenvectors of the 1 1 matrix
1
.T
A
(I)
A
(I)0
. and
^

(I)
=
1
T
A
0
^
1
(I)
.
In the (/) t/ expectation step, we form a balanced panel A
(I+1)
:
A
(I+1)
it
=
_
A
it
. if 1
it
= 1
^
`
(I)0
i
^
,
(I)
t
. if 1
it
= 0
.
We keep iterating until some convergence criteria are met, say max
fi,t:1
it
=0g

A
(I+1)
it
A
(I)
it

<
10
7
. This EM algorithm is applied to the log dividend-price ratio data to impute the missing
16
values. Then I calculate the variance explained by the rst ve leading principal components:
Eigenvalue Variance explained
13.8766 96.77%
0.3448 99.18%
0.0442 99.48%
0.0081 99.54%
0.0052 99.58%
Although the Bai and Ng (2002) information criteria chose three factors, we see from the
above table that assuming two factors will be equally well in terms of tting the data. Take
the annual consumption data of US from 1929 to 2006
5
and use the two-step approach, I
obtain the following estimates:
0 10 20 30 40 50 60 70 80
-0.2
0
0.2
LLR component of consumption growth
dc
x
hat
(pca)
0 10 20 30 40 50 60 70 80
-0.02
0
0.02
x
hat
(pca)
0 10 20 30 40 50 60 70 80
0
2
4
x 10
-3
Conditional Variance

2
The consumption process is given by
c
t+1
= 0.0325 +r
t
+o
t
j
t+1
r
t+1
= 0.9015 r
t
+ 0.0029 o
t
c
t+1
o
2
t+1
8 10
5
= 0.9771 (o
2
t
8 10
5
) + 0.00015 n
t+1
5
I use the real personal consumption expenditure measured in chained-2005 dollars series. Data source:
Federal Reserve Bank of St. Louis. Other measures of the consumption can also be used to address the same
issue in the same manner.
17
The magnitude of the parameter estimates matches that of the calibrated values in Bansal
and Yaron (2004), though the specic values could be much dierent. A remarkable feature of
the above esimates is that the factor-augmented MLE is able to distangle both the small long-
run component of consumption growth and its slow-moving stochastic volatility. I expect
that a more detailed empirical work will oer further insight in justifying the existence of a
small and persistent long-run risk component.
5 Conclusion
Using both Monte Carlo experiments and a preliminary empirical study, I demostrate the
usefulness of the factor-augmented MLE method in the estimation of the state sequence
that lies in the heart of dynamic macro models. Dierent from small-model approach, the
realizations of the state variables can be consistently estimated. The proposed two-step ap-
proach has good nite sample properties in estimating the entire time series of the hidden
states. The empirical study also shows the potential power of the proposed approach in jus-
tifying the existence of an otherwise statistically insignicant component in the consumption
growth. By allowing the cross-sectional information to tend to innity, the signicance of the
long-run component can be studied using the large-(`. 1) asymptotics. A more extensive
empirical study needs to be considered for the evaluation of the asset pricing implications
of the estimated consumption process. Potentially, the two-step approach can be applied to
oer empirical support of the long-run risk explanation of the equity-premium puzzle and
the risk-free rate puzzle. A quantitative study is expected to be conducted to compare the
implications of the model with estimated LLR with that of Bansal and Yaron (2004) where
the LLR is within the agents information set.
References
Amengual, D., and M. Watson (2007): Consistent estimation of the number of dynamic
factors in large N and T panel, Journal of Business and Economic Statistics, 25(1), 9196.
Anderson, H., and F. Vahid (2007): Forecasting the volatility of Australian stock re-
turns: Do common factors help?, Journal of the American Statistical Association, 25(1),
7590.
Anderson, T. (1984): An Introduction to Multivariate Statistical Analysis. New York:
Wiley.
18
Andrews, D. W. K. (2005): Cross-section regression with common shocks, Economet-
rica, 73, 15511585.
Backus, D. K., B. R. Routledge, and S. E. Zin (2007): Asset prices in business cycle
analysis, Working paper.
Bai, J. (2003): Inferential theory for factor models of large dimensions, Econometrica,
71(1), 135172.
(2004): Estimating cross-section common stochastic trends in non-stationary panel
data, Journal of Econometrics, 122, 137183.
(2009): Panel Data Models with Interactive Fixed Eects, Econometrica, 77(4),
12291279.
Bai, J., and S. Ng (2006a): Condence intervals for diusion index forecasts and inference
for factor-augmented regressions, Econometrica, 74(4), 11331150.
(2006b): Determining the number of factors in approximate factor models, Econo-
metrica, 70(1), 191221.
(2006c): Evaluating latent and observed factors in macroeconomics and nance,
Journal of Econometrics, 113(1-2), 507537.
Bansal, R., R. F. Dittmar, and C. T. Lundblad (2005): Consumption, Dividends,
and the Cross Section of Equity Returns, The Journal of Finance, LX(4), 16391672.
Bansal, R., R. Gallant, and G. Tauchen (2007): Rational Pessimism, Rational
Exuberance, and Asset Pricing Models, Review of Economic Studies, 74, 10051033.
Bansal, R., D. Kiku, and A. Yaron (2007): Risks For the Long Run: Estimation and
Inference, Working paper.
(2009): An Empirical Evaluation of the Long-Run Risks Model for Asset Prices,
Working paper.
Bansal, R., and A. Yaron (2004): Risks for the Long Run: A Potential Resolution of
Asset Pricing Puzzles, The Journal of Finance, 59(4), 14811509.
Bernanke, B., and J. Boivin (2003): Monetary policy in a data rich environment,
Journal of Monetary Economics, 50(3), 525546.
19
(2005): Measuring monetary policy: a factor augmented vector autoregressive
(FAVAR) approach, Quarterly Journal of Economics, 120(1), 387422.
Boivin, J., and M. Giannoni (2006): DSGE models in a data-rich environment, Work-
ing paper.
Boivin, J., and S. Ng (2005): Are more data always better for factor analysis?, Journal
of Econometrics, 132, 169194.
(2007): Sticky prices and monetary policy: evidence from disaggregated U.S.
data, Forthcoming in The American Economic Review.
Chamberlain, G., and M. Rothschild (1983): Arbitrage, factor structure and mean-
variance analysis in large asset markets, Econometrica, 51, 12811304.
Chen, X., L. P. Hansen, and J. Scheinkman (2009): Nonlinear principal components
and long-run implications of multivariate diusions, Annals of Statistics, 37(6B), 4279
4312.
Doz, C., D. Giannone, and L. Reichlin (2007): A quasi-maximum likelihood approach
for large approximate dynamic factor models, European Central Bank Working Paper
Series 674.
Forni, M., D. Giannone, M. Lippi, and L. Reichlin (2003): Opening the black box:
Identifying shocks and propagation mechanisms in VAR and factor models, Working
paper.
Forni, M., M. Hallin, M. Lippi, and L. Reichlin (2000): The generaized dynamic
factor model: identication and estimation, Review of Economics and Statistics, 82(4),
540554.
(2001): Do nancial variables help in forecasting ination and real activity in the
euro area, Working paper.
(2004): The generalized factor model: consistency and rates, Journal of Econo-
metrics, 119, 231255.
(2005): The generalized dynamic factor model, one sided estimation and forecast-
ing, Journal of the American Statistical Association, 100, 830840.
Geweke, J. F., and K. J. Singleton (1981): Maximum likelihood Conrmatory
Factor Analysis of Economic Time Series, International Economic Review, 22(1), 3754.
20
Hamilton, J. (1994): Time Series Analysis. Princeton: Princeton University Press.
Hansen, L. P., and J. Scheinkman (2009): LONG-TERM RISK: AN OPERATOR
APPROACH, Econometrica, 77(1), 177234.
Kim, C.-J., and C. Nelson (1999): State Space Models With Regime Switching: Classical
and Gibbs Sampling Approaches With Applications. Massachusetts: MIT Press.
Lettau, M., S. Ludvigson, and J. A. Wachter (2008): The declining equity premium:
what role does macroeconomic risk play?, The Review of Financial Studies, 21(4), 1653
1687.
Ludvigson, S., and S. Ng (2005): Macro factors in bond risk premia, NBER Working
Paper 11703.
(2007): The empirical risk return relation: A factor analysis approach, Journal
of Financial Economics, 83, 171222.
Marcellino, M., J. H. Stock, and M. Watson (2003): Macroeconomic forecasting
in the Euro area: country specic versus Euro wide information, European Economic
Review, 47, 118.
McLachlan, G. J., and T. Krishnan (1996): The EM Algorithm and Extensions. Wiley-
Interscience.
Mumtaz, H., and P. Surico (2009): The Transmission of International Shocks: A Factor-
Augmented VAR Approach, Journal of Money, Credit and Banking, 41(s1), 71100.
Ng, S., and S. Ludvigson (2007): The Empirical Risk-Return Tradeo: A Factor Analy-
sis Approach, Journal of Financial Economics, 83, 171222.
(2009): Macro Factors in Bond Risk Premia, Review of Financial Studies, 22(12),
50275067.
Onatski, A. (2005): Determining the number of factors from empirical distribution of
eigenvalues, Department of Economics, Columbia University, Discussion Paper 0405-19.
(2006): A formal statistical test for the number of factors in approximate factor
models, Department of Economics, Columbia University, Unpublished Manuscript.
Quah, D., and T. Sargent (1992): A dynamic index model for large cross sections,
Federal Reserve Bank of Minneapolis, Discussion Paper 77.
21
Sargent, T., and C. Sims (1977): Business cycle modelling without pretending to have
too much a priori economic theory, in In: C. Sims (ed.): New Methods in Business Cycle
Research. Minneapolis: Federal Reserve Bank of Minneapolis.
Stock, J. H., and M. W. Watson (1988): Testing for common trends, Journal of the
American Statistical Association, 83, 10971107.
(1998): Diusion Indexes, NBER Working Paper 6702.
(2002a): Forecasting using principal components from a large number of predic-
tors, Journal of the American Statistical Association, 97, 11671179.
(2002b): Macroeconomic forecasting using diusion indexes, Journal of Business
and Economic Statistics, 20(2), 147162.
(2005): Implications of dynamic factor models for VAR analysis, NBER Working
Paper 11467.
(2006): Forecasting with many predictors, in Handbook of Economic Forecasting.
North Holland: Elsevier.
22

Вам также может понравиться