Sei sulla pagina 1di 12

An Unobserved-Component Model With

Switching Permanent and Transitory Innovations

Chung-Ming K UAN
Institute of Economics, Academia Sinica, Taipei, Taiwan ( )

Yu-Lieh H UANG
Department of Quantitative Finance, National Tsing-Hua University, Hsin-Chu, Taiwan ( )

Ruey S. T SAY
Graduate School of Business, University of Chicago, 5807 S. Woodlawn Avenue, Chicago, IL 60637
( )

This article proposes an unobserved-component model in which component innovations are governed by
a state variable that follows a Markov process. The proposed model is capable of describing both sta-
tionary and nonstationary behaviors of real data and allows the random innovations to have permanent
and transitory effects in different periods. The model also permits a deterministic trend with or without
breaks and hence bridges the gap between the trend-stationary model and a random walk with drift. For
ease in presentation and in application, our discussion focuses on the model consisting of a random-walk
component and a stationary autoregressive moving average component. However, the proposed model is
much more flexible. We investigate properties of the proposed model and derive an estimation algorithm.
We also propose a simulation-based test to distinguish between the proposed model and an autoregressive
integrated moving average model. For application, we apply the model to U.S. quarterly real gross domes-
tic product and find that unit-root nonstationarity is likely to be the prevailing dynamic pattern in more
than 80% of the sample periods. Because nonstationarity (stationarity) periods match the National Bureau
of Economic Research dating of expansions (recessions) closely, our result suggests that the innovations
in expansion (recession) are more likely to have a permanent (transitory) effect.

KEY WORDS: Innovation regime-switching model; Markov trend; Permanent innovation; Transi-
tory innovation; Trend stationarity; Unit-root nonstationarity; Unobserved-component

1. INTRODUCTION in the article is motivated by two leading models in time se-

ries econometrics, the unit-root model and the trend-stationary
It has been well documented in the literature that many eco- model. Since the seminal work of Nelson and Plosser (1982),
nomic and financial datasets exhibit different characteristics unit-root nonstationarity has been widely accepted as a stylized
(or regimes) over time. Econometric models that can accommo- fact of many macroeconomic time series (see also Campbell
date multiple regimes include structure-change models, thresh- and Mankiw 1987). Yet some researchers believe that trend-
old models, and Markov switching models. In these models it stationary models with or without breaks are better model-
is typical to postulate one function for the data and characterize ing tools (see, e.g., Blanchard 1981; Clark 1987; Perron 1989;
different regimes by distinct parameter values of the function. Rappoport and Richelin 1989). Thus it is helpful to consider a
Although this modeling strategy is convenient and useful in model that can accommodate both dynamic patterns.
some applications, it is rather restrictive because it only permits For simplicity, we focus on the special model consisting
similar dynamic patterns in different regimes. A more general of a random-walk component and a stationary autoregressive
approach is to use different functions for different data regimes. moving average (ARMA) component with a Markov process
Although some models are constructed along this line (e.g., governing the state transition. This simple model permits a de-
Evans and Wachtel 1993; Engle and Smith 1999), this general terministic trend with or without breaks and hence can serve as
approach has not received much attention in the literature. an intermediate case between the trend-stationary model and a
In this article we follow the latter approach and propose random walk with drift. It can also serve as a bridge between
an unobserved-component model with regime switching. The the ARMA model and a random walk without drift. If the com-
model consists of two (or more) unobserved components and ponents are state independent, then the model reduces to one of
a state variable linked specifically to innovations. Depending the two extreme cases. The proposed model is in sharp contrast
on the value of the state variable, each innovation excites only with the commonly used time series models and the regime-
switching model of Hamilton (1989), in that it is capable of
one of the components while the other component continues
describing both stationary and nonstationary dynamics and al-
to evolve without any innovation. As such, there is only a sin-
lows its random innovations to have both permanent and tran-
gle innovation for the observed process at each time point.
sitory effects. Details are given in Section 2. Moreover, trend
The prevailing component in turn determines the characteris-
tics of the innovation and hence the new data dynamics. By
© 2005 American Statistical Association
assigning distinct functions to different components, the pro- Journal of Business & Economic Statistics
posed model is able to completely characterize different dy- October 2005, Vol. 23, No. 4
namic structures in different regimes. Our choice of functions DOI 10.1198/073500105000000054

444 Journal of Business & Economic Statistics, October 2005

breaks are endogenous in the sense that the trend function shifts component y1,t is activated and the effect of υt will be prop-
when permanent shocks are present. Compared with the model agated through the function g(·), whereas the second compo-
of Evans and Wachtel (1993) and McCulloch and Tsay (1994), nent y0,t evolves according to h without any innovation. When
the Markovian state variable here is attached to innovations st = 0, υt excites y0,t but does not enter the first component y1,t .
and entails different dynamic behaviors. This model also re- Furthermore, because both st υt and (1 − st )υt are present in (1),
sults in an ARMA representation with random moving-average the state variable st does not affect yt . But it does determine
(MA) coefficients and hence is quite different from the more fa- how the innovation υt affects the subsequent yt+j for j > 0. Sec-
miliar, random autoregressive (AR)-coefficient models, such as ond, the dynamic of the observed process yt is a mixture of two
those of McCabe and Tremayne (1995) and Granger and Swan- components governed by the state variable st . For distinct func-
son (1997). tions g(·) and h(·), the dynamic of each component evolves over
We also derive an estimation algorithm for the proposed time with inputs {y1,t−j } and {y0,t−j }. In contrast, the regime-
model using a state–space representation and propose a simula- switching models of Hamilton (1989, 1994) and Tong (1990)
tion-based test to distinguish between the model and a pure au- typically contain a single component with state-dependent pa-
toregressive integrated moving-average (ARIMA) process. As rameter θ st ,
an application, we apply the proposed model to U.S. quarterly  
real gross domestic product (GDP) and find that unit-root non- yt = g yt−1 , . . . , yt−p ; θ st + υt .
stationarity is likely to be the prevailing dynamic pattern for Given that there is only one function g(·) in the model, the
about 80% of the sample periods, whereas the remaining peri- dynamic structures in different regimes are similar in essence.
ods are likely to be in the stationarity regime. This finding is Also, the input of the g(·) function is {yt−j } for both regimes.
quite different from the inference drawn from either a unit-root Third, for the proposed model, the state variable st+j ( j > 0)
or a trend-break model. Furthermore, we observe that the non- determines the allocation of the innovation υt+j , but the effect
stationarity (stationarity) periods match the National Bureau of of υt on yt+j is controlled by the state st . This special character-
Economic Research (NBER) dating of expansions (recessions) istic enables us to classify the innovation υt as a permanent or
closely. These results suggest that the innovations in expansion transitory shock. In contrast, many econometric models avail-
(recession) are more likely to have a permanent (transitory) ef- able in the literature classify the impact of υt on yt+j based on
fect. That the shocks in expansion are more persistent than those the state variable st+j and can lead to different classifications
in recession is compatible with the conclusion of Beaudry and for the same innovation υt when j changes.
Koop (1993). Our empirical results thus provide an alternative Although it is possible to express (1) as a special case of
view of the characteristics of U.S. real GDP. a general model with state-dependent coefficients [e.g., the
This article is organized as follows. Section 2 introduces the ARMA model of (4) and the state–space model of (8)], this
proposed unobserved-component model and compares it with two-component structure with the state variable linked to in-
some existing models. Section 3 derives some statistical prop- novations is convenient for researchers to specify the dynamic
erties of the proposed model. Section 4 discusses the estima- patterns that they would like to characterize.
tion algorithm and hypothesis testing, and Section 5 presents
the empirical analysis of U.S. real GDP based on the proposed
2.1 The Proposed Model
model. Section 6 concludes.
Clearly, the model in (1) is not operational unless the func-
2. AN INNOVATION SWITCHING MODEL tions g and h are specified. Many selections are possible. Our
choice of the functions is motivated by two important empiri-
Consider the simple case of two components. Suppose that cal characteristics commonly observed in application, unit-root
the observed process yt is the sum of two components, namely nonstationarity and (trend) stationarity. There has been a long
yt = y1,t + y0,t , (1) history of debate regarding whether an economic time series
should be modeled using a unit-root model or a trend-stationary
where model. These two models generate distinct dynamic behaviors
y1,t = g( y1,t−1 , . . . , y1,t−p ; θ 1 ) + st υt and bear quite different economic interpretations. When a time
series contains a unit root, its innovations all have a permanent
and effect, and its time path exhibits large swings. When a time se-
y0,t = h( y0,t−1 , . . . , y0,t−q ; θ 0 ) + (1 − st )υt , ries is trend-stationary, its innovations have a transitory effect
and induce only short-run fluctuations around the trend. In this
where g and h are two possibly different functions, p and q article we propose a model based on (1) that can accommodate
are positive integers, θ 1 and θ 0 are parameter vectors, st is the these two distinctive dynamic patterns.
state variable at time t that assumes the value 1 or 0, and {υt } is We propose an unobserved-component model that consists of
a sequence of independent and identically distributed random a random-walk component and a stationary ARMA component,
variables. The model is readily generalized to have more than that is,
two components.
The proposed model in (1) has some unique characteristics. yt = y1,t + y0,t (2)
First, each innovation υt excites only one component deter-
mined by the state variable, st . As such, st determines the dy-
namic effect of υt on the observed series. When st = 1, the first (1 − B)y1,t = α0 + st υt = (α0 + st α1 ) + st εt
Kuan, Huang, and Tsay: Unobserved-Component Model 445

(B)y0,t = (B)(1 − st )υt
= (B)(1 − st )α1 + (B)(1 − st )εt ,
where (B) = 1 − ψ1 B − · · · − ψm Bm and (B) = 1 − ϕ1 B −
· · · − ϕn Bn are finite-order polynomials of the backshift oper-
ator B such that they have no common factors and their roots
are all outside the unit circle, and υt = α1 + εt with {εt } a white
noise with mean 0. Note that allowing υt to have a (possibly)
non-0 mean adds more flexibility to the model; see further dis-
cussion below after (3). To avoid identification problems, we
do not include a state-independent constant term in the second
component, y0,t . In what follows we postulate that st follows a
Figure 1. Simulated Markov Trend ( ) and Smooth Markov Trend
first-order Markov chain (as in Hamilton 1989).
( )( st = 1).
Similar to the work of Clark (1987) and Cochrane (1988),
here the yt of the proposed model is the sum of a random walk
and a stationary component. A major difference is that the in- be seen that when st follows the first-order Markov chain,
novation in our model excites only one component in a given the first two components of (3) are just the “Markov trend”
time period. When st = 1, the first component is excited and of Hamilton (1989). While the Markov trend is kinked, the
evolves like a random walk; the associated innovation εt has a third component involves a weighted sum of 1 − st−j and in-
permanent effect on yt+j ( j > 0). When st = 0, εt excites the duces smooth transitions between trend segments. Figure 1 il-
stationary ARMA component and has a transitory effect on fu- lustrates a Markov trend and a smooth trend function of (3),
ture yt+j . Hence the effect of an innovation may alternate from where the smooth trend function is generated by setting α0 = 1,
time to time, depending on the state variable st . Many exist- α1 = 2, (B) = 1, (B) = 1 + .4B + .2B2 , and st = 1 for
ing models, in contrast, postulate that there are permanent and t = 1, . . . , 5, 16, . . . , 25, 36, . . . , 40, as indicated by the shaded
transitory innovations at each time index. As such, these in- boxes on the horizontal axis, and st = 0 otherwise. The differ-
novations together still have a permanent effect and generate ence between
unit-root nonstationarity for all time indices. This setup may  these trend behaviors is evident. The fourth com-
ponent, ti=1 si εi , admits only those εi with si = 1 (permanent
not always be appropriate, however. First, it is conceivable that shocks) and can be interpreted as a flexible stochastic trend. The
random shocks may be fundamentally different in different pe- last component, (B)−1 (B)(1 − st )εt , is a weakly stationary
riods. For example, Beaudry and Koop (1993) showed that pos- process generated by transitory innovations and gives rise to
itive shocks to U.S. GDP are more persistent than negative short-run fluctuations.
shocks (see also Bradley and Jansen 1997). Second, a time se- The features of this model are now clear. First, it admits both
ries does not always exhibit unit-root nonstationarity. For ex- deterministic and stochastic trends. Second, apart from the de-
ample, an asset price may have large changes and behave like terministic trend, it exhibits both stationary and nonstationary
a unit-root process when the market is inundated with vital in- patterns over different time periods. Third, it allows for en-
formation, yet it may fluctuate only slightly and behave like a dogenous trend breaks, in the sense that breaks are due to the
stationary process when the market is tranquil. By allowing the
presence of permanent shocks. Because different regimes in this
innovation to excite only one component, the proposed model
model are linked to the effects of innovations, (2) with α1 = 0
can easily distinguish between the innovation effects and al-
and a first-order Markov chain st is called an innovation regime-
lows for distinct (unit-root and stationary) dynamics in different
switching (IRS) model of orders 1 and (m, n), or IRS(1; m, n)
for short, with a smooth Markov trend. When α1 = 0, it is an
Setting y0 = 0 and εi = 0 for i ≤ 0, we have the following
IRS(1; m, n) model with a linear trend; when both α0 = α1 = 0,
it is simply an IRS(1; m, n) model without trend.

Note that if st = 1 for all t, then
y t = α0 t + α 1 si + α1 (B)−1 (B)(1 − st )
yt = (α0 + α1 )t + εi = (α0 + α1 ) + yt−1 + εt ,

+ si εi + (B) (B)(1 − st )εt . (3)
i=1 which is a random walk with the drift term α0 + α1 . If st = 0
The first component of this representation, α0 t, is a determin- for all t, then yt is a trend-stationary process without break,
istic trend of the model. It is the time-trend of the model when yt = α1 (1)−1 (1) + α0 t + (B)−1 (B)εt .
st = 0 for all t. This
 trend function is augmented by the second
component, α1 ti=1 si , if α1 = 0 and P(st = 1) = 0. In par- Model (2) thus constitutes intermediate cases between the
ticular, the deterministic trend is subject to a level shift when trend-stationary model (with or without breaks) and a random
there is a one-time permanent shock, and its slope changes walk with drift.
to α0 + α1 when permanent shocks are present consecutively. Model (2) can also be written as a special case of a general
Of course, there will be no trend break if α1 = 0. It can also dynamic model with state-dependent coefficients. In particular,
446 Journal of Business & Economic Statistics, October 2005

Table 1. MA Representation of the IRS(1; 1, 0) Model (5) With

it can be shown that (2) has an ARMA representation with ran-
(s1 , . . . , s7 ) = (0, 0, 1, 1, 1, 0, 0)
dom MA coefficients,
t y1, t y0, t
(B)(1 − B)yt
1 0 ε1

r+1 2 0 ψ1 ε1 + ε2
= α0 (1) + ξi,st−i (α1 + εt−i ) + (α1 + εt ), (4) 3 ε3 ψ12 ε1 + ψ1 ε2
i=1 4 ε3 + ε4 ψ13 ε1 + ψ12 ε2
where r = max{m, n}, 5 ε3 + ε4 + ε5 ψ14 ε1 + ψ13 ε2
 6 ε3 + ε4 + ε5 5
ψ1 ε1 + ψ14 ε2 + ε6
−ψ1 if st−1 = 1
ξ1,st−1 = 7 ε3 + ε4 + ε5 ψ16 ε1 + ψ15 ε2 + ψ1 ε6 + ε7
−1 − ϕ1 otherwise, .. ..
. .
and ∞
 ∞ i =1 si εi 0
−ψi if st−i = 1
ξi,st−i =
ϕi−1 − ϕi otherwise
for i = 2, . . . , r, and the last coefficient is of past innovations are not altered. By letting
 t tend to infinity,
 only the random-walk component ( y1,∞ = ∞ i=1 si εi ) remains;
0 if st−r−1 = 1 see the last row of Table 1.
ξr+1,st−r−1 =
ϕr otherwise; The IRS(1; 1, 0) model (5) is conceptually different from that
ψi = 0 for i > m, and ϕi = 0 for i > n. From (4), we can see of Evans and Wachtel (1993) and McCulloch and Tsay (1994),
that only the past state variables st−1 , . . . , st−r−1 can affect yt which also allows for switching between a random walk and an
and, as stated before, the concurrent state st is irrelevant to yt , AR(1) process,
because both st εt and (1 − st )εt are present at time t. Thus un- yt = st y1,t + (1 − st )y0,t ,
der the proposed model, the effect of a state variable st can only
where y1,t = y1,t−1 + vt and y0,t = ψ1 y0,t−1 + ut with |ψ1 | < 1,
be revealed in subsequent periods t + j with j > 0. Moreover,
and {ut } and {vt } are two sequences of white noise random
the model (2) has a state–space representation with switching
variables. This model differs from the proposed model in its
parameters [see (8)] and hence is a special case of the Markov-
switching mechanism. Here yt switches between two processes,
switching state–space model considered by Kim (1994) and so that all past innovations must change accordingly. Such a
Kim and Nelson (1999). switching mechanism affects the past dynamics but not future
Although (2) is a specialized model, its two-component dynamics. It also induces drastic changes ( jumps) in the time
structure is useful in practice, because it allows researchers to path of yt , especially when t gets larger, given that a random
explicitly specify the dynamic patterns of their choice. Note walk wanders off quickly. It seems implausible to have a time
that (2) is just one possibility of modeling economic time se- series exhibiting frequent jumps while the magnitude of jump
ries under the general framework of (1). Many other interest- is increasingly large. In contrast, the st of the proposed model
ing models can also be constructed as special cases of (1). determines the effect of εt on future observations but has no
For example, one may set yt as the sum of a fractionally inte- influence whatsoever on past innovations. Thus yt does not
grated component and a weakly stationary ARMA component, have sudden jumps, because the new dynamics result from new
so that yt may exhibit both long- and short-range dependence. shocks, rather than from the entire history of past innovations.
It is worth noting that when a fractionally integrated compo- Note also that because the Evans–Wachtel model contains two
nent is present, there will be no finite-dimensional state–space sets of innovations, it is not easy to explain why one set pre-
representation. It is also possible to specify a different switch- vails in some periods but does not play any role in the others.
ing mechanism for st , compare with the threshold-disturbance The proposed model does not have this problem.
moving-average model of Elwood (1998). From the random MA-coefficient representation (4), (5) can
be expressed as
2.2 Comparison With Existing Models yt = (1 + ψ1 )yt−1 − ψ1 yt−2 + ξst−1 εt−1 + εt ,
To compare with the existing models that allow for different with ξst−1 = −ψ1 if st−1 = 1 and ξst−1 = −1 otherwise. This
structures, we consider the IRS(1; 1, 0) model without trend, model is thus different from the standard random AR-coefficient
which consists of a random-walk component and a stationary model, such as that of McCabe and Tremayne (1995). A simple
AR(1) component with α0 = α1 = 0, that is, yt = y1,t + y0,t random AR-coefficient model can be expressed as
with yt = at yt−1 + ut ,
y1,t = y1,t−1 + st εt , where at are random variables, typically assumed to be exoge-
y0,t = ψ1 y0,t−1 + (1 − st )εt , |ψ1 | < 1. nous. The stochastic unit-root model of Granger and Swanson
(1997) assumes at = exp(αt ), with αt being a weakly stationary
Table 1 gives the MA representation of y1,t and y0,t for process with mean 0. Setting the initial value y1 = u1 , we can
t = 1, . . . , 7 when {s1 , . . . , s7 } = {0, 0, 1, 1, 1, 0, 0} and yi,t = 0 write
for t ≤ 0 and i = 0, 1. Here yt is a stationary AR(1) process  i−1 

at the beginning and starts to evolve like a random walk when yt = ut + at−j ut−i .
st = 1. It can be seen that when st switches, the classifications i=1 j=0
Kuan, Huang, and Tsay: Unobserved-Component Model 447

Similar to the Evans–Wachtel model, this model may not be chain with transition matrix
easy to interpret when the product i−1

j=0 at−j switches from the P(st = 0|st−1 = 0) P(st = 1|st−1 = 0) p p01
stable region to the explosive region. = 00 .
P(st = 0|st−1 = 1) P(st = 1|st−1 = 1) p10 p11
A similar but fundamentally different model is the STOP-
BREAK model of Engle and Smith (1999). The MA represen- Let St = {st , st−1 , . . .} denote the collection of all state vari-
tation of the simplest STOPBREAK model is ables up to time t. Also assume that {εt } is a sequence of

random variables such that E(εt |St ) = 0, var(εt |St ) = σε2 , and
 E(εt εt−i |St ) = 0 for all i > 0. By invoking the law of iterated
yt = qt−i εt−i + εt , (6)
expectations, it is easy to verify that {εt } is a white noise. Also,
E(st εt ) = E[st E(εt |St )] = 0,
where qt−i = εt−i2 /(γ + ε 2 ) with the parameter γ ≥ 0. When
t−i var(st εt ) = E s2t E(εt2 |St ) = σε2 P(st = 1),
γ is very small, as in the empirical example of Engle and Smith
(1999), qt ≈ 1, so that the resulting process is close to a pure and cov(st εt , st−i εt−i ) = E[st st−i E(εt εt−i |St )] = 0. Similarly,
random walk. When εt has a continuous distribution, qt = 0 oc- (1 − st )εt also has mean 0 and variance [1 − P(st = 1)]σε2 and
curs only with probability 0; that is, the STOPBREAK model are serially uncorrelated. These two series are white noise when
cannot exhibit stationary behavior. Because 0 < qt ≤ 1 with P(st = 1) is a constant π0 . Moreover,
probability 1, this process is in effect an I(1) process with non-  
cov st εt , (1 − st−i )εt−i = E st (1 − st−i )E(εt εt−i |St ) = 0
linear MA terms.
To illustrate, we simulate a random walk, the STOPBREAK for i ≥ 0, so that st εt and (1 − st )εt are mutually uncorrelated at
process (6) with γ = 1 and process (5) with ψ1 = 0. In our all leads and lags.
simulations, these processes are generated from the same set of We first consider model (2) with α1 = 0. By (3),
innovations; process (5) is such that st = 1 for those t indicated

by the shaded boxes on the horizontal axis and st = 0 otherwise. y t = α0 t + si εi + (B)−1 (B)(1 − st )εt , (7)
The simulated paths are plotted in Figure 2, where the thick i=1
line is the random walk, the line with “+” is the STOPBREAK
process, and the thin line is process (5). These figures show that which is the sum of uncorrelated components. Then E( yt ) =
the STOPBREAK process always mimics the random walk, α0 t and
yet process (5) exhibits flexible dynamic patterns. In particu- 
lar, process (5) may be quite different from the random walk var( yt ) = σε2 P(si = 1) + σε2 (ψi∗ )2 [1 − P(si = 1)],
and the STOPBREAK process, or it may move closely with the i=1 i=1
random walk and the STOPBREAK process. where ψi∗ are the coefficients of (B)∗ = (B)−1 (B) =
1 − ψ1∗ B − ψ2∗ B2 − · · · . When ψi∗ are square-summable, the
3. PROPERTIES OF THE PROPOSED IRS MODEL second term on the right side converges. Thus var( yt ) would
grow without bound if the partial sum ti=1 P(si = 1) diverges.
To derive properties of the yt process generated from (2), we In particular, when P(si = 1) is a constant π0 > 0, var( yt ) will
maintain the assumption that st follows a first-order Markov grow linearly with t as σε2 π0 t, which is proportional to that of

Figure 2. Examples of Simulated Random Walk ( ), STOPBREAK Process ( ), and IRS(1; 1, 0) ( ) Process.
448 Journal of Business & Economic Statistics, October 2005

a pure random walk. In contrast, when ti=1 P(si = 1) con- P(st = 1) is a constant π0 , it is now easy to see that the
verges, the Borel–Cantelli lemma implies that P(si = 1 infi- autocovariance-generating function of zt is
nitely often) = 0; that is, the event {si = 1} occurs for at most
finitely many i with probability 1. It follows that yt is eventually g(a) = gz1 (a) + gz2 (a)
a stationary ARMA process, even though it may be nonstation- = π0 σε2 + (1 − π0 )(1 − a)(1 − a−1 )
ary during any finite time period. 
α1 = 0, E( yt ) = α0 t + α1 ti=1 P(si = 1) + α1 × × (a)−1 (a−1 )−1 (a)(a−1 )σε2 .
t ∗
i=1 ψi [1 − P(si = 1)]. It is, however, more cumbersome to For example, for the IRS(1; 1, 0) model considered in Sec-
calculate var( yt ) because the components of (3) are no longer
tion 2.2, zt has mean 0 and
mutually uncorrelated. To see this, first note that
1 − ψ1 2
cov(st , st−j ) = P(st = 1 and st−j = 1) − P(st = 1)P(st−j = 1) var(zt ) = σε2 + (1 − π0 )(1 − ψ12t−2 ) σ .
1 + ψ1 ε
= P(st−j = 1) P(st = 1|st−j = 1) − P(st = 1) .
The autocovariances are
By the Markovian property, P(st = 1|st−j = 1) is the (2, 2) el-
1 − ψ1 2
ement of the jth power of the transition matrix for j ≥ 1. Thus cov(zt , zt−1 ) = −(1 − π0 )(1 + ψ12t−1 ) σ
st are serially correlated and st and (1 − sτ ) are not mutually 1 + ψ1 ε
uncorrelated. Moreover, cov(st , st−j εt−j ) are non-0 when j ≥ 1. and
To ease our analysis of var( yt ), we assume for the time being
that E(εt |Sτ ) = 0 for τ > t. Then cov(st , st−j εt−j ) = 0, so that cov(zt , zt−i ) = ψ i−1 cov(zt , zt−1 ), i ≥ 2.
 t   t 
   t These autocovariances depend only on i as t goes to infinity,

var( yt ) = α1 var
si + 2α1 cov
si , ψi (1 − si ) demonstrating that zt is asymptotically a covariance stationary
i=1 i=1 i=1 process. When π0 = 1, zt is simply a white noise. Note also
  that these autocovariances agree with those of a noninvertible

+ α12 var ψi∗ (1 − si ) + σε2 P(si = 1) ARMA(1, 1) process if and only if π0 = 0. Thus zt would be
i=1 i=1 invertible provided that π0 > 0.

+ σε2 (ψi∗ )2 [1 − P(si = 1)].
Note that var( ti=1 si ) is bounded when the partial sum
i=1 P(si = 1) converges; the same conclusion also holds for
4.1 Estimation of the State–Space Representation
the second and third terms on the right side of the prior equa-
tion. This shows that the stationarity of yt still depends essen- There are different ways to estimate the proposed IRS(1;
 m, n) model. Following the suggestion of a referee, we adopt
tially on the convergence of ti=1 P(si = 1), as in the case
where α1 = 0. the approach of Kim (1994). We first put the proposed model
From (3), it is clear that the impulse response function of the into a Markov-switching state–space form and then derive an
proposed model depends on the realization of si . Let F t denote algorithm to compute the approximate quasi-maximum likeli-
the information set up to time t and hood estimates (QMLEs).
From the ARMA representation (4), we observe that the
∂E( yt+k |F t ) past r + 1 state variables affect yt , where r = max(m, n). De-
δt ≡ lim
k→∞ ∂εt fine s∗t−1 as the new state variable such that each of its val-
denote the long-run effect of εt on the optimal forecast of yt+k . ues (in {1, 2, . . . , 2r+1 }) represents a particular realization of
Then δt = st , which is 1 or 0 and changes from time to time. (st−1 , . . . , st−r−1 ). For example, when r = 2, s∗t−1 has eight pos-
Recall that the long-run effect δt is 1 for a random walk and sible values: s∗t−1 = 1 if st−1 = st−2 = st−3 = 0, s∗t−1 = 2 if
0 for a weakly stationary process. Note also that for the simplest st−1 = 0, st−2 = 0, and st−3 = 1, . . . , s∗t−1 = 8 if st−1 = st−2 =
STOPBREAK process (6), st−3 = 1. It can be seen that s∗t also forms a first-order Markov
chain with transition matrix
∂qt 2γ εt2  
δt = qt + εt = qt + , P00 0
∂εt (γ + εt2 )2
 0 P10 
where qt = εt2 /(γ + εt2 ). When εt has a continuous distribu- P∗ = 
 P01
0 
tion, εt is 0 with probability 0. Thus δt is positive with probabil-
0 P11
ity 1, demonstrating that the innovations of the STOPBREAK
process must have a permanent effect. with Pij (i, j = 0, 1) a 2r−1 × 2r block diagonal matrix given by
For the property of zt = (1 − B)yt , we consider the case p p
where α1 = 0. Write zt as zt = z1,t + z2,t , where z1,t = α0 + ij ij 0 0 ··· 0 0 
st εt and z2,t = (1 − B)(B)−1 (B)(1 − st )εt . Then zt is the  0 0 pij pij · · · 0 0 
Pij = 
 .. .. .. .. .. .. .. 
sum of two uncorrelated components. Let gz1 and gz2 denote . . . . . . .
the autocovariance-generating functions of z1 and z2 . When 0 0 0 0 · · · pij pij
Kuan, Huang, and Tsay: Unobserved-Component Model 449

To keep the exposition simple, we assume that the row j, column 4.2 Hypothesis Testing
i element of P∗ is p∗ij = P(s∗t = j|s∗t−1 = i). Also, let ξ t−1,s∗ =
(ξ1,st−1 , ξ2,st−2 , . . . , ξr+1,st−r−1 ) denote the collection of the ran- It is well known that many economic and financial time se-
dom MA coefficients in (4). When s∗t−1 = j, we write ξ t−1,s∗ ries contain a unit root. However, an interesting hypothesis is
t−1 whether the data follow a random walk. This amounts to test-
as ξ t−1,j , so that
ing the null hypothesis H0 : p11 = 1 under the proposed model,

r+1 that is, st = 1 almost surely. Under this null hypothesis, the
ξ t−1,j 1 = ξj,st−j , stationary component does not enter the model, so its para-
j=1 meters [i.e., those of (B) and (B)] are not identified. In
this case, standard likelihood-based tests, such as the Wald,
where 1 is the vector of 1’s and the realizations of st−1 , . . . , Lagrange multiplier (LM), and likelihood ratio tests, are not
st−r−1 are such that s∗t−1 = j. When r = 2 and j = 3, for ex- applicable (see Davies 1977, 1987; Hansen 1996). The problem
ample, the realization of (st−1 , st−2 , st−3 ) is (0, 1, 0), so that that certain parameters are not identified under the null hypoth-
ξ t−1,3 1 = −(1 + ϕ1 ) − ψ2 + ϕ3 . esis also arises in other regime-switching models. In contrast
It is now easy to show that the proposed IRS(1; m, n) model with Hamilton’s model, whether α1 = 0 is not of primary con-
can be expressed as a state–space model with the following cern here. Once we exclude the possibility that the process is
measurement and transition equations: a random walk, hypothesis testing on other parameters is stan-
zt = µ0 + µ1 ξ t−1,s∗ 1 + Hs∗t−1  t , dard and can be done using likelihood-based tests. Therefore,
(8) we focus on the null hypothesis of p11 = 1.
 t = F t−1 + ε t , Because the data follow a random walk when p11 = 1, it is
of interest to study the performance of the Dickey–Fuller (DF)
where µ0 = α0 + α1 (1)−1 and µ1 = α1 (1)−1 are two con- test (Dickey and Fuller 1979). We simulate yt according to (3)
stant terms, Hs∗t−1 = (1 ξ t−1,s∗ ) is an (r + 2)-dimensional row with α0 = α1 = 0, σε2 = 1, (B) = 1 − .5B, (B) = 1, and var-
vector, ious combinations of the transition probabilities p11 and p00 .
 
ψ1 ψ2 · · · ψ r 0 0 In the simulation, the nominal size is 5%, the sample size
 1 0 ··· 0 0 0 is 120, and the number of replications is 5,000. The resulting
 
 0 1 ··· 0 0 0 rejection frequencies of the DF test are plotted in Figure 3(a).
F= .
 .. .
.. .
.. .
.. . . 
.. ..  We see that the DF test is not powerful against the alterna-
 
 0 0 ··· 1 0 0
tive (3) except when p11 is small and p00 is large. For example,
given p11 = .9, when p00 = .8 and .2, the powers are 14.9%
0 0 ··· 0 1 0
and 7.6%; given p11 = .1, when p00 = .8 and .2, the powers are
is an (r + 2) × (r + 2) matrix such that ψi = 0 for i > m, 55.8% and 23.7%. A detailed table of rejection frequencies is
and ε t = (εt , 0, . . . , 0) denotes the shock of the model. This available on request.
is also a Markov-switching state–space model discussed by As shown in Figure 3(b), the KPSS test of Kwiatkowski,
Kim (1994). Phillips, Schmidt, and Shin (1992) is more powerful against (3)
We therefore follow Kim (1994) to derive an estimation except when p00 is close to 1. The rejection frequencies are
algorithm that involves the Kalman filter and the Hamilton typically around 70% when p11 and p00 are between .1 and .9.
filter (see also Kim and Nelson 1999 for a thorough discus- When the KPSS test rejects the null of stationarity, it is still dif-
sion). The derivation of our estimation algorithm is similar, but ficult to judge whether the series being tested is a random walk
not identical, to that of Kim (1994), because here yt depends or a process generated from the proposed model. More testing
only on the past (not on the concurrent) state variables; a de- results are needed to support the proposed model.
tailed description of this algorithm is given in the Appendix. In this article we use a simulation-based test to distinguish
From the recursions of the derived algorithm, we obtain the fil- an ARIMA model from the proposed IRS model. For a given
tering probabilities, P(st = 1|Z t ; θ); the smoothing probabili- series, we first estimate an array of ARIMA( p, 1, q) models
ties, P(st = 1|Z T ; θ ); and an approximate quasi–log-likelihood and choose an appropriate specification based on an infor-
function, mation criterion [e.g., Schwarz information criterion (SIC) or
Akaike information criterion (AIC)]. The selected model is de-

ln L = ln f (zt |Z t−1 ), noted as ARIMA( p∗ , 1, q∗ ). Similarly, we also estimate an ar-
t=1 ray of IRS(1; m, n) models and denote the selected model by
IRS(1; m∗ , n∗ ) and the estimated transition probability by p̂∗11 .
where zt = (1 − B)yt and Z t = {z1 , . . . , zt }. The approximate The selected ARIMA( p∗ , 1, q∗ ) model is then taken as the
QMLE, data-generating process to generate simulated samples. For
θ̂ = (α̂0 , α̂1 , ψ̂1 , . . . , ψ̂m , ϕ̂1 , . . . , ϕ̂n , σ̂ε , p̂00 , p̂11 ) , each simulated sample, we reestimate the IRS(1; m∗ , n∗ ) model
and obtain an estimate of p11 , denoted by p̂11 . Replicating
can then be found using a numerical search method. Our pro- this procedure many times yields a finite-sample reference
gram is written in GAUSS, which uses the BFGS (Broyden– distribution of p̂11 on which we can compute the p-value
Fletcher–Goldfarb–Shanno) search algorithm. Plugging θ̂ into of p̂∗11 . We reject the null hypothesis that the series follows
the formulas of P(st = 1|Z T ; θ), we obtain the estimated the ARIMA( p∗ , 1, q∗ ) model if the p-value of p̂∗11 is small
smoothing probabilities. (say <5%). Note that this test does not solve all of the problems
450 Journal of Business & Economic Statistics, October 2005

(a) (b)

Figure 3. Empirical Powers of the Dickey–Fuller Test (a) and KPSS Test (b).

when there are unidentified nuisance paramameters under the with σu = .0097, where yt = yt − yt−1 . We then reestimate
null, but it is used to provide some justification of the proposed the IRS(1; 2, 2) model using the data generated from (9) and
model. A better testing procedure would be highly desirable but obtain p̂11 . Using 1,000 replications, we obtain a finite-sample
is beyond the scope of this article. reference distribution of p̂11 . The p value of p̂∗11 = .91 based on
this simulated distribution is about .034, and hence we reject
5. EMPIRICAL STUDY the model in (9) at the 5% significance level.
In addition, we also take the random-walk model as the null
To demonstrate the applicability of the proposed model, we
hypothesis and note that zt = yt − yt−1 should be uncorre-
apply the model (2) with a smooth Markov trend to U.S. quar-
lated with all past zt−i under the null. We then regress zt on
terly real GDP. Leading models for GDP or gross national
zt−1 , . . . , zt−k for k = 1, . . . , 4 with a constant term and check
product (GNP) include the trend-stationary models, unit-root
the joint significance of the coefficients of zt−i using the Wald
models, and regime-switching models. For example, Blanchard
(1981), Kydland and Prescott (1980), and Clark (1987) sug- test. Note that there will be no unidentified nuisance parame-
gested that the logarithm of real GNP is trend stationary, ters under this framework. A similar approach was also taken
whereas Nelson and Plosser (1982) and Campbell and Mankiw by Tsay (1989) to test for threshold AR models. The resulting
(1987) argued that real GNP contains a unit root. In contrast, Wald statistics are 29.77, 30.88, 32.07, and 35.91, which are all
Hamilton (1989), Lam (1990), and Kim and Nelson (1999) significant at the 1% level under the χ 2 (k) distribution. We thus
adopted a Markov switching model to describe GNP or GDP. reject the null hypothesis that the data series is a pure random
Because the proposed model constitutes intermediate cases walk.
between these two models, it would be interesting to know Given that the data are neither an ARIMA process nor a pure
whether it is capable of accounting for the fluctuations of U.S. random walk, we now proceed to test other hypotheses by the
real GDP. Wald test. In particular, as discussed by Engel and Hamilton
The data used are seasonally adjusted, quarterly U.S. real (1990), the proposed model would be a simple mixture model
GDP from 1947:I–2002:I with 221 observations. The dataset if the probability that st = 0 or 1 is independent of the previous
is taken from the AREMOS databank of the Taiwan Ministry state. This amounts to testing the null hypothesis p00 + p11 = 1.
of Education. We take log GDP as yt and estimate an array
of IRS(1; m, n) models with a Markov trend and 0 ≤ m, n ≤ 4.
Table 2. QMLEs of the Proposed IRS Model
The parameters are estimated using the algorithm described in
Section 4.1 and the Appendix. This algorithm is initialized by Parameter Estimate Standard error t-statistic
a broad range of random initial values. The covariance matrix α̂0 −.00832 .00306 −2.718∗
of θ̂ is −H(θ̂)−1 , where H(θ̂ ) is the Hessian matrix of the log- α̂1 .01989 .00216 9.200∗
likelihood function evaluated at the QMLE θ̂ . ψ̂1 .24265 .24931 .973
Among all the models considered, both AIC and SIC select ψ̂2 −.51115 .19204 −2.662∗
ϕ̂1 −.28466 .13282 −2.143∗
the IRS(1; 2, 2) model. The estimation results are summarized ϕ̂2 −.63296 .22149 −2.858∗
in Table 2; the estimated transition probabilities are p̂∗11 ≈ .91 σ̂ε .00815 .00244 3.340∗
and p̂∗00 ≈ .61. We first apply the simulation-based approach in p̂00 .60647 .11364
p̂11 .90882 .03479
Section 4.2 to test p11 . We estimate an array of ARIMA( p, 1, q)
Log-likelihood = 700.336
models with p and q no greater than 4; both AIC and SIC select AR roots: .121 ± .704i AIC = −1,382.67
the ARIMA(1, 1, 0) model, MA roots: −.142 ± .783i SIC = −1,352.13
yt = .008 + .347yt−1 + ut , (9) NOTE: t -statistics with an asterisk are significant at the 5% level.
Kuan, Huang, and Tsay: Unobserved-Component Model 451

The Wald statistic of this hypothesis is 18.79, which is also sig- From Table 2, we see that the estimated quarterly growth
nificant at the 1% level under the χ 2 (1) distribution. The rejec- rates of U.S. real GDP are α0 = −.83% during the state of
tion of this hypothesis may justify our Markovian specification transitory shocks (recessions) and (α0 + α1 ) = 1.15% during
of the state variable. We also conduct some diagnostic checks the state of permanent shocks (expansion). The expected du-
on the estimated model, including the Q test of Ljung and Box rations of recession and expansion can be calculated from the
(1978) on serial correlations and the LM test of Engle (1982) for transition probabilities, 1/(1 − .61) = 2.6 quarters for reces-
autoregressive conditional heteroscedasticity (ARCH) effects. sion and 1/(1 − .91) = 11 quarters for expansion. According
These tests are applied to the residual series ε̂t . The statis- to NBER dating, the average growth rates for recession and ex-
tics are Q(20) = 17.244, Q(30) = 25.490, ARCH(2) = 2.049, pansion are −.35% and 1.10%, and the average durations are
and ARCH(4) = 2.882. They are all insignificant even at the 3.4 and 18.4 quarters. Compared with NBER dating, our results
10% level, under the χ 2 (20), χ 2 (30), χ 2 (2), and χ 2 (4) distri- lead to a shorter expected duration for expansion and “deeper”
butions. These tests suggest that there are no significant serial recessions. We also apply the model of Hamilton (1989) to
correlations or conditional heteroscedasticity in residuals. The the dataset. The estimation Gauss program is taken from
proposed model thus fits the GDP data well. C. R. Nelson’s website and is initialized by 100 initial values
Figure 4 plots the estimated smoothing probabilities of (see
st = 0, where the shaded areas signify the recession peri- and the program HMT4_KIM.OPT). Unfortunately, the esti-
ods identified by NBER and the solid (dashed) lines de- mation results fail to provide reasonable parameter estimates
note the peaks (troughs). We find that there are 33 periods for the data. Kim and Nelson (1999, p. 78) also reported a sim-
(about 15.2% of the sample) with estimated smoothing prob- ilar problem when a different dataset was used. This may not
ability P(st = 0|Z T ; θ̂ ) > .5. This shows that unit-root nonsta- be surprising, because Boldin (1996) noticed that Hamilton’s
tionarity is more likely to prevail in about 85% of the sample result is sensitive to the sample period.
periods, yet stationarity dominates in the remaining periods. We also compute the expected trend line (smooth Markov
We also observe that the nonstationarity (stationarity) periods trend) as
match the NBER dating of expansions (recessions) closely. It

is also worth mentioning that the proposed model successfully α̂0 t + α̂1 P(si = 1|Z T ; θ̂ )
identifies the recession period starting in March 2001, which is i=1
at the end of the sample span. Hence the innovations in expan-
sion (recession) are more likely to have a permanent (transitory) ˆ
+ α̂1 (B)−1 ˆ
(B) 1 − P(st = 1|Z T ; θ̂)
effect. These results together suggest the following features of 
and the expected stochastic trend component as ti=1 P(si = 1|
real GDP. First, the nonstationary characteristic and permanent
shocks do not appear all of the time, in contrast with the re- Z T ; θ̂)ε̂i , both based on the estimated smoothing probabili-
sult of unit-root models. Second, permanent shocks occur more ties P(si = 1|Z T ; θ̂ ). These two expected trend components to-
frequently than the assertion of trend-break models (cf. Perron gether capture the behavior of log(GDP) quite well, as can be
1989; Balke and Fomby 1991). Third, the shocks in the ex- seen in Figure 5.
pansion periods generate nonstationary patterns and hence are
more persistent than those in the recession periods. This is com- 6. CONCLUSIONS
patible with the conclusion of Beaudry and Koop (1993), who
found that positive shocks to GDP are more persistent than neg- In this article we have proposed a class of unobserved-
ative shocks. component models with switching permanent and transitory
innovations. The model has several interesting features. First,

Figure 4. Estimated Smoothing Probabilities of st = 0 for U.S. Quar-

terly Real GDP From 1947:I to 2002:I. Figure 5. The Expected Trend Line (- - - - -) in U.S. Real GDP.
452 Journal of Business & Economic Statistics, October 2005

it admits both deterministic and stochastic trends. Second, it can The Kalman filter for (8) consists of the following updating
describe both stationary and nonstationary characteristics over equations:
different time periods. The long-term effects of corresponding (i,j) (i)
innovations thus may alternate from time to time. Third, it al-  t|t−1 = F t−1|t−1 ,
lows for endogenous breaks in the trend function such that dif-
ϒ t|t−1 = Fϒ t−1|t−1 F + Q∗ ,
(i,j) (i)
ferent trending patterns are directly linked to the shocks with
distinct effects. Fourth, when there are trend breaks, the tran- ηt|t−1 = zt − µ0 − µ1 ξ t−1,j 1 − Hj  t|t−1 ,
(i,j) (i,j)
sitions between trend segments are smooth. Thus the proposed (A.1)
ft|t−1 = Hj ϒ t|t−1 H j ,
(i,j) (i,j)
models bridge the gap between trend-stationary and unit-root
models and is able to accommodate both trend-reverting and (i,j) −1 (i,j)
 t|t =  t|t−1 + ϒ t|t−1 H j ft|t−1 ηt|t−1 ,
(i,j) (i,j) (i,j)
trend-disturbing behaviors.
Application of the proposed model to U.S. quarterly real  (i,j) −1  (i,j)
ϒ t|t = I − ϒ t|t−1 H j ft|t−1 Hj ϒ t|t−1 ,
(i,j) (i,j)
GDP suggests that it is a useful analytical tool in describing
the data characteristics. In particular, it shows that unit-root where Q∗ is the variance–covariance matrix of ε t that involves
nonstationarity is more likely to prevail in more than 80% of the unknown parameter σε2 . Note that this algorithm calculates
the sample periods and that these periods closely match the ex- (i,j) (i,j)
(2r+1 × 2r+1 ) forecasts of ( t|t , ϒ t|t ) for each date t, corre-
pansion periods dated by NBER. Thus the shocks in expansion
sponding to every possible value for i and j.
are more likely to be permanent. This result differs from that ( j) ( j)
The (2r+1 ) elements of ( t|t , ϒ t|t ) can be obtained by taking
of unit-root (trend-stationary) models in that the shocks may
weighted averages over all state values at t − 2,
not always be permanent (transitory). The fact that permanent
shocks occur frequently is also different from the assertion of r+1
( j) P(s∗t−2 = i, s∗t−1 = j|Z t ) (i,j)
trend-break models. The proposed model thus may serve as an  t|t ≈  t|t ,
alternative for modeling economic time series. P(s∗t−1 = j|Z t )
( j) P(s∗t−2 = i, s∗t−1 = j|Z t )
P(s∗t−1 = j|Z t )
This article was originally titled “A Component-Driven  (i,j)  ( j) (i,j)  ( j) (i,j)  
× ϒ t|t +  t|t −  t|t  t|t −  t|t ,
Model for Regime Switching and Its Empirical Evidence.” The
authors thank Y. T. Chen, R. Engle, C. W. J. Granger, C. C. Hsu, where, by Bayes’ theorem, P(s∗t−2 = i, s∗t−1 = j|Z t ) can be cal-
C. H. Huang, C. J. Kim, C. S. Mao, K. Tanaka, and T. Teräsvirta culated as
for valuable suggestions on early drafts of the article. They are
also indebted to the associate editor and two anonymous ref- P(s∗t−2 = i, s∗t−1 = j|Z t )
erees, whose comments led to a much improved version of P(s∗t−2 = i, s∗t−1 = j|Z t−1 )f (zt |s∗t−2 = i, s∗t−1 = j, Z t−1 )
the article. All remaining errors are the authors’. Kuan grate- = .
f (zt |Z t−1 )
fully acknowledges research support from the National Science
Council of the Republic of China (grants NSC-89-2415-H-001- (A.3)
037 and -071). The work of Tsay is supported by the National The first term of the numerator in (A.3) can be easily computed
Science Foundation and the Graduate School of Business, Uni- as
versity of Chicago.
P(s∗t−2 = i, s∗t−1 = j|Z t−1 ) = p∗ij P(s∗t−2 = i|Z t−1 ), (A.4)

APPENDIX: ESTIMATION ALGORITHM the second term of the numerator is

f (zt |s∗t−2 = i, s∗t−1 = j, Z t−1 )
To estimate the state–space representation in (8) of the pro-
posed model, we derive an algorithm following Kim (1994).  (i,j) −1/2 1 (i,j) 2 (i,j) −1
= (2π)−T/2  ft|t−1  exp − ηt|t−1 ft|t−1 ,
We denote the expectation of the variable Xt , conditional on the 2
information available up to time s and the realized state values
and the denominator can be expressed as
s∗t−2 = i and s∗t−1 = j, as Xt|s . For example, conditional on the

information up to time t, the mean and variance of the transition r+1 2r+1
component  t are f (zt |Z t−1 ) = f (zt |s∗t−2 = i, s∗t−1 = j, Z t−1 )
i=1 j=1
 t|t = E( t |s∗t−1 = j, s∗t−2 = i, Z t )
× P(s∗t−2 = i, s∗t−1 = j|Z t−1 ).
Finally, summing over the values of s∗t−2 , the filter in (A.3) be-
 (i,j)  (i,j)  
= E  t −  t|t  t −  t|t s∗t−1 = j, s∗t−2 = i, Z t .
ϒ t|t comes
( j)
We also let Xt|t denote the expectation of Xt , conditional on the P(s∗t−1 = j|Z ) =
P(s∗t−2 = i, s∗t−1 = j|Z t ). (A.5)
information up to time t and the realized state value s∗t−1 = j. i=1
Kuan, Huang, and Tsay: Unobserved-Component Model 453

Thus, with the initial values ( r|r , ϒ r|r ) and P(s∗r−1 = i|Z r ),
(i) (i)
Using the filtering probability P(s∗T = i|Z T ) as the initial value,
we can iterate (A.1)–(A.3) to obtain P(s∗t−1 = i, s∗t = j|Z t ) we can iterate equations (A.5)–(A.7) backward for t = T −
from (A.4) for t = r, r + 1, . . . , T. Then for each t, the desired 1, . . . , p + 1. Consequently, the desired smoothing probability
filtering probability is for each t is

r+1 P(st = 1|Z T ) = P(s∗t = i|Z T ),
P(st = 1|Z ) = t
P(s∗t−1 = i, s∗t = j|Z t ), (A.6) i
j i=1 where the summation is taken over all i’s associated with st = 1.
where the first summation is taken over all j’s associated with We also have that P(st = 0|Z T ) = 1 − P(st = 1|Z T ).
st = 1. Clearly, P(st = 0|Z t ) = 1 − P(st = 1|Z t ).
Following Kim and Nelson (1999), we set the initial val- [Received April 2003. Revised March 2005.]
(i) (i)
ues ( r|r , ϒ r|r ) to the unconditional mean and variance,
 r|r = 0 and REFERENCES
 (i)  Balke, N., and Fomby, T. B. (1991), “Shifting Trends, Segmented Trends, and
vec ϒ r|r = (I − F ⊗ F)−1 vec(Q∗ ), Infrequent Permanent Shocks,” Journal of Monetary Economics, 28, 61–85.
Beaudry, P., and Koop, G. (1993), “Do Recessions Permanently Change Out-
where the “vec” operator stacks the columns of a matrix into put?” Journal of Monetary Economics, 31, 149–163.
a vector, I is the identity matrix, and ⊗ denotes the Kronecker Blanchard, O. J. (1981), “What Is Left of the Multiplier-Accelerator?” Ameri-
product. We also follow Hamilton (1989, 1994) and set the ini- can Economic Review Proceedings, 71, 150–154.
Boldin, M. D. (1996), “A Check on the Robustness of Hamilton’s Markov
tial value P(s∗r−1 |Z r ) to its limiting unconditional counterpart, Switching Model Approach to the Economic Analysis of the Business Cy-
the (2r+1 + 1)st column of the matrix (A A)−1 A , where cle,” Studies in Nonlinear Dynamics and Econometrics, 1, 35–46.

Bradley, M. D., and Jansen, D. W. (1997), “Nonlinear Business Cycle Dy-
I − P∗ namics: Cross-Country Evidence on the Persistence of Aggregate Shocks,”
A= Economic Inquiry, 35, 495–509.
1 Campbell, J. H., and Mankiw, N. G. (1987), “Are Output Fluctuations Transi-
tory?” Quarterly Journal of Economics, 102, 857–880.
and 1 is the 2r+1 -dimensional vector of 1’s (see Hamilton 1994, Clark, P. K. (1987), “The Cyclical Component of U.S. Economic Activity,”
p. 684, for details). Quarterly Journal of Economics, 102, 798–814.
Cochrane, J. H. (1988), “How Big Is the Random Walk in GNP?” Journal of
We also follow Kim (1994) to calculate the smoothing proba- Political Economy, 96, 893–920.
bilities P(st |Z T ) for t ≤ T, which are the optimal forecasts of st Davies, R. B. (1977), “Hypothesis Testing When a Nuisance Parameter Is
based on all information in the sample. Observe that Present Only Under the Alternative,” Biometrika, 64, 247–254.
(1987), “Hypothesis Testing When a Nuisance Parameter Is Present
P(s∗t = i|s∗t+1 = j, Z T ) Only Under the Alternative,” Biometrika, 74, 33–43.
Dickey, D. A., and Fuller, W. A. (1979), “Distribution of the Estimators for

= P(s∗t = i|s∗t+1 = j, Z t+1 ) Autoregressive Time Series With a Unit Root,” Journal of the American Sta-
tistical Association, 74, 427–431.

× P(zT , . . . , zt+2 |s∗t = i, s∗t+1 = j, Z t+1 )
Elwood, S. K. (1998), “Is the Persistence of Shocks to Output Asymmetric?”
Journal of Monetary Economics, 41, 411–426.
 −1 Engel, C., and Hamilton, J. D. (1990), “Long Swings in the Dollar: Are They
× P(zT , . . . , zt+2 |s∗t+1 = j, Z t+1 ) . in the Data and Do Markets Know It?” American Economic Review, 80,
In the current context, Engle, R. F. (1982), “Autoregressive Conditional Heteroskedasticity With Esti-
mates of the Variance of U.K. Inflation,” Econometrica, 50, 987–1008.
P(zT , . . . , zt+2 |s∗t = i, s∗t+1 = j, Z t+1 ) Engle, R. F., and Smith, A. D. (1999), “Stochastic Permanent Breaks,” The Re-
view of Economics and Statistics, 25, 475–511.
= P(zT , . . . , zt+2 |s∗t+1 = j, Z t+1 ), Evans, M., and Wachtel, P. (1993), “Inflation Regimes and the Sources of Infla-
tion Uncertainty,” Journal of Money, Credit, and Banking, 25, 475–511.
so that P(s∗t = i|s∗t+1 = j, Z T ) = P(s∗t = i|s∗t+1 = j, Z t+1 ). Note, Granger, C., and Swanson, N. R. (1997), “An Introduction to Stochastic Unit-
Root Processes,” Journal of Econometrics, 80, 35–62.
however, that the foregoing condition does not hold in the work Gray, S. F. (1996), “Modeling the Conditional Distribution of Interest Rates as
of Kim (1994). It follows that a Regime-Switching Process,” Journal of Financial Economics, 42, 27–62.
Hamilton, J. D. (1989), “A New Approach to the Economic Analysis of Non-
P(s∗t = i|Z T ) stationary Time Series and the Business Cycle,” Econometrica, 57, 357–384.
(1994), Time Series Analysis, Princeton, NJ: Princeton University
2 Press.
= P(s∗t+1 = j|Z T )P(s∗t = i|s∗t+1 = j, Z t+1 ) Hansen, B. E. (1996), “Inference When a Nuisance Parameter Is not Identified
Under the Null Hypothesis,” Econometrica, 64, 413–430.
j=1 Kim, C. J. (1994), “Dynamic Linear Models With Markov-Switching,” Journal
of Econometrics, 60, 1–22.
2 Kim, C.-J., and Nelson, C. R. (1999), State Space Models With Regime Switch-
= P(s∗t+1 = j|Z T ) ing, Classical and Gibbs Sampling Approaches With Applications, Cam-
bridge, MA: MIT Press.
j=1 Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., and Shin, Y. (1992), “Testing
the Null Hypothesis of Stationarity Against the Alternative of a Unit Root,”
P(s∗t+1 = j|s∗t = i, Z t+1 )P(s∗t = i|Z t+1 ) Journal of Econometrics, 54, 159–178.
P(s∗t+1 = j|Z t+1 ) Kydland, F. E., and Prescott, E. C. (1980), “A Competitive Theory of Fluctua-
tions and the Feasibility and Desirability of Stabilization Policy,” in Rational
Expectations and Economic Policy, ed. S. Fischer, Chicago: University of
p∗ij P(s∗t+1 = j|Z T )
Chicago Press, pp. 169–198.
= P(s∗t = i|Z t+1
) . (A.7) Lam, P. S. (1990), “The Hamilton Model With a General Autoregressive Com-
P(s∗t+1 = j|Z t+1 ) ponent,” Journal of Monetary Economics, 26, 409–432.
454 Journal of Business & Economic Statistics, October 2005

Ljung, G. M., and Box, G. E. P. (1978), “On a Measure of Lack of Fit in Time Perron, P. (1989), “The Great Crash, the Oil Price Shock, and the Unit-Root
Series Models,” Biometrika, 65, 297–303. Hypothesis,” Econometrica, 57, 1361–1401.
McCabe, B. P. M., and Tremayne, A. R. (1995), “Testing a Time Series for Rappoport, P., and Richelin, L. (1989), “Segmented Trends and Non-Stationary
Difference Stationarity,” The Annals of Statistics, 23, 1015–1028. Time Series,” Economic Journal (supplement 1989), 168–177.
McCulloch, R. E., and Tsay, R. S. (1994), “Bayesian Inference of Trend- and
Difference-Stationarity,” Econometric Theory, 10, 596–608. Tong, H. (1990), Nonlinear Time Series, Oxford, U.K.: Oxford University
Nelson, C. R., and Plosser, C. I. (1982), “Trends and Random Walks in Macro- Press.
economic Time Series: Some Evidence and Implications,” Journal of Mone- Tsay, R. S. (1989), “Testing and Modeling Threshold Autoregressive
tary Economics, 10, 139–162. Processes,” Journal of the American Statistical Association, 84, 231–240.