Trend Cycle
Trend Cycle
Eric Zivot
April 22, 2005
1 Introduction
A convenient way of representing an economic time series yt is through the so-called
trend-cycle decomposition
yt = T Dt + Zt (1)
where T Dt represents the deterministic trend and Zt represents the stochastic, and
possibly cyclic, noise component. For simplicity, the deterministic trend takes the
form of a simple time trend
T Dt = κ + δt (2)
and the noise component has a finite order ARIMA(p,d,q) representation
Zt = φ(L)−1 θ(L)εt
= ψ(L)εt
P
where ψ(L) = φ(L)−1 θ(L) = ∞ k
k=0 ψ k L with ψ 0 = 1 and ψ(1) 6= 0. Alternatively, if
yt is difference stationary then φ(L) can be factored as φ(L) = (1 − L)φ∗ (L) where
1
φ∗ (z) = 0 has all p − 1 roots outside the unit circle. In this case, ∆Zt has the
stationary ARMA(p − 1, q) representation
Let φ(L)Zt = εt with φ(L) = 1 − φ1 L − φ2 L2 . Assume that φ(z) = 0 has one root
equal to unity and the other root real valued with absolute value less than 1. Factor
φ(L) so that
φ(L) = (1 − φ∗ L)(1 − L) = φ∗ (L)(1 − L)
where φ∗ (L) = 1 − φ∗ L with |φ∗ | < 1. Then
yt = yt−1 + ∆yt
2
Similarly, the level at time t + h may be represented as
∂yt+h X ∗
∞
lim = ψj = ψ∗ (1).
h→∞ ∂εt
j=1
1. Since ∂y t
∂εt
= 1 it follows that ψ∗ (1) can also be interpreted as the long-run effect
of a shock relative to the immediate effect of a shock.
3. If ψ∗ (1) = 0 then Zt ∼ I(0). To see this suppose Zt ∼ I(0) and has the
Wold representation Zt = ψ(L)εt with ψ(1) 6= 0. Then ∆Zt = (1 − L)Zt =
(1 − L)ψ(L)εt = ψ∗ (L)εt where ψ∗ (L) = (1 − L)ψ(L). It follows that ψ∗ (1) =
(1 − 1)ψ(1) = 0.
Then
Notice that forecasting an I(1) process proceeds from the most recent observation.
3
Example 2 Forecasting from an AR(1) model for ∆yt
where |φ| < 1. Using the chain-rule of forecasting, the h-step ahead forecast of
∆yt+h based on information at time t is
X
h
yt+h|t = yt + [µ + φs (∆yt − µ)] (6)
s=1
X
h
= yt + hµ + (∆yt − µ) φh
s=1
Zt = T St + Ct (7)
The stochastic trend, T St , captures shocks that have a permanent effect on the level
of yt and the stationary component, Ct , captures shocks that only have a temporary
effect on the level of yt . Given the decomposition of Zt , the representation for yt
becomes
yt = T Dt + T St + Ct (8)
where T Dt + T St measures the overall or total trend and Ct represents the deviations
about the trend.
The decomposition of Zt in (7) is not unique. In fact, there are an infinite number
of such combinations depending on how T St and Ct are defined. Two decompositions
have been popular in the empirical literature: the Beveridge-Nelson (BN) decom-
position; and the orthogonal unobserved components (UC0) decomposition. Both
decompositions define T St as a pure random walk. They primarily differ in how
they model the serial correlation in ∆Zt . The BN decomposition uses an unrestricted
ARMA(p, q) model for ∆Zt , and the UC0 model uses a restricted ARMA(p,q) for
∆Zt .
4
2.1 The Beveridge-Nelson Decomposition
Beveridge and Nelson (1980) proposed a definition of the permanent component of
an I(1) time series yt with drift µ as the limiting forecast as horizon goes to infinity,
adjusted for the mean rate of growth over the forecast horizon,
The stochastic part of the permanent component (11), BNt , is referred to as the BN
trend. The implied cycle at time t is then
Ct = yt − T Dt − BNt
The derivation of the BN trend (11) relies on the following algebraic result.
P∞
Proposition 3 Let ψ(L) = k=0 ψk Lk with ψ0 = 1. Then
e
ψ(L) = ψ(1) + (1 − L)ψ(L), (12)
X∞
ψ(1) = ψk ,
k=0
X
∞ X
∞
e
ψ(L) = ψ ej = −
e j Lj , ψ ψk .
j=0 k=j+1
P P∞ e
In addition, if ∞ k=0 k|ψ k | < ∞ (1-summability) then k=0 |ψ k | < ∞. 1-summability
is satisfied by all covariance stationary ARMA(p,q) processes. For an algebraic proof,
see Hamilton (1993) pages 534 and 535.
Now, consider the Wold representation for ∆yt given in (10). Then by recursive
substitution
X
t
∗
yt = y0 + δt + ψ (L) εj
j=1
5
Applying (12) to ψ∗ (L)
³ ∗
´X
t
yt ∗ e
= y0 + δt + ψ (1) + (1 − L)ψ (L) εj
j=1
X
t
= y0 + δt + ψ∗ (1) εj + e
εt − e
ε0
j=1
= T Dt + T St + Ct
X
t
∗
yt+h|t = y0 + δ(t + h) + ψ (1) εj + e
εt+h|t
j=1
The limiting forecast as horizon goes to infinity, adjusted for mean growth, is
X
t
∗
lim yt+h|t − δh = y0 + δt + ψ (1) εj + lim e
εt+h|t
h→∞ h→∞
j=1
X
t
∗
= y0 + δt + ψ (1) εj
j=1
= T Dt + BNt
as limh→∞ e
εt+h|t = 0 since e
εt+h is a mean-zero stationary process. We have just proved
∗ Pt
that ψ (1) j=1 εj is the BN trend, and that the BN trend follows a pure random
walk.
6
For the MA(1) model, the Wold representation for ∆yt has the simple form ∆yt =
δ + ψ∗ (L)εt where ψ∗ (L) = 1 + ψ∗1 L and ψ∗1 = 0.03. Straightforward calculations give
ψ∗ (1) = 1 + ψ∗1 = 1.03
∗ X∞
e
ψ0 = − ψ∗j = −ψ∗1 = −0.03
j=1
X∞
e∗ = −
ψ ψ∗j = 0, j = 1, 2 . . .
j
j=k+1
X
t
= y0 + 0.008t + 1.3 εj − 0.3εt
j=1
so that
T Dt = y0 + 0.008t
X t
BNt = 1.3 εj
j=1
Ct = −0.3εt
∂yt ∂yt+s
Note that ∂εt
= 1 and ∂εt
= 1.3 for s > 0.
7
To compute an estimate of ψ∗ (1) from the ARMA(p,q) model, solve for the wold
representation
φ(L)∆yt = θ(L)εt
⇒ ∆yt = φ(L)−1 θ(L)εt = ψ∗ (L)εt
X
t
T Dt + BNt = y0 + 0.816t + 1.276 ε̂j
j=1
The previous example shows that the computation of the BN decomposition from
an estimated ARMA(p,q) model is straightforward but somewhat tedious. The fol-
lowing example shows that the computation of the BN decomposition from an AR(1)
model is simple and elegant.
From the h−step ahead forecast for yt given in (6), it is easy to compute the BN
trend for yt :
¡ ¢
T Dt + BNt = lim yt+h|t − hδ
h→∞
X
h
= yt + (∆yt − δ) lim φh
h→∞
s=1
φ
= yt + (∆yt − δ)
1−φ
The cycle component is then
Ct = yt − T Dt − BNt
φ
= (∆yt − µ)
1−φ
Notice that the cycle inherits the behavior of the AR(1) model for ∆yt .
8
Morley (2002) shows how the BN decomposition for an AR(1) model for ∆yt may
be extended to any model for ∆yt that can be represented in state space form. In
particular, suppose ∆yt − µ is a linear combination of the elements of the m × 1 state
vector αt : £ ¤
∆yt − µ = z1 z2 · · · zm αt
where zi (i = 1, . . . , m) is the weight of the ith element of αt in determining ∆yt − µ.
Suppose further that
such that all of the eigenvalues of T have modulus less than unity, and T is invertible.
Then, Morley shows that
£ ¤
T Dt + BNt = yt + z1 z2 · · · zm T(Im −T)−1 at|t (14)
£ ¤
Ct = yt − T Dt − BNt = − z1 z2 · · · zm T(Im −T)−1 at|t
where at|t denotes the filtered estimate of αt from the Kalman filter recursions.
yt = µt + Ct (15)
µt = α + µt−1 + εt , εt ∼ iid(0, σ 2ε )
φ(L)Ct = θ(L)η t , η t ∼ iid(0, σ 2η )
9
Clark considered the UC-ARIMA(2,0) model
yt = µt + Ct (16)
µt = α + µt−1 + εt , εt ∼ iid(0, σ 2ε )
Ct = φ1 Ct−1 + φ2 Ct−2 + η t , η t ∼ iid(0, σ 2η )
cov(εt , η t ) = 0
The model may be put in state space form in several way. For example, one may
define the transition equation for the state vector αt = (µt , Ct , Ct−1 )0 as
µt 1 0 0 µt−1 α 1 0 µ ¶
Ct = 0 φ1 φ2 Ct−1 + 0 + 0 1 εt
ηt
Ct−1 0 1 0 Ct−2 0 0 0
a1,0 = 0
var(α1,0 ) = κ · 106
The distribution for α2,0 may be based on its unconditional stationary distribution.
Therefore, the initial state distribution is characterized by
a0 = 0
µ ¶
κ · 106 0
P0 =
0 P2
10
Defining the state variables as αt = (Ct , Ct−1 )0 , the transition equation is
µ ¶ µ ¶µ ¶ µ ¶
Ct φ1 φ2 Ct−1 ηt
= +
Ct−1 1 0 Ct−2 0
The corresponding measurement equation is
∆yt = ( 1 −1 )αt + εt
Since αt is covariance stationary, the distribution of the initial state vector α0 may
be determined in the usual way.
Proposition 8 Any UC-ARIMA model of the form (15) with a specified correlation
between εt and η t is observationally equivalent to an ARMA model for ∆yt with non-
linear restrictions on the parameters. The restricted ARMA model for ∆yt is called
the reduced form of the UC-ARIMA model.
Proposition 9 (Lippi and Reichlin (1992) JME) ψ∗ (1) computed from the reduced
form ARMA model for ∆yt based on a UC-ARIMA model for yt is always less than
1.
11
where the ARMA(0,1) parameters θ and σ 2ς are nonlinearly related to the UC-ARIMA
parameters σ 2η and σ 2ε , and the error term ς t encompasses the structural shocks η t
and εt .
Notice that the reduced form ARMA(0,1) model has two parameters θ and σ 2ς ,
and the structural random walk plus noise model also has two parameters σ 2ε and σ 2η .
Since the reduced form and structural models have the same number of parameters the
order condition for identification is satisfied. If εt and η t were allowed to be correlated
then there would be three structural parameters (two variances and a covariance) and
only two reduced form parameters, and the order condition for identification would
not be satisfied. Hence, setting cov(εt , η t ) = 0 is an identifying restriction in this
model.
The mapping from UC-ARIMA parameters to the reduced form parameters is de-
termined as follows. The first order autocorrelation for the reduced form ARMA(0,1)
is
θ
ρ∗1 =
1 + θ2
and for the UC-ARIMA it is
γ∗ −1
ρ1 = 1∗ =
γ0 q+2
Setting ρ∗1 = ρ1 and solving for θ gives
p
−(q + 2) ± (q + 2)2 − 4
θ=
2
The invertible solution is
p
−(q + 2) + q2 + 4q
θ= , θ < 0.
2
Notice that the MA coefficient θ for the reduced form ARMA(0,1) model is restricted
to be negative. If q = 0, then θ = −1. Similarly, matching variances for the two
models gives
σ 2η
σ 2ς = .
−θ
Finally, note that since θ < 0 it follows that ψ∗ (1) = 1 + θ < 1.
12
variance of the kth difference of GDP to the variance of the first difference. to be more
specific, let yt denote the log-level of real GDP, µ = E[∆yt ] and define the variance
of the kth difference as
1
Vk = var(yt+k − yt − kµ)
k
Cochrane’s variance ratio statistic is then defined as
Vk
Rk = .
V1
Example 11 Pure random walk model
yt = yt−1 + εt , εt ∼ iid(0, σ 2ε ).
P P Pt+k
Then yt = y0 + tj=1 εj , yt+k = y0 + t+k j=1 εt and yt+k − yt = j=t+1 εj so that
Vk = V1 = σ 2ε and Rk = 1 for all values of k.
To interpret the limiting form of the variance ratio statistic we require the following
results.
13
P∞
Proposition 13 Let ∆yt = ψ∗ (L)εt , εt ∼ iid(0, σ 2ε ) and j=0 j|ψ∗j | < ∞. Then as
T →∞ √ d
T ∆yt → N(0, σ 3ε ψ∗ (1)2 )
√
where σ 3ε ψ∗ (1)2 is the asymptotic or long-run variance of T ∆yt .
Combining the results in Propositions, the limiting form of the variance ratio
statistic, R, may be re-expressed as
σ 3ε ψ∗ (1)2
R =
γ ∗0
var(∆T St )
=
var(∆yt )
The expression above shows that R may be interpreted as the ratio of the variance
of the stochastic trend or permanent shock to the variance of the total change in yt .
Remarks
1. The variance ratio statistics are usually computed for various values of k and
reported in graphical form.
5. Campbell and Mankiw (1986) derive the following relationship between R and
ψ∗ (1). Define
σ2 σ2
ρ2 = 1 − 2ε = 1 − ∗ε
σ ∆y γ0
14
which is a measure of the predictability of ∆yt . Then it is easy to show that
s
R
ψ∗ (1) =
1 − ρ2
√
and so R is a lower bound for ψ∗ (1).
4 References
1. Clark, P. K. (1987). “The Cyclical Component of U.S. Economic Activity,” The
Quarterly Journal of Economics, 102, 797-814.
5. Morley, J., C.R. Nelson, and E. Zivot (2003). “Why are Beveridge-Nelson and
Unobserved-Components Decompositions of GDP so Different?”, The Review
of Economics and Statistics, 85(2), 235-243.
6. Stock, J.H. and M.W. Watson (1988). “Variable Trends in Economic Time
Series,” Journal of Economics Perspectives, 2, 147-174.
15