Notes on Trend and Cycle Components

Garey Ramey
University of California, San Diego

February 2015

1 Denitions
Suppose the series fYt g has a trend component fYttr g. Write
Yt = Yttr
which denes the cycle component as

Yt
= Yttr Ybt ;
Yttr

Take logs:

Yt
Ybt = tr :
Yt

Now dene

ln Yt = ln Yttr + ln Ybt :
yt = ln Yt ;

Then we may write:

yttr = ln Yttr ;

y^t = ln Ybt = ln Y

ln Yttr :

yt = yttr + y^t :
Note that y^t approximates the percentage deviation of output from trend. To see this,
consider the Taylor expansion around the trend component:
ln Yt ' ln Yttr +

d ln Yt
dYt

(Yt
Yt =Yttr

Yttr ) = ln Yttr +

which implies
ln Yt

ln Yttr = y^t '

1

Yt

Yttr
Yttr

1
(Yt
Yttr

Yttr );

This is a good approximation given that values of y^t are small empirically.

2 Modeling the trend component

a Polynomial trend
The trend component is assumed to be a polynomial in t:

yttr =
Example: Log-linear trend (K = 1):
yttr =

XK

kt

k=0

1t

Yttr = Y0 t ;

where
= e 1:

Y0 = e 0 ,

Note that a polynomial can be used to approximate any deterministic trend. Moreover,
estimation and inference are straightforward.
b Hodrick-Prescott (HP) lter
For a given series fyt gTt=1 , the trend component fyttr gTt=1 is chosen to solve
min

fyttr gT
t=1

XT

t=1

[(yt

tr
yttr )2 + ((yt+1

yttr )

(yttr

yttr 1 ))2 ]:

The rst term captures how closely trend tracks the data, while the second term captures
smoothness of the trend in terms of second dierences. The smoothing parameter
the weight assigned to tracking vs. smoothness in the minimization problem. As

governs
rises,

the trend becomes smoother, and correspondingly more variations are assigned to the cycle
component. As

! 1,positive second dierences are not allowed, and yttr becomes linear

in t.
The choice

= 1600 is reasonable for business cycle analysis using quarterly data, as it

works to remove cyclical movements with periods in excess of 4-6 years.

2

The HP lter is very exible, and takes no stand on the form of the trend. It can induce
spurious volatility into the cycle component, however.
c Stochastic trend
The trend component is assumed to be a random process.
Example: Suppose fyt g is a random walk with drift:
yt = yt

+ "t ;

where "t is white noise. This is a "unit root" specication. Dene:

yttr = yt

+ ;

y^t = "t :

In this case the trend is the linear projection, and the cycle is the innovation. Moreover,
innovations to output have a permanent eect on the trend.
To obtain a stationary representation of fyt g, take the rst dierence:
yt = yt
Note that

yt

+ "t :

yt approximates the growth rate of output:

yt = ln Yt

ln Yt

'

Yt

Yt
Yt

Thus, for a unit root specication, we consider output growth rates rather than log levels.
There are numerous other approaches to modeling trend and cycle components, e.g.,
state-space and frequency domain methods.