11-FinEcon-Time Series Page 310 Wednesday, February 4, 2004 12:58 PM
310 The Mathematics of Financial Modeling and Investment Management
Zt =∆Yt =∆(∆Xt) =(Xt – Xt – 1 )–(Xt – 2 – Xt – 3 )
=Xt –Xt – 1 – Xt – 2 +Xt – 3
Differencing can be written in terms of the lag operator as
∆Xtd =( 1 – L)dXt
A difference stationary series is a series that is transformed into a
stationary series by differencing. A difference stationary series can be
written as
∆Xt μ ε+ = ()t
Xt = Xt – 1 ++με()t
where ε(t) is a zero-mean stationary process and μis a constant. A trend
stationary series with a linear trend is also difference stationary, if spac-
ings are regular. The opposite is not generally true. A time series is said
to be integrated of order n if it can be transformed into a stationary
series by differencing n times.
Note that the concept of integrated series as defined above entails
that a series extends on the entire time axis. If a series starts from a set
of initial conditions, the difference sequence can only be asymptotically
stationary.
There are a number of obvious differences between trend stationary
and difference stationary series. A trend stationary series experiences
stationary fluctuation, with constant variance, around an arbitrary
trend. A difference stationary series meanders arbitrarily far from a lin-
ear trend, producing fluctuations of growing variance. The simplest
example of difference stationary series is the random walk.
An integrated series is characterized by a stochastic trend. In fact, a
difference stationary series can be written as
t – 1
Xt =μt + ∑ε() s +ε()t
s + 0
The difference Xt –X*t between the value of a process at time t and
the best affine prediction at time t – 1 is called the innovation of the pro-
cess. In the above linear equation, the stationary process ε(t) is the inno-
vation process. A key aspect of integrated processes is that innovations