11-FinEcon-Time Series Page 303 Wednesday, February 4, 2004 12:58 PM
Financial Econometrics: Time Series Concepts, Representations, and Models 303
∞ ∞
adj[A ()L ]
--------------------------B ()Lεεεεt = ∑HiLiεt, with ∑Hi absolutely summable
det[A ()L] i = 1 i = 1
which demonstrates that the process xt is stationary.^1 As in the univari-
ate case, if one would consider the equations in 1/z, the same reasoning
applies but with roots strictly inside the unit circle.
A stationary ARMA(p,q) process is an autocorrelated process. Its
time-independent autocorrelation function satisfies a set of linear differ-
ence equations. Consider an ARMA(p,q) process which satisfies the fol-
lowing equation:
A 0 xt + A 1 xt – 1 + ...+ APxtP– = B 0 εεεεt+ B 1 εεεεt – 1 + ...+ Bqεεεεtq–
where A 0 = I. By expanding the expression for the autocovariance func-
tion, it can be demonstrated that the autocovariance function satisfies
the following set of linear difference equations:
0if hq>
qh
A 0 ΓΓΓΓh + A 1 ΓΓΓΓh – 1 + ...+ APΓΓΓΓhp=
∑BjhΩΩΩΩH′ j
- j = 0
where ΩΩΩΩand Hi are, respectively, the covariance matrix and the Markov
coefficients of the process in its infinite moving-average representation:
∞
xt = ∑Hiεεεεti–
i = 0
From the above representation, it is clear that if the process is purely MA,
that is, if p = 0, then the autocovariance function vanishes for lag h > q.
It is also possible to demonstrate the converse of this theorem. If a
linear stationary process admits an autocovariance function that satis-
fies the following equations,
A 0 ΓΓΓΓh + A 1 ΓΓΓΓh – 1 + ...+ APΓΓΓΓhp– = 0 if h > q
(^1) Christian Gourieroux and Alain Monfort, Time Series and Dynamic Models (Cam-
bridge: Cambridge University Press, 1997).