Suppose we have MA(1) model (watch the sign!) Xt=ϵt+θ1ϵt−1 This is already a causal representation.
We can rewrite the equation and get Xt−θ1ϵt−1=ϵt.
That means we can do the same for Xt−1, and write Xt−1−θ1ϵt−2=ϵt−1 We can substitute this into ϵt−1 above.
Substituting this expression of ϵt−1 into the first equation, Xt−θ1ϵt−1=ϵt.Xt−θ1(Xt−1−θ1ϵt−2)=ϵtXt−θ1Xt−1−θ21ϵt−2=ϵt
Repeat this n times, we get Xt−θ1Xt−1−⋯−θn−11Xt−n+1−θn1et−n=ϵt If |θ1|<1, then we can write ϵt=∞∑i=0θi1Xt−i This is called invertible representation.
ϵt=∞∑i=0πiXt−i=(π0+π1B+π2B2+π3B3+⋯)Xt=Π(B)Xt
So for our MA(q) model Xt=Θ(B)ϵtΠ(B)Xt=ϵt
Very similar to causal representation in AR(p), we have identity, Π(z)=1Θ(z) (π0+π1z+π2z2+π3z3+⋯)=1(1+θ1z+θ2z2+⋯+θqzq)
We can calculate coefficients by moving all the polynomials to the left and matching (1−θ1z−θ2z2−⋯−θqzq)(π0+π1z+π2z2+π3z3+⋯)=1
For example, if we have MA(1), (1−θ1z)(π0+π1z+π2z2+π3z3+⋯)=1
To find out the coeffeicnets πi, we can match the order of z. π0=1( coefficient without z )−θ1π0+π1=0( coefficient of z)−θ1π1+π2=0( coefficient of z2)−θ1π2+π3=0( coefficient of z3)⋮
π0=1π1=θ1π0π2=θ1π1π3=θ1π2⋮
π0=1π1=θ1π2=θ21π3=θ31⋮
Now we get invertible representation for MA(1), ϵt=∞∑i=0πiXt−i=∞∑i=0θi1Xt−i.
To write the today’s error (innovation) ϵt, we need infinetely many past observation Xt−i.
When can we do this?
If all roots of characteristic polynomial Θ(z)=1+θ1z+θ2z2+⋯+θqzq is outside of the unit circle, then MA(q) admits a invertible representation.
Same as causal condition in AR(p).
Watch the sign! We have same sign as Φ(z) in AR(p), because we started with Brockwell’s convention; Xt=ϵt+θ1ϵt−1.
We assume all MA(q) we deal with
are invertible.
If we have MA(3) process, Xt=ϵt−ˆθ1ϵt−1−ˆθ2ϵt−2−ˆθ3ϵt−3 with estimated parameter ˆθi, we want to get residuals ˆϵt, and check their randomness.
In AR(p), this was intuitive. e.g. if we had AR(2), residual was Xt−ˆϕ1Xt−1−ˆϕ2Xt−2=ˆϵt
In MA(q), we must use invertible
expression, ˆϵt=ˆπ0Xt+ˆπ1Xt−1+ˆπ2Xt−2+⋯
We haven’t talked about forecasting, but invertibility will be used in forecasting pretty much the same way as in the residuals.
MA(q) process is defiend with equation Xt=Θ(B)ϵt you can check the roots of the characteristic polynomial Θ(z)=1+θ1z+θ2z2+⋯+θqzq. If all root are outside of the (complex) unit circle, then MA(q) admits a invertible representation.
Inverted representation allows to write today’s error using infinite sum of past observations. ϵt=∞∑i=0Xt−i
Bcause we can’t directly observe ϵt, invertibility is importatnt in calculating residuals and forecasts.