Processing math: 100%

Canonical decomposition of ARIMA models

We consider in the description of SEATS a simplified model with two components (the signal and the noise):

yt=st+nt ϕ(B)Δ(B)yt=θ(B)ϵt ϕs(B)Δs(B)st=θs(B)ϵst ϕn(B)Δn(B)nt=θn(B)ϵnt

The variances of the innovations are respectively σ2, σ2s and σ2n

The aggregation constraint yields:

ϕ(B)=ϕs(B)ϕn(B) Δ(B)=Δs(B)Δn(B)

σ2θ(B)θ(F)=σ2sϕn(B)Δn(B)θs(B)ϕn(F)Δn(F)θs(F) +σ2nϕs(B)Δs(B)θn(B)ϕs(F)Δs(F)θn(F)

The estimator of the signal is obtained as

ˆst=ksΨs(B)Ψs(F)Ψ(B)Ψ(F)yt