Lecture Notes 10: Introduction to VARs

Wold Decomposition Theorem 1 14.384 Time Series Analysis Lecture 10 Introduction to VARs Wold Decomposition Theorem Theorem 1 (Wold decomposition). Let yt be a 2nd order stationary process with Eyt = 0. Then yt could be written as yt = vt + c(L)et , where 1. et : Eet = 0, Eet e0t = Ω, and et ∈ span{yt , yt−1 , ...} = It ( i.e. et are fundamental) 2. Eet e0s = 0 for all t 6= s P∞ 2 3. j=0 kcj k < ∞ 4. vt is deterministic, that is vt ∈ I−∞ = ∪∞ j=0 It−j Remark 2. The span{yt , yt−1 , ...} = It is the space of all linear combinations of {yt , yt−1 , ...}. Proof. The theorem is stated for vector processes, but we’ll just prove it in the scalar case. We will prove the theorem by constructing et and showing that it satisfies the conditions stated. Let ˆ t |It−1 ) = yt − a(L)yt−1 , where E(y ˆ t |It−1 ) is the best linear forecast of yt from past y. Note et = yt − E(y that the form of this forecast (a(L)) does not depend on t. This is because yt is second order stationary, and the best linear forecast only depends on the first and second moments of yt . By best linear forecast, we mean that a(L) solves: min E(yt − {aj } ∞ X aj yt−j )2 j=1 The first order conditions are ∞ X ∂ : E[yt−j (yt − aj yt−j )] = 0 ∂aj j=1 E[yt−j et ] = 0 So et is uncorrelated with yt−j ∀j > 0. This implies that It = It−1 ⊕ {et }. We now verifty that et satisfies conditions 1-4. 1. Eyt = 0 ∀t implies that Eet = 0. Also, Var(et ) = E(yt − a(L)yt )2 is just a function of the covariances of yt and aj , so Var(et ) = σ 2 is constant. 2. Eet yt−j = 0 ∀j ≥ 1, and et−j = yt−j − a(L)yt−j implies Eet et−j = 0. 3. By definition of et , yt =et + ∞ X k=1 ak yt−k Discussion 2 Repeatedly substituting for yt−k gives yt =et + ∞ X ak yt−k k=1 =et + a1 (et−1 + =et + a1 et−1 + =et + a1 et−1 + .. . = k X ∞ X ak yt−1−k ) + k=1 ∞ X ∞ X ak yt−k k=2 ãk yt−k−1 k=1 vt1 cj et−j + vtk j=0 where vtk ∈ span{yt−k−1 , yt−k−2 , ...}. Consider: E(yt − k X cj et−j )2 =Eyt2 + j=0 k X cj2 σ 2 − 2 j=0 =Eyt2 + k X =Eyt2 − cj E[yt et−j ] j=0 c2j σ 2 − 2 j =0 k X k X k X c2j σ 2 j=0 c2j σ 2 j=0 Pk where we used the fact that E [yt et−j ] = E[( l=0 cl et−l + vtk )et−j ] = cl σ 2 . Now we know that E(yt − Pk Pk P∞ Ey 2 2 Therefore, it must be that j =0 c2j ≤ σ2t and j=0 c2j ≤ ∞. j=0 cj et−j ) ≥ 0.P ∞ This implies that, j=0 cj et−j is well-defined. P∞ 4. Finally, vt = yt − j=0 cj et−j ∈ I−∞ . Discussion Definition 3. If vt = 0, the yt is called linearly regular. • Wold decomposition (an MA representation satisfying conditions (1)-(4) of the theorem)is “unique.” We put unique in quotes, because it is only unique up to a rotation of et . Let R be a rotation. et = Rut , then yt = c(L)Rut = c̃(L)ut , where c̃j = cj R. What is really unique is the space spanned by {et }. Structural VARs are about how we pick a particular rotation of the, and then attach interpretations to the components of et . • It’s true that any 2nd order stationary process has an MA representation, but that doesn’t mean that it’s always a good idea to study the MA representation. ( P (yt = 1|yt−1 = 1) = p Example 4. Two-state Markov chain: . This is 2nd order stationary and P (yt = 0|yt−1 = 0) = q Discussion 3 has the following MA representation: ( p E(yt |yt−1 ) = 1−q yt−1 = 1 yt−1 = 0 =1 − q + yt−1 (q − 1 + p) so yt =1 − q + yt−1 (q − 1 + p) + et ∞ X 1−q = + (p + q − 1)j et−j 2 − p − q j =0 where  prob p | yt−1 = 1  1 − p  −p prob 1 − p | yt−1 = 1 et = −(1 − q) prob q | yt−1 = 0    q prob 1 − q | yt−1 = 0 The MA representation is not the best description of the dynamics in this case. This example is simple, but the point is more general. There are macro models with simple state-space representations and complicated MA representations. • There are many MA representations. The Wold representation is only unique up to rotation when you require the MA to be fundamental. Example 5. Suppose yt = et + θet−1 , Var(et ) = σ 2 . Then γ0 = (1 + θ2 )σ 2 , γ1 = θσ 2 , γj = 0 ∀j > 1. Consider yt = ηt + θ1 ηt−1 with Varηt = θ2 σ 2 . It is easy to see that this process has the same covariances. However, only the first representation is fundamental. yt =(1 + θL)et et =(1 + θL)−1 yt = ∞ X (−θ)j yt−j j=0 For the second process, first define F = L−1 . yt =(θF + 1)ηt−1 ∞ X ηt−1 =θ (−θ)j F j yt j=0 Remark 6. It is possible that your model has a non-fundamental MA representation. A fundamental representation requires that your shocks are spanned by the set of variables in your VAR. For example, if you’re looking at shocks to money supply in a VAR with money and output, it is very likely that the Fed looks at variables other than output and money and when choosing policy, so shocks to the money supply may not be fundamental. One solution might be to include more variables in your VAR. Relatedly, you might use a factor-augmented VAR. Remark 7. Relation to spectrum decomposition: there are three ways to represent a process: the Wold decomposition, the covariances, and the spectrum. There is a 1-1 mapping among these representations, in the following sense: there is a one-to-one mapping between the spectrum and the set of auto-covariances; there is also a 1-1 mapping between autocovariances and the set of coefficients of the fundamental MA representation.