258. VAR

Vector autoregression

VAR(𝑝)𝚽(𝐵)𝒙𝑡=𝒄+𝜺𝑡𝚽(𝐵)=𝑰𝚽1𝐵𝚽𝑝𝐵𝑝

i.e.,

𝒙𝑡=𝒄+𝚽1𝒙𝑡1++𝚽𝑝𝒙𝑡𝑝+𝜺𝑡

Each component depends on 𝑝 lags of all 𝐾 series.

Parameters: 𝚽1,,𝚽𝑝 (𝐾2 each), 𝒄 (𝐾), 𝚺 (𝐾(𝐾+1)/2)
Orders: 𝑝, 𝐾 (series)

Example: VAR(1), 𝐾=2

Given

  • Order: 𝑝=1, 𝐾=2
  • Coefficient matrix:

    𝚽1=[0.40.20.20.4]
  • Intercept: 𝒄=[00]
  • Two series stacked into a vector 𝒙𝑡=[𝑥1,𝑡𝑥2,𝑡]:
𝑡12345678910111213141516
x_(1\,)𝑡121081114129131614111518161317
𝑡12345678910111213141516
x_(2\,)𝑡89761011981213111014151312

Step 1 — formula

Substitute 𝑝=1 into the VAR(𝑝) recursion:

𝒙𝑡=𝒄+𝚽1𝒙𝑡1+𝜺𝑡

Forecast (set 𝜺𝑡=𝟎):

𝒙̂𝑡=𝒄+𝚽1𝒙𝑡1

Componentwise (write out the matrix-vector product):

[𝑥̂1,𝑡𝑥̂2,𝑡]=[𝑐1𝑐2]+[0.40.20.20.4][𝑥1,𝑡1𝑥2,𝑡1]𝑥̂1,𝑡=𝑐1+0.4𝑥1,𝑡1+0.2𝑥2,𝑡1𝑥̂2,𝑡=𝑐2+0.2𝑥1,𝑡1+0.4𝑥2,𝑡1

Innovation:

𝜺𝑡=𝒙𝑡𝒙̂𝑡

Step 2 — apply at 𝑡=2

Plug in 𝒄=𝟎, 𝚽1, 𝒙1=[128]:

𝑥̂1,2=0.412+0.28=4.8+1.6=6.4𝑥̂2,2=0.212+0.48=2.4+3.2=5.6𝜺2=[109][6.45.6]=[3.63.4]

Step 3 — iterate

𝑡𝒙𝑡𝑥̂1,𝑡=0.4𝑥1,𝑡1+0.2𝑥2,𝑡1𝑥̂2,𝑡=0.2𝑥1,𝑡1+0.4𝑥2,𝑡1𝜺𝑡
2[109]0.4(12)+0.2(8)=6.40.2(12)+0.4(8)=5.6[3.63.4]
3[87]0.4(10)+0.2(9)=5.80.2(10)+0.4(9)=5.6[2.21.4]
4[116]0.4(8)+0.2(7)=4.60.2(8)+0.4(7)=4.4[6.41.6]
5[1410]0.4(11)+0.2(6)=5.60.2(11)+0.4(6)=4.6[8.45.4]
6[1211]0.4(14)+0.2(10)=7.60.2(14)+0.4(10)=6.8[4.44.2]
7[99]0.4(12)+0.2(11)=70.2(12)+0.4(11)=6.8[22.2]
8[138]0.4(9)+0.2(9)=5.40.2(9)+0.4(9)=5.4[7.62.6]
9[1612]0.4(13)+0.2(8)=6.80.2(13)+0.4(8)=5.8[9.26.2]
10[1413]0.4(16)+0.2(12)=8.80.2(16)+0.4(12)=8[5.25]
11[1111]0.4(14)+0.2(13)=8.20.2(14)+0.4(13)=8[2.83]
12[1510]0.4(11)+0.2(11)=6.60.2(11)+0.4(11)=6.6[8.43.4]
13[1814]0.4(15)+0.2(10)=80.2(15)+0.4(10)=7[107]
14[1615]0.4(18)+0.2(14)=100.2(18)+0.4(14)=9.2[65.8]
15[1313]0.4(16)+0.2(15)=9.40.2(16)+0.4(15)=9.2[3.63.8]
16[1712]0.4(13)+0.2(13)=7.80.2(13)+0.4(13)=7.8[9.24.2]
170.4(17)+0.2(12)=9.20.2(17)+0.4(12)=8.2