背景介绍
自回归模型:用同一变量的历史值预测未来值。
应用背景:弱平稳时间序列
{
x
t
}
\{x_t\}
{xt}(即满足:
E
(
x
t
)
=
μ
,
C
o
v
(
x
t
,
x
t
−
l
)
=
γ
l
E(x_t)=\mu,Cov(x_t,x_{t-l})=\gamma_l
E(xt)=μ,Cov(xt,xt−l)=γl性质的时间序列)
模型
{ x t = φ 0 + ∑ i = 1 p φ i x t − i + ε t φ p ≠ 0 E ( ε t ) = 0 , v a r ( ε t ) = σ 2 , E ( ε t ε s ) = 0 , s ≠ t E ( x s ε t ) = 0 , ∀ s < t (1) \left\{\tag{1} \begin{aligned} & x_t = \varphi_0 + \sum_{i=1}^p \varphi_i x_{t-i} + \varepsilon_t \\ & \varphi_p \neq 0 \\ & E(\varepsilon_t)=0,var(\varepsilon_t)=\sigma^2,E(\varepsilon_t \varepsilon_s)=0,s\neq t \\ & E(x_s \varepsilon_t)=0,\forall s < t \end{aligned} \right. ⎩⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎧xt=φ0+i=1∑pφixt−i+εtφp=0E(εt)=0,var(εt)=σ2,E(εtεs)=0,s=tE(xsεt)=0,∀s<t(1)
限制条件
- φ p ≠ 0 \varphi_p \neq 0 φp=0,确保模型的最高阶数为 p p p;
- E ( ε i ) = 0 , V a r ( ε i ) = σ ε 2 , E ( ε t ε s ) = 0 , s ≠ t E(\varepsilon_i) = 0,Var(\varepsilon_i) = \sigma_{\varepsilon}^2,E(\varepsilon_t \varepsilon_s)=0,s\neq t E(εi)=0,Var(εi)=σε2,E(εtεs)=0,s=t,要求随机干扰序列 { ε t } \{\varepsilon_t\} {εt}为零均值白噪声序列;
- E ( x s ε t ) = 0 , ∀ s < t E(x_s \varepsilon_t)=0,\forall s < t E(xsεt)=0,∀s<t,当期的随机干扰与过去的序列值无关。
通常默认限制条件,把
A
R
(
p
)
AR(p)
AR(p)模型简记为:
x
t
=
φ
0
+
∑
i
=
1
p
φ
i
x
t
−
i
+
ε
t
(2)
x_t = \varphi_0 + \sum_{i=1}^p \varphi_i x_{t-i} + \varepsilon_t \tag{2}
xt=φ0+i=1∑pφixt−i+εt(2)
统计性质
均值
由
{
x
t
}
\{x_t\}
{xt}的弱平稳性知,
E
(
x
t
)
=
μ
E(x_t)=\mu
E(xt)=μ
E
(
x
t
)
=
φ
0
+
∑
i
=
1
p
φ
i
E
(
x
t
−
i
)
+
E
(
ε
t
)
(3)
E(x_t) = \varphi_0 + \sum_{i=1}^p \varphi_i E(x_{t-i}) + E(\varepsilon_t) \tag{3}
E(xt)=φ0+i=1∑pφiE(xt−i)+E(εt)(3)
μ = φ 0 1 − ∑ i = 1 p φ i (4) \mu = \frac{\varphi_0}{1-\sum_{i=1}^p \varphi_i} \tag{4} μ=1−∑i=1pφiφ0(4)
对于中心化 A R ( p ) AR(p) AR(p)模型,有 E ( x t ) = 0 E(x_t)=0 E(xt)=0,显然对于任意 A R ( p ) AR(p) AR(p)模型都可以通过减去自身均值 μ \mu μ从而变为中心化 A R ( p ) AR(p) AR(p)模型。
自协方差函数
含义:度量变量过去的行为对现在的影响。
定义1:对于时间序列
{
x
t
,
t
∈
T
}
\{x_t,t \in T\}
{xt,t∈T},任取
t
,
s
∈
T
t,s\in T
t,s∈T,
γ
t
,
s
\gamma_{t,s}
γt,s为序列
{
x
t
}
\{x_t\}
{xt}的自协方差函数
γ
t
,
s
=
E
(
x
t
−
μ
t
)
(
x
s
−
μ
s
)
(5)
\gamma_{t,s}=E(x_t-\mu_t)(x_s-\mu_s) \tag{5}
γt,s=E(xt−μt)(xs−μs)(5)
对于中心化
A
R
(
p
)
AR(p)
AR(p)模型而言,
μ
t
=
μ
s
=
0
\mu_t=\mu_s=0
μt=μs=0,因此
γ
t
,
s
=
E
(
x
t
x
s
)
(6)
\gamma_{t,s}=E(x_t x_s) \tag{6}
γt,s=E(xtxs)(6)
定义2:
γ
k
\gamma_k
γk为时间序列
{
x
t
}
\{x_t\}
{xt}的延迟
k
k
k阶自协方差函数
γ
k
=
γ
t
,
t
−
k
=
E
(
x
t
,
x
t
−
k
)
\gamma_k=\gamma_{t,t-k}=E(x_t,x_{t-k})
γk=γt,t−k=E(xt,xt−k)。
在中心化
A
R
(
p
)
AR(p)
AR(p)模型
x
t
=
∑
i
=
1
p
φ
i
x
t
−
i
+
ε
t
x_t=\sum_{i=1}^p \varphi_i x_{t-i} + \varepsilon_t
xt=∑i=1pφixt−i+εt等式两边同乘
x
t
−
k
(
k
≥
1
)
x_{t-k}(k \ge 1)
xt−k(k≥1),再求期望得
E
(
x
t
x
t
−
k
)
=
∑
i
=
1
p
φ
i
E
(
x
t
−
i
x
t
−
k
)
+
E
(
ε
t
x
t
−
k
)
,
∀
k
≥
1
(7)
E(x_t x_{t-k}) = \sum_{i=1}^p \varphi_i E(x_{t-i}x_{t-k})+E(\varepsilon_t x_{t-k}),\forall k \ge 1 \tag{7}
E(xtxt−k)=i=1∑pφiE(xt−ixt−k)+E(εtxt−k),∀k≥1(7)
由
A
R
(
p
)
AR(p)
AR(p)模型的条件三知
E
(
ε
t
x
t
−
k
)
=
0
E(\varepsilon_t x_{t-k})=0
E(εtxt−k)=0,所以
γ
k
=
∑
i
=
1
p
φ
i
γ
k
−
i
(8)
\gamma_k = \sum_{i=1}^p \varphi_i \gamma_{k-i} \tag{8}
γk=i=1∑pφiγk−i(8)
自相关系数
ρ k = γ t , t − k v a r ( x t ) v a r ( x t + k ) = γ k γ 0 (9) \rho_{k} = \frac{\gamma_{t,t-k}}{\sqrt{var(x_t)var(x_{t+k})}}=\frac{\gamma_k}{\gamma_0} \tag{9} ρk=var(xt)var(xt+k)γt,t−k=γ0γk(9)
偏自相关系数
定义
对于平稳序列
{
x
t
}
\{x_t\}
{xt},滞后
k
k
k偏自相关系数就是在给定时间序列中间
k
−
1
k-1
k−1个随机变量
x
t
−
1
,
x
t
−
2
,
⋯
,
x
t
−
k
+
1
x_{t-1},x_{t-2},\cdots,x_{t-k+1}
xt−1,xt−2,⋯,xt−k+1的条件下,
x
t
−
k
x_{t-k}
xt−k对
x
t
x_t
xt相关影响的度量。
ρ
x
t
,
x
t
−
k
∣
x
t
−
1
,
⋯
,
x
t
−
k
+
1
=
E
[
(
x
t
−
E
^
x
t
)
(
x
t
−
k
−
E
^
x
t
−
k
)
]
E
[
(
x
t
−
k
−
E
^
x
t
−
k
)
2
]
(10)
\rho_{x_t,x_{t-k}|x_{t-1,\cdots,x_{t-k+1}}} = \frac{E[(x_t - \hat{E}x_t)(x_{t-k} - \hat{E}x_{t-k})]}{E[(x_{t-k}-\hat{E}x_{t-k})^2]} \tag{10}
ρxt,xt−k∣xt−1,⋯,xt−k+1=E[(xt−k−E^xt−k)2]E[(xt−E^xt)(xt−k−E^xt−k)](10)
式中,
E
^
x
t
=
E
[
x
t
∣
x
t
−
1
,
⋯
,
x
t
−
k
+
1
]
\hat{E}x_t=E[x_t|x_{t-1,\cdots,x_{t-k+1}}]
E^xt=E[xt∣xt−1,⋯,xt−k+1],
E
^
x
t
−
k
=
E
[
x
t
−
k
∣
x
t
−
1
,
⋯
,
x
t
−
k
+
1
]
\hat{E}x_{t-k}=E[x_{t-k}|x_{t-1,\cdots,x_{t-k+1}}]
E^xt−k=E[xt−k∣xt−1,⋯,xt−k+1]
计算方式
考虑
{
x
t
}
\{x_t\}
{xt}为中心化平稳序列,用过去的
k
k
k期序列值
x
t
−
1
,
x
t
−
2
,
⋯
,
x
t
−
k
x_{t-1},x_{t-2},\cdots,x_{t-k}
xt−1,xt−2,⋯,xt−k对
x
i
x_i
xi作
k
k
k阶自回归拟合,即
x
t
=
φ
k
1
x
t
−
1
+
φ
k
2
x
t
−
2
+
⋯
+
φ
k
k
x
t
−
k
+
ε
k
(11)
x_t = \varphi_{k1}x_{t-1} + \varphi_{k2}x_{t-2}+\cdots+\varphi_{kk}x_{t-k} + \varepsilon_k \tag{11}
xt=φk1xt−1+φk2xt−2+⋯+φkkxt−k+εk(11)
取期望有:
E
^
x
t
=
∑
i
=
1
k
−
1
φ
k
i
x
t
−
i
+
φ
k
k
E
^
(
x
t
−
k
)
(12)
\hat{E}x_t = \sum_{i=1}^{k-1} \varphi_{ki}x_{t-i} +\varphi_{kk} \hat{E}(x_{t-k}) \tag{12}
E^xt=i=1∑k−1φkixt−i+φkkE^(xt−k)(12)
x
t
−
E
^
x
t
=
φ
k
k
(
x
t
−
k
−
E
^
x
t
−
k
)
+
ε
t
(13)
x_t - \hat{E}x_t = \varphi_{kk}(x_{t-k} - \hat{E}x_{t-k}) + \varepsilon_t \tag{13}
xt−E^xt=φkk(xt−k−E^xt−k)+εt(13)
等式两边同乘
x
t
−
k
−
E
^
x
t
−
k
x_{t-k} - \hat{E}x_{t-k}
xt−k−E^xt−k,并求期望有:
E
[
(
x
t
−
E
^
x
t
)
(
x
t
−
k
−
E
^
x
t
−
k
)
]
=
φ
k
k
E
[
(
x
t
−
k
−
E
^
x
t
−
k
)
2
]
(14)
E[(x_t - \hat{E}x_t) (x_{t-k} - \hat{E}x_{t-k})] = \varphi_{kk}E[(x_{t-k} - \hat{E}x_{t-k})^2] \tag{14}
E[(xt−E^xt)(xt−k−E^xt−k)]=φkkE[(xt−k−E^xt−k)2](14)
φ
k
k
=
E
[
(
x
t
−
E
^
x
t
)
(
x
t
−
k
−
E
^
x
t
−
k
)
]
E
[
(
x
t
−
k
−
E
^
x
t
−
k
)
2
]
(15)
\varphi_{kk} = \frac{E[(x_t - \hat{E}x_t) (x_{t-k} - \hat{E}x_{t-k})]}{E[(x_{t-k} - \hat{E}x_{t-k})^2]} \tag{15}
φkk=E[(xt−k−E^xt−k)2]E[(xt−E^xt)(xt−k−E^xt−k)](15)
滞后
k
k
k偏自相关系数实际上就等于
k
k
k阶自回归模型第
k
k
k个回归系数
φ
k
k
\varphi_{kk}
φkk的值。可以证明平稳
A
R
(
p
)
AR(p)
AR(p)模型的偏自相关系数具有
p
p
p步截尾性,即
φ
k
k
=
0
(
∀
k
>
p
)
\varphi_{kk}=0(\forall k > p)
φkk=0(∀k>p)。
定阶
A R AR AR模型可使用偏自相关函数 P A C F PACF PACF截尾(相关系数快速收敛到等于0的置信区间内)定阶。
预测
不妨设预测原点为 h h h, F h F_h Fh表示在预测原点已知信息, x ^ h ( l ) \hat{x}_h(l) x^h(l)为序列 { x t } \{x_t\} {xt}的以 x h x_h xh为预测原点的向前 l l l步预测(即:在预测原点 h h h已知信息 F h F_h Fh的条件下, x h + l x_{h+l} xh+l的条件均值)。
1步预测
x
h
+
1
=
φ
0
+
∑
i
=
1
p
φ
i
x
h
+
1
−
i
+
ε
h
+
1
x
^
h
(
1
)
=
E
(
x
h
+
1
∣
F
h
)
=
φ
0
+
∑
i
=
1
p
φ
i
x
h
+
1
−
i
(16)
\begin{aligned} \tag{16} x_{h+1} &=\varphi_0+\sum_{i=1}^p \varphi_i x_{h+1-i}+\varepsilon_{h+1}\\ \hat{x}_h(1) &= E(x_{h+1}|F_h)=\varphi_0+\sum_{i=1}^p \varphi_i x_{h+1-i} \end{aligned}
xh+1x^h(1)=φ0+i=1∑pφixh+1−i+εh+1=E(xh+1∣Fh)=φ0+i=1∑pφixh+1−i(16)
注意:
E
(
ε
h
+
1
)
=
0
E(\varepsilon_{h+1})=0
E(εh+1)=0,是因为随机误差是白噪声序列,
E
(
x
h
+
1
−
i
)
=
x
h
+
1
−
i
E(x_{h+1-i})=x_{h+1-i}
E(xh+1−i)=xh+1−i是因为预测原点
h
h
h的前
p
p
p个历史值已然发生,是确定的数值。
1步预测误差:
ε
h
(
1
)
=
x
h
+
1
−
x
h
(
1
)
=
ε
h
+
1
(17)
\varepsilon_h(1)=x_{h+1}-x_h(1)=\varepsilon_{h+1} \tag{17}
εh(1)=xh+1−xh(1)=εh+1(17)
1步预测误差方差:
v
a
r
(
ε
h
(
1
)
)
=
v
a
r
(
ε
h
+
1
)
=
σ
ε
2
(18)
var(\varepsilon_h(1)) = var(\varepsilon_{h+1})=\sigma_{\varepsilon}^2 \tag{18}
var(εh(1))=var(εh+1)=σε2(18)
2步预测
x
h
+
2
=
φ
0
+
∑
i
=
1
p
φ
i
x
h
+
2
−
i
+
ε
h
+
2
x
^
h
(
2
)
=
E
(
x
h
+
2
∣
F
h
)
=
φ
0
+
φ
1
x
^
h
+
1
+
∑
i
=
2
p
φ
i
x
h
+
2
−
i
(19)
\begin{aligned}\tag{19} x_{h+2} &=\varphi_0+\sum_{i=1}^p \varphi_i x_{h+2-i}+\varepsilon_{h+2}\\ \hat{x}_h(2) &= E(x_{h+2}|F_h)=\varphi_0+\varphi_1 \hat{x}_{h+1} + \sum_{i=2}^p \varphi_i x_{h+2-i} \end{aligned}
xh+2x^h(2)=φ0+i=1∑pφixh+2−i+εh+2=E(xh+2∣Fh)=φ0+φ1x^h+1+i=2∑pφixh+2−i(19)
注意:因为在预测原点
h
h
h对
h
+
2
h+2
h+2时刻的序列值做出预测时,
h
+
1
h+1
h+1时刻的序列值并未发生,所以需要首先计算
x
^
h
+
1
\hat{x}_{h+1}
x^h+1作为
h
+
1
h+1
h+1时刻序列值的近似替换。
2步预测误差
ε
h
(
2
)
=
x
h
+
2
−
x
^
h
(
2
)
=
φ
1
(
x
h
+
1
−
x
^
h
(
1
)
)
+
ε
h
+
2
=
ε
h
+
2
+
φ
1
ε
h
+
1
(20)
\tag{20} \begin{aligned} \varepsilon_h(2) &= x_{h+2}-\hat{x}_h(2)= \varphi_1(x_{h+1}-\hat{x}_h(1)) + \varepsilon_{h+2}\\ &= \varepsilon_{h+2} + \varphi_1 \varepsilon_{h+1} \end{aligned}
εh(2)=xh+2−x^h(2)=φ1(xh+1−x^h(1))+εh+2=εh+2+φ1εh+1(20)
2步预测误差方差
v
a
r
(
ε
h
(
2
)
)
=
(
1
+
φ
1
)
σ
ε
2
(21)
var(\varepsilon_h(2)) =(1+\varphi_1)\sigma_{\varepsilon}^2 \tag{21}
var(εh(2))=(1+φ1)σε2(21)
显然
v
a
r
(
ε
h
(
2
)
)
>
v
a
r
(
ε
h
(
1
)
)
var(\varepsilon_h(2)) > var(\varepsilon_h(1))
var(εh(2))>var(εh(1)),这意味着预测步长的增加会使预测的不确定性也增加。
向前多步预测( l ≥ 2 l\ge2 l≥2)
x h + l = φ 0 + ∑ i = 1 p φ i x h + l − i + ε h + l x ^ h ( l ) = E ( x h + l ∣ F h ) = φ 0 + ∑ i = 1 l − 1 φ i x ^ h ( i ) + ∑ i = l p φ i x h + l − i (22) \tag{22} \begin{aligned} x_{h+l} &=\varphi_0+\sum_{i=1}^p \varphi_i x_{h+l-i}+\varepsilon_{h+l}\\ \hat{x}_h(l) &= E(x_{h+l}|F_h)=\varphi_0+ \sum_{i=1}^{l-1} \varphi_i \hat{x}_h(i) + \sum_{i=l}^p \varphi_i x_{h+l- i} \end{aligned} xh+lx^h(l)=φ0+i=1∑pφixh+l−i+εh+l=E(xh+l∣Fh)=φ0+i=1∑l−1φix^h(i)+i=l∑pφixh+l−i(22)
优缺点
优点:简单,仅用自身变量数列就可进行预测。
缺点:变量必须具有自相关,只适用于预测与自身前期相关的经济现象,对社会因素影响较大的经济现象,不宜采用自回归。
王燕.应用时间序列分析[M].中国人民大学出版社.201907