Time Series Analysis
author:zoxiii
序列趋势分析
-
1、线性拟合
-
2、曲线拟合
-
3、移动平均法
-
-
(1)基本思想
-
(2)n期中心移动平均法:分析趋势
-
(3)n期移动平均法:预测
-
(4)n期中心移动平均法—> 提取低阶趋势
-
-
对
xt
=
a
+
b
t
+
ε
t
x_t=a+bt+\varepsilon_t
x
t
=
a
+
b
t
+
ε
t
进行
n=
2
k
+
1
n=2k+1
n
=
2
k
+
1
期的中心移动平均
-
对
xt
=
a
+
b
t
+
c
t
2
+
ε
t
x_t=a+bt+ct^2+\varepsilon_t
x
t
=
a
+
b
t
+
c
t
2
+
ε
t
进行
n=
2
k
+
1
n=2k+1
n
=
2
k
+
1
期的中心移动平均
-
对
xt
=
a
+
b
t
+
c
t
2
+
ε
t
x_t=a+bt+ct^2+\varepsilon_t
x
t
=
a
+
b
t
+
c
t
2
+
ε
t
进行
n=
2
k
n=2k
n
=
2
k
期的中心移动平均
-
-
-
4、指数平滑法
【参考文献】王燕. 应用时间序列分析-第5版[M]. 中国人民大学出版社, 2019.
1、线性拟合
(1)基本思想
当序列的时序图出现显著的线性特征时,可使用线性模型去拟合
(2)公式
x
t
=
a
+
b
t
+
I
t
x_t=a+bt+I_t
x
t
=
a
+
b
t
+
I
t
E
(
I
t
)
=
0
,
V
a
r
(
I
t
)
=
σ
2
E(I_t)=0,Var(I_t)=\sigma^2
E
(
I
t
)
=
0
,
V
a
r
(
I
t
)
=
σ
2
其中随机波动:
{
I
t
}
\{I_t\}
{
I
t
}
消除随机波动的影响之后该序列的长期趋势:
T
t
=
a
+
b
t
T_t=a+bt
T
t
=
a
+
b
t
2、曲线拟合
(1)基本思想
当序列的时序图出现非线性特征时,可使用曲线模型去拟合
(2)二次型拟合公式
x
t
=
a
+
b
t
+
c
t
2
+
I
t
或
x
t
=
a
+
c
t
2
+
I
t
x_t=a+bt+ct2+I_t~~或~~x_t=a+ct2+I_t
x
t
=
a
+
b
t
+
c
t
2
+
I
t
或
x
t
=
a
+
c
t
2
+
I
t
t
2
=
t
2
t2=t^2
t
2
=
t
2
E
(
I
t
)
=
0
,
V
a
r
(
I
t
)
=
σ
2
E(I_t)=0,Var(I_t)=\sigma^2
E
(
I
t
)
=
0
,
V
a
r
(
I
t
)
=
σ
2
(3)指数型拟合公式
T
t
=
a
b
t
T_t=ab^t
T
t
=
a
b
t
取对数,令
T
t
′
=
l
n
T
t
,
a
′
=
l
n
a
,
b
′
=
l
n
b
,
T_t’=lnT_t,a’=lna,b’=lnb,
T
t
′
=
l
n
T
t
,
a
′
=
l
n
a
,
b
′
=
l
n
b
,
得到
T
t
′
=
a
′
+
b
′
t
T_t’=a’+b’t
T
t
′
=
a
′
+
b
′
t
3、移动平均法
(1)基本思想
用一定时间间隔之间的平均值作为某一期的估计值
如何确定n?
- 考虑n=周期长度,如4、12
- 考虑平滑性,n越大拟合曲线越平滑
- 考虑趋势近期敏感程度,n越小趋势对近期变化越敏感
(2)n期中心移动平均法:分析趋势
n为奇数
x
~
t
=
1
n
(
x
t
−
n
−
1
2
+
x
t
−
n
−
1
2
+
1
+
.
.
.
+
x
t
+
n
−
1
2
)
\widetilde x_t=\frac{1}{n}(x_{t-\frac{n-1}{2}}+x_{t-\frac{n-1}{2}+1}+…+x_{t+\frac{n-1}{2}})
x
t
=
n
1
(
x
t
−
2
n
−
1
+
x
t
−
2
n
−
1
+
1
+
.
.
.
+
x
t
+
2
n
−
1
)
n为偶数
x
~
t
=
1
n
(
1
2
x
t
−
n
2
+
x
t
−
n
2
+
1
+
.
.
.
+
x
t
−
n
2
−
1
+
1
2
x
t
+
n
2
)
\widetilde x_t=\frac{1}{n}(\frac{1}{2}x_{t-\frac{n}{2}}+x_{t-\frac{n}{2}+1}+…+x_{t-\frac{n}{2}-1}+\frac{1}{2}x_{t+\frac{n}{2}})
x
t
=
n
1
(
2
1
x
t
−
2
n
+
x
t
−
2
n
+
1
+
.
.
.
+
x
t
−
2
n
−
1
+
2
1
x
t
+
2
n
)
(3)n期移动平均法:预测
x
~
t
=
1
n
(
x
t
+
x
t
−
1
+
.
.
.
+
x
t
−
n
+
1
)
\widetilde x_t=\frac{1}{n}(x_t+x_{t-1}+…+x_{t-n+1})
x
t
=
n
1
(
x
t
+
x
t
−
1
+
.
.
.
+
x
t
−
n
+
1
)
(4)n期中心移动平均法—> 提取低阶趋势
对
x
t
=
a
+
b
t
+
ε
t
x_t=a+bt+\varepsilon_t
x
t
=
a
+
b
t
+
ε
t
进行
n
=
2
k
+
1
n=2k+1
n
=
2
k
+
1
期的中心移动平均
x
~
t
=
1
2
k
+
1
∑
i
=
−
k
k
x
t
+
i
=
1
2
k
+
1
∑
i
=
−
k
k
(
a
+
b
t
+
b
i
+
ε
t
+
i
)
=
a
+
b
t
\widetilde x_t \\ =\cfrac{1}{2k+1}\sum_{i=-k}^{k}x_{t+i} \\ =\cfrac{1}{2k+1}\sum_{i=-k}^{k}(a+bt+bi+\varepsilon_{t+i}) \\ =a+bt
x
t
=
2
k
+
1
1
i
=
−
k
∑
k
x
t
+
i
=
2
k
+
1
1
i
=
−
k
∑
k
(
a
+
b
t
+
b
i
+
ε
t
+
i
)
=
a
+
b
t
对
x
t
=
a
+
b
t
+
c
t
2
+
ε
t
x_t=a+bt+ct^2+\varepsilon_t
x
t
=
a
+
b
t
+
c
t
2
+
ε
t
进行
n
=
2
k
+
1
n=2k+1
n
=
2
k
+
1
期的中心移动平均
x
~
t
=
1
2
k
+
1
∑
i
=
−
k
k
x
t
+
i
=
1
2
k
+
1
∑
i
=
−
k
k
(
a
+
b
t
+
b
i
+
c
(
t
+
i
)
2
+
ε
t
+
i
)
=
a
+
b
t
+
c
t
2
+
c
k
(
k
+
1
)
3
\widetilde x_t \\ =\cfrac{1}{2k+1}\sum_{i=-k}^{k}x_{t+i} \\ =\cfrac{1}{2k+1}\sum_{i=-k}^{k}(a+bt+bi+c(t+i)^2+\varepsilon_{t+i}) \\ =a+bt+ct^2+\cfrac{ck(k+1)}{3}
x
t
=
2
k
+
1
1
i
=
−
k
∑
k
x
t
+
i
=
2
k
+
1
1
i
=
−
k
∑
k
(
a
+
b
t
+
b
i
+
c
(
t
+
i
)
2
+
ε
t
+
i
)
=
a
+
b
t
+
c
t
2
+
3
c
k
(
k
+
1
)
- 可以完整地提取二次趋势信息,但拟合序列和原序列会有一个截距的小偏差
对
x
t
=
a
+
b
t
+
c
t
2
+
ε
t
x_t=a+bt+ct^2+\varepsilon_t
x
t
=
a
+
b
t
+
c
t
2
+
ε
t
进行
n
=
2
k
n=2k
n
=
2
k
期的中心移动平均
x
~
t
=
1
2
k
[
∑
i
=
−
k
k
x
t
+
i
−
1
2
(
x
t
−
k
+
x
t
+
k
)
]
=
1
2
k
[
∑
i
=
−
k
k
(
a
+
b
t
+
b
i
+
c
(
t
+
i
)
2
+
ε
t
+
i
)
−
1
2
(
a
+
b
t
−
b
k
+
c
(
t
−
k
)
2
)
=
ε
t
−
k
+
a
+
b
t
+
b
k
+
c
(
t
+
k
)
2
+
ε
t
+
k
]
=
1
2
k
[
2
k
a
+
2
k
b
t
+
2
k
c
t
2
+
2
c
k
2
(
k
+
1
)
3
−
1
2
(
2
a
+
2
b
t
+
2
c
t
2
+
2
c
k
2
)
]
=
1
2
k
[
(
2
k
−
1
)
a
+
(
2
k
−
1
)
b
t
+
(
2
k
−
1
)
c
t
2
+
2
3
c
k
3
+
2
3
c
k
2
−
c
k
2
]
=
1
2
k
[
(
2
k
−
1
)
a
+
(
2
k
−
1
)
b
t
+
(
2
k
−
1
)
c
t
2
+
c
k
2
(
2
k
−
1
)
3
]
=
2
k
−
1
2
k
(
a
+
b
t
+
c
t
2
+
c
k
2
3
)
\widetilde x_t \\ =\cfrac{1}{2k}[\sum_{i=-k}^{k}x_{t+i}-\cfrac{1}{2}(x_{t-k}+x_{t+k})] \\ =\cfrac{1}{2k}[\sum_{i=-k}^{k}(a+bt+bi+c(t+i)^2+\varepsilon_{t+i})-\cfrac{1}{2}(a+bt-bk+c(t-k)^2)=\varepsilon_{t-k}+a+bt+bk+c(t+k)^2+\varepsilon_{t+k}] \\ =\cfrac{1}{2k}[2ka+2kbt+2kct^2+\cfrac{2ck^2(k+1)}{3}-\cfrac{1}{2}(2a+2bt+2ct^2+2ck^2)] \\ =\cfrac{1}{2k}[(2k-1)a+(2k-1)bt+(2k-1)ct^2+\cfrac{2}{3}ck^3+\cfrac{2}{3}ck^2-ck^2] \\ =\cfrac{1}{2k}[(2k-1)a+(2k-1)bt+(2k-1)ct^2+\cfrac{ck^2(2k-1)}{3}] \\ =\cfrac{2k-1}{2k}(a+bt+ct^2+\cfrac{ck^2}{3})
x
t
=
2
k
1
[
i
=
−
k
∑
k
x
t
+
i
−
2
1
(
x
t
−
k
+
x
t
+
k
)
]
=
2
k
1
[
i
=
−
k
∑
k
(
a
+
b
t
+
b
i
+
c
(
t
+
i
)
2
+
ε
t
+
i
)
−
2
1
(
a
+
b
t
−
b
k
+
c
(
t
−
k
)
2
)
=
ε
t
−
k
+
a
+
b
t
+
b
k
+
c
(
t
+
k
)
2
+
ε
t
+
k
]
=
2
k
1
[
2
k
a
+
2
k
b
t
+
2
k
c
t
2
+
3
2
c
k
2
(
k
+
1
)
−
2
1
(
2
a
+
2
b
t
+
2
c
t
2
+
2
c
k
2
)
]
=
2
k
1
[
(
2
k
−
1
)
a
+
(
2
k
−
1
)
b
t
+
(
2
k
−
1
)
c
t
2
+
3
2
c
k
3
+
3
2
c
k
2
−
c
k
2
]
=
2
k
1
[
(
2
k
−
1
)
a
+
(
2
k
−
1
)
b
t
+
(
2
k
−
1
)
c
t
2
+
3
c
k
2
(
2
k
−
1
)
]
=
2
k
2
k
−
1
(
a
+
b
t
+
c
t
2
+
3
c
k
2
)
4、指数平滑法
(1)简单指数平滑
基本思想
适用于既无长期趋势,又无季节效应的序列
对序列修匀
x
~
t
=
α
x
t
+
α
(
1
−
α
)
x
t
−
1
+
α
(
1
−
α
)
2
x
t
−
2
+
.
.
.
\widetilde x_t=\alpha x_t+\alpha(1-\alpha)x_{t-1}+\alpha(1-\alpha)^2x_{t-2}+…
x
t
=
α
x
t
+
α
(
1
−
α
)
x
t
−
1
+
α
(
1
−
α
)
2
x
t
−
2
+
.
.
.
x
~
t
=
α
x
t
+
(
1
−
α
)
x
~
t
−
1
\widetilde x_t=\alpha x_t+(1-\alpha)\widetilde x_{t-1}
x
t
=
α
x
t
+
(
1
−
α
)
x
t
−
1
平滑系数
0
<
α
<
1
0 < \alpha < 1
0
<
α
<
1
指定
x
~
0
=
x
1
\widetilde x_0=x_1
x
0
=
x
1
未来预测
根据预测公式:
x
^
T
+
1
=
x
~
T
\hat x_{T+1}=\widetilde x_T
x
^
T
+
1
=
x
T
x
^
T
=
x
~
T
−
1
\hat x_T=\widetilde x_{T-1}
x
^
T
=
x
T
−
1
1期预测值:
x
^
T
+
1
=
x
~
T
=
α
x
T
+
α
(
1
−
α
)
x
T
−
1
+
α
(
1
−
α
)
2
x
T
−
2
+
.
.
.
=
α
x
T
+
(
1
−
α
)
x
^
T
\hat x_{T+1} \\ =\widetilde x_T \\ =\alpha x_T+\alpha(1-\alpha)x_{T-1}+\alpha(1-\alpha)^2x_{T-2}+… \\ =\alpha x_T+(1-\alpha)\hat x_T
x
^
T
+
1
=
x
T
=
α
x
T
+
α
(
1
−
α
)
x
T
−
1
+
α
(
1
−
α
)
2
x
T
−
2
+
.
.
.
=
α
x
T
+
(
1
−
α
)
x
^
T
2期预测值:
x
^
T
+
2
=
α
x
T
+
1
+
α
(
1
−
α
)
x
T
+
α
(
1
−
α
)
2
x
T
−
1
+
.
.
.
=
α
x
^
T
+
1
+
(
1
−
α
)
x
^
T
+
1
=
x
^
T
+
1
\hat x_{T+2} \\ =\alpha x_{T+1}+\alpha(1-\alpha)x_T+\alpha(1-\alpha)^2x_{T-1}+… \\ =\alpha \hat x_{T+1}+(1-\alpha)\hat x_{T+1} \\ =\hat x_{T+1}
x
^
T
+
2
=
α
x
T
+
1
+
α
(
1
−
α
)
x
T
+
α
(
1
−
α
)
2
x
T
−
1
+
.
.
.
=
α
x
^
T
+
1
+
(
1
−
α
)
x
^
T
+
1
=
x
^
T
+
1
即未来预测的值都等于序列平滑的最后一期的值:
x
^
T
+
l
=
x
^
T
+
1
=
x
~
T
,
l
≥
2
\hat x_{T+l}=\hat x_{T+1}=\widetilde x_T,l \ge 2
x
^
T
+
l
=
x
^
T
+
1
=
x
T
,
l
≥
2
(2)Holt两参数指数平滑
基本思想
适用于有长期趋势,无季节效应的序列
原始数据序列
{
x
t
}
\{x_t\}
{
x
t
}
,
趋势序列
{
r
t
}
\{r_t\}
{
r
t
}
x
^
t
=
x
t
−
1
+
r
t
−
1
\hat x_t=x_{t-1}+r_{t-1}
x
^
t
=
x
t
−
1
+
r
t
−
1
过去拟合:对原始数据序列和趋势序列修匀
x
~
t
=
α
x
t
+
(
1
−
α
)
(
x
~
t
−
1
+
r
t
−
1
)
\widetilde x_t=\alpha x_t+(1-\alpha)(\widetilde x_{t-1}+r_{t-1})
x
t
=
α
x
t
+
(
1
−
α
)
(
x
t
−
1
+
r
t
−
1
)
r
t
=
γ
(
x
~
t
−
x
~
t
−
1
)
+
(
1
−
γ
)
r
t
−
1
r_t=\gamma(\widetilde x_t-\widetilde x_{t-1})+(1-\gamma)r_{t-1}
r
t
=
γ
(
x
t
−
x
t
−
1
)
+
(
1
−
γ
)
r
t
−
1
平滑系数
0
<
α
,
γ
<
1
0 < \alpha,\gamma < 1
0
<
α
,
γ
<
1
指定
x
~
0
=
x
1
,
r
0
=
x
n
+
1
−
x
1
n
\widetilde x_0=x_1,r_0=\frac{x_{n+1}~~-x_1}{n}
x
0
=
x
1
,
r
0
=
n
x
n
+
1
−
x
1
未来预测
向前
l
l
l
步的预测值为:
x
^
T
+
l
=
x
~
T
+
l
∗
r
T
\hat x_{T+l}=\widetilde x_T+l*r_T
x
^
T
+
l
=
x
T
+
l
∗
r
T
(3)Holt-Winters三参数指数平滑
基本思想
适用于一定有季节效应,但长期趋势可有可无的序列