关于 Stata 时间序列的处理,我前面写过两篇文章分享了一些基础知识和概念。
(1)盲区行者王:【Stata时间序列】时间序列定义 - Defining Time-Series in Stata
(2) 盲区行者王:【Stata时间序列】AR/MA/ARIMA模型及定阶问题
接下来,我们将使用真实的数据建立ARIMA模型,并基于该模型进行预测和评估。
<hr/>1. ARIMA(p, d, q)模型简介
- AR(p),变量的自回归部分,即用变量的p个滞后期对当期进行建模
- I(d),d阶差分
- MA(q),残差的q个滞后期加入模型(当期的残差也保留在模型)
- 以上3个都是平稳过程。所以MA(q) 残差序列是个白噪声序列
- 所谓平稳:单位根的值全部在单位圆之外
- p和q定阶:可通过AIC/BIC。且通常p+q <= 5
- 模型估计后,若残差不是白噪声,应考虑增加阶数
<hr/>2. 导入数据和数据探索
该数据是美国1960q1~1990q1的批发价格指数的时间序列。
**A导入数据
use &#34;D:\Stata\wpi.dta&#34;, clear
*数据整理
tsset t
gen lnwpi = ln(wpi)
tsline wpi lnwpi
tsline wpi lnwpi
从图中可以看出,对数化后的lnwpi已经比较平稳,适合用来建模。
<hr/>3. ARIMA定阶
ac lnwpi, xline(10)
pac lnwpi
ac lnwpi, xline(10)
从图中可以看出,AC拖尾。所以q我们可以定在2,最终的阶数换成AIC/BIC来界定。
pac lnwpi
PAC:从PAC图中可以看出,我们可以p把定在2阶。
<hr/>4. ARIMA建模
我们可以先尝试ARIMA(2, 0, 2),然后再对比更低阶的组合。
. arima lnwpi, arima(2, 0, 2)
(setting optimization to BHHH)
Iteration 0: log likelihood = -1283.2761
Iteration 1: log likelihood = -103.11292
Iteration 2: log likelihood = 341.39364
Iteration 3: log likelihood = 358.77798
Iteration 4: log likelihood = 367.6482
(switching optimization to BFGS)
Iteration 5: log likelihood = 369.83366
Iteration 6: log likelihood = 372.04944
Iteration 7: log likelihood = 373.17791
Iteration 8: log likelihood = 377.9509
Iteration 9: log likelihood = 379.67962
Iteration 10: log likelihood = 379.81083
Iteration 11: log likelihood = 379.82898
Iteration 12: log likelihood = 379.83179
Iteration 13: log likelihood = 379.83388
Iteration 14: log likelihood = 379.83542
(switching optimization to BHHH)
Iteration 15: log likelihood = 379.83566
Iteration 16: log likelihood = 379.83572
Iteration 17: log likelihood = 379.83573
Iteration 18: log likelihood = 379.83573
Iteration 19: log likelihood = 379.83574
(switching optimization to BFGS)
Iteration 20: log likelihood = 379.83574
Iteration 21: log likelihood = 379.83574
Iteration 22: log likelihood = 379.83574
Iteration 23: log likelihood = 379.83574
Iteration 24: log likelihood = 379.83574
Iteration 25: log likelihood = 379.83574
ARIMA regression
Sample: 1960q1 - 1990q4 Number of obs = 124
Wald chi2(4) = 700423.10
Log likelihood = 379.8357 Prob > chi2 = 0.0000
------------------------------------------------------------------------------
| OPG
lnwpi | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lnwpi |
_cons | 4.155654 .5365695 7.74 0.000 3.103998 5.207311
-------------+----------------------------------------------------------------
ARMA |
ar |
L1. | 1.948445 .0453341 42.98 0.000 1.859592 2.037298
L2. | -.9490278 .0453945 -20.91 0.000 -1.038 -.8600562
|
ma |
L1. | -.5230613 .0867427 -6.03 0.000 -.6930739 -.3530488
L2. | -.0215154 .0940398 -0.23 0.819 -.2058301 .1627992
-------------+----------------------------------------------------------------
/sigma | .0108905 .0004617 23.59 0.000 .0099855 .0117955
------------------------------------------------------------------------------
Note: The test of the variance against zero is one sided, and the two-sided
confidence interval is truncated at zero.
. estat ic //AIC/BIC
Akaike&#39;s information criterion and Bayesian information criterion
-----------------------------------------------------------------------------
Model | N ll(null) ll(model) df AIC BIC
-------------+---------------------------------------------------------------
. | 124 . 379.8357 6 -747.6715 -730.7498
-----------------------------------------------------------------------------其中ma的二阶项不显著。
我们转向(2,0,1)。
. arima lnwpi, arima(2, 0, 1)
(setting optimization to BHHH)
Iteration 0: log likelihood = -527.79871
Iteration 1: log likelihood = 55.48948
Iteration 2: log likelihood = 304.69711
Iteration 3: log likelihood = 359.9311
Iteration 4: log likelihood = 369.31001
(switching optimization to BFGS)
Iteration 5: log likelihood = 369.50296
Iteration 6: log likelihood = 376.48981
Iteration 7: log likelihood = 378.49434
Iteration 8: log likelihood = 379.46257
Iteration 9: log likelihood = 379.75527
Iteration 10: log likelihood = 379.78866
Iteration 11: log likelihood = 379.79723
Iteration 12: log likelihood = 379.80608
Iteration 13: log likelihood = 379.80917
Iteration 14: log likelihood = 379.80949
(switching optimization to BHHH)
Iteration 15: log likelihood = 379.80954
Iteration 16: log likelihood = 379.80954
Iteration 17: log likelihood = 379.80954
Iteration 18: log likelihood = 379.80954
Iteration 19: log likelihood = 379.80955
(switching optimization to BFGS)
Iteration 20: log likelihood = 379.80955
Iteration 21: log likelihood = 379.80955
Iteration 22: log likelihood = 379.80955
ARIMA regression
Sample: 1960q1 - 1990q4 Number of obs = 124
Wald chi2(3) = 668241.19
Log likelihood = 379.8095 Prob > chi2 = 0.0000
------------------------------------------------------------------------------
| OPG
lnwpi | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lnwpi |
_cons | 4.155434 .5353873 7.76 0.000 3.106094 5.204774
-------------+----------------------------------------------------------------
ARMA |
ar |
L1. | 1.946284 .0246227 79.04 0.000 1.898024 1.994544
L2. | -.9468694 .0243307 -38.92 0.000 -.9945568 -.8991821
|
ma |
L1. | -.5325685 .0817276 -6.52 0.000 -.6927516 -.3723853
-------------+----------------------------------------------------------------
/sigma | .0108925 .0004577 23.80 0.000 .0099954 .0117895
------------------------------------------------------------------------------
Note: The test of the variance against zero is one sided, and the two-sided
confidence interval is truncated at zero.
. estat ic
Akaike&#39;s information criterion and Bayesian information criterion
-----------------------------------------------------------------------------
Model | N ll(null) ll(model) df AIC BIC
-------------+---------------------------------------------------------------
. | 124 . 379.8095 5 -749.6191 -735.5177
-----------------------------------------------------------------------------
Note: BIC uses N = number of observations. See [R] BIC note.AIC和BIC都变小了。模型更优。
接下来尝试(1,0,1)。
. arima lnwpi, arima(1, 0, 1)
(setting optimization to BHHH)
Iteration 0: log likelihood = 260.46185
Iteration 1: log likelihood = 335.28097
Iteration 2: log likelihood = 338.35393
Iteration 3: log likelihood = 342.81478
Iteration 4: log likelihood = 343.05124
(switching optimization to BFGS)
Iteration 5: log likelihood = 343.1412
Iteration 6: log likelihood = 344.24394
Iteration 7: log likelihood = 344.40456
Iteration 8: log likelihood = 344.4141
Iteration 9: log likelihood = 344.42378
Iteration 10: log likelihood = 344.42747
Iteration 11: log likelihood = 344.42763
Iteration 12: log likelihood = 344.42763
Iteration 13: log likelihood = 344.42763
ARIMA regression
Sample: 1960q1 - 1990q4 Number of obs = 124
Wald chi2(2) = 29666.47
Log likelihood = 344.4276 Prob > chi2 = 0.0000
------------------------------------------------------------------------------
| OPG
lnwpi | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lnwpi |
_cons | 4.091086 .6539676 6.26 0.000 2.809333 5.372838
-------------+----------------------------------------------------------------
ARMA |
ar |
L1. | .9994616 .0059189 168.86 0.000 .9878607 1.011062
|
ma |
L1. | .4902354 .0855512 5.73 0.000 .3225582 .6579126
-------------+----------------------------------------------------------------
/sigma | .0145748 .0006139 23.74 0.000 .0133716 .015778
------------------------------------------------------------------------------
Note: The test of the variance against zero is one sided, and the two-sided
confidence interval is truncated at zero.
. estat ic
Akaike&#39;s information criterion and Bayesian information criterion
-----------------------------------------------------------------------------
Model | N ll(null) ll(model) df AIC BIC
-------------+---------------------------------------------------------------
. | 124 . 344.4276 4 -680.8553 -669.5741
-----------------------------------------------------------------------------
Note: BIC uses N = number of observations. See [R] BIC note.综合3个模型的AIC和BIC,可以看出ARIMA(1, 0, 1)的模型最优,且其中p+q<=5。
<hr/>5. 模型预测
. **predict
. predict lnwpi_hat, xb
. predict r, residual
. list lnwpi lnwpi_hat r in 1/10
+---------------------------------+
| lnwpi lnwpi_hat r |
|---------------------------------|
1. | 3.424263 4.091085 -.6668228 |
2. | 3.427515 3.424463 .0030513 |
3. | 3.424263 3.429078 -.0048152 |
4. | 3.424263 3.422366 .0018966 |
5. | 3.427515 3.425542 .0019728 |
|---------------------------------|
6. | 3.417727 3.428837 -.0111098 |
7. | 3.417727 3.412646 .0050805 |
8. | 3.421 3.42058 .0004204 |
9. | 3.424263 3.421567 .0026959 |
10. | 3.421 3.425943 -.0049434 |
+---------------------------------+最终,要对模型的残差进行检验,看是否已经是白噪声。
. wntestq r
Portmanteau test for white noise
---------------------------------------
Portmanteau (Q) statistic = 1.0207
Prob > chi2(40) = 1.0000
. wntestb r
wntestb r
wntestq和wntestb检验的p值都非常大,所以我们不能拒绝残差是平稳序列的原假设。
因此,原模型ARIMA(1, 0, 1)拟合效果非常好。
<hr/>6. 进一步探索
- 考虑到wpi是季节数据,我们可以在MA中加入四阶的之后项。
对应命令:arima lnwpi, ar(1) ma(1 4)
- 或者,对lnwpi进一步检验,如果它不平稳考虑进行一阶差分
对应命令:arima lnwpi, ar(1) d(1) ma(1 4) <hr/>References
- 陈强,2015,《计量经济学和Stata应用》
- 杨维忠,2020,《Stata统计分析与实验指导》
<hr/>更多Stata/R语言和Python的数据分析经验分享,可移步我的个人资深主页:
-----全文结束----- |