原文地址:怎么样用AIC和SC准则判断滞后阶数
作者:syzbbw66
怎么样用AIC和SC准则判断滞后阶数
ADF Test Statistic -0.480303
1%
Critical Value*
-4.7315
5%
Critical Value
-3.7611
10% Critical Value
-3.3228
*MacKinnon critical values for rejection of hypothesis of a unit root.
Augmented Dickey-Fuller Test Equation
Dependent Variable: D(Y)
Method: Least Squares
Date: 06/19/07
Time: 21:45
Sample(adjusted): 1991 2005
Included observations: 15 after adjusting endpoints
Variable Coefficient Std. Error t-Statistic Prob.
Y(-1) -0.090587 0.188604 -0.480303 0.6404
D(Y(-1)) 0.915838 0.295206 3.102368 0.0101
C -434.3619 2725.411 -0.159375 0.8763
@TREND(1989) 1076.342 1624.113 0.662726 0.5212
R-squared 0.791251
Mean dependent var
11015.85
Adjusted R-squared 0.734320
S.D. dependent var
6398.144
S.E. of regression 3297.871
Akaike info criterion
19.26312
Sum squared resid 1.20E+08
Schwarz criterion
19.45193
Log likelihood -140.4734
F-statistic
13.89831
Durbin-Watson stat 1.217269
Prob(F-statistic)
0.000464
首先在回答你的问题前,我想说你的模型是不平稳的,有单位根!
怎么用用AIC和SC准则判断滞后阶数?
多取几次滞后建立模型,比如分别建立一阶、二阶、……模型,各模型都会有一个AIC和SC统计量,取最小的统计量所对应的阶数。
只需考虑AIC和SC统计量之中的一个就可以了!通常用AIC!
Akaike info criterion 19.26312这个就是AIC的值是吗?Schwarz criterion
19.45193这个就是SC的值是吗?
AIC和SC准则并不是严格越小越好
Eviews3.1里面是没有自动提供推荐滞后阶数,需要人工输入;但是在Eviews6里面确实自动提供的,但是并不是按照最小AIC或SC准则的,AIC或SC准则越小,可能要求滞后P值越大,但是P值越大,导致数据序列损失度越大,所以肯定Eviews6做了处理,就是不知道做了什么处理。
而且我研究发现Eviews早期版本,ADF检验过程中AIC准则和SC准则都与Eviews6版本结果有些差距,问题出在模型形式的选择上(3种,带常数项、趋势项、不带等等),也就是AIC或者SC准则定义中参数k,代表要估计参数个数上,早期Eviews版本没考虑这些多余的参数,导致结果有少许差别,但不影响P的选择。
至于Eviews6中阶数是自动推荐的,但是有个Maximum选项可供输入(程序会提供一个默认值),其中Maximum=[12*(T/100)^0.25],在Maximum值变动以后,程序推荐的P值也会变动,但是这个P值绝对不是按照AIC或者SC准则推荐的在小于Maximum范围内的最小值对应的P值,这里我就一直感到困惑,希望谁能给指导一下!
我的结果在这:
Maximum=10默认,选择带常数项和趋势项情况:
AIC准则下Eviews6:
Null Hypothesis: CCC has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 5 (Automatic based on AIC, MAXLAG=10)
t-Statistic
Prob.*
Augmented Dickey-Fuller test statistic
-3.208100
0.0941
Test critical values: 1% level -4.144584
5% level
-3.498692
10% level
-3.178578
*MacKinnon (1996) one-sided p-values.
Augmented Dickey-Fuller Test Equation
Dependent Variable: D(CCC)
Method: Least Squares
Date: 08/22/08
Time: 10:46
Sample (adjusted): 7 58
Included observations: 52 after adjustments
Coefficient Std. Error t-Statistic Prob.
CCC(-1) -0.039917 0.012443 -3.208100 0.0025
D(CCC(-1)) 2.234078 0.143978 15.51677 0.0000
D(CCC(-2)) -2.163752 0.359186 -6.024037 0.0000
D(CCC(-3)) 1.220026 0.454195 2.686126 0.0102
D(CCC(-4)) -0.406455 0.359970 -1.129135 0.2650
D(CCC(-5)) 0.187012 0.161019 1.161427 0.2517
C 4.338536 1.364059 3.180607 0.0027
@TREND(1) -0.001814 0.001293 -1.402702 0.1677
R-squared 0.976187
Mean dependent var
0.005962
Adjusted R-squared 0.972398
S.D. dependent var
0.512216
S.E. of regression 0.085098
Akaike info criterion
-1.949385
Sum squared resid 0.318635
Schwarz criterion
-1.649193
Log likelihood 58.68401
Hannan-Quinn criter.
-1.834299
F-statistic 257.6749
Durbin-Watson stat
1.735674
Prob(F-statistic) 0.000000
SC准则,Evews6:
Null Hypothesis: CCC has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 3 (Automatic based on SIC, MAXLAG=10)
t-Statistic
Prob.*
Augmented Dickey-Fuller test statistic
-4.135397
0.0101
Test critical values: 1% level -4.137279
5% level
-3.495295
10% level
-3.176618
*MacKinnon (1996) one-sided p-values.
Augmented Dickey-Fuller Test Equation
Dependent Variable: D(CCC)
Method: Least Squares
Date: 08/22/08
Time: 10:46
Sample (adjusted): 5 58
Included observations: 54 after adjustments
Coefficient Std. Error t-Statistic Prob.
CCC(-1) -0.033616 0.008129 -4.135397 0.0001
D(CCC(-1)) 2.187745 0.091287 23.96556 0.0000
D(CCC(-2)) -1.959517 0.160505 -12.20843 0.0000
D(CCC(-3)) 0.798387 0.102512 7.788261 0.0000
C 3.641515 0.890272 4.090339 0.0002
@TREND(1) -0.001171 0.000964 -1.213975 0.2307
R-squared 0.977158
Mean dependent var
0.033519
Adjusted R-squared 0.974778
S.D. dependent var
0.522151
S.E. of regression 0.082925
Akaike info criterion
-2.037325
Sum squared resid 0.330073
Schwarz criterion
-1.816327
Log likelihood 61.00779
Hannan-Quinn criter.
-1.952095
F-statistic 410.6713
Durbin-Watson stat
1.932021
Prob(F-statistic) 0.000000
可以看出,明显AIC值:推荐P=5,Schwarz criterion -1.649193
Akaike info criterion
-1.949385
SC推荐P=3,Schwarz criterion -1.816327
Akaike info criterion
-2.037325
如果按照AIC准则,明显P=3时的值比P=5时的值小,但是Eviews6选择的是P=5;
同理按照SC准则,明显P=3时的值比P=5时的值小,这里还算正确,但是我算出其它还有比P=3时SC准则值更小的P值,例如P=5或者6,问题就在这里。
ADF Test Statistic -0.480303
首先在回答你的问题前,我想说你的模型是不平稳的,有单位根!
怎么用用AIC和SC准则判断滞后阶数?
多取几次滞后建立模型,比如分别建立一阶、二阶、……模型,各模型都会有一个AIC和SC统计量,取最小的统计量所对应的阶数。
只需考虑AIC和SC统计量之中的一个就可以了!通常用AIC!
Akaike info criterion
AIC和SC准则并不是严格越小越好
Eviews3.1里面是没有自动提供推荐滞后阶数,需要人工输入;但是在Eviews6里面确实自动提供的,但是并不是按照最小AIC或SC准则的,AIC或SC准则越小,可能要求滞后P值越大,但是P值越大,导致数据序列损失度越大,所以肯定Eviews6做了处理,就是不知道做了什么处理。
而且我研究发现Eviews早期版本,ADF检验过程中AIC准则和SC准则都与Eviews6版本结果有些差距,问题出在模型形式的选择上(3种,带常数项、趋势项、不带等等),也就是AIC或者SC准则定义中参数k,代表要估计参数个数上,早期Eviews版本没考虑这些多余的参数,导致结果有少许差别,但不影响P的选择。
至于Eviews6中阶数是自动推荐的,但是有个Maximum选项可供输入(程序会提供一个默认值),其中Maximum=[12*(T/100)^0.25],在Maximum值变动以后,程序推荐的P值也会变动,但是这个P值绝对不是按照AIC或者SC准则推荐的在小于Maximum范围内的最小值对应的P值,这里我就一直感到困惑,希望谁能给指导一下!
我的结果在这:
Maximum=10默认,选择带常数项和趋势项情况:
AIC准则下Eviews6:
Null Hypothesis: CCC has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 5 (Automatic based on AIC, MAXLAG=10)
Augmented Dickey-Fuller test statistic
Test critical values: 1% level
*MacKinnon (1996) one-sided p-values.
Augmented Dickey-Fuller Test Equation
Dependent Variable: D(CCC)
Method: Least Squares
Date: 08/22/08
Sample (adjusted): 7 58
Included observations: 52 after adjustments
CCC(-1) -0.039917 0.012443 -3.208100 0.0025
D(CCC(-1)) 2.234078 0.143978 15.51677 0.0000
D(CCC(-2)) -2.163752 0.359186 -6.024037 0.0000
D(CCC(-3)) 1.220026 0.454195 2.686126 0.0102
D(CCC(-4)) -0.406455 0.359970 -1.129135 0.2650
D(CCC(-5)) 0.187012 0.161019 1.161427 0.2517
C 4.338536 1.364059 3.180607 0.0027
@TREND(1) -0.001814 0.001293 -1.402702 0.1677
R-squared 0.976187
Adjusted R-squared 0.972398
S.E. of regression 0.085098
Sum squared resid 0.318635
Log likelihood 58.68401
F-statistic 257.6749
Prob(F-statistic) 0.000000
SC准则,Evews6:
Null Hypothesis: CCC has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 3 (Automatic based on SIC, MAXLAG=10)
Augmented Dickey-Fuller test statistic
Test critical values: 1% level
*MacKinnon (1996) one-sided p-values.
Augmented Dickey-Fuller Test Equation
Dependent Variable: D(CCC)
Method: Least Squares
Date: 08/22/08
Sample (adjusted): 5 58
Included observations: 54 after adjustments
CCC(-1) -0.033616 0.008129 -4.135397 0.0001
D(CCC(-1)) 2.187745 0.091287 23.96556 0.0000
D(CCC(-2)) -1.959517 0.160505 -12.20843 0.0000
D(CCC(-3)) 0.798387 0.102512 7.788261 0.0000
C 3.641515 0.890272 4.090339 0.0002
@TREND(1) -0.001171 0.000964 -1.213975 0.2307
R-squared 0.977158
Adjusted R-squared 0.974778
S.E. of regression 0.082925
Sum squared resid 0.330073
Log likelihood 61.00779
F-statistic 410.6713
Prob(F-statistic) 0.000000
可以看出,明显AIC值:推荐P=5,Schwarz criterion
SC推荐P=3,Schwarz criterion
如果按照AIC准则,明显P=3时的值比P=5时的值小,但是Eviews6选择的是P=5;
同理按照SC准则,明显P=3时的值比P=5时的值小,这里还算正确,但是我算出其它还有比P=3时SC准则值更小的P值,例如P=5或者6,问题就在这里。