Download RATS Programming Manual, By Walter Enders

Transcript
IF Statements and Monte Carlo Experiments 149
Statistics on Series F
Observations
50000
Sample Mean
1.6524588263
Standard Error 1.1741364996
t-Statistic
314.70023
Skewness
1.73656
Kurtosis
5.01475
Jarque-Bera
77521.44422
Variance
SE of Sample
Signif Level
Signif Level
Signif Level
Signif Level
Minimum
01-%ile
0.0013651902
0.1977081049
Maximum
99-%ile
14.4743025649
5.6967829958
05-%ile
0.3681509883
95-%ile
3.9233591336
10-%ile
25-%ile
Median
0.5025774653
0.8187794721
1.3579015615
90-%ile
75-%ile
3.1857908038
2.1691250789
Mean
(Mean=0)
(Sk=0)
(Ku=0)
(JB=0)
1.378597
0.005251
0.00000000
0.00000000
0.00000000
0.00000000
Suppose that you estimated a series as a threshold process:
∆yt = Itρ1yt-1 + (1 – It)ρ2yt-1 + εt
 1 if y t −1 ≥ 0
where: I t = 
 0 if y t −1 < 0
If the sample F-statistic of the null hypothesis ρ1 = ρ2 = 0 was 3.5, you would be able to reject
the null hypothesis at the 10% significance level but not the 5% level. You can modify the
program by (i) including an INFOBOX, (ii) obtaining the critical values for additional sample
sizes, and (iii) obtaining the critical values when Chan’s (1993) method (see Section 3.1 in this
chapter) is used to estimate the threshold.
5.5 Inference in a Cointegrated System
This is one of my favorite programs. It illustrates a number of RATS advanced features and the
folly of using traditional distribution theory to perform hypothesis tests on a cointegrating vector.
Suppose that xt and yt are two non-stationary time-series variables that are cointegrated of order
1. The error-correction representation of the system is:
yt = yt-1 - α1 [ β0 + yt-1 - β1xt-1 ] + A11(L)∆yt-1 + A12(L)∆xt-1 + e1t
xt = xt-1 + α2[ β0 + yt-1 - β1xt-1 ] + A21(L)∆yt-1 + A22(L)∆xt-1 + e2t
It is assumed that e1t and e2t are serially uncorrelated but the covariance Ee1t e2t need not be zero.
As in Chapter 2, if the variances and covariance are time-invariant, we can write the
variance/covariance matrix as: