RSTAR package (Version 0.1-1)
搜不到,应该没发布.
利用R data lynx,在s-plus运行ESTAR MODEL
可得结果如下:
#################in s-plus
> lynx.estar$parameters

[1] -0.4641629
3.1881329
> exp(lynx.estar$parameters[1])/var(lynx.RHS[,2])

[1] 2.015791

这个结果跟在R运行estar.R的结果接近
estar.R是我修改的R code
#################in R
ESTAR
Coefficients:
Low regime:
   phi1.0    phi1.1    phi1.2
1.585903  1.588434 -1.132034
High regime:
    phi2.0     phi2.1     phi2.2
-1.0445247 -0.5018974  0.8767483
Smoothing parameter: gamma = 1.998
Threshold
Variable: Z(t) = + (1) X(t) + (0) X(t-1)
Value: 3.189
##########
#s-plus code
ESTAR.res = function(theta, g.scale, x, y, q)
{
k = ncol(x)
G = 1 - exp( - exp(theta[1])/g.scale * (q - theta[2])^2)
X = cbind(x * (1 - G), x * G)
m = crossprod(t(backsolve(chol(crossprod(X)), diag(2 * k))))
beta = m %*% t(X) %*% y
y - X %*% beta
}
source("lynx.R")
lynx.LHS = log10(lynx)[3:length(lynx)]
lynx.RHS = cbind(1, tslag(log10(lynx), 1:2, trim=T))
lynx.estar = nlregb(length(lynx)-2,
start=c(0,mean(lynx.RHS[,2])),
residuals=ESTAR.res,
lower=c(-Inf, min(lynx.RHS[,2])),
upper=c( Inf, max(lynx.RHS[,2])),
g.scale=var(lynx.RHS[,2]),
x=lynx.RHS, y=lynx.LHS, q=lynx.RHS[,2])
lynx.estar$parameters
exp(lynx.estar$parameters[1])/var(lynx.RHS[,2])

##############
#R code
#don't need library(tsDyn)
source("estar.R")
lynx.estar <- estar(log10(lynx), m=2,d=1, control=list(maxit=3000))
lynx.estar

estar.R

estar.rar
大小:(3.17 KB)

 马上下载

本附件包括：

• estar.R

谢谢你了啊，我刚把教程详细的看了一遍，打算用S-PLUS来做，可是在论坛里找到的那个版本我下载了，但是老是解压不成功！
下面是地址http://www.pinggu.org/bbs/viewthread.php?tid=296390&extra=&highlight=Finmetrics&page=26
你那里有SPLUS吗？有的话可以发给我个下载地址吗？

http://www.pinggu.org/bbs/dispbb ... d=290684&star=3

谢谢你了啊，我刚安装好了S-PLUS，可是又碰到新问题了，是不是我没有调用函数包啊？
> x<-c(127.4,128.4,127.1,126.6,124.8,122.8,119.4,115.3,111.5,108.6,107.4,106.6,104.3,104.4,103.4,103.2,102.7,101.1,101.1,102.5,102.9,103.1,103.7,104.3,102.2,101,101.6,101.3,103.6);
> nonlinearTest(x, method="STAR-LM", p=2, d=1:2)
Problem: Couldn't find a function definition for "nonlinearTest"
Use traceback() to see the call stack
> traceback()
4: eval(action, sys.parent())
3: doErrorAction("Problem: Couldn't find a function definition for \"nonlinearTest\"", 1000)
2: eval(expression(nonlinearTest(x, method = "STAR-LM", p = 2, d = 1:2)))
1:
Message: Problem: Couldn't find a function definition for "nonlinearTest"
>

9# epoh
谢谢你了啊！

The LM test for STAR nonlinearity can be performed in S+FinMetrics
所以首先要载入FinMetrics

1.file\load module\finmetrics  or
2.module(finmetrics)
#########
module(finmetrics)
ndx.ret2 = getReturns(ndx.dat[,"Close"])^2
ndx.rvol = sqrt(aggregate(ndx.ret2, FUN=sum, by="weeks",week.align=1))
nonlinearTest(log(ndx.rvol), method="STAR-LM", p=2, d=1:2)

谢谢你了啊，我试了你给我的程序可以运行，我自己再好好学习一下，有问题再向你请教啊！ 11# epoh

您好！又有问题请教你了，我把数据导入了object explorer，如果我想在commands框（编程窗口）调用这个数据集中的某一个变量该如何编程呢，谢谢你了啊！例如在我导入了名为CPI的excel格式的文件（含有TIME和CPI两个变量），现在想在程序窗口调用CPI变量该如何编程序呢？ 11# epoh

#file name: data.xls
class(data)
#[1] "data.frame"
names(data)
#[1] "TIME" "CPI"
cpi=data$CPI
cpi

您好，我在运行程序时出现了这个问题，麻烦您帮我看下问题出在那里好吗？我运行你发给我的那个程序就不会出问题，谢谢您了啊！

> #file name: y.xls
> class(y)
[1] "data.frame"
> names(y)
[1] "x"
> x
NULL
> x=y$x
> x
  [1] 127.4 128.4 127.1 126.6 124.8 122.8 119.4 115.3 111.5 108.6 107.4 106.6
[13] 104.3 104.4 103.4 103.2 102.7 101.1 101.1 102.5 102.9 103.1 103.7 104.3
[25] 102.2 101.0 101.6 101.3 103.6 104.4 104.7 104.9 104.5 104.8 104.4 104.5
[37] 105.5 105.3 105.3 107.1 104.7 104.8 105.2 105.8 107.5 107.9 108.2 108.8
[49] 110.3 110.5 112.2 112.6 114.0 115.1 116.2 116.0 115.7 115.9 116.7 118.8
[61] 121.1 123.2 122.4 121.7 121.3 122.6 124.0 125.8 127.3 127.7 127.5 125.5
[73] 124.1 122.4 121.3 120.7 120.3 118.2 116.7 114.5 113.2 112.1 111.2 110.1
[85] 109.0 109.3 109.8 109.7 108.9 108.6 108.3 108.1 107.4 107.0 106.9 107.0
[97] 105.9 105.6 104.0 103.2 102.8 102.8 102.7 101.9 101.8 101.5 101.1 100.4
[109] 100.3  99.9 100.7  99.7  99.0  98.7  98.6  98.6  98.5  98.9  98.8  99.0
[121]  98.8  98.7  98.2  97.8  97.8  97.9  98.6  98.7  99.2  99.4  99.1  99.0
[133]  99.8 100.7  99.8  99.7 100.1 100.5 100.5 100.3 100.0 100.0 101.3 101.5
[145] 101.2 100.0 100.8 101.6 101.7 101.4 101.5 101.0  99.9 100.2  99.7  99.7
[157]  99.0 100.0  99.2  98.7  98.9  99.2  99.1  99.3  99.3  99.2  99.3  99.6
[169] 100.4 100.2 100.9 101.0 100.7 100.3 100.5 100.9 101.1 101.8 103.0 103.2
[181] 103.2 102.1 103.0 103.8 104.4 105.0 105.3 105.3 105.2 104.3 102.8 102.4
[193] 101.9 103.9 102.7 101.8 101.8 101.6 101.8 101.3 100.9 101.2 101.3 101.6
[205] 101.9 100.9 100.8 101.2 101.4 101.5 101.0 101.3 101.5 101.4 101.9 102.8
[217] 102.2 102.7 103.3 103.0 103.4 104.4 105.6 106.5 106.2 106.5 106.9 106.5
[229] 107.1 108.7 108.3 108.5 107.7 107.1 106.3 104.9 104.6 104.0 102.4 101.2
[241] 101.0  98.4  98.8  98.5  98.6  98.3  98.2  98.8  99.2  99.5 100.6 101.9
[253] 101.5 102.7 102.4 102.8 103.1
> module(finmetrics)
> nonlinearTest(x, method="STAR-LM", p=2, d=1:2)
Problem in solve.qr(a, b): apparently singular matrix
Use traceback() to see the call stack
> nonlinearTest(log(x), method="STAR-LM", p=2, d=1:2)
Problem in solve.qr(a, b): apparently singular matrix
Use traceback() to see the call stack
>
14# epoh

这跟你的数据有关
因为求解过程会用到
solve(crossprod(rr), diag(k.restrict))

###数据是lynx
>crossprod(rr)
          lag1      lag2      lag1      lag2     lag1      lag2
lag1 12.575130  6.348474 17.784222  9.529057 20.35809 11.688822
lag2  6.348474  7.611890  9.019838  9.198657 10.29759  9.553279
lag1 17.784222  9.019838 25.795302 14.036471 29.87362 17.455131
lag2  9.529057  9.198657 14.036471 12.234994 16.28877 13.568947
lag1 20.358092 10.297585 29.873624 16.288774 34.79839 20.406622
lag2 11.688822  9.553279 17.455131 13.568947 20.40662 15.717428

>solve(crossprod(rr))
          lag1       lag2      lag1      lag2       lag1       lag2
lag1  110.74121  -69.78052 -218.7741  128.9377  123.80462  -69.03512
lag2  -69.78052   65.23911  138.2469 -123.7797  -79.76322   69.12993
lag1 -218.77408  138.24688  442.1080 -259.3908 -253.83493  141.18214
lag2  128.93773 -123.77965 -259.3908  239.9478  151.78494 -136.80259
lag1  123.80462  -79.76322 -253.8349  151.7849  147.34293  -84.03038
lag2  -69.03512   69.12993  141.1821 -136.8026  -84.03038   79.79755

###数据是cpi
>crossprod(rr):
           lag1       lag2       lag1       lag2       lag1       lag2
lag1 0.01564787 0.01475953 0.03040995 0.02874525 0.04403419 0.04171723
lag2 0.01475953 0.01445768 0.02868977 0.02813183 0.04155108 0.04078790
lag1 0.03040995 0.02868977 0.05918261 0.05595727 0.08581856 0.08132714
lag2 0.02874525 0.02813183 0.05595727 0.05482213 0.08116009 0.07960549
lag1 0.04403419 0.04155108 0.08581856 0.08116009 0.12461643 0.11812597
lag2 0.04171723 0.04078790 0.08132714 0.07960549 0.11812597 0.11576519

solve(crossprod(rr))
Problem in solve.qr(a): apparently singular matrix

谢谢啊，有问题再向您请教！ 16# epoh

您好！下面是我用S-PLUS估计的STAR模型结果，麻烦您帮我看一下后面给我的警告是Choleski decomposition时chol(crossp\ rod(x$gradient))不是满秩矩阵造成的吧？是数据本身的原因吧？
> ndx.lstar = STAR(y, p=1, d=2)
>  summary(ndx.lstar)
Call:
STAR(X = y, p = 1, d = 2)
Parameter estimates:
                 Values     Std.Error Std.Error.white
    gamma 1.525360e+002 1.314598e+018            NA  
threshold 8.624000e+000 1.529253e+013            NA  

Coefficient estimates and standard errors:
Lower regime:
                 Values Std.Error Std.Error.white
intercept(lower) 1.432  0.146     0.032         
     lag1(lower) 1.107  0.006     0.000         
Upper regime:
                 Values Std.Error Std.Error.white
intercept(upper) -2.204  0.913     0.228         
     lag1(upper)  0.107  0.013     0.000         
Std. Errors of Residuals:
[1] 3.91
Information Criteria:
     logL      AIC      BIC       HQ
  -81.475  170.950  176.555  172.743

Degrees of freedom:
total residuals
    30        26
Warning messages:
1: Choleski decomposition not of full rank in: chol(crossp\
rod(x$gradient))
2: Choleski decomposition not of full rank in: chol(crossp\
rod(x$gradient))
还有如果我想在S-PLUS里调用转换函数做转换函数和转换变量的散点图应该如何写程序啊？谢谢您了！还有如果我想做LSTAR和ESTAR的转换函数的图形，转换变量可以任意取值（即可以从负无穷取到正无穷），应该如何写程序作图呢？谢谢您了啊！
16# epoh

你是要划page 677的图形?
FIGURE 18.9. Logistic and exponential transition functions

是的，677页的我已经解决。现在我想做出转换变量和转换函数的散点图，想问下您模型估计后可不可以直接调用计算过程中的变量来做出这个散点图! 19# epoh

module(finmetrics)
ndx.ret2 = getReturns(ndx.dat[,"Close"])^2
ndx.rvol = sqrt(aggregate(ndx.ret2, FUN=sum, by="weeks",week.align=1))
ndx.lstar = STAR(log(ndx.rvol), p=2, d=1)
summary(ndx.lstar)

n=length(ndx.rvol)          #302
z=log10(ndx.rvol)[2:(n-1)]  #300
g=ndx.lstar$gamma
th=ndx.lstar$threshold
gfun <- function(y, g, th)      1 / (1 + exp(-g*(y-th)))  
G=gfun(z,g,th)
par(mfrow=c(2,1))
plot(G)

G.zt=as.vector(G)
zt=as.vector(z)
plot(G.zt,zt)

请问前辈有没有三机制或者四机制的LSTAR或者ESTAR的教程和程序吗？有的话麻烦您发给我好吗？谢谢您了！谢谢您最近一直帮助我， 21# epoh

抱歉!我没这方面资料.

你的三机制,四机制,指的可是
multiple regime smooth transition autoregressive

对，就是这方面的资料。 24# epoh

multiple regime smooth transition autoregressive
可以应用package "tsDyn" function star()的引数noRegimes
  function star() can fit with automatic selection
  of the number of regimes based on LM tests.

数据lynx,ndx.rvol都不适合 Add regime
因为 Regime 3 is NOT accepted (p-Value = 0.8678928 )

在此我提供另一个数据svpdx.dat供练习:
The data consist of a time series of daily
  Pound/Dollar exchange rates
  from 01/10/81 to 28/6/85.

library(tsDyn)
svpdx=read.table("svpdx.dat")
mod1.star <- star(svpdx,noRegimes=3, m=2,d=1, control=list(maxit=3000))
mod1.star

Using default threshold variable: thDelay=0
Testing linearity...   p-Value =  0.01685132
The series is nonlinear. Incremental building procedure:
Building a 2 regime STAR.
Using default threshold variable: thDelay=0
Performing grid search for starting values...
Starting values fixed: gamma =  17 , th =  -0.8636345 ; SSE =  467.8267
Optimization algorithm converged
Optimized values fixed for regime 2  : gamma =  17.00000 , th =  -0.9237364
Testing for addition of regime 3.
  Estimating gradient matrix...
  Done. Computing the test statistic...
  Done. Regime 3 is needed (p-Value =  0.004081644 ).
Adding regime  3 .
Fixing good starting values for regime  3 ...
Reordering regimes...
Estimating parameters of regime 3 ...
Optimized values fixed for regime  3 : gamma =  156.3195 , th =  1.685289
Optimization algorithm converged
Optimized linear values:
0.7151267 0.5046379 -0.1193560
-0.7685278 -0.3892954 0.1769597
2.446755 -0.881913 -0.2143251
Ok.
Testing for addition of regime  4 .
  Estimating gradient matrix...
  Computing the test statistic...
Regime  4  is NOT accepted (p-Value =  0.1171196 ).
Finished building a MRSTAR with  3  regimes

> mod1.star
Non linear autoregressive model
Multiple regime STAR model
Regime  1 :
    Linear parameters: 0.7151267, 0.5046379, -0.119356
Regime  2 :
    Linear parameters: -0.7685278, -0.3892954, 0.1769597
    Non-linear parameters:
    22.7263412, -0.9325508
Regime  3 :
    Linear parameters: 2.4467549, -0.8819129, -0.2143251
    Non-linear parameters:
    156.3195267, 1.685289


svpdx.dat

svpdx.rar
大小:(8.77 KB)

 马上下载

本附件包括：

• svpdx.dat

Smooth Transition Autoregressive Models

Smooth Transition Autoregressive Models.pdf
大小:(657.91 KB)

 马上下载

谢谢您啊！ 26# epoh

您好！麻烦您帮我看一下下面的程序，我在R中估计LSTAR模型时，估计的很顺利，但是我想查看中间过程中产生的多有变量结果老是错误：
> sum.s<-summary(ndx.lstar)
错误于summary.lstar(ndx.lstar) : dims [product 48]与对象长度[0]不匹配
> summary(ndx.lstar)
错误于summary.lstar(ndx.lstar) : dims [product 48]与对象长度[0]不匹配
请问您一下这是什么原因呢？希望您给予指导，期待您的回复，谢谢了！



> getwd()
[1] "d:/我的文档"
> svpdx <- read.table("data.txt", header = TRUE);
> svpdx
    cpi  rpi          m2
1   3.4  2.9  0.26524514
2   6.4  5.4  0.31278198
3  14.7 13.2  0.37310154
4  24.1 21.7  0.34529154
5  17.1 14.8  0.29467111
6   8.3  6.1 -0.03867458
7   2.8  0.8  0.34668927
8  -0.8 -2.6  0.17245702
9  -1.4 -3.0  0.14411363
10  0.4 -1.5 -0.31433668
11  0.7 -0.8  0.87572994
12 -0.8 -1.3  0.16337324
13  1.2 -0.1  0.19679602
14  3.9  2.8  0.18161695
15  1.8  0.8  0.14387323
16  1.5  1.0  0.18886459
17  4.8  3.8  0.15787746
18  5.9  5.9  0.18875571
19 -0.7 -1.2  0.18736354
> x <- svpdx$cpi
> x
[1]  3.4  6.4 14.7 24.1 17.1  8.3  2.8 -0.8 -1.4  0.4  0.7 -0.8  1.2  3.9  1.8  1.5  4.8  5.9 -0.7
> y<-svpdx$rpi
> y
[1]  2.9  5.4 13.2 21.7 14.8  6.1  0.8 -2.6 -3.0 -1.5 -0.8 -1.3 -0.1  2.8  0.8  1.0  3.8  5.9 -1.2
> z<-svpdx$m2
> z
[1]  0.26524514  0.31278198  0.37310154  0.34529154  0.29467111 -0.03867458  0.34668927  0.17245702  0.14411363
[10] -0.31433668  0.87572994  0.16337324  0.19679602  0.18161695  0.14387323  0.18886459  0.15787746  0.18875571
[19]  0.18736354
> library(tsDyn)
> ndx.lstar <- lstar(x, m=3,d=1, thVar=z,control=list(maxit=3000));
Using maximum autoregressive order for low regime: mL = 3
Using maximum autoregressive order for high regime: mH = 3
Using only first 16 elements of thVar
Performing grid search for starting values...
Starting values fixed: gamma =  40 , th =  0.3291658 ; SSE =  118.9413
Optimization algorithm converged
Optimized values fixed for regime 2  : gamma =  48.01837 , th =  0.3223614
> rdx.lstar <- lstar(y, m=3,d=1, thVar=z,control=list(maxit=3000));
Using maximum autoregressive order for low regime: mL = 3
Using maximum autoregressive order for high regime: mH = 3
Using only first 16 elements of thVar
Performing grid search for starting values...
Starting values fixed: gamma =  40 , th =  0.352213 ; SSE =  119.8971
Optimization algorithm converged
Optimized values fixed for regime 2  : gamma =  45.87497 , th =  0.3458736
> ndx.lstar
Non linear autoregressive model
LSTAR model
Coefficients:
Low regime:
    phi1.0     phi1.1     phi1.2     phi1.3
-0.2840885  1.9360245 -1.4690397  0.4925477
High regime:
    phi2.0     phi2.1     phi2.2     phi2.3
4.3974268 -2.3388445  2.2482280 -0.9932285
Smoothing parameter: gamma = 48.02
Threshold
Variable: external
Value: 0.3224
> rdx.lstar
Non linear autoregressive model
LSTAR model
Coefficients:
Low regime:
    phi1.0     phi1.1     phi1.2     phi1.3
-0.4204974  1.8394945 -1.4018456  0.4320727
High regime:
   phi2.0    phi2.1    phi2.2    phi2.3
3.939309 -3.911762  3.639241 -1.489407
Smoothing parameter: gamma = 45.87
Threshold
Variable: external
Value: 0.3459
> sum.s<-summary(ndx.lstar)
错误于summary.lstar(ndx.lstar) : dims [product 48]与对象长度[0]不匹配
> summary(ndx.lstar)
错误于summary.lstar(ndx.lstar) : dims [product 48]与对象长度[0]不匹配

26# epoh

这不是数据的问题
因为package自带的数据lynx,也发生同样问题
library(tsDyn)
mod.lstar <- lstar(log10(lynx), m=2, d=1, control=list(maxit=3000))
summary(mod.lstar)
#dims [product 224] do not match the length of object [0]
这是因为source code lstar.R 中的function summary.lstar
sX <- object$mod$thVar
应该是sX <- object$model.specific$thVar
order.L <- object$mod$mL
应该是order.L <- object$model.specific$mL
....
....
我初步修改为lstar_new.R后可运行
但需再花点时间再检查.
source("lstar_new.R")
mod.lstar <- lstar(log10(lynx), m=2, d=1, control=list(maxit=3000))
mod.lstar
summary(mod.lstar)

#####
> summary(mod.lstar)
Non linear autoregressive model

LSTAR model
Coefficients:
Low regime:
    phi1.0     phi1.1     phi1.2
0.4055717  1.2378257 -0.3265076

High regime:
    phi2.0     phi2.1     phi2.2
0.8187712  0.3087342 -0.6418812

Smoothing parameter: gamma = 40

Threshold
Variable: Z(t) = + (1) X(t) + (0) X(t-1)

Value: 2.566

Residuals:
      Min        1Q    Median        3Q       Max
-0.596805 -0.133230  0.034821  0.139065  0.471333

Fit:
residuals variance = 0.04104,  AIC = -348, MAPE = 6.003%

Non-linearity test of full-order LSTAR model against full-order AR model
F = 8.529 ; p-value = 0.00036569

Threshold
Variable: Z(t) = + (1) X(t) + (0) X(t-1)

其实你也可以使用function star
因为star 也是使用Logistic transition function
底下两者结果相同
mod.star <- star(log10(lynx), m=2, d=1, control=list(maxit=3000))
mod.star
summary(mod.star)

mod.lstar <- lstar(log10(lynx), m=2, d=1, control=list(maxit=3000))
mod.lstar

但是summary(mod.star)无误

我修改了source code并重新compile

1.移除原来的package "tsDyn"
2.From menus\Packages\install package(s) from local zip-files\tsDyn_0.7-40.zip


library(tsDyn)
mod.lstar <- lstar(log10(lynx), m=2, d=1, control=list(maxit=3000))
summary(mod.lstar)
#################
Non linear autoregressive model
LSTAR model
Coefficients:
Low regime:
    phi1.0     phi1.1     phi1.2
0.4055717  1.2378257 -0.3265076
High regime:
    phi2.0     phi2.1     phi2.2
0.8187712  0.3087342 -0.6418812
Smoothing parameter: gamma = 40
Threshold
Variable: Z(t) = + (1) X(t) + (0) X(t-1)
Value: 2.566
Residuals:
      Min        1Q    Median        3Q       Max
-0.596805 -0.133230  0.034821  0.139065  0.471333
Fit:
residuals variance = 0.04104,  AIC = -348, MAPE = 6.003%
Non-linearity test of full-order LSTAR model against full-order AR model
F = 8.529 ; p-value = 0.00036569
Threshold
Variable: Z(t) = + (1) X(t) + (0) X(t-1)


tsDyn_0.7-40.zip

tsDyn_0.7-40.zip
大小:(1.55 MB)

 马上下载