HoltWinters fitted level parameter not bounded between 0 and 1 (PR#11469)

Hi,

Thanks for the quick response. I upgraded by version of R on Windows to the
latest (2.7.0) and re-ran the analysis and get the same result of 48.87989.

The original time series was a non-regular zoo() object and I converted it
to a time series with daily granularity by setting the values of the
undefined dates to zero. I initially wondered if the zeros were maybe an
issue, i.e., I know multiplicative Holt-Winters requires non-negative
values, so I tried shifting the values, which should change the level,
though the fitted alpha values should be the same:

delta <- 10 ^ seq(-3, 0);
sapply(delta, function(delta) HoltWinters(x + delta, beta = 0, gamma =

0)$alpha);
     alpha      alpha      alpha      alpha
48.8798912 48.8798912 48.8798912  0.6881547

Note that by applying a shift,  I also obtained varying values of alpha,
similar to that of what David achieved by changing the initial conditions.

Thanks
-John

On Fri, May 16, 2008 at 4:20 AM, David Meyer <david.meyer at wu-wien.ac.at>
wrote:

I get John's value (48.8789) in 2.7.0 and R-devel (both on Ubuntu). Really
seems to be a numeric issue:

HoltWinters(x, beta = 0, gamma = 0)$alpha

  alpha
48.87989

HoltWinters(x * 1.0000000001, beta = 0, gamma = 0)$alpha

   alpha
0.6881547

HoltWinters(x * 1.00000000001, beta = 0, gamma = 0)$alpha

  alpha
48.87989

Providing starting values seems to help, but not always:

HoltWinters(x, beta = 0, gamma = 0, l.start = 0.00001)$alpha

  alpha
48.88999

HoltWinters(x, beta = 0, gamma = 0, l.start = 0.0001)$alpha

   alpha
0.6880582

Yes, it's easy to use optimize() instead of optim() in the univariate
cases, will do.

David.


Prof Brian Ripley wrote:

It doesn't do it on my system (I get a value of about 0.688 in R 2.7.0
patched on Linux), and 2.5.1 is not current.  Does a better starting value
help?

However, HoltWinters is using optim() in a case it is not designed for
(one-dimensional optimization): see the note on its help page.  I think this
could easily be changed, but as HoltWinters is contributed code I am Cc:ing
the author for comment.

On Fri, 16 May 2008, john.bodley at gmail.com wrote:

 Full_Name: John Bodley

Version: 2.5.1 (2007-06-27)
OS: Windows XP
Submission from: (NULL) (12.144.182.66)


I was fitting a number of time series in R using the stats::HoltWinters
method
to define a single exponential smoothing model, i.e., beta = gamma = 0.

I came across an example where the fitted value of alpha was not defined
in the
[0, 1] interval which seems to violate the lower and upper bound
constraints
used for the optim method. On my computer the following code returns a
value of
48.87989.

R code:

x <- c(
0,
0.000843170320404722,
0,
0,
0,
0.0103773584905660,
0.00832466181061394,
0.0038560411311054,
0,
0,
0.00484966052376334,
0,
0,
0,
0.00274348422496571,
0,
0,
0,
0,
0,
0.0207064555420219,
0.0334975369458128,
0.0334975369458128,
0.00338983050847458,
0.00483675937122128,
0,
0,
0.00224971878515186,
0,
0,
0,
0.00135685210312076,
0,
0,
0,
0.0035377358490566,
0.0035377358490566,
0.00501002004008016,
0.0107632093933464,
0,
0,
0.0143329658213892,
0.0330459770114943,
0,
0,
0,
0,
0.0109890109890110,
0,
0.00118623962040332,
0.007380073800738,
0.00695410292072323,
0.0104895104895105,
0.00278551532033426,
0.00278551532033426
);

# Single exponential smoothing
m <- stats::HoltWinters(x, beta = 0, gamma = 0);
m$alpha

--
Dr. David Meyer
Department of Information Systems and Operations

Vienna University of Economics and Business Administration
Augasse 2-6, A-1090 Wien, Austria, Europe
Tel: +43-1-313 36 4393
Fax: +43-1-313 36 90 4393
HP:  http://wi.wu-wien.ac.at/~meyer/ <http://wi.wu-wien.ac.at/%7Emeyer/>