[R-SIG-Mac] NaN bug with arima() on Mac

[Kuschnig, Nikolas]

Hey Ken and Simon,

Good point on the 'bug', I guess it's rather unexpected behaviour since it does
not occur on other systems. The Hessian issue causing system dependency makes a
lot of sense. 
In that case I guess the system differences can't feasibly be eliminated. Should
I document this system difference as a 'bug' or the like somewhere? In my case
I'm now just converting the warning to a message (just need sigma + users can
provide values instead).

Thanks for the help!
-- 
Nikolas


On Tue, 2021-11-23 at 10:19 +1100, Ken Beath wrote:
> Looking at the returned object there are problems with some of the standard
> errors which are NaN for both aima and arima0. This usually indicates that the
> Hessian doesn’t have the required form. This can be very dependent on the
> system. It also probably means that the model you are fitting is over
> parameterised.
> 
> Ken
> 
> > On 23 Nov 2021, at 9:21 am, Simon Urbanek <simon.urbanek using R-project.org>
> > wrote:
> > 
> > 
> > Nikolas,
> > 
> > what makes you think it is a bug? It is an optimization process, so the
> > objective function is free to produce NaN for unsuitable parameters so it
> > can steer the optimization away from such areas.
> > 
> > Cheers,
> > Simon
> > 
> > 
> > > On Nov 23, 2021, at 2:29 AM, Kuschnig, Nikolas <nikolas.kuschnig using wu.ac.at>
> > > wrote:
> > > 
> > > Dear all,
> > > 
> > > There seems to be a small bug with `arima()` on Mac. For some very
> > > specific data
> > > I get the following warning:
> > >         Warning message:
> > >         In log(s2) : NaNs produced
> > > This seems to stem from `.Call(C_ARIMA_Like, x, Z, 0L, FALSE)` in the
> > > `armafn()`
> > > objective function. This warning does not occur on Linux or Windows
> > > systems or
> > > when using the `arima0()` function.
> > > Below is an example that I hope reproduces (please excuse the dependency,
> > > I only
> > > encountered it with this data, partly due to my lack of a Mac). Note that
> > > the
> > > warning in question occurs during estimation, printing always induces a
> > > warning.
> > > 
> > > 
> > > # install.packages("BVAR")
> > > library("BVAR")
> > > 
> > > x <- fred_qd[1:243, 
> > > c("GDPC1", "PCECC96", "GPDIC1", "HOANBS", "GDPCTPI", "FEDFUNDS")]
> > > x <- fred_transform(x, codes = c(4, 4, 4, 4, 4, 1))
> > > Y <- as.matrix(x)[6:nrow(x), ]
> > > 
> > > ar <- apply(Y, 2, arima, order = c(5, 0, 0)) # Warning only on Mac
> > > ar0 <- apply(Y, 2, arima0, order = c(5, 0, 0)) # No warning
> > > 
> > > 
> > > Does anyone have any more information on this? Thanks for the help!
> > > 
> > > Best,
> > > -- 
> > > Nikolas Kuschnig
> > > 
> > > _______________________________________________
> > > R-SIG-Mac mailing list
> > > R-SIG-Mac using r-project.org
> > > https://stat.ethz.ch/mailman/listinfo/r-sig-mac
> > 
> > _______________________________________________
> > R-SIG-Mac mailing list
> > R-SIG-Mac using r-project.org
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mac
> 

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 695 bytes
Desc: This is a digitally signed message part
URL: <https://stat.ethz.ch/pipermail/r-sig-mac/attachments/20211123/9549003c/attachment.sig>