R auto.arima "No ARIMA model able to be estimated"

Question

I have been using auto.arima with a lot of success in the past. However, I started running into an error that I have had difficulties troubleshooting. The error is :

    Error in search.arima(x, d, D, max.p, max.q, max.P, max.Q, max.order,  : 
    No ARIMA model able to be estimated

This is the code that I used; unfortunately I cannot share the data as they are proprietary.

            auto.arima(myts
                      ,max.p=5, max.d=5, max.q=5, max.P=50, max.D=5, max.Q=5
                      ,ic="aicc"
                      ,seasonal=FALSE, allowdrift =FALSE, allowmean=TRUE, stationary=FALSE
                      ,test="kpss"
                      ,stepwise=FALSE, approximation=FALSE
                      ,lambda=NULL
                      ,xreg=historic_xreg
                      )

The way I am currently treating this is to gradually reduce the exogenous variables in xreg (which start from 10 down to 1 in a loop), but even this fails. This might be because I have fewer than 30 data points.

I looked into the source code but as I am not an experienced programmer I had trouble pinning down the reason of the error. I know it has to do with the functions's inability to find the bestfit; or better put I might be expecting too much given the data that I have. The data is yearly and there is no seasonality.

https://github.com/robjhyndman/forecast/blob/master/R/arima.R

My question is: What adjustments do I need to do to allow auto.arima to provide me with a fitted model which then I can assess on its forecasting performance?

This is likely a data issue given what you've said (30 points isn't many with 10 exogenous vars). Even if you can't provide the actual data, can you provide example data using dput()? — JohnSG
Thank you for commenting. Even if I only drop all exogenous variables I get the same error. — J. Doe.
almost impossible to answer without a reproducible example. If n<30, and no metadata/information about where data come from and what they represent, might it be safe to post them? You could also try strategies like multiplying them by an arbitrary integer, add small amounts of random numeric "fuzz", etc. (all of which would be likely to leave the problematic behaviour intact but make it harder to guess the original information: stackoverflow.com/questions/10454973/… — Ben Bolker
Thank you for your response and you are correct. I think I actually deserve downvoting for this one. — J. Doe.
well, you get credit for a clearly phrased question and for documenting your efforts to solve your own problem ... — Ben Bolker

J. Doe. J. Doe. · Accepted Answer · 2016-10-05T15:07:20

OK, I have found my error. Looking at the myts and the way I construct it, I have characters instead of numeric format. So apologies for taking up your time with such a noob-ish question.

For the record, these are the steps I took with the hopes of helping others stumbling into the "“No ARIMA model able to be estimated”" error:

After I removed the exogenous variables I was still getting the error. I created some dummy data to feed into auto.arima and it worked.

I then tried to fit myts to Arima and it was not working - which meant that the problem was with my data. I looked up myts and the data frame it was generated from and lo and behold, the values were stored as characters.

R auto.arima "No ARIMA model able to be estimated"

1 Answers