#68 arima 1 0 0 fails

closed
None
5
2008-07-19
2008-07-14
No

With current cvs on Kubuntu Linux 8.04:

open data3-6.gdt
arima 1 0 0 ; Ct

gives an error

? arima 1 0 0 ; Ct
Not a Number in calculation

Discussion

  • Riccardo "Jack" Lucchetti

    Logged In: YES
    user_id=1368575
    Originator: NO

    The technical reason why this happens is that the likelihood is maximised at phi=0.999996. When we attempt to compute the Hessian numerically, we try to evaluate the loglikelihood for phi>1 and things break (understandably).

    I can't think right now of a quick way to avoid this problem, which certainly isn't nice to have. I suppose that for the moment we could state somewhere in the docs that this may happen if the estimated parameters are very close to their boundary.

    The only clean way out of this would be coding analytical derivatives for ARMA models, which would have a number of other advantages in terms of speed and accuracy, but is a very time consuming and error-prone (not to say rather boring) task. Is it really worth the effort?

     
  • Allin Cottrell

    Allin Cottrell - 2008-07-19
    • assigned_to: nobody --> allin
    • status: open --> closed
     
  • Allin Cottrell

    Allin Cottrell - 2008-07-19

    Logged In: YES
    user_id=330339
    Originator: NO

    I've arranged for a somewhat more informative error message to be printed
    in this sort of case. I agree with Jack: this is not really a bug as
    such; it's a difficult case which we can't handle without extensive
    changes.

    If you think it's worth coding analytical derivatives that could be
    entered as a feature request.

     

Log in to post a comment.