Dear all,

Thanks for finding "armax" package useful. I eventually got time to solve this beta release. I've just released a patch in which "occasional armax aborting" should be  fixed. Please give a trial.

Yi-Nung Yang (楊奕農)
Associate Professor
Department of IB, CYCU, Taiwan.


2014-05-21 19:07 GMT+08:00 <oleg_komashko@ukr.net>:
For armax "diggers"
Data sets for testing and a try to explain occasional armax aborting

Data file attached
# Script to select lag length in regression of d_u on lagged g
# u-unimployment (%) g-gdp rate (%)

## begin script
# "Okun law" artifical data similar to POE-4, ch.9
open mod2data.gdt
include armax.gfn
diff u1
# max lag length =4
lags 4; g1
# subsample for equal data length whlile choosing
smpl 1995:1 2013:4

list arg4 = 0 g1 g1_1 g1_2 g1_3 g1_4
list arg4a = g1 g1_1 g1_2 g1_3 g1_4

# uncomment for lag length selection
# list arg3 = g1 g1_1 g1_2 g1_3
# list arg2 = g1 g1_1 g1_2
# list arg1 = g1 g1_1 
# list arg0 = 0 g1

# works ok on my system, ARMA(0,1) by all ic 
armax(2, 2, d_u1, arg4, 1, 1, 1, 1, 0)

# gives "Failed to compute numerical Hessian"
armax(2, 2, d_u1, arg4a, 1, 1, 1, 1, 0)


# armax(2, 2, d_u1, arg3, 1, 1, 0, 1, 0)
# armax(2, 2, d_u1, arg2, 1, 1, 0, 1, 0)
# armax(2, 2, d_u1, arg1, 1, 1, 0, 1, 0)
# armax(2, 2, d_u1, arg0, 1, 1, 0, 1, 0)

## end script

GUI's peculiar features:
With arg4 and include constant chequed results are 
  the same as for script armax(2, 2, d_u1, arg4, 1, 1, 1, 1, 0)
With arg4 and incl. const unchequed 
? armax(2, 2, d_u1, arg4, 0, 1, 0, 1, 0)
Failed to compute numerical Hessian
With arg4a and include constant chequed
? armax(2, 2, d_u1, arg4a, 1, 1, 1, 1, 0)
Failed to compute numerical Hessian
With arg4a and include constant unchequed
? armax(2, 2, d_u1, arg4a, 1, 1, 0, 1, 0)
Failed to compute numerical Hessian

Further diagnostics:
armax(1, 2, d_u1, arg4a, 1, 1, 1, 1, 0) works ok
arma 2 1 ; d_u1 arg4a Failed to compute numerical Hessian
arma 2 1 ; d_u1 arg4a --opg works ok
From Command Reference(p.7) "Note that failure of the numerical
Hessian computation is generally an indicator of a misspecified model".
From output:
AR
    Root  1           1.0001
So, nothing enigmatic.
Armax_auto would be more convenient if it scipped such models
instead of stopping the cycle.
!

_______________________________________________
Gretl-users mailing list
Gretl-users@lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users