near singular matrix

[nizar10]

I tried to estimate a SARIMAX equation with monthly data from 1993M01 to 2010M12 and to forecast it from 2011M01 to 2012M12. The X letter of Sarimax reffers to an external variable which helps us to estimate and forcast better. My program works perfectely. In order to know approximativelly the error of my model i’ve then taken only data from 1993M01 to 2008M12 and tried to forecast it from 2009M01 to 2010M12 in order to compare the value forcasted to the reel value between 2009M01 to 2010M12. The problem i have is that when i change my sample from 93-2011 to 93-2009 i have this message : near singular matrix in… The only solution a found is to drop the external variable and apparentely it works but i don’t have the real error of my equation because the error matchs with the Sarima equation and not the sarimaX.

Any ideas or thoughts on what I need to do will be much appreciated.

Thanks
Nizar

[EViews Gareth]

You've probably run into perfect collinearity problems.

[nizar10]

Thank you for your answer.

it's weird that it works from 01/93 to 12/2010 and not for 01/93 to 12/2008... Do you have any suggestion?

[EViews Gareth]

Best guess would be that there is a dummy variable (or a variable that doesn't change much), that in the smaller sample has zero variation, thus becoming perfectly collinear with the constant.

[Rosa.hh]

Dear users,

I am having a similar problem obtaining the 'near singular error'. I'm estimating an EGARCH model of the change of an interest rate using calendar dummy variables, my objective is to compare my results using different subsamples, pre/post crisis. The model is working correctly when I use a dataset for the period 2005-2011 (with 1677 obs). However, it's not working when I enlarge the timeframe to 1999-2011 (3209 obs), not even when I restrict the sample to 2005-2011 using the larger database.

I have followed previously posted recommendations : tried not including the constant, deleting some of my dummy variables as well as excluding some explanatory variables, however I continue to get the same error.

Do you have any idea of what could I do to be able to calculate my model?

This is my equation:
equation eq01.arch(1,1,egarch,m=1000,h) d_rate c d_rate(-1) d_rate(-2) d_rate(-3) d_rate(-4) d_rate(-5) d_refi bf_1dhol af_1dhol bf_2dhol af_2dhol bf_3dhol af_3dhol bf_4dhol af_4dhol mp_du1 mp_dd1 mp_dd2 mp_dd3 mp_dd4 mp_dd5 w_du5 m_du1 bf_em m_dd1 q_dy_up q_dd1 bf_endyear af_endyear @ bf_1dhol af_1dhol bf_2dhol af_2dhol bf_3dhol af_3dhol bf_4dhol af_4dhol mp_du1 mp_dd1 mp_dd2 mp_dd3 mp_dd4 mp_dd5 w_du5 m_du1 bf_em m_dd1 q_dy_up q_dd1 bf_endyear af_endyear

Thanks a lot,

[EViews Gareth]

Could you provide the workfile?

[Rosa.hh]

Yes, sorry for the late reply.

[EViews Gareth]

Your variables af_4dhol and bf_4dhol have no variance (they are always equal to 0). Thus they are perfectly colinear with each other, and the constant.

[Rosa.hh]

Dear Gareth,
For some reason my dummy vars changed when expanding the sample period. I've have corrected them and the model is now working, thank you!