I have a problem where I am using auto.arima with multiple exogenous regressors (thus xreg being a matrix). My dependent variable as well as all of the exogenous regressors have NA values at some time points. Thus, my y:X matrix (the dependent variable in the first column and all the regressors in the other columns) have some NA rows. If I try running auto.arima on this kind of data, it breaks down, seemingly thinking it is facing multicollinearity among regressors or something. Namely, the code of auto.arima includes this bit:
if (min(sv)/sum(sv) < .Machine$double.eps) { stop("xreg is rank deficient") }
Thus, auto.arima throws an error message if xreg contains several rows of NAs (or maybe even a single row of NAs? I have not checked).
Intuitively, I do not see why some NA rows should be a big problem for a Kalman filter (?) running in the background of auto.arima. Indeed, if I disable this if statement in the code, I am able to run my model with (the modified) auto.arima successfully. At least it gives some results; I cannot guarantee they are always meaningful.
So is it really a problem if the xreg matrix suffers from row-wise rank deficiency rather than column-wise rank deficiency? And if this row-wise rank deficiency is actually harmless, should the code of auto.arima be updated accordingly to allow it?
Edit: While this is the case for my data, I cannot replicate the behavior using simulated data. So probably the problem is not just the presence of several rows of NAs. But still, disabling the if statement gives OK results for my data. I am puzzled...
Referred here by Forecasting: Principles and Practice, by Rob J Hyndman and George Athanasopoulos