Not reported but evident here: http://nbviewer.ipython.org/cb6e9b476a41586958b5
Reported here: http://stackoverflow.com/questions/18721547/armax-model-forecasting-leads-to-valueerror-matrices-are-not-aligned-when-pas
y = np.random.random(100)
x = np.random.random(100)
mod = sm.tsa.ARMA(y, (2,1), x).fit()
newx = np.random.random(12)
This breaks because we need to ensure that exog is 2d.
mod.forecast(steps=12, alpha=.05, exog=newx)
This gets further but falls down later
mod.forecast(steps=12, alpha=.05, exog=newx[:,None])
Need to investigate what's going on.
It falls down because you need to pass in past data too. If you want to predict 12 steps ahead with an ARMAX(2,q) model then you need to pass in 14 exogenous observations. I'm going to fix the dims check in exog and add this to the docs.
BUG: Ensure 2d for conformability. Closes #1076.
Backport PR #1077: BUG: Allow 1d exog in ARMAX forecasting.
Closes #1076. Another candidate for backporting.