looking at PR #1018 : demeaned xm is not used
unit test for pacf_yw only uses 1 or 2 decimals.
It looks like I get much higher agreement with reference after demeaning. (checked only with pdb so far.)
Aside: Where do the reference numbers come from? looks like matlab but it doesn't say explicitly.
False alarm, yule_walker demeans by default
just to clarify: Skipper had added tests for pac_yw comparing to Stata at 8 decimals.
My original unit test compared it with the ols version (matlab version IIUC) which only agrees at 1 decimal because of the different treatment of initial observations.