Hello, I did some digging:
> Maybe it's a bug in python-tz? Most likely: Pandas uses this code to get the time offset for the local time in tslib.pyx: cpdef _get_utcoffset(tzinfo, obj): try: return tzinfo._utcoffset except AttributeError: return tzinfo.utcoffset(obj) Now with tz = pytz.timezone('Asia/Tokyo') dt = datetime.datetime(2011,1,1) I get In[76]: tz.utcoffset(dt) Out[76]: datetime.timedelta(0, 32400) In [77]: tz._utcoffset Out[77]: datetime.timedelta(0, 33540) which is exactly the 19 min difference we see, and the result depends on the code path taken in _get_utoffset. Best, Gert