BUG: there is no year 0

[jbrockmendel]

ts = pd.Timestamp("001-01-01")
td = pd.Timedelta(days=1).as_unit("s")

>>> ts - td
Timestamp('0-12-31 00:00:00')
>>> (ts - td).year
0

We document in a few places that we are using the Proleptic Gregorian calendar which does not have a year 0. The year before 1AD is 1BC in this calendar.

np.datetime64 objects have the same behavior (cc @sberg)

Expected Behavior

Refuse to parse year 0, create a Timestamp with year=0, subtraction skips year 0.

Or document the current behavior.

[Kman303]

I'd like to try and fix this bug, if I could get the issue assigned to me

[jbrockmendel]

I'd advise against working on this until there is consensus on how to address the issue.

[jorisvandenbossche]

FWIW, Arrow has the same behaviour as numpy:

In [9]: pc.subtract(pa.array([pd.Timestamp("001-01-01")]), pa.array([pd.Timedelta(days=1)]))
Out[9]: 
<pyarrow.lib.TimestampArray object at 0x7fd268b57e80>
[
  0000-12-31 00:00:00.000000
]

[jorisvandenbossche]

Quoting the linked wiki page:

For these calendars one can distinguish two systems of numbering years BC. Bede and later historians did not enumerate any year as zero (nulla in Latin; see Year zero); therefore the year preceding AD 1 is 1 BC. In this system the year 1 BC is a leap year (likewise in the proleptic Julian calendar). Mathematically, it is more convenient to include a year 0 and represent earlier years as negative numbers for the specific purpose of facilitating the calculation of the number of years between a negative (BC) year and a positive (AD) year. This is the convention in astronomical year numbering and the international standard date system, ISO 8601. In these systems, the year 0 is a leap year.[4]

So this mentions that the "the international standard date system, ISO 8601" uses a year zero. And in the end, one could also say that what we see in the output above is an ISO8601 formatted timestamp?

[Kman303]

take

[jorisvandenbossche]

@Kman303 as @jbrockmendel mentioned above, there is not yet a clear resolution about what exactly needs to be done

[jbrockmendel]

I like Joris's current-behavior-is-fine approach.

[miccoli]

As already commented by @jorisvandenbossche current behaviour is expected.

Numpy docs ¹ clarify datetime64 semantics listing the conventions adopted, namely

• proleptic Gregorian calendar,
• astronomical year numbering
• 86400 seconds per day (POSIX semantics)

It seems to me that it would be safe to simply make a reference in the pandas docs to the numpy ones.

Footnotes

1. https://numpy.org/doc/stable/reference/arrays.datetime.html#datetime64-conventions-and-assumptions ↩