ENH: support for non-nanosecond precision in pd.Timedelta

[veenstrajelmer]

Feature Type

• Adding new functionality to pandas

• Changing existing functionality in pandas

• Removing existing functionality in pandas

Problem Description

When converting matlab datenums, I would expect this should be easily possible with pd.Timedelta(), however, several approaches give different errors:

import pandas as pd
# raises OutOfBoundsTimedelta: seconds=62703763200000000000, milliseconds=0, microseconds=0, nanoseconds=0
pd.Timedelta(days=725738)
# raises OutOfBoundsTimedelta: Cannot cast 725738 from D to 'ns' without overflow.
pd.to_timedelta(725738, unit='days')

I expect this is because of the nanosecond accuracy. Matlab dates are in days (dtype float) and I would be happy with microsecond (or less) precision. pd.date_range() and pd.Timestamp() now support non-nanosecond accuracy. Would this also be possible with pd.Timedelta()?

Feature Description

Add support for non-nanosecond accuracy for pd.Timedelta()

Alternative Solutions

It is possible to convert matlab datenums to pandas datetimes with the following approach from https://stackoverflow.com/a/49135037/5578371:

import pandas as pd
import numpy as np
time_wl = pd.to_datetime(np.array([725738])-719529, unit='D').round('s')

Gives: DatetimeIndex(['1987-01-01'], dtype='datetime64[ns]', freq=None)

Additional Context

No response

[rhshadrach]

Thanks for the request. For pd.to_timedelta, I believe this is a duplicate of #49060.

@jbrockmendel - should pd.Timedelta also be added to #49060?

[veenstrajelmer]

Thanks for the reply. I actually found that issue in my search, but since it was so old I wrongly assumed it was pre 2.0. I will read better in the future. Either way, feel free to close this issue if needed.

[jbrockmendel]

I dont think we should add the keyword to pd.Timedelta, no. The particular construction from the OP is a tough case; something like pd.Timedelta(np.timedelta64(N, "D")) should work as expected