BUG: negating pd.Na and None

[attack68]

• I have checked that this issue has not already been reported.
• I have confirmed this bug exists on the latest version of pandas.
• I have confirmed this bug exists on the master branch of pandas.

I don't really know how to describe this or if it is as expected but it seems strange behaviour to me:

>>> df = pd.DataFrame({"a": [pd.NA, -1, None], "b": [np.nan, -1, 1]})
      a     b
0  <NA>   NaN
1    -1  -1.0
2  None   1.0

>>> print(df.dtypes)
a    object
b   float64

>>>print(df * -1)
     a    b
0  NaN  NaN
1    1  1.0
2  NaN -1.0

>>> print((df * -1).dtypes
a    object
b   float64

>>> print(-df)
TypeError: bad operand type for unary -: 'NoneType'

I guess there are two things here:

1. (-df) is not the same as (df * -1)
2. the operation keeps dtype as object but the <NA> and None are converted to NaN display

[KrishnaSai2020]

In df *-1 it defaults to NaN which is expected as None * -1 is invalid but surely it should do the same for -df. but maybe because there's no * sign DataFrame.multiply doesn't register it. So then Python kicks in and since in Python Multiplication of NaN and -1 is invalid it raises an error.

[attack68]

In df *-1 it defaults to NaN which is expected as None * -1 is invalid but surely it should do the same for -df. but maybe because there's no * sign DataFrame.multiply doesn't register it. So then Python kicks in and since in Python Multiplication of NaN and -1 is invalid it raises an error.

ok, but why is pd.NA * -1 = np.nan and what is the value of -pd.NA?

[KrishnaSai2020]

for consistency maybe ie to make all values floats as pd.NA stores an integer value and np.nan registers as a float. with regards to -pd.NA it comes up as <NA> which it should.

[KrishnaSai2020]

It seems strange but makes sense. Sort of anyway.

[mzeitlin11]

I think currently pd.NA in object type data is just not well-defined (xref #33066 and #32931). In general, I think the current behavior for nullable types makes sense under the semantics that pd.NA just signifies an unknown value:

In [4]: ser = pd.Series([1, None], dtype="Int64")

In [5]: ser * -1
Out[5]:
0      -1
1    <NA>
dtype: Int64

So IMO object type data should match this and just propagate NA

[attack68]

So IMO object type data should match this and just propagate NA

I agree, I think pd.NA = -pd.NA = pd.NA [*/+/-/**] float/int/complex

[KrishnaSai2020]

+1 to redefining pd.NA

[mzeitlin11]

+1 to redefining pd.NA

To be clear on the scope here, I'm not recommending a redefinition of pd.NA, this is how it already works. There are just problems that are specific to object type data