BUG: read_csv() into FloatingArrays converts np.nan into pd.NA, even with distinguish_nan_and_na=True

[carlocastoldi]

Pandas version checks

• 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 main branch of pandas.

Reproducible Example

import io
import pandas as pd
import numpy as np

csv = """,x
a,1.0
b,
c,nan"""

with pd.option_context("future.distinguish_nan_and_na", True):
    s = pd.read_csv(io.StringIO(csv), index_col=0, dtype={0: str, 1: pd.Float64Dtype()},
                     keep_default_na=False, na_values=[""])["x"]
s

Issue Description

Working with data having both pd.NA and np.nan is possible and, at least until pandas2, it was possible to save such dataframes as CSV in a format that keeps a distinction between the two (see #65227).

However, in the example above pandas fails with:

ValueError: Unable to parse string "nan" at position 2

The current workaround consists in reading the CSV as string, and manually map its value to floats:

import io
import pandas as pd
import numpy as np

csv = """,x
a,1.0
b,
c,nan"""


def convert_nan(val: str):
    match val:
        case "":
            return pd.NA
        case _:
            return np.float64(val) # includes np.nan

with pd.option_context("future.distinguish_nan_and_na", True):
    s1 = pd.read_csv(io.StringIO(csv), index_col=0, dtype=str, keep_default_na=False)["x"]
    s2 = s1.map(convert_nan)
    s3 = s2.convert_dtypes()
s1,s2,s3

returns:

(a    1.0
 b       
 c    nan
 Name: x, dtype: str,
 a     1.0
 b    <NA>
 c     NaN
 Name: x, dtype: object,
 a     1.0
 b    <NA>
 c     NaN
 Name: x, dtype: Float64)

Expected Behavior

I would expect read_csv() to distinguish between NA and NaN in the columns with dtype=Float64. Even more so if distinguish_nan_and_na=True.

 a     1.0
 b    <NA>
 c     NaN
 Name: x, dtype: Float64

Installed Versions

Details

INSTALLED VERSIONS

commit : ab90747
python : 3.13.13
python-bits : 64
OS : Linux
OS-release : 6.17.0-20-generic
Version : #20~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Thu Mar 19 01:28:37 UTC 2
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : en_GB.UTF-8
LOCALE : en_GB.UTF-8

pandas : 3.0.2
numpy : 2.4.4
dateutil : 2.9.0.post0
pip : 26.0.1
Cython : None
sphinx : None
IPython : 9.5.0
adbc-driver-postgresql: None
adbc-driver-sqlite : None
bs4 : 4.13.5
bottleneck : None
fastparquet : None
fsspec : 2025.7.0
html5lib : None
hypothesis : None
gcsfs : None
jinja2 : 3.1.6
lxml.etree : None
matplotlib : 3.10.6
numba : 0.65.0
numexpr : 2.11.0
odfpy : None
openpyxl : None
psycopg2 : None
pymysql : None
pyarrow : 21.0.0
pyiceberg : None
pyreadstat : None
pytest : 8.4.1
python-calamine : None
pytz : None
pyxlsb : None
s3fs : None
scipy : 1.16.1
sqlalchemy : None
tables : 3.10.2
tabulate : None
xarray : None
xlrd : None
xlsxwriter : None
zstandard : None
qtpy : None
pyqt5 : None

[jorisvandenbossche]

@carlocastoldi thanks for the report. Just to be clear, this also didn't work before with pandas 2.3, right? (but I agree it should work, and map NaN to NaN or NA depending on the option)

[carlocastoldi]

I confirm this was the behaviour in pandas 2, too. Unlike #65227, this issue is not a regression.
I opened it now because it wasn't unilt Pandas3 that the support for NA/NaN distinction is ~somewhat official