Negative NaN (-nan) in CSV input treated as object

[Bklyn]

The value "-nan" (yes, this is a valid type of NaN!) in a CSV input file causes that column to be treated as 'object' instead of float64.

pd.read_csv(StringIO.StringIO('a,b\n1,2.0\n2,nan\n3,-nan')).b
Out[15]: 
0     2.0
1     NaN
2    -nan
Name: b, dtype: object

pd.read_csv(StringIO.StringIO('a,b\n1,2.0\n2,nan\n')).b
Out[16]: 
0     2
1   NaN
Name: b, dtype: float64

When the file is sufficiently large, the following error is generated:

In [3]: pd.read_csv('big.bad.csv')
/usr/local/lib/python2.7/dist-packages/pandas/io/parsers.py:1033: DtypeWarning: Columns (58,64) have mixed types. Specify dtype option on import or set low_memory=False.
  data = self._reader.read(nrows)

If the string "-nan" is replaced with "nan" all is well. I don't really need to distinguish negative NaN from NaN but would like to be able to read my data files w/o having to pre-process them to scrub all the '-nan's.

[jreback]

try specifying it as a na_value see here: http://pandas.pydata.org/pandas-docs/dev/io.html#na-values
lmk if this works; what version of pandas?

[jreback]

pd.read_csv(StringIO.StringIO('a,b\n1,2.0\n2,nan\n3,-nan'), na-values=['-nan']).b

works

[jreback]

I don't have a problem with making -nan a default nan value; if you would like to do a PR would be great