Pandas: Why is default column type for numeric float?

Question

I am using Pandas 0.18.1 with python 2.7.x. I have an empty dataframe that I read first. I see that the types of these columns are object which is OK. When I as

Answer 1

The why is almost certainly to do with flexibility and speed. Just because Pandas has only seen an integer in that column so far doesn't mean that you're not going to try to add a float later, which would require Pandas to go back and change the type for all that column. A float is the most robust/flexible numeric type.

There's no global way to override that behaviour (that I'm aware of), but you can use the astype method to modify an individual DataFrame.

http://pandas.pydata.org/pandas-docs/stable/generated/pandas.DataFrame.astype.html

Answer 2

It's not possible for Pandas to store NaN values in integer columns.

This makes float the obvious default choice for data storage, because as soon as missing value arises Pandas would have to change the data type for the entire column. And missing values arise very often in practice.

As for why this is, it's a restriction inherited from Numpy. Basically, Pandas needs to set aside a particular bit pattern to represent NaN. This is straightforward for floating point numbers and it's defined in the IEEE 754 standard. It's more awkward and less efficient to do this for a fixed-width integer.

Update

Exciting news in pandas 0.24. IntegerArray is an experimental feature but might render my original answer obsolete. So if you're reading this on or after 27 Feb 2019, check out the docs for that feature.

Answer 3

If you are reading an empty dataframe, you can explicitly cast the types for each column after reading it.

dtypes = {
    'bbox_id_seqno': object,
    'type': object,
    'layer': object,
    'll_x': int,
    'll_y': int,
    'ur_x': int,
    'ur_y': int,
    'polygon_count': int
}


df = pd.read_csv('foo.csv', engine='python', keep_default_na=False)

for col, dtype in dtypes.iteritems():
    df[col] = df[col].astype(dtype)

df.loc[0] = ['a', 'b', 'c', 1, 2, 3, 4, 5]

>>> df.dtypes
bbox_id_seqno    object
type             object
layer            object
ll_x              int64
ll_y              int64
ur_x              int64
ur_y              int64
polygon_count     int64
dtype: object

If you don't know the column names in your empty dataframe, you can initially assign everything as an int and then let Pandas sort it out.

for col in df:
    df[col] = df[col].astype(int)