将浮动转换为pandas中的字符串时精度损失

Question

I am converting in Python some floats (some shorter some longer) to strings and getting unexpected (?) results:

Case 1

pd.options.display.float_format = '{:.2f}'.format
pd.DataFrame({'x': [12345.67]})

           x
0   12345.67

Case 2

pd.DataFrame({'x': [1234589890808980.67]})

                      x
0   1234589890808980.75

Case 3

pd.DataFrame({'x': [1234589890878708980.67]})

                         x
0   1234589890878708992.00

I even tried dtypes np.float128 and np.longdouble , but without avail.

Can somebody please explain what is happenning here and is "proper" conversion posiible in Cases 2 and 3?

Thanks!

Answer 1

I'm afraid this "issue" happens in the Python (instead of pandas) side. When you have some instant values like 1234589890878708980.67 it's recognized as float and loses precision instantly, eg:

>>> 1234589890878708980.67
1.234589890878709e+18
>>> 1234589890878708980.67 == 1234589890878708980.6712345
True

You might try something like decimal.Decimal :

>>> import decimal
>>> pd.DataFrame({'x': [decimal.Decimal('1234589890808980.67')]})
                     x
0  1234589890808980.67

EDITED:

OP's added a few questions in the comment.

However, do I understand it right that for this method work correctly the value should be string in the first place?

Yes :)

What if it's float read from csv file?

AFAIK Python's csv reader shall not have performed any type conversion, and you'll get strings that could be later converted freely. Otherwise if you're using pandas.read_csv , you could try setting the dtype and float_precision arguments (you could also ask pandas to load plain strings, and have the values converted later yourself).