Linear regression with pandas dataframe

Instead of replacing '#DIV/0!' by hand, force the data to be numeric. This does two things at once: it ensures that the result is numeric type (not str), and it substitutes NaN for any entries that cannot be parsed as a number. Example:

In [5]: Series([1, 2, 'blah', '#DIV/0!']).convert_objects(convert_numeric=True)
Out[5]: 
0     1
1     2
2   NaN
3   NaN
dtype: float64

This should fix your error. But, on the general subject of fitting a line to data, I keep handy two ways of doing this that I like better than polyfit. The second of the two is more robust (and can potentially return much more detailed information about the statistics) but it requires statsmodels.

from scipy.stats import linregress
def fit_line1(x, y):
    """Return slope, intercept of best fit line."""
    # Remove entries where either x or y is NaN.
    clean_data = pd.concat([x, y], 1).dropna(0) # row-wise
    (_, x), (_, y) = clean_data.iteritems()
    slope, intercept, r, p, stderr = linregress(x, y)
    return slope, intercept # could also return stderr

import statsmodels.api as sm
def fit_line2(x, y):
    """Return slope, intercept of best fit line."""
    X = sm.add_constant(x)
    model = sm.OLS(y, X, missing='drop') # ignores entires where x or y is NaN
    fit = model.fit()
    return fit.params[1], fit.params[0] # could also return stderr in each via fit.bse

To plot it, do something like

m, b = fit_line2(x, y)
N = 100 # could be just 2 if you are only drawing a straight line...
points = np.linspace(x.min(), x.max(), N)
plt.plot(points, m*points + b)