Categorical features

Categorical features can't be used by sklearn models without some kind of transformation. Because there are a number of different methods for ([reference](https://en.wikipedia.org/wiki/Categorical_variable#Categorical_variables_and_regression) and [an explanation using patsy](http://statsmodels.sourceforge.net/devel/contrasts.html))

We should encode categorical features at model time, for example using scikit learn's [OneHotEncoder](http://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OneHotEncoder.html).

Even if no other codings are used other than onehot, implementing categorical features in this way we are not reimplementing what already exists in sklearn/patsy by hand.