XGBoost: Feature Names Mismatch

I'm struggling big-time to get my XGBoost model to predict an article's engagement time from its text. First, I get a dataframe representing the features I extracted from the article like this:

article_features = pd.concat([tfidf_df, numeric_df_normalized], axis=1)

I then train my model and get the relevant correct columns (features):

with open('correct_columns') as fp:
        correct_columns = pickle.load(fp)

Then I go through all of the required features and set them to 0.0 if they're not already in article_features:

for col in correct_columns:
        if col not in article_features.columns:
            article_features[col] = 0.0

Finally, I delete features that were extracted from this article that don't exist in the training data:

for col in article_features:
    if col not in correct_columns:
        del article_features[col]

So now article_features has the correct number of features. I try to run:

model.predict(article_features)

And I get:

ValueError: feature_names mismatch:...

So I Google around and try converting my dataframe to :

model.predict(article_features.as_matrix())

But I get the same error.

I was then worried about order of columns in article_features not being the same as correct_columns so I did:

article_features.sort_index(axis=1, inplace=True)

But got the same error.

Any idea how to fix?

Thanks!