Reader Comments

Post a new comment on this article

Testing nowcasting performance on held out data

Posted by lampos on 09 Jan 2015 at 14:50 GMT

After a quick read, it is not clear to me from the paper's text whether testing has been performed on held-out data. In that case, I could not find the actual training / testing periods --- could you please point me to the part of the text that resolves my question? Also, did you perform the 10-article selection (via Pearson correlation) using all data or just the training data?

No competing interests declared.

RE: Testing nowcasting performance on held out data

gfairchild replied to lampos on 04 Feb 2015 at 04:23 GMT

We did not do formal cross-validation when building and testing our models. We discussed this, but since the goal of our paper was to show the feasibility of this new unstudied data source, we decided not to. In fact, in another study we've worked on that's under review (available on arXiv at http://arxiv.org/abs/1410...), we much more rigorously study the quality of the Wikipedia data in the context of influenza forecasting. The 10-article selection was done using all the data as well.

Competing interests declared: I am the second author of the paper.