Hi, thanks for your comment:

I opened an issue on Github. This didn't seem obvious to many people that read the article, so it's nice that now we can keep this in mind while using the model:

https://github.com/dmlc/xgboost/issues/4069