The prediction of a h2o xgboost model using h2o.predict() is very different from what I get by aggregating the predict value of each tree. Why?

[flippercy]

I built an xgboost model using h2o.xgboost(). For a new dataset, I can score it using either h2o.predict() or for the perspective of implementation, I extracted the predict value of each tree, added them up and then converted it into probability of one. However, the results got by these two methods were very different. The second method returns extremely small probability of one. Why? What I did wrong?