These values are predicted after xgboost::xgb.train:
247367.2 258693.3 149572.2 201675.8 250493.9 292349.2 414828.0 296503.2 260851.9 190413.3

These values are predicted after xgboost::xgb.save and xgboost::xgb.load of the previous model:
247508.8 258658.2 149252.1 201692.6 250458.1 292313.4 414787.2 296462.5 260879.0 190430.1

They are close, but not the same. The differences between these two predictions range from -1317.094 to 1088.859 on a set of 25k samples. When comparing with true labels, then the MAE/RMSE of these two predictions do not differ much.

So I suspect that this has to do with rounding errors during load/save since the MAE/RMSE do not differ as much. Still, I find this strange since binary storing the model should not introduce rounding errors?

Anyone a clue?

PS Uploading and documenting the training process seems not important to me here. I could provide details if necessary, or make a simulation with dummy data to prove the point.