Hey Kevin, thanks for the comment, that's a creative idea!

If by the trend of the data you mean the first derivative or a difference feature, this approach might improve the predictions, but the issue of extrapolation still remains.

XGBoost, by its nature, cannot extrapolate target values outside of those in the training data particularly well. Even with a trend feature, XGBoost would be unable to fit a function with a monotonically increasing or decreasing trend.