Yes, that’s the weak spot of tree-based models. They technically can’t extrapolate on higher/lower feature values than seen in training data. Therefore, it is hard for them to predict target values outside the seen interval.

However, there are some workarounds possible using target transformations, which may help in some cases. One way is to take delta from the previous day as the target. This way out tree-based model will predict delta from the previous day and we can re-construct the real target values one by one, adding predicted deltas to the previous. In this setup tree-based model will be able to predict values higher than the highest in the training data.

Software Architect, Data Scientist, Kaggle Grandmaster (https://www.kaggle.com/alijs1). Sometimes I write about Python, Data Science and Machine Learning…

Software Architect, Data Scientist, Kaggle Grandmaster (https://www.kaggle.com/alijs1). Sometimes I write about Python, Data Science and Machine Learning…