Is time series data normalization for xgboost required?

Consider a stock that values

  • 10,00 USD in 2010
  • 75,00 USD in 2015
  • 150,00 USD in 2020
  • and it continues to grow by this day.

Given that decision tree based algorithms like xgboost are generating the tree (splitting the values) based on the ranges, I don’t understand how the tree built on the past data (e.g. years 2000 - 2015) could be in any form applicable for the future price predictions (e.g. years 2015 - 2080).

Could somebody confirm that that feature normalization is truly not required for data that grows beyond the original(/fit/train) range with time?

Do I need to run the raw stock price through some log or sigmoid function before training or is xgboost actually smart enough to deal with this kind of data automatically?

Thank you :slight_smile: