I had some general questions:
- Is this (rolling cross validation) the most "rigorous" way to test the performance of time series models? I am thinking that this type of cross validation approach will provide more realistic forecast error estimates compared to other cross validation approaches?
- Once we have completely finished training a time series model and want to use it in the real world - how often should we re-fit the model to newly available data?
Thank you so much!
PS: In another question, I am learning how to compare the performance of ARIMA models as they scale to larger datasets: Correctly Using Microbenchmark in R