I am working on an energy demand forecasting project, where I am using the Facebook Prophet model. I have used 3 years (2017-2020) of training data to forecast energy demand for a week at an hourly interval. The model performs best during summer and worst during winter.

If I want to compare and evaluate the model performance during the 2019 summer, I Have to reduce the training period (2017 - June 2019). Is this a statistically correct way to do this?

In general, if I want to see how seasonality affects the forecasts, is it fair to change the training period of the model?

  • $\begingroup$ I have not seen any statistical theory for training models period. I think these are rules of thumb based on experience rather than being theory based. If you are going to predict seasonality you need enough data to do that. Will you have enough data if you reduce your training data, remember it is catching the seasonality not the raw number of time periods that is key here. $\endgroup$ – user54285 Apr 13 at 2:43
  • $\begingroup$ See robjhyndman.com/hyndsight/tscv $\endgroup$ – kjetil b halvorsen Apr 14 at 15:33

If you have enough data (and since it is an hourly period you do) there should be no problem. You can try some kind of time series cross validation where each resample goes to a different period as long as each resample has sufficient sample sufficiency. This way you will be able to compare


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.