Peter has another question. Can anyone help answer his question about the

add_history

parameter? Thank you so much!

I work at a fintech company (Tradesmith.com) am currently testing timeGPT for stock forecasting with some of our proprietary features.

Does the add_history=True param really get you the correct “would have been” projection at each point in time? The documentation confused me on that because it says the historical projections are not affected by the horizon setting. I don’t know how to interpret that if my horizon is 21 and the dataset is daily.

Using

add_history = True

will indeed provide projections at each point, using available data at a past moment to make forecasts and append forecasts together. However, for historical forecasts, the horizon is predetermined. For instance, if`h = 21`, the

historical_forecast

will make a 21-step forecast into the future, but the past data will still be forecasted with a predefined horizon.