The docs mention the historical forecast horizon != horizon; instead, fixed based on freq. Is there a mapping somewhere to see what each freq is mapped to? This would be extremely helpful in establishing benchmarks. Thx!

Not sure I understand the question - which specific part of the documentation are you referring to?

@Olivier - sure, a lot of these things can be difficult to explain. From: https://docs.nixtla.io/docs/tutorials-historical_forecast In Section 3:

Currently, the historical forecasts are not affected by

h

, and have a fix horizon depending on the frequency of the data.

So basically, would like to understand where the fixed horizons are defined. And a related question: Let's say we have 5min freq, and the fixed h = 12 (1hr). After the historical forecasts are back-propagated, what's the correct way to interpret; the TimeGPT value for a given timestamp reflect the forecast horizon? e.g. TimeGPT value @ 08:00 = 100. meaning, at 07:00, the forecast of 100 was made for 08:00? Ultimately trying to understand if add_history could be used as a reasonable backtest (at least for that fixed horizon), or if we'd need to create our own history by requesting forecasts across a historical dataset

So basically, would like to understand where the fixed horizons are defined.

It's fixed, so it can't be changed. Re. the other question: your interpretation is correct, we pick a fixed number of historical datapoints, and make a forecast

h

steps ahead, and concatenate all those forecasts. I think for backtesting it makes much more sense to use cross-validation, because what you are interested in is to get a notion of the performance on historical data in a manner most reflecting the actual set-up.

That's ok, I'm not asking if we can change it. The docs seem pretty clear on that:

Currently, the historical forecasts are not affected by h

The question, in other words: Can I see the list of fixed mappings? If my freq='5min' - what

h

will the model use for history? Without knowing what the forecast horizon was for the historical forecasts, we don't know whether they're relevant for us or not.