Basic question about

add_history=True

. I have a dataframe,

df_red

, with daily data rows, and I'm calling the forecast function as shown below with a forecast horizon,

h=5

and

add_history=True

.

df_fcst = nixtla_client.forecast(df_red, h=5, freq = 'B', time_col='target_date', target_col='today_close', add_history=True)

This returns the

df_fcst

dataframe (screenshot below), which has a

TimeGPT

forecast column. For each date/row in the

TimeGPT

forecast column, does the value represent the forecast for

h

periods (days in my case) from the date in the same row? E.g. Does the

TimeGPT

value for 2023-08-23 represent the forecasted value for 2023-08-30 (in

h=5

business days from 2023-08-23). I had assumed that it does (logical) but then came across something in the docs that seems to negate my assumption.

Hello! The values in the column

TimeGPT

are aligned with the date. This means that the value for 2023-08-23 is the value returned by TimeGPT for the date 2023-08-23. The

add_history

parameter simply returns the fitted values of TimeGPT. It's basically the insample values. If set to

False

, then only the predictions are returned. Let me know if that answers your question!

Thanks Marco. So, is the value shown on 2023-08-23 the forecasted value for that date, where that forecasted value was created h days prior, on 2023-08-18. Sorry, that's just a rephrasing/date-shift of my original question. Alternatively, was the forecasted value shown in the 2023-08-23 row created 1-day prior? Basically, there has to be some h-period value associated with each forecasted value, and I'm trying to figure out what it is.

TimeGPT takes an input sequence and generates an output sequence. So the predictions simply come from the default input size. The predictions are not made one at a time.

So, if I need a historical record of 5-day forecasts from each date row in my dataframe, would I have to iterate through all dates/rows one and call the forecast method for each row and then capture the fifth row in the forecast dataframe?

This can be done with cross-validation. You can follow this guide for a detailed tutorial on cross-validation with TimeGPT.

Okay, great. I started playing around with it. I'll look into it more. Thanks for forwarding the tutorial.