Right now, I am performing time-series multi-step forecasting and I am facing a problem in creating lag features for the testing data.
Here is an example dataset I am working with, it contains weekly data from June 2021 - Dec 2021:
Name Sales
Datetime
2021-06-06 Amy 10000
2021-06-13 Amy 15000
...
2021-12-26 Amy 14500
2022-01-02 Amy 10000
2021-06-06 Zoe 12500
2021-06-13 Zoe 13500
...
2021-12-26 Zoe 12000
2022-01-02 Zoe 9000
And I decided to split the data where data from Jun 2021 to Nov 2021 is the training dataset and the leftover will be the testing data.
I did some searching on the web that creating lag features would be a help in performing forecasts, therefore I tried to create lag features of 1,2,3,4 from reading the pacf plot. It is fine with the training dataset, however, I got stuck in the testing dataset.
With the forecast horizon set to be 4, is it not make sense to create lag feature of 1,2,3?
As in the reality, the sales value in the testing dataset will not be known before December. Without any values, how do we know the lag values?
For example, the values in 2021-12-05 and 2021-12-12 will be empty, so when the lag 1 value of 2021-12-12 will be NaN as we do not know the value in 2021-12-05 yet.
Therefore, should I set the minimum value of lag to be 4? Since I aim to forecast 4 steps ahead.
Hopes someone could tell me the correct concept. Thanks!