This message was deleted.

Hey. • If you already know the values of the 9th to 30th you can retrain it or provide the full dataframe (train + the new values) through the

df

argument of predict. Otherwise you can set h=8 and ignore the first 4 values. • If you have historic and future values of a feature you have to add it to

futr_exog_list

when instantiating the model and then provide it through

futr_df

when predicting. The difference between historic and future exogenous is how many samples we provide the model (for historic it's input_size and for future is input_size + h).

Hi @José Morales, thanks for your clarifications and suggestions. On point 1 and a little bit more background on what I am doing. I got a df say containing actual data with exogenous features for WK1 to WK35. Horizon is 10 WKS. I used WK1 to WK30 for training. I then use the trained model to predict WK31 to WK35 (unseen data) and check the predictions with the actuals for these weeks. I am happy with the model and now I want to use this model to predict WK36 to WK40. Based on one of your suggestions and for the

df

in the predict argument, I will need to provide data in

df

containing train (WK1 to WK35) + unseen (WK31 to WK35) + horizon (WK36 to WK40)? Thanks.

I believe you need to provide WK1 to WK35 through

df

and WK36 to WK40 through

futr_df

but how do I deal with WK30 to WK35 which I do not want to be used for training if I am to provide them through

df

in the

nfc.fit()

argument?

They should be provided to the predict method (after you've trained). This is so that the model considers from that point onwards to predict

Oh, so the steps and code should be,

1. Training ->

nfc.fit(df=df_WK1_WK30)

2. Prediction ->

nfc.predict(df=df_WK1_WK35, futr_df=df_WK36_WK40)

?

exactly

Will try this out and revert on outcome. Thanks a million @José Morales