This message was deleted Nixtla Community #general

Join Slack

This message was deleted.

# general

Slackbot

08/17/2023, 8:37 AM

This message was deleted.

👀 1

Matej

08/17/2023, 12:31 PM

(am just a new user as well) imho, you do not need to retrain everytime clearly for instance the cross_validate method

Copy code

crossvalidation_df = sf.cross_validation(
    df = Y_df,
    h = 3,
    n_windows=...
    fitted=True,
    refit= False
  )

has refit=False option hence it definitely predicts without retrain. Tho, I wonder how I would go about setting up for example a weekly refit over the data for example. Currently, It seems quite tricky to assure that the Y_df does not leak into the Y_df_test that is derived from the n_windows and h parameters.

Gerard Conangla Planes

08/17/2023, 12:45 PM

Maybe I'm missing something. If I fit a model with StatsForecast over some timeseries data, and then I want to make predictions over a different timeseries dataset (for instance, future data that I didn't have at the moment of training), intuition tells me that I would need to use the https://nixtla.github.io/statsforecast/src/core/core.html#satstforecast.predict predict method. However, there is no y time series data as an argument of that method. So where do I put this?

Gerard Conangla Planes

08/17/2023, 12:47 PM

In the forward method that I was referring to before (https://nixtla.github.io/statsforecast/src/core/models.html#autoarima.forward) this is more clear, but this functionality seems hidden with the StatsForecast wrapper. Perhaps you need to update the

df

attribute of the StatsForecast object? In any case I find it could be made more clear

Gerard Conangla Planes

08/21/2023, 7:30 AM

Hi @Mariana Menchero , I have a couple of questions 1. I've seen your other reply to Akash. That seems to confirm that you need to "train every time" if you, for instance, update the data (for instance because you got more recent time series). Is that so? 2. The

forward

method appears in the underlying models but apparently not in the

StatsForecast

object. Could it be made public? That would be a strong reason to use it! Thanks 😉

Mariana Menchero

08/21/2023, 5:21 PM

Hi @Gerard Conangla Planes 1. Yes, if your data increases or changes in some way, you’ll need to train again the models from statsforecast. 2. We’re actively working on this. You might also be interested in our TimeGPT. It’s still in beta but maybe it can help your use case. See details here: https://www.nixtla.io/ and here https://docs.nixtla.io/

✅ 1

4 Views

Open in Slack

Previous Next