This message was deleted Nixtla Community #statsforecast

Join Slack

This message was deleted.

# statsforecast

Slackbot

11/29/2023, 1:21 PM

This message was deleted.

José Morales

11/29/2023, 3:08 PM

Hey. What's the

that you're using for the prediction intervals?

Sergey

11/29/2023, 3:13 PM

I use

h = 56 = forecasting horizon

José Morales

11/29/2023, 3:15 PM

The intervals are computed for each step, so it's possible that in those 5 windows the error became smaller as time went on, thus the intervals would be narrower

Sergey

11/29/2023, 3:23 PM

OK, I get it. So I should use a larger number of intervals? For example

n_windows = 10

or more? The problem is that in such a case I need more than

h*n_windows = 56*10 = 560

days in the training set. And such data is not always available.

Sergey

11/29/2023, 3:29 PM

I think it would be nice to be able to use overlapping windows for the conformal prediction intervals. In the code the

step_size

is hard coded (link) to be equal to

. Maybe you should consider making a

step_size

a parameter? Or are there any thoughts on why this is a bad idea?

José Morales

11/29/2023, 3:34 PM

5 should be fine. If you're able to provide a minimum example where this happens I can help investigate further

Sergey

11/29/2023, 3:43 PM

Okay, I will prepare a Jupyter notebook with code and data to reproduce such behaviour and post it here later.

Sergey

12/01/2023, 8:55 PM

Here is a minimal example to reproduce the discussed behaviour. Any comments about the code are welcome )).

nixtla_example.ipynb

Open in Slack

Previous Next