https://www.nixtla.io logo
Join Slack
Powered by
This message was deleted.
# neural-forecast
s
This message was deleted.
c
Hi @Farzad E! The 
Auto
model have the input size as a multiplier of the horizon rather than seasonality. This is how the NBEATS and NHITS paper define it. But this are general rule of thumbs, it can be the case that for your application a longer input size will work better.
One drawback of a longer input is that you will have less training windows, so you need longer historic data. Also the increase in performance due to more input data varies by model (in principle, transformers are better in extracting relevant parts on longer inputs due to the attention mechanism).
f
@Cristian (Nixtla) Thanks a lot for the quick response. So basically it is a trade off between wanting to include more data (i.e., more information) into each pass which means longer chunks governed by input_size but also wanting to have many more sampling from across the series which is what the windows_batch_size covers, right?
c
No problem! Yes, as with most cases the model's performance as a function of 
input_size
will have the classic U-shape behaviour. There is a sweet-spot given the amount of data/architecture.
Another important thing to consider is that you can pass exogenous variables that might help to capture some long-term relations and seasonalities outside the 
input_size
. For example calendar features and holidays dummies.
f
Yeah that's a really good point on exogenous features. I pass them to the model as a habit but don't usually think much about their importance especially that they can compensate for shorter input length. While we are talking about model params I may as well ask another question I have. If the number of series in your data increases, would you increase the batch size? I know there is also a trade off between small batch size taking a longer time to go through all batches and large batch size finishing more quickly but then you end up updating the model fewer times. Have you seen the optimal batch size to be a function of total number of time series in the data?
3 Views
Open in Slack
PreviousNext
Powered by