https://www.nixtla.io logo
Join Slack
Powered by
This message was deleted.
# neural-forecast
s
This message was deleted.
c
Can you provide more details of your setting? How many time series do you have? Only the one in the plot? One quick thing you can try is using the 
Auto
version of the models, like the 
AutoNHITS
. It will perform hyperparameter selection by itself on a default or custom search space. More details here: https://nixtlaverse.nixtla.io/neuralforecast/examples/automatic_hyperparameter_tuning.html
w
Hi Christian, Yes it's the one in the plot...I can increase but there will be null values in the target. Is it ok if I just exclude these null values but then there will be breaks in the timestamp. I also want to confirm if I am right to say that the auto models do not support exogenous variables? I also noted that the parameters/arguments supported is quite different (much lesser) compared to the non-auto models, is that correct?
c
You can impute missing values with 0 (or any other method and value) and use the 
available_mask
column in the dataframe marking those rows with 0.
Auto models do allow for exogenous variables and all the native parameters of each model. But they need to be specified in the 
config
dictionary like in the documentation I sent above. The idea is to set ranges and options to some hyperparameters and the auto model will tune then automatically
However, I think your main limitation is the amount of data. With only one rather short time series you can try diminishing the size of the models and train them for less iterations, or other regularization methods. Set 
val_size
to something larger than 0 so early stopping is enabled.
3 Views
Open in Slack
PreviousNext
Powered by