This message was deleted.
# neural-forecasts
Slackbot
12/22/2023, 10:59 AMThis message was deleted.
c
Cristian (Nixtla)
12/26/2023, 4:00 PMHi @Aman Singh. The distribution is usually given by the domain and characteristics of the underlying process of the time series. For most cases Gaussian distribution works. If you are forecasting positive sales of sparse items with low sales maybe a Poisson will work better
Cristian (Nixtla)
12/26/2023, 4:02 PMWe are currently fixing the same loss train and validation loss for all configurations in the same auto model (loss and valid_loss params are in the class, not in the config dictionary). You will need to have two different AutoDeepAR models with different losses and compare results afterwards.