I have a problem of 20 timeseries groups. I need to forecast them for a horizon of 180 days. I am having a context length of 365 days as I only have 4 years of data for them.
I coded timeseriesdataset, dataloaders etc. However I see that the validation loss and accuracy of 16 timeseries is not at par. In general some pbservation is that few of the series that are low in volume around 20,000 as value , is under performing. Series where values are 100,000 etc are good.
How do I improve few series that underperform. ChatGPT suggests that I give weights map and multiply loss with those weights, so higher validation loss for under performing series so model will learn more on it. I am unsure how to use weights. Please help me with some code example.
Any other ideas as model is a black box to understand.
Also, I am running 10 or so epochs as I used early stopping with patience 2 and I see from learning curve that val loss was least in 5th or 6th epoch. If I am increasing the patience, the resultant model reduces even further in accuracy.
Some more useful information: Hidden layer 64 and rnn 3.few things tried includes adjusting hidden layer, loss rank, gradient clip etc. I have strictly positive values so using soft max transformation on target and normal distribution loss.
