Includes "deepvar" and "gpvar" models in GluonTS.

[mreismendes]

Hi!

Could someone include DeepVar and GPVar models in GluonTS?

What do you think about that?

Thank you.

[winedarksea]

These are both models added to GluonTS after I built the initial implementation here... I am happy to include them (at some point, soon, my current concern is revamping some of the transformers).
I see it says that DeepVar is "multivariate" - I am not sure if that affects the implementation at all. I don't suppose you know anything more about those new models, any tips or tricks I should keep in mind if I add them in?

[winedarksea]

Example found here: awslabs/gluonts#1095

[mreismendes]

I did some experiments with them, Gluonts docs are very confusing at least to me! But after long research, I discovered I should group the targets.

And then, training!

[mreismendes]

@winedarksea I think your example it's a lot better than mine!

[winedarksea]

Thanks for the examples, that will help!

[mreismendes]

Usually, I work with commodities time-series very often. In those data, I noticed that when you apply at least a year of context improves a lot the validation metrics. Usually, those commodities here have a cycle of one year, like cattle!

I don't have a theoretical explanation of the reason!

Example 01:

freq='M',
prediction_length=3
prediction_length=12

Example 02:

freq='W',
prediction_length=10
prediction_length=52

The drawback is that usually takes a lot more time to train!