https://github.com/nixtla logo
#general
Title
# general
a

Arthur

07/01/2023, 5:02 PM
I might be overlooking it, but is there documentation or a tutorial with multiple univariate series in one (global) model to capture the correlations between series?
k

Kevin Kho

07/01/2023, 8:23 PM
Maybe you can check Hierarchical Forecasting? https://github.com/Nixtla/hierarchicalforecast
a

Arthur

07/01/2023, 10:44 PM
Thanks, but I'm not sure the focus of the Hierarchical Forecasting is to capture the interdependencies. Moreover, there's no hierarchy in the scenario I'm talking about. If you consider a product portfolio, I'm talking about multiple products (with interdependencies) but no categories, countries or other groupings.
c

Cristian (Nixtla)

07/02/2023, 2:51 PM
Hi @Arthur, we added a graph multivariate model in Neuralforecast, the
stemgnn
model. We still dont have a specific tutorial for it, but the syntax is the same as other models.
With that said, we have observed that in most cases univariate models tend to perform better. I strongly suggest you to start with simpler univariate models.
a

Arthur

07/03/2023, 8:10 AM
Ok, thanks for the info Cristian!
t

Tyler Blume

07/03/2023, 3:15 PM
If you have product names or descriptions you could use a llm(or really any method but llms are all the rage) to generate embeddings and pass that to MLForecast, alternatively you could generate time series features and pass that as well. Probably won’t give you too much over stat but you could give it a shot.
👍 1
a

Arthur

07/03/2023, 4:39 PM
I was mostly curious if there was a simple way to explore training on multiple series as Darts provides: https://unit8co.github.io/darts/examples/01-multi-time-series-and-covariates.html#Training-on-Both-Air-Traffic-and-Milk-Series
Or this is a more comprehensive example: https://unit8.com/resources/training-forecasting-models/