In the stallion tutorial, is it possible to predict on data for unknown agencies or skus?

In the Predict on new data section of the stallion tutorial, it generates future data to predict. This data is tied to each of the existing {agency, sku} groups.

Suppose we have a new agency and a new sku, but we have 6 months of data on it. Is it possible to predict future months volume on this data? This is a case where there may not necessarily be enough data to re-train with this new agency/sku group. However, the hope would be that a model can learn trends from other groups that can then be applied to this data.

In other words, can it predict on a short timeseries with unknown group IDs? I saw this as a (less strict) variant of "Cold-Start" listed in the Model Architecture section, which has an X for the TemporalFusionTransformer, the model used in the stallion tutorial.

If it cannot predict for unknown groups, is there a model that is more appropriate for this? -- Thank you.

Start with the stallion tutorial's “Predict on new data” section and the model architecture documentation linked in the issue. Investigate how unknown agency and SKU group IDs are handled, then document whether short series can be predicted and which model, if any, supports this cold-start case.