Support cold-start forecasting for Torch models without target history

Is your feature request related to a current problem? Please describe. I would like to use Darts' Torch forecasting models for cold-start forecasting.

A typical digital marketing example is forecasting traffic, such as page views or video plays, for newly published content (e.g., a new article series or video release) before its release date. A global model can be trained on historical data from existing content, while the new content has:

• no historical target values,
• future covariates such as planned ad spending, promotion schedules, and calendar features,
• static covariates such as content category, creator attributes, and target demographic.

The goal is to forecast traffic from the release date onward, before the content is released, without requiring past target observations for the new content.

Current limitation

From the documentation and previous issues such as #1289 and #2151, my understanding is that Torch forecasting models currently require input_chunk_length >= 1, so this type of cold-start prediction is not supported.

Describe proposed solution The Torch model can forecast with input_chunk_length = 0

Questions for confirmation

• Does #2151 also cover cold-start prediction for new series with zero historical target values?
• Since #2151 was opened in 2024, has there been any progress or updated roadmap for this feature?