Migrating Forecasting Models to Darts

[dominiquekleeven]

Draft PR – Migrating Forecasting Models to Darts

Closes #29
Branched from #43

This PR introduces a major structural change to both the back end and front end by migrating native forecasting model usage and bindings to the Darts framework.

Changes

• Migrates the existing Prophet model to DartsProphet.
• Adds a new XGBoost model with its own configuration.
• Models are now stored in .pkl format, the default serialization used by Darts for trained/fitted models. The model_storage_service now uses Darts bindings to save and load models.
• Updates terminology for consistency:
  • In general, regressors are now referred to as covariates.
  • In the context of Prophet, the term regressor is still used, as it remains technically accurate.
• Adds support for both past and future covariates:
  • Future covariates require both historical and predicted data.
  • Past covariates require only historical data.
  • For reference: a Prophet regressor is essentially a future covariate, while XGBoost and other models also support past covariates.
• Introduces a dedicated model registry on the front end, which registers model configs and makes them available to the editor. Each forecast model now has its own config containing base settings, handlers, and templates.

Status

Work in progress.