| English | 中文 |
The tracked model uses a target history with historical exogenous series for forecasting. Distinguish information available at the forecast origin from future observations. Specify sampling interval, history and horizon, target/covariate roles and whether the goal is retrospective evaluation or deployment. Use chronological development and a separately protected final test; compare relevant persistence or seasonal references.
Scientific background: time-series design.
AirMosaic currently provides an external Python/PyTorch reference, not a controlled training/inference adapter. Future-known covariates, multiple entities and history-only inference require explicit adaptation. Configuration, code calls and current checks belong to the matching skill.
Use the shared evaluation design and submit a model plan before generating experiment code.