forecasting-ch14-global-neurallisted

# Chapter 14: The Globalizer ## Scope and intake Use when many related short histories may benefit from a global probabilistic neural model. Do not use for claiming zero-observation transfer or calling the demonstrated Gaussian MLP DeepAR. Ask only for unresolved material inputs: Which entities are related and which will be unseen at deployment? How much recent context exists? Is the outcome continuous, count or nonnegative? What horizon and distributional loss matter? ## Input contract and additional evidence CSV series_id,timestamp,target with regular ordered observations per entity. Config context,horizon,season,epochs. The CLI holds out the last quarter of sorted entity groups; use notebook adaptation for a different declared entity split. The Gaussian MLP assumes a continuous response; test entities still need observed context. Minimal **format illustration**, not sufficient training data: ```csv series_id,timestamp,target A,2026-01-01,20 A,2026-01-02,22 B,2026-01-01,60 ``` ## Executable interface Exact CLI columns: `series_id,timestamp,target`. Supported method controls: `as_of, context, epochs, frequency, horizon, season, seed`. At least four aligned related series are required. The last quarter of sorted entity groups is held out, not a user-configurable split. The actual Gaussian MLP uses 100 recursive paths for nominal 80% bands. It does not implement DeepAR or tune hyperparameters on held-out entities. ## Applied procedure 1. Define time and entity hold