forecasting-ch15-foundation-modelslisted

# Chapter 15: The Foundation ## Scope and intake Use when a pretrained time-series model is being evaluated on a new series under a fixed model revision. Do not use for assuming a foundation model is automatically superior or diagnosing training contamination from pattern similarity alone. Ask only for unresolved material inputs: Which checkpoint and revision will be fixed? When did the data become public? Are private later data available? What context, horizon and baseline are appropriate? Can the environment run the actual dependency? ## Input contract and additional evidence A regular series with at least three seasons of history and 30 observations beyond the horizon. `checkpoint` is `tiny` (cached, pinned revision) by default; `mini` (about 80 MB), `small` (about 185 MB) and `base` (about 800 MB) download weights on first use, so state the size and get the reader's agreement before choosing one. Minimal **format illustration**, not sufficient training data: ```csv timestamp,target 2010-01-01,41.3 2010-02-01,41.8 2010-03-01,44.9 ``` ## Executable interface Exact CLI columns: `timestamp,target`. Supported method controls: `as_of, checkpoint, frequency, horizon, origins, quantiles, samples, season, seed`. The tool runs Chronos-T5 zero-shot in a separate worker process (so it can share a session with LightGBM), sampling `samples` paths and reporting the requested `quantiles` at `origins` expanding origins plus the untouched final holdout: median MAE, coverage of th