Glossary

Definitions of core forecasting, orchestration, and Google Cloud terms used across scale-forecasting code, configurations, and documentation.

---

Term	Definition
Backtest fold (Rolling-origin window)	A historical evaluation split (backtest.n_windows, backtest.step_size) where the model trains on data up to a cutoff timestamp and predicts the next horizon steps against held-out actuals.
BigLake Iceberg table	A BigQuery table (source_series_iceberg) backed by open Apache Iceberg metadata and Parquet files in Cloud Storage (gs://<project>-warehouse/warehouse), accessed via a BigQuery Cloud Resource connection.
Calibration (Conformal prediction)	Post-processing step (calibration.py) that adjusts prediction intervals (lower_80, upper_80, lower_95, upper_95) using empirical out-of-fold residual quantiles so empirical coverage matches the nominal level.
Cell ((ts_id, model_type))	The atomic unit of execution and persistence in forecast_metadata: one time series (ts_id) evaluated and forecast by one model (model_type) within a run_id.
Ensemble stacker	Post-processing model (ensembler.py) that blends base model forecasts using out-of-fold backtest accuracy (mean, median, trimmed_mean, inverse_metric, nnls, ridge, or xgb).
Family (model_family)	One of the five execution groups (statistical, ml, deep_learning, automl, native) that the DAG router dispatches as an independent parallel compute job.
FPP3	Forecasting: Principles and Practice (3rd ed.) by Hyndman & Athanasopoulos — the canonical textbook reference for several statistical baselines (naive, snaive, drift, tslm, stl_ets, harmonic_arima) and evaluation metrics (mase, rmsse, winkler).
Horizon (data.horizon)	Number of future time steps ($H$) predicted beyond the end of each series (or beyond each backtest fold cutoff).
HPO (Hyperparameter optimization)	Automated Optuna search (hpo.py) over model hyperparameters (hpo.enabled, hpo.n_trials) minimizing backtest validation error prior to final refit.
Local vs. global (panel) training mode	Local (training_mode = "local"): fits an independent parameter set per ts_id. Global / panel (training_mode = "global"): trains a single cross-series model (lightgbm_global, PyTorch deep_learning, or vertex_automl) across all series in the panel simultaneously.
LPT scheduling (order_chunks_lpt)	Longest Processing Time first chunk ordering in vertex_engine.py, which schedules the largest series chunks first across worker threads/processes to minimize tail stragglers.
Out-of-fold (OOF) predictions	Historical predictions produced on held-out backtest folds and stored in backtest_oof, used to compute leakage-free leaderboard metrics, conformal intervals, and learned ensemble weights.
Reconciliation (Hierarchical forecasting)	Post-processing step (reconciliation.py) that adjusts forecasts across a hierarchy (such as total -> category -> store_item) using bottom_up, top_down, or mint_shrink (Minimum Trace shrinkage) so child forecasts sum exactly to parent totals.
Registry	The five native BigQuery output tables (run_registry, run_jobs, forecast_metadata, forecast_predictions, backtest_oof) and five analytical SQL views (v_model_leaderboard, v_model_leaderboard_comparable, v_backtest_coverage, v_run_summary, v_run_jobs) defined in registry/ddl.py and registry/views.py.
run_id	Deterministic identifier (<slug>-<12hex>) computed by registry/ids.py from the semantic RunConfig specification, excluding operational plumbing (project_id, region, "auto" machine sizing).
Smoke configuration	One of the 42 fast, end-to-end cloud verification configs in configs/smokes/ (01–42) that exercises a specific runtime, model family, hardware accelerator, or post-processing feature on 50–100 series.
Tier 1 & Tier 2 feature attributions	Tier 1 (fit_diagnostics["feature_attributions"] in forecast_metadata): series-level global feature importance. Tier 2 (explanations JSON in forecast_predictions): per-horizon-step local feature contributions produced by BaseModel.explain().
Validation ledger	docs/validation.md — the audited record linking every smoke config (01–42), root demo config (20), and notebook (11) to its live run_id, execution date, and verified architecture axes, enforced by tests/unit/test_validation_ledger.py.
Warehouse	The Cloud Storage bucket (gs://<project>-warehouse/warehouse) and BigQuery dataset (scale_forecasting) storing the source time-series panels and serialized run artifacts.
Worker barrier	Synchronization step in multi-worker Vertex AI / GKE runs (_should_use_worker_barrier and _wait_for_worker_pool_barrier in vertex_engine.py) that ensures replica workers finish writing before the primary worker finalizes the run header.