NEWS.md
In CRAFT: Conditional Regime Analog Forecasting with Trajectories

CRAFT 0.4.0

Replaced the original craft_autotune() with the accelerated implementation. Staged screening, scoring-only backtests, bounded caches, dynamic parallel scheduling, and deferred full fitting are now available under this one name.
Removed the separate craft_autotune_fast() function and export. Existing calls to that name should be changed to craft_autotune().
The returned class is craft_autotune. To evaluate every candidate at the final budget, set stage_budgets = n_testing.
Consolidated source, reference documentation, README, and vignette; updated regression tests and the benchmark for the single public tuning interface.

Added craft_autotune_fast() with scoring-only backtests that skip forecast distributions and correlation diagnostics, then fit selected models in full.
Added staged candidate screening with stable seeds and reuse of completed origins, plus a common backtest-origin grid across candidate configurations.
Added bounded caches for trajectory embeddings and deterministic regime fits, including their SVD computations. Stochastic and plotting regime calls bypass regime caching. Cache entries never cross search calls or training prefixes.
Added dynamically scheduled parallel scoring with one history transfer per worker, compact scoring results, and no nested parallelism.
Added regression tests, usage documentation, and a reproducible tuner benchmark.
Preserved craft_fit() and craft_autotune() public interfaces.

Added craft_autotune() random-search hyperparameter tuning for craft_fit().
Added scalar, choice, integer, numeric, log-scale, and function search spaces.
Added backtest metric and composite ranking, retained top fits, candidate failure diagnostics, and optional parallel search with bounded worker counts.
Added tuning examples, reference documentation, and regression tests.
Corrected stale license and maintainer information in the README.

Initial CRAN-preparation release.
Added trajectory embedding helpers for lag and lead cumulative changes.
Added SVD changepoint regime detection and transition-based regime forecasting through craft_fit() and craft_predict().
Added smooth empirical forecast distribution fitting.
Added tests, manual pages, README, and vignette.