Changelog

2.1.0 — 2026-11-01

Added

• Merged ICAIF v9 causal features, three-seed training entrypoint, portfolio adapter and documented experiment provenance into the default branch.
• Included completed, unpromoted FOMC calendar research and its tests.
• Added a final handoff guide, offline replay evidence and a frozen tensor regression fixture.

Fixed

• Restored the frozen feature arithmetic as the default to preserve bitwise compatibility with trained ICAIF inputs. The running-sum implementation is retained explicitly as enhanced_features_fast for research.
• Aligned package, module and citation version metadata with the repository’s v2 release series.

Validation and scope

• Unit/runtime tests, lint, compilation and installed dependency checks pass.
• All 138 sibling competition tests pass; frozen model lineage and offline forecasts, ensemble and execution weights are verified.
• This is a source release checkpoint, not a claim of guaranteed returns, completed organizer review or universal cross-platform reproducibility.
• Frozen submitted source/model materials remain unchanged.

Earlier architecture work — included in this release

Added

• Maintained quant_forecast_lab research utility package.
• Naive-relative forecast evaluation.
• Expanding-window validation primitives.
• Conformal interval utilities.
• Cost-aware directional backtest diagnostics.
• CSV benchmark CLI.
• Unit tests and GitHub Actions CI.
• Contributor and issue templates.

Changed

• Repository layout cleaned so generated experiments no longer dominate the project root.
• Current source distribution uses Apache-2.0. Earlier CC0 distributions remain governed by their original terms.
• Documentation reorganised around reproducibility and out-of-sample evaluation.

Compatibility

• Historical root training and inference scripts are retained.
• Existing committed experiment artifacts remain available in Git history.