macroforecast

What you can do: run reproducible macro-forecasting experiments with custom preprocessing DAGs, 35+ models, statistical tests, importance interpretation, and FRED-SD geographic visualization. One YAML recipe → bit-exact replicable manifest, replayable bit-for-bit by macroforecast.replicate(...).

Pick your path

If you want to …	Start here
Run a forecast on FRED data and read the results	Researchers
Write a custom recipe with your model / preprocessor (incl. partial-layer debugging and custom hooks)	Recipe authors
Modify the package source / contribute	Contributors
Understand the design — why layers / sub-layers / contracts are split this way (read top-to-bottom)	Architecture — design narrative, ~18 prose pages
Look up an option — what does apply_official_tcode do, what models are L7 shap_tree compatible with, etc. (search / browse)	Encyclopedia — auto-generated reference, 189 pages
See replication studies (bundled examples + 3 research walkthroughs today, 4+ more in v0.9.1)	Replications
Visually explore the layer / DAG topology	Navigator
Hit an error or something doesn’t work	Troubleshooting & FAQ

Architecture vs Encyclopedia: same 12-layer system, two angles. Architecture is prose — “why is L2 separated from L3”, “how does L7 read L4 sinks”, “what are the cross-layer references”. Encyclopedia is lookup — one page per axis with every option’s definition, when to use, when NOT, references, related options. Architecture is hand-written; encyclopedia is auto-generated from LayerImplementationSpec + OPTION_DOCS and locked by the ci-docs drift gate.

Architecture overview

12-layer canonical design — see architecture. The full 4-part design lives under plans/design/ in the repo.

L0 -> L1 -> L2 -> L3(DAG) -> L4(DAG) -> L5 -> L6 -> L7(DAG) -> L8
        |      |      |       |
       L1.5   L2.5   L3.5    L4.5 diagnostics

Install

pip install macroforecast

See install for extras and source install.

License

MIT