macroforecast 0.9.5a0

Pandas-first macro forecasting workflow package

macroforecast

macroforecast is being rebuilt as a pandas-first macro forecasting workflow package. The current public surface is intentionally small:

• macroforecast.meta: package-wide defaults such as random seed and worker count.
• macroforecast.data: canonical date-indexed panels, metadata, FRED/custom loaders, and study data specs.
• macroforecast.preprocessing: direct pandas preprocessing callables.
• macroforecast.data_summary: one-panel summary tables.
• macroforecast.data_analysis: before/after preprocessing analysis.
• macroforecast.metrics: scoring metrics and forecast-table ranking.
• macroforecast.tests: forecast-comparison tests and diagnostics.
• macroforecast.evaluation: namespace wrapper for metrics and tests.

The old YAML/runtime implementation is no longer part of the clean importable package. A reference copy is preserved on the legacy-runtime-reference branch.

Install

pip install -e ".[dev]"

Torch is not installed by default in this rebuild.

Quick Use

import macroforecast as mf

mf.configure(random_seed=42, n_jobs=1)

bundle = mf.data.load_custom_csv(
    "panel.csv",
    date="date",
    dataset="my_panel",
    frequency="monthly",
)

data_spec = mf.data.spec(
    bundle,
    target="INDPRO",
    horizons=[1, 3, 6],
    start="1990-01-01",
    end="2024-12-01",
)

processed = mf.preprocessing.reprocess(
    data_spec,
    transform="custom",
    transform_codes={"INDPRO": 5},
    outliers="iqr",
    impute="em_factor",
)

summary = mf.data_summary.summarize_data(processed.panel)
analysis = mf.data_analysis.analyze_data(bundle.panel, processed.panel)

Data Shape

The standard panel is a pandas.DataFrame with:

• a DatetimeIndex named date
• one macro series per column
• numeric values in the cells
• dataset metadata stored separately and mirrored in panel.attrs["macroforecast_metadata"]

macroforecast.data.load_*() returns a DataBundle(panel, metadata). macroforecast.data.spec(...) attaches target, horizon, sample-window, and predictor choices to that panel.

Documentation

Function-level documentation lives under docs/reference/.

License

MIT