FairModelJAX

Ray C. Fair's US and Multi-Country macroeconomic models, ported from FORTRAN 77 + the Fair-Parke DSL to Python + JAX. Fair is great; the goal here is to play with his models in a modern stack — autodiff, JIT, vmap — instead of feeding FP.EXE a deck.

Upstream: https://fairmodel.econ.yale.edu/.

Install

git clone https://github.com/mvazcar/FairModelJAX.git
cd FairModelJAX
pip install -e .[dev]

Python 3.13+. CPU JAX by default; for GPU, pick the right jaxlib from JAX's install matrix.

Run

python run.py --model is                       # 2-equation tutorial
python run.py --model is --data-source fair_2013   # reproduces Fair (2013) coefficients
python run.py --model us                       # full US model
python run.py --model mc                       # 36-country MC model
python run.py --model mc --mc-country CA       # one country only

Every step caches to output/ as parquet. --resume N skips to step N by reading from cache; --force ignores cache.

Update the data

# IS model (FRED, default source) — refresh from the FRED CSVs:
python build_is_dat_from_fred.py && python run.py --model is --force

# US model — drop a new FMFP.ZIP into raw_source/03_us_model/, then:
python run.py --model us --force

# MC model — drop a new mcj.zip extract into raw_source/04_mc_model/mcj_extracted/, then:
python run.py --model mc --force

Fair re-estimates the US model every ~90 days; MC ~yearly. raw_source/SOURCES.md has the permanent download URLs. See IS_DATA_SOURCES.md for why the FRED vintage differs from Fair (2013).

Layout

src/pyfair/
  us/         model.py, solve.py            — 24 stochastic eqs + 95 identities + forecast
  mc/         model.py, solve.py, shr.py    — 36 countries + 1,686 trade-share eqs
  core/       estimate.py, solver.py, readers.py
  pipeline/   is_pipeline.py, mc_pipeline.py
  tools/      fpexe.py                      — FP.EXE shell-out for golden-file tests
raw_source/   Fair's originals, bundled
tests/        461 tests (+1 slow, gated by --runslow)

Test

python -m pytest
python -m pytest --runslow   # adds the FP.EXE golden-file regeneration (~5 min)

References

The model — every line of the original economics is Ray C. Fair's. The book-length versions that document each iteration:

• Fair, A Short-Run Forecasting Model of the United States Economy, D.C. Heath, 1971.
• Fair, A Model of Macroeconomic Activity, Vol. I (1974) and Vol. II (1976), Ballinger.
• Fair, Specification, Estimation, and Analysis of Macroeconometric Models, Harvard UP, 1984. (Introduces the Fair-Parke Program — the FORTRAN engine this repo ports.)
• Fair, Testing Macroeconometric Models, Harvard UP, 1994.
• Fair, Estimating How the Macroeconomy Works, Harvard UP, 2004.
• Fair, Macroeconomic Modeling: The Cowles Commission Approach, M.I.T. Press, 2024.

Primary references bundled as PDF in raw_source/07_docs/.

The methodology — if you want the history of macroeconometric models and why the Cowles tradition and DSGE diverged:

• Fernández-Villaverde, Horizons of Understanding: A Review of Ray Fair's Estimating How the Macroeconomy Works, Journal of Economic Literature, 46(3), September 2008, 685–703. doi:10.1257/jel.46.3.685 · PDF.

  Good story, short, reads the lineage from Tinbergen (1937) → Klein–Goldberger (1955) → the 1970s macroeconometric crisis (Nelson 1972; Lucas 1976) → Fair and DSGE as the two surviving branches.

Development

Written with Claude Code (Claude Opus 4.7 Max).

License

The Python code under src/ and tests/ is released into the public domain — do whatever you want with it. Fair's bundled originals under raw_source/ keep his copyright. See LICENSE.