🇬🇧 English · 🇪🇸 Español

Independent evidence and structural estimation

These research workflows connect independent validation and structural estimation. Each can run offline from an installed package with the standard numerical dependencies.

The real-data study adds an original-data Romer–Romer (2010) baseline replication against independent software and published rounding targets. The synthetic NK example below remains a parameter-recovery test; the empirical study has its own data, assumptions and conclusions.

Task	Entry point	Guide
Inspect independent numerical/data evidence	puremacro.validation.run_research_benchmarks	Research benchmarks
Fit structural models to correlated empirical moments	puremacro.structural.fit_structural and lp_moment_targets	Empirical-to-structural bridge

1. Run the independent benchmarks

from puremacro.validation import run_research_benchmarks

report = run_research_benchmarks()
print(report.passed)  # False: the documented RR2010 t-rounding comparison fails.
report.write("research_output/benchmarks")

The dossier distinguishes analytical solutions, authenticated external Dynare output and the original Romer–Romer baseline with independent software and published targets. It records every observed/reference value, source, units, tolerance, limitation and negative-control result. Four of five cases pass; the original-data RR2010 t statistic misses the paper's rounding interval by about 0.000003925. That failed comparison is retained, as explained in the real-data guide. A pass establishes only its stated scope.

2. Run a complete empirical-to-structural study

python -m puremacro.examples.structural_irf_matching --output research_output/structural

The application independently simulates a linear three-equation New Keynesian economy with a persistent policy disturbance and measurement errors. It estimates output/inflation LP responses, constructs their full joint HAC covariance, and fits risk aversion, Phillips-curve slope and shock persistence by solving the model's equations through load_mod. Model responses and targets use a 0.01 quarterly-rate innovation (one percentage point).

Horizons 0–4 are fitted; horizons 5–8 are descriptive out-of-fit checks using the same sample. The application exports the simulated observations, moment vector, joint covariance, fitted parameters, held-out comparisons, a figure and a manifest containing the model, data hash and random seed. This is a transparent synthetic recovery exercise, not a historical monetary-policy replication.

For an empirical study replace the simulator with your transformed data and identified shocks, or provide your own MomentTargets. The same model callback can wrap a DSGE, VFI or HANK solve. Model equations, moment definitions, shock units, frequency and cumulative/level conventions must agree. The bridge does not supply missing identification assumptions or HA equilibrium derivatives.