Skip to content

Releases

v6.6.0

Choose a tag to compare

@jalonso1979 jalonso1979 released this 16 Nov 00:56

Full Changelog: v6.5.0...v6.6.0

v6.5.0

Choose a tag to compare

@jalonso1979 jalonso1979 released this 15 Nov 02:29

Full Changelog: v6.4.0...v6.5.0

puremacro 6.4.0: own-data fiscal brief, connector freshness, LP contracts, and correctness fixes

Choose a tag to compare

@jalonso1979 jalonso1979 released this 12 Nov 22:10

puremacro 6.4.0 adds reusable own-data quarterly fiscal policy briefs, bounded connector provenance and freshness reporting, local-projection chronological and specification contracts with two-way fixed-effects panel IV, and eleven numerical and data-provenance correctness fixes across DSGE, sovereign DSA, weak-IV, SMC, proxy-SVAR, and trade.

  • Reusable own-data quarterly fiscal policy brief (examples.first_brief_fiscal: FiscalBriefConfig, prepare_fiscal_data, parse_quarterly_index, and CLI flags) supporting custom CSV/DataFrame inputs, Gordon–Krenn scaling, explicit external-instrument vs opt-in Blanchard–Perotti identification, lagged state regimes, short-sample/short-regime guards, and portable bilingual (en/es) SHA-256 output bundles while preserving exact reproduction of Ramey–Zubairy (2018).
  • Bounded connector provenance and freshness reporting (fetch.freshness: connector_freshness_report, write_freshness_report, RateBudget, and python -m puremacro.fetch.freshness) across cache, recorded, and rate-budgeted live modes with domain-identity checks (LF = E + U on BLS LAUS, nominal/real/deflator consistency on OECD QNA, and [0, 100]% non-fixture share bounds on energy transition), plus ten registered cross-country panel builders in fetch.registry.
  • Local-projection contracts across lp_hac, lp_iv, estimate_distributional_lp, and estimate_quantile_lp enforcing chronological regularity on date indices, rejecting ±np.inf and conflicting canonical/alias arguments, recording specification provenance and weak-IV warnings, and estimating genuine horizon-matched two-way fixed-effects 2SLS with structural residuals and entity-clustered standard errors in panel_lp_iv.
  • Eleven wrong-number or silent-substitution conditions fixed and documented in the correctness advisories: deterministic .mod space-separated signed shock lists, var.identify.sign_zero orientation and joint restriction verification, BLS LAUS measure codes 04/06, explicit failure/warning instead of fixture substitution in fetch.energy, third-order Sylvester (1, 1, 2) mixed real/complex Schur blocks, quarterly compounding in sovereign.dsa.simulate_stochastic_dsa, multivariate Kleibergen–Paap rk Wald $F$ and FWL-consistent weak_iv_summary, exact [lb, ub]-truncated SMC priors, Jentsch–Lunsford (2019, 2022) centered moving-block bootstrap in proxy_svar, baseline-anchored Geary–Khamis welfare accounting in trade.optimal_tariffs, and non-fallback live snapshot accounting in connector_health.
  • Lazy module loading in inference and models decoupling import puremacro.lp from DSGE and trade graphs, plus arch>=9.0 lowercase Literal compatibility across GARCH wrappers.

Install with python -m pip install puremacro==6.4.0 on Python 3.11 or later. Read the fiscal brief guide https://jalonso1979.github.io/puremacro/docs/first_brief_fiscal/, connector freshness guide https://jalonso1979.github.io/puremacro/docs/CONNECTOR_HEALTH/, local-projections guide https://jalonso1979.github.io/puremacro/docs/lp/, correctness advisories https://jalonso1979.github.io/puremacro/docs/ADVISORY/, and full changelog https://github.com/jalonso1979/puremacro/blob/v6.4.0/CHANGELOG.md.

Own-data fiscal estimates require user-supplied causal identification; opt-in shock_mode="blanchard_perotti" imposes recursive within-quarter timing and should be defended explicitly for the application. Short samples (< 120 effective quarters) and small state regimes (< 40 observations) emit explicit asymptotic warnings rather than blocking estimation. Connector freshness reports inspect bounded metadata, cached series, recorded wire fixtures, or rate-budgeted live probes and do not certify permanent upstream API stability. Previously saved BLS LAUS panels, energy-transition tables, third-order complex-root DSGE solutions, and quarterly stochastic DSA paths produced with versions up to 6.3.0 should be rebuilt or re-run per docs/ADVISORY.md.

Validation on 4be8a2ffc3cb8be447453916bdf971c0921e40d9: 24,044 local tests passed, 89 skipped, and all seven local release gates passed. The main CI matrix https://github.com/jalonso1979/puremacro/actions/runs/43270156065 and tag release workflow https://github.com/jalonso1979/puremacro/actions/runs/43293497320 passed, including nine OS/Python targets, native and Windows typing on Python 3.11/3.12 (arch 9.0.0, mypy 2.4.1), strict documentation builds, 134 opt-in slow/reference/replication tests (19 skipped), and installed-wheel examples. The real Pyodide 0.28.3 kernel passed 121 smoke tests. Downloaded PyPI artifact hashes and all 1,052 packaged wheel files (1,058 entries including .dist-info) and 1,588 repository sdist files were verified byte-for-byte against v6.4.0, followed by a fresh --no-cache PyPI installation and isolated execution of the four base workflows, own-data fiscal brief CLI, and 9-connector freshness CLI. These cohorts overlap. See the attached release-validation.json for artifact hashes and scope.

The examples gallery retains the documented Romer–Romer (2010) published t-statistic rounding mismatch; it is not relabeled as passing. Timing checks use the repository's CI scale of 10 and do not certify unscaled wall-clock deadlines.

The pinned course edition keeps its own package version and validation record. This library release does not upgrade or requalify those course environments.

Español. puremacro 6.4.0 incorpora informes fiscales trimestrales reutilizables con datos propios (examples.first_brief_fiscal), informes acotados de procedencia y frescura de conectores (fetch.freshness), contratos cronológicos y de especificación en proyecciones locales con IV de panel de efectos fijos bidireccionales (lp.panel_lp_iv), y once correcciones numéricas y de procedencia de datos documentadas en los avisos de corrección. Consulte la guía del informe fiscal https://jalonso1979.github.io/puremacro/docs/es/first_brief_fiscal/, la salud de conectores https://jalonso1979.github.io/puremacro/docs/es/CONNECTOR_HEALTH/, proyecciones locales https://jalonso1979.github.io/puremacro/docs/es/lp/ y los avisos de corrección https://jalonso1979.github.io/puremacro/docs/es/ADVISORY/.

Las estimaciones fiscales con datos propios requieren justificar explícitamente la identificación causal, y las muestras cortas emiten advertencias asintóticas explícitas. Los paneles BLS LAUS, tablas de transición energética, soluciones DSGE de tercer orden con raíces complejas y trayectorias DSA trimestrales generadas con versiones hasta 6.3.0 deben recalcularse según docs/es/ADVISORY.md. La galería de ejemplos conserva la discrepancia de redondeo publicada de Romer–Romer (2010) sin reclasificarla, y la edición fija del curso mantiene su entorno y validación independientes.

puremacro 6.3.0: research APIs and second-order HANK

Choose a tag to compare

@jalonso1979 jalonso1979 released this 10 Nov 20:08

puremacro 6.3.0 adds research APIs for solving equilibria, accounting for simulation uncertainty, and auditing policy models. It merges #232 and #233 while preserving existing entry points and their approximation orders.

  • Equilibrium derivatives, reusable dense factorizations, and pseudo-arclength continuation with explicit fold and failure diagnostics.
  • Reproducible simulation-aware SMM, separate sampling/simulation uncertainty, and identification-robust tests and confidence sets within documented model classes.
  • Bounded research models for trade reallocation, heterogeneous expectations, sovereign banking, and sovereign adaptation.
  • Real-Schur HANK solution and experimental second-order full-basis one-asset policies, anticipated-risk terms, pruned simulation, and independent nonlinear audits.
  • Shared validation and ownership contracts, immutable retained arrays, public API snapshots, and executable English/Spanish guides.
  • The bilingual NUTS example now uses four chains and 1,000 retained draws per chain, with the same convergence and divergence checks.

Install with python -m pip install puremacro==6.3.0 on Python 3.11 or later. Read the research API guide, second-order HANK guide, and full changelog.

The new economic models remain experimental. The second-order HANK benchmark accepts 22 of 24 signed-amplitude cases: the 20-asset negative shocks -0.001 and -0.0005 fail the unchanged distribution audit. Eight of ten finite nonlinear reference solves meet their tolerance; two coarse-grid failures remain recorded. This evidence does not establish household-grid convergence, infinite-horizon/global accuracy, welfare, or reduced-basis/two-asset/capital support. Inspect each returned audit before interpreting a path. The model modules have separate economic closures; combining their outputs requires an explicitly specified joint model.

Validation on bf3e2d320e024e832b11a4b32b8d649c78a1e481: 23,902 local tests passed, 91 skipped, and all seven local release gates passed. The main CI matrix and tag release workflow passed, including nine OS/Python targets, typing/docs, and installed-wheel examples. The opt-in and showcase run passed: 134 opt-in tests, 19 skipped, plus the complete showcase notebook execution. The real Pyodide kernel passed 120 smoke tests. Installed-package checks covered eight new research CLI entry points, four numerical research examples, the second-order HANK guide, and four base workflows. These cohorts overlap. Downloaded PyPI artifact hashes and all 1,051 packaged files were verified against the tagged source, followed by a fresh PyPI installation and executed examples. See the attached release-validation.json for the artifact hashes and scope.

The examples gallery retains the documented Romer–Romer (2010) published t-statistic rounding mismatch; it is not relabeled as passing. Timing checks use the repository's CI scale of 10 and do not certify unscaled wall-clock deadlines.

The pinned course edition keeps its own package version and validation record. This library release does not upgrade or requalify those course environments.

Español. puremacro 6.3.0 incorpora derivadas y continuación de equilibrios, SMM con incertidumbre de simulación, inferencia robusta a identificación y modelos acotados de reasignación comercial, expectativas heterogéneas, banca soberana y adaptación. Incluye HANK experimental de segundo orden con un activo y bases completas, simulación podada y auditorías no lineales. Las API existentes conservan su orden de aproximación. Consulte las API de investigación y la guía HANK.

Los límites numéricos se conservan: 22/24 casos de amplitud HANK superan la auditoría y 8/10 referencias no lineales convergen a la tolerancia declarada. Los casos rechazados no se reclasifican como éxitos. No se certifican convergencia de malla, exactitud global o de horizonte infinito, bienestar ni modelos de dos activos. La edición fija del curso mantiene su entorno y sus comprobaciones independientes.

puremacro 6.2.0

Choose a tag to compare

@jalonso1979 jalonso1979 released this 06 Nov 04:20

Fiscal rules & debt-sustainability analysis, state-dependent LP & weak-IV HAC diagnostics, sequence-space optimal policy, SVAR identification, GIV, and investment-research workflows.

Added

  • Experimental investment-research kernels in asset_pricing: lagged,
    self-financing long-only backtests with cash, costs, eligibility and JSON replay;
    Ledoit–Wolf covariance, bounded mean/minimum-variance allocation and frontier;
    causal technical indicators, momentum policies and a block-bootstrap max test;
    availability-dated fundamentals, financial ratios, FCFF/FCFE/dividend valuation
    and sensitivity; Brock–Hommes affine-belief dynamics with explicit fitness timing.
    No new runtime dependencies or bundled market data. Bilingual guides and
    executed notebooks document numerical checks, timing and unsupported extensions.
  • cumulative=True on lp_iv and lp_state_dep_iv estimates integral
    multipliers directly: the outcome and the endogenous regressor become
    window sums over 0..h, the instrument stays dated at t. Results carry
    the flag is_cumulative and "(cumulative)" method labels. Window sums need
    a regular index, so an irregular date or period index raises an error.
    (#217)
  • python -m puremacro.examples.first_brief_fiscal [--lang es] writes a
    bilingual fiscal policy brief on the Ramey-Zubairy historical data, with
    the pages docs/first_brief_fiscal.md and docs/es/first_brief_fiscal.md.
    It reports linear and slack/no-slack cumulative spending multipliers at 2
    and 4 years, quotes weak cells by Anderson-Rubin sets, runs the
    experiment without WWII under the RZ rule and reports a concentration
    diagnostic. The page numbers come from a frozen record, refreshed with
    tools/reference_validation/refresh_fiscal_brief.py and checked by the tests.
  • RZ_CONTROLS is public in the Ramey-Zubairy example; _CTL stays as an
    alias.
  • olea_pflueger_f, anderson_rubin_test, and msw_bands in
    puremacro.inference.weak_iv accept lags: int | None = None for
    Bartlett-HAC long-run covariance estimation (matching lp_iv when
    lags=h+1), and olea_pflueger_f accepts controls so raw x and z can
    be residualized inside the helper. (#217)
  • instrument_influence(z, controls=None, weights=None) in
    puremacro.inference.weak_iv (also re-exported from puremacro.inference)
    computes per-observation shares of residualized instrument variation
    $\tilde{z}_t^2 / \sum_s \tilde{z}_s^2$, effective sample size
    $1 / \sum_t w_t^2$, and max share / top indices to flag outlier-driven first
    stages. (#217)
  • lp_state_dep, lp_state_dep_iv, and lp_garch_state accept
    state_lag: int = 0 to shift the regime state column by state_lag periods
    before evaluating transition weights $F(s_{t-\text{state_lag}})$.
    lp_state_dep_iv also reports Sanderson–Windmeijer (2016) conditional
    first-stage F statistics (cond_f_H, cond_f_L) and uses them for
    first_stage_f_H / first_stage_f_L when transition="logistic".
    (#217)
  • Archived Arellano (2008) calibration: grid convergence established.
    check_arellano_grids --calibration archive keeps the author's exclusion
    timing, Fortran default-real literals and midpoint-refined 200-node grid.
    Under the public ladder's pre-registered tolerances and precision rule, 797 to
    6,369 debt nodes and 321/641 income states pass both final refinements. The
    standard design left one bound unclassified (debt/output, income direction);
    it is kept in the record. On converged grids spread volatility is about
    2.9 pp against the published 6.36, and the trade-balance–spread correlation
    about 0.82 against 0.43. Spread volatility still falls along the income
    direction: extrapolation places its limit 0.07–0.14 pp lower. Nothing is
    recalibrated. Record: tests/fixtures/arellano_2008/archive_grid_convergence.json;
    guide.
  • puremacro.sovereign.FiscalRule and solve_fiscal_sovereign_default(rule=...):
    debt ceilings, deficit limits, escape clauses and soft (penalty) rules,
    priced by lenders. Ceilings, escapes and penalties use the certified
    monotone search; deficit limits use the new policy_search="exhaustive"
    (or "dense"). States a hard rule leaves without an admissible choice raise
    ValueError naming them, or default for sure with infeasible="default".
  • Predictions about the rules were registered before any rule solve
    (tests/fixtures/sovereign_fiscal/rule_predictions_protocol.json) and
    measured on a coarse 401 x 21 grid by
    tools/reference_validation/check_fiscal_rule_predictions.py; the guide
    reports each verdict, including the one that did not hold.
  • Debt-sustainability analysis (puremacro.sovereign):
    simulate_stochastic_dsa (reduced-form debt fan charts with a
    FiscalReactionFunction), estimate_fiscal_reaction_function and
    fit_stochastic_dsa_from_data (Bohn-style reaction estimates feeding the
    simulation), and simulate_fiscal_dsa / compare_fiscal_rules_dsa
    (transitions of solved fiscal models from high initial debt, with common
    random numbers across rules). StochasticDSAResult and
    FiscalReactionEstimateResult follow the result-object conventions. The
    structural trajectories are illustrative 401 × 21 calculations, not
    grid-converged.
  • calibrate_fiscal_sovereign_smm (SovereignSMMResult): simulated method of
    moments for the fiscal sovereign model through structural.fit_structural,
    with structural.ags_sensitivity (AGSSensitivityResult) for
    Andrews–Gentzkow–Shapiro sensitivity. Local inference is withheld, with a
    warning, when the moment Jacobian is rank-deficient or a parameter sits on a
    bound.
  • Sequence-space policy (puremacro.dsge): ssj_ramsey, ssj_osr and
    ssj_welfare_decomposition (linear-quadratic Ramsey and optimal simple rules
    on supplied sequence-space Jacobians); solve_risky_steady_state and
    solve_precautionary_obc (Coeurdacier–Rey–Winant risky steady state and a
    precautionary occasionally binding constraint).
  • Policy counterfactuals (puremacro.lp): mckay_wolf_counterfactual and
    barnichon_mesters_opp, from estimated policy-shock impulse responses.
  • SVAR identification (puremacro.var): svar_heteroskedastic,
    svar_non_gaussian, svar_proxy and svar_sign_narrative (SVARResult),
    and estimate_functional_var (FunctionalVARResult) for yield curves and
    distributions.
  • puremacro.inference.estimate_giv (GIVResult): granular instrumental
    variables.
  • puremacro.portfolio: covariance shrinkage, mean-variance and target-return
    optimization, efficient frontiers, hierarchical risk parity, Black–Litterman,
    macro-quadrant classification and an all-weather allocation, factor
    attribution, rebalancing backtests with transaction costs and Monte Carlo
    wealth glidepaths.
  • Bilingual guides for these additions and showcase notebook 76
    (notebooks/76_bilingual_macro_showcase.py and _es.py).

Changed

  • lp_state_dep and lp_state_dep_iv take interact={"all", "intercept", "none"}, default "all": the constant and every control (lags and
    contemporaneous controls) are interacted with the regime weight, as in
    Ramey–Zubairy (2018, eq. 2) and the direct projections of
    Auerbach–Gorodnichenko (2013). With a threshold transition the point
    estimates equal one regression per regime. Behaviour change: 6.1.0 and
    earlier pooled the intercept and the control coefficients across regimes;
    interact="none" reproduces those results (relative difference below
    1e-9 on the bundled Ramey–Zubairy snapshot). lp_garch_state inherits the
    new default. (#217)
  • Both functions also return diff_HL, se_diff_HL and p_diff_HL (HAC test
    of β_H = β_L) and the effective regime sizes n_eff_H, n_eff_L.
  • The Ramey-Zubairy example: the headline is now at h = 7 and 15 (the 8- and
    16-quarter integrals), which reproduce Table 1. The state layer is one-step
    lp_state_dep_iv, with standard errors. Weak-IV diagnostics use the HAC
    effective F and the HAC Anderson-Rubin set, so the example no longer uses
    the non-HAC helpers of inference.weak_iv. Teaching assertions that did not hold under HAC statistics were
    replaced: the slack F does not collapse at long horizons (55.9 at h = 20),
    and the robust set is wider than the conventional band in the no-slack
    regime (1.28 times at h = 7), not in the slack regime (1.02). The figure is
    regenerated. The linear and no-slack headlines, whose instrument is weak
    (effective F 16.0 / 12.0 and 7.3 / 10.8), are reported with their
    Anderson-Rubin 90% sets. Compatibility: scripts that read
    run_rz_replication()'s output must be updated. The robust dict's keys
    are now h, regime, beta, se, conv_lo, conv_hi, ar_lo, ar_hi,
    ar_set_type and f_eff; msw_lo, msw_hi, beta_2sls, se_2sls,
    n_slack, ar_at_point and ar_at_zero are gone. ar_horizon now
    defaults to 7, in the no-slack regime (it was 12, in the slack regime).
    linear is now an LPResult indexed by h. New keys: state_dep (the
    lp_state_dep_iv result) and expansion_iv (lp_iv on the no-slack
    quarters, with Anderson-Rubin sets).

Fixed

  • Pooled controls biased the regime coefficients. On the bundled
    Ramey–Zubairy snapshot, lp_state_dep_iv erased the slack/no-slack gap once
    World War II is excluded (0.70/0.69 against 1.07/0.49 with everything
    interacted) and returned explosive estimates on the post-1947 sample
    (+2538 at 16 quarters).
  • lp_state_dep_iv and lp_smooth_transition_irf no longer swallow estimation
    errors: a horizon that cannot be estimated emits a RuntimeWarning naming
    the horizon...
Read more

puremacro 6.1.0 — sovereign grid convergence and certified monotone search

Choose a tag to compare

@jalonso1979 jalonso1979 released this 03 Nov 18:17

6.1.0 (2026-11-03)

Reliability & Scale Release: Sovereign grid convergence (25,601x1,281), certified fast Arellano monotone solver, HANK with productive capital & matrix-free Newton-GMRES, DSGE solution workbench, and household trade adjustment.

This release establishes full sovereign grid convergence through 25,601 × 1,281
uniform nodes and curved grid refinements, provides an exhaustively certified
monotone search solver for Arellano models, introduces the immutable DSGE
solution workbench with .mod replay, adds matrix-free Newton-GMRES and
productive capital to two-asset HANK, and delivers empirical household trade
incidence analysis on Mexican ENIGH 2022 microdata.

Added

  • solve_sovereign_default(debt_grid=...) accepts explicit nonuniform debt
    coordinates with an exact zero. Default-grid equilibria and random paths
    retain pinned legacy parity; supplied coordinates survive maturity
    restructuring, and their price tangents use nonuniform PCHIP weights.
    A separate fixed-support curved-grid experiment evaluates the high-reentry
    brief scenario with unchanged simulation design and moment tolerances.

  • Bounded the short-term monotone policy search's workspace by batching
    independent income columns. Values, choices, equilibrium metadata and random
    paths retain exact parity with the pinned original algorithm. Extended
    sovereign grid convergence studies through 25,601 × 1,281 uniform nodes and
    curved grid refinements (tests/fixtures/sovereign_brief/sensitive_grid_convergence.json
    and curved_grid_convergence.json), retaining all failed crossed refinements
    through 12,801 × 1,281 without relaxing moment tolerances or rewriting
    historical source provenance.

  • DSGE workbench with immutable experiments, explicit .mod replay, Python/HTML
    export, unit/timing-aware comparisons and separate accuracy evidence. The IRF
    widget clears stale curves and exports after failed or unapplied edits, reports
    actionable failures and supports batched changes with Apply mode. Deterministic
    Krylov runs expose cooperative cancellation and accepted-path recovery.
    Guide.

  • Reduced household marginal-value controls join distribution compression, so
    Jacobians and Schur systems are built directly in reduced coordinates.
    Full-grid audits expose omitted equations and insufficient ranks. Conditional
    nonlinear corrections re-solve after each observed shock and separately
    check horizon and branching refinements. Productive-capital two-asset HANK
    adds investment, installation costs and endogenous returns with independent
    market/resource accounts and an author-code capital-block reference. These
    remain local/finite-tree research solvers, with household grid errors reported
    separately. Reduction,
    continuation,
    capital.

  • Solved local stochastic HANK policies with constraint- and market-moment-
    preserving distribution reduction, Schur determinacy checks and held-out
    full-distribution audits. Nonlinear finite trees can use that solved
    first-order stochastic terminal policy. A two-asset consol economy clears
    both asset markets and accounts for coupons, capital gains and portfolio
    adjustment resources. Failed compression audits remain visible; these are
    local solutions, not global or second-order aggregate-risk calculations.
    Guide.

  • HANKEquilibriumWorkspace and solve_hank_equilibrium: matrix-free
    Newton-GMRES solves the coupled one-asset finite-history household and market
    equations with analytic EGM/distribution tangents, a reusable sparse
    preconditioner and explicit warm starts. Numerical failures retain the last
    feasible iterate; the dense reference remains unchanged. This extends the
    computational backend, not the model's terminal or stochastic closure.
    Guide.

  • Observed ENIGH 2022 household policy studies retain source hashes, survey
    weights, zero-spending exclusions and household-level welfare. Explicit
    budget-neutral compensation supports signed per-household transfers.
    Robust policy comparisons report dominance and minimax regret over complete
    assumption grids, with no implicit probabilities or failed-cell deletion.

  • Conditional dynamic CES sourcing accounts for procurement and quadratic
    adjustment resources. Finite household consumption paths report CES/CRRA
    consumption equivalents under explicit financing. These are conditional
    incidence experiments, not national dynamic GE or causal tariff estimates.

  • Finite-tree aggregate-risk HANK solves household choices before future shocks
    are revealed, with one distribution per history and explicit firm valuation
    and terminal closures. Deterministic paths reduce to the existing published
    one-asset model; this does not certify stationary stochastic HANK.

  • calibrate_one_asset_hank and solve_one_asset_hank: the published ABRS
    one-asset equilibrium with endogenous labor, debt revaluation, asset/labor
    clearing and nonlinear Phillips curve. Independent author-code fixtures
    check the published 500×7 calibration and both shock signs. A fixed-tolerance
    study extends income refinement to 31 states and checks assets and horizons;
    failed intermediate refinements remain visible. The existing reduced HANK
    API keeps its original closure. Guide.

  • Added puremacro.inference.gmm, a public generic GMM engine: user moment
    functions, two-step, iterated, identity or user weighting, Bartlett HAC
    long-run covariance (package default bandwidth), analytic or numerical
    Jacobians, sandwich standard errors, Hansen J and Wald tests. Returns
    GMMFitResult. Linear IV estimates, standard errors and J match
    linearmodels IV2SLS/IVGMM and inference.hansen_j to 1e-8.
    Optimizer stopping is invariant to common changes of moment units, and a
    rank check rejects unidentified parameter combinations before inference.

  • Added puremacro.asset_pricing (macro-finance): to_returns,
    excess_returns, compound and annualization helpers; risk_summary
    (Sharpe ratio with a Lo 2002 HAC standard error, Sortino, moments,
    drawdowns); value_at_risk / expected_shortfall (historical, normal,
    filtered GARCH) and var_backtest (Kupiec, Christoffersen);
    time_series_test (GRS and HAC Wald); fama_macbeth (Fama-MacBeth,
    Shanken and Newey-West se); sdf_gmm and hj_distance (Jagannathan-Wang
    p-value); ccapm (Hansen-Singleton) with FRED helpers
    fred_consumption_growth and fred_risk_free. Inputs are refused when they
    contain NaN/inf, look percent-scaled or mix frequencies. Estimates match
    statsmodels and linearmodels to 1e-8 where definitions coincide; the
    remaining differences are documented. Multi-period historical VaR/ES compound
    simple returns; SDF/HJ reject rank-deficient systems before inference. Bilingual guide docs/asset_pricing.md.

  • Added puremacro.fetch.kenfrench_factors.fetch for the Fama-French three-
    and five-factor, momentum and 25 size/book-to-market files. Data are
    downloaded into the user cache and never bundled (registry entry
    kenfrench.factors, not redistributable). A network test reproduces the
    Fama and French (2015) GRS of the three-factor model on the 25 portfolios
    (3.57 against the published 3.62).

  • puremacro.sovereign.solve_fiscal_sovereign_default: a fiscal block on the
    short-term Arellano solver. A fixed tax share tau funds public spending
    g = tau*y - b + q(b', y) b', valued as psi * v(g) next to private
    consumption (1 - tau) y. FiscalSovereignResult adds spending,
    primary-balance and debt-to-output policies, private consumption, and a
    simulate() built on a vectorized path engine with common random numbers.
    The certified monotone search is the default. With tau=1 and private
    consumption off the solver is bit-identical to solve_sovereign_default.
    tau and psi have no defaults; tau=0.2, psi=1 is an illustrative
    calibration, not fitted to any target. Guide:
    fiscal block.

  • check_arellano_grids.paired_comparison accepts a tolerances mapping;
    its default is unchanged.

  • A bilingual sovereign-risk policy brief,
    docs/brief_sovereign_risk.md (Spanish:
    docs/es/brief_sovereign_risk.md). It shows how the market-reentry
    probability, income volatility and income persistence move average debt,
    spreads, default frequency and the value of market access in the validated
    short-term Arellano model. These are illustrative comparative statics of the
    public calibration, not a claim about any country. Every number is shown with
    its 95% interval across 16 seeds and its paired 3,201 -> 6,401 grid change.
    The original frozen sweep, tests/fixtures/sovereign_brief/sweep.json,
    retains two scenarios that fail its grid criterion; the brief reports their
    further refinements separately. Quarterly-income debt ratios divide by four
    when expressed relative to approximate annual GDP.

  • puremacro.sovereign.ergodic_moments(result, *, periods=500_000, seeds=range(1979, 1995), burn=1_000): raw ergodic debt (% of mean quarterly
    income), mean spread,
    annual default-entry frequency and spread SD, per seed and with 95%
    t-intervals. It reproduces SovereignDefaultResult.simulate draw for draw
    and runs 16 x 500,000 quarters in seconds. Continuing exclusion is not counted
    as default.

  • puremacro.sovereign.value_of_market_access(result): the
    consumption-equivalent value (in %) of market access at zero debt, compared
    with permanent autarky without a default cost.

  • Example python -m puremacro.examples.first_brief_sovereign:
    --scenario theta=0.15 (one scenario on the baseline's 6,401 x 641 grid),
    --sweep --coarse (all seven scenarios at 801 x 161 in minutes, with a
    coarse-grid caveat)...

Read more

Puremacro 5.7.1 — nonlinear DSGE workflows and bilingual solver guides

Choose a tag to compare

@jalonso1979 jalonso1979 released this 24 Oct 04:06

Puremacro 5.7.1

Puremacro 5.7 lets a supported .mod file or Python equilibrium callback define
one nonlinear model for steady states, local stochastic approximations,
deterministic transitions and explicitly mapped global policy problems.
Loading no longer requires a local solution: read_mod compiles the model;
the caller chooses the solution method.

The 5.7.0 tag was not published to PyPI: its release gate exposed a course-test
bootstrap mismatch. This patch fixes the test environment selection, preserves
the frozen 5.5.0 course edition, and refreshes the eight updated notebook pins.
The numerical solver implementation is unchanged from the 5.7.0 source tag.

Models, transitions and policy accuracy

  • Ordered deterministic .mod execution preserves initialization, history,
    terminal values, steady-state solves, dated shocks and repeated setup/solver
    snapshots. Unsupported commands and options fail explicitly.
  • Nonlinear transitions add adaptive continuation, economic domains and scaling,
    analytic derivatives, explicit announcement dates, terminal policies and
    complementarity mappings, with finite solve and horizon budgets.
  • Global time iteration supports explicit state/control mappings, Chebyshev,
    Smolyak and linear interpolation, quadrature and constrained policies.
    Independent off-grid Euler/KKT, feasibility and domain-exit audits allow
    comparisons with perturbation orders 1–3 and controlled policy refinement.
  • The one-asset EGM compiler now validates its affine budget contract
    structurally; unsupported nonlinear budgets cannot pass just because they
    happen to agree at grid nodes.

This release also includes the 5.6.0 development work, which was not published
separately to PyPI: DP .mod files, reproducible study archives, editable HANK aggregate
equations, rank-normalized MCMC diagnostics, additional result audits and
explicit experimental dependence-bootstrap references. See the
complete changelog
for their assumptions and limits.

Documentation and MAV notebooks

Four English/Spanish guide pairs explain mathematical timing and array shapes,
method selection, compatibility, diagnostics and reproducibility, with 29
Python examples per language:

T02_A–D in both languages now pin code, helpers and models to v5.7.1.
All eight completed a fresh local execution: 120 code cells and 72 figures.
The stochastic Euler equations retain future marginal utility inside the
expectation. The original course-2026A-5.5.0 edition and its evidence remain
unchanged. Follow the notebook setup
and execution record.

Scope

Dynare compatibility is a documented subset. Global policies require explicit
states, controls and domains. Passing finite residual/refinement checks does
not establish global existence or uniqueness. Krylov currently rejects MCP and
terminal-policy callbacks. Course evidence covers local Python; the Pyodide
smoke tests do not certify every browser, Colab or Juno notebook session.

Known higher-order accuracy issue

The existing higher-order Sylvester solver rounds its LU-cache keys to 12
decimal places. Distinct Schur products can therefore reuse a factorization.
In a determinate test with roots near 0.9999998, the second- and third-order
policy coefficients differ from their analytic values by up to approximately
3.8e-7 in relative terms. The existing near-unit-root warning does not quantify
this coefficient error. The attached near_unit_sylvester_reproducer.py uses
the public model API and an analytic reference. This predates the 5.7 changes;
exact cache keys and independent operator checks are targeted for 5.7.2, not
included here. There is no general drop-in public-API workaround: validate
affected coefficients against an independent analytic/dense reference rather
than assuming full floating-point precision near unit roots. This example is
not evidence of a general unit-invariance failure.

The library supports Python 3.11+; the MAV notebooks use Python 3.12+.

python -m pip install "puremacro==5.7.1"

Verification of the published release

  • The exact-tag release workflow
    passed all nine Linux/macOS/Windows and Python 3.11–3.13 combinations, the
    opt-in suite (130 passed, 19 skipped), artifact checks and trusted PyPI publishing.
  • All six local release gates passed: 21,255 tests passed, 85 skipped and
    223 deselected. The stored gallery retained its documented Romer–Romer 2010
    rounding difference; there were no new gallery failures.
  • The actual PyPI wheel and source archive were downloaded and checked against
    PyPI's SHA-256 values. An isolated PyPI installation matched all 871 runtime
    source files
    in the accepted notebook record and executed seven examples
    from the nonlinear/global guides.
  • The clean local wheel passed 61 Pyodide 0.28.3 smoke tests. This is separate
    from the eight local notebook runs and does not certify full browser sessions.

The attached puremacro-5.7.1-validation.json records the commit, workflow,
artifact hashes and validation scope. The release's wheel and source archive
assets are the distributions downloaded from PyPI, not rebuilt substitutes.

Puremacro 5.5.0 — model inspection, result audits, and verified course edition

Choose a tag to compare

@jalonso1979 jalonso1979 released this 23 Oct 01:43
367d956

Puremacro 5.5.0 adds model compatibility inspection and numerical result audits, corrects steady-state trajectory references, and consolidates the reviewed development backlog.

  • Inspect a .mod file with puremacro.dsge.inspect_mod or puremacro-dynare inspect before solving it. Reports preserve source locations, macro/include provenance and unsupported commands.
  • Opt into ordered execution for the supported fixed-calibration command subset. Repeated simulations, smoothers and decompositions retain separate options and results; unsupported programs fail preflight. The legacy runner remains the default.
  • Audit supported household, first-order DSGE and dynamic-trade results with puremacro.validation.audit_result. Convergence, equations, Euler/KKT residuals, distribution checks and explicit refinement experiments are separate evidence. Unsupported or unchecked cases are named explicitly.
  • Correct STEADY_STATE() references in nonlinear trajectories and derivatives, retain literal terminal assignments unless a terminal equilibrium is requested, and reject exogenous variables inside the operator. Parsed initial/terminal declarations now agree across model-loading, builder and command-dispatch routes; explicit Python overrides retain precedence. Re-run affected nonlinear calculations; see the correctness advisory.
  • Recognize supported two-asset household budgets structurally. Safely migrate flat caches, improve quarterly projection and repeated bootstrap/keyword work, preserve seeded results, and improve keyboard and CSV accessibility.

The release includes the features previously shipped in 5.4.2, including supported two-asset DP bridges and analytic DSGE likelihood scores. Tags 5.3.0, 5.4.0 and 5.4.1 did not publish packages to PyPI; their status is documented in the changelog.

Compatibility inspection is not a numerical certificate. Audit tolerances apply to individual checks; solver convergence does not establish grid accuracy or economic validity. Ordered execution and household adapters have explicit supported scopes. The documented RR2010 published-statistic discrepancy remains unchanged.

Complete changelog · Inspection guide · Result audits

Validation at source 2db695c passed 20,672 local baseline tests and all six release gates, Python 3.11 and Windows-target typing, strict documentation, clean package checks, and all 37 real-Pyodide tests. The examples gate found 84 passing examples, 8 intentional skips, and the unchanged documented RR2010 mismatch. The 72 course notebooks executed 1,020 code cells against candidate wheel SHA-256 148a8892070509773c930efb9b18be6ca9795a2e39215bf096778b9b50676576; all 544 frozen inputs were preserved, required DP/text sections ran, and no network attempts or language-model cache misses occurred. All 49 integrated course regression tests also passed.

Get the bilingual 2026A course edition, which pins puremacro==5.5.0 and requires Python 3.12 or later. Its notebooks and execution record are versioned in the separate course-2026A-5.5.0 tag; the library itself continues to support Python 3.11+.

The exact-tag release workflow passed all nine platform targets, the extended suite (129 passed, 19 skipped), package checks, and trusted PyPI publishing. The course CI also passed all nine targets; its merged tree is identical to the validated revision.

The official PyPI 5.5.0 wheel has SHA-256 7b88efd5bac614c2640c031ecbc2742a7247cf39935d0e5c54c048b3629d78cb. A fresh no-cache installation from official PyPI verified all 912 installed runtime files against the tagged source, two ordered invocations, equation/Jacobian audits, and CLI inspection. The attached content-equivalence evidence records complete runtime-file and dependency-metadata equality between the public wheel and the candidate used for the course execution. The archives have different ZIP hashes; this is content equivalence, not a claim that the public archive was separately used to rerun the notebooks.

puremacro 5.2.0

Choose a tag to compare

@jalonso1979 jalonso1979 released this 17 Oct 01:16

puremacro.trade is back. 5.0 and 5.1 shipped without it; this release brings the multi-country, multi-sector trade library back with consistent producer/purchaser accounting as its only accounting, a new bundled OECD table built by a script that ships with it, and no built-in defaults for the table: every dimension, country and sector comes from the table or from an argument. Code written for the 4.x trade API needs the changes listed under "Removed relative to 4.8.1" and "Changed relative to 4.8.1".

Added

  • puremacro.trade with the static model (calibrate_trade_model, solve_trade_equilibrium, solve_equilibrium, TariffScenario, run_scenario, run_scenario_batch, Geary-Khamis PPP), Hicksian welfare (compute_hicksian_welfare), audited policy recovery (solve_policy_equilibrium), optimal tariffs and Nash tariff games (optimal_tariffs, game), Caliendo-Parro (2015) exact hat algebra (CaliendoParroModel), the flexible CGE (solve_flexible_trade_equilibrium: nested CES technology, Stone-Geary demand, Armington sourcing, Atkeson-Burstein markups, capacity margins), policy analytics, the exact nested-CES block Newton (ces_newton), household demand systems and distributional incidence (household, distributional), audited parameter continuation (continuation), the reduced local-stability diagnostic (stability, experimental), native MRIO tables (mrio), the condensed one-factor Leontief tariff model (condensed), the perfect-foresight dynamic MRIO (dynamic, experimental), plots and the optional torch/MLX solvers (gpu). models.TradePolicySimulator returns with it.
  • A bundled table, puremacro/trade/_datafiles/icio_oecd2023_2019_isic_g20.npz: the OECD Inter-Country Input-Output tables, 2023 edition, reference year 2019, summed into the 19 G20 member countries plus a rest of the world and the 20 sections of ISIC Rev.4 (shape 403 x 460). load_icio_data() reads it by default (source="oecd_isic") and labels it with the table's own codes. puremacro/trade/_datafiles/SOURCES.md gives the OECD citation, the terms of use, the concordance and the checksums of the source and of the bundled file. If you publish results computed from it, cite the OECD.
  • tools/build_icio_isic.py rebuilds that table byte for byte from the OECD file; --download fetches the OECD's 2016-2020_SML.zip (about 85 MB) once into ~/.cache/puremacro/oecd_icio (or $PUREMACRO_OECD_ICIO_DIR). tools/reference_validation/validate_oecd.py takes the same option.
  • puremacro.trade.data.synthetic_icio_table(nc, ns, *, nfd=3, seed=0) builds a seeded synthetic table of any size in the library's layout, with nonnegative final-use shares and positive investment for every seed, and synthetic_codes(nc, ns) gives its labels (C01..., S01...). The tests run on it instead of a real table.
  • Docs (EN/ES): trade_accounting, trade_welfare, trade_policy, trade_continuation, trade_stability, trade_household, trade_distributional, trade_ces_newton, trade_condensed, trade_mrio, trade_dynamic, trade_gpu; the trade sections of spatial_and_trade_ge, policy_simulators, STRUCTURAL_VALIDATION_STATUS and ADVISORY. Notebooks 55 and 61-65 (EN/ES), the example puremacro/examples/distributional_trade_ge.py, the OECD/accounting/welfare/Hicksian validators in tools/reference_validation/ and three trade cases in the validation gallery.

Removed relative to 4.8.1

  • The legacy accounting. accounting="legacy" (the 4.x default), tariff_revenue_mode="legacy_national" and the precedence and compatibility switches that went with it are gone from every entry point, gpu included; consistent accounting is the only one, so the same call can give different numbers than in 4.x. accounting="legacy" raises ValueError saying so; a removed keyword raises TypeError naming it.
  • Named scenarios and the scenario registry (SCENARIOS, the canonical-scenario constants and lookup functions, ExtendedTariffScenario, run_tariff_scenario). run_scenario, solve_keller_pac and policy_analytics take a TariffScenario or explicit tariff arrays; a string raises TypeError("named scenarios were removed; build a TariffScenario").
  • The 4.x bundled table, the bundled reference solutions and results workbook, their loaders and constants (the reference-solution and workbook loaders, available_reference_scenarios, bundled_icio_path, bundled_reference_path, bundled_workbook_path, load_raw_45sector_icio, BUNDLED_ICIO_PROVENANCE, OECD2020_ICIO_PROVENANCE), and that table's code lists (CANONICAL_COUNTRY_CODES, CANONICAL_SECTOR_CODES, CANONICAL_SECTOR_NAMES, get_country_codes, get_sector_codes, get_sector_names). load_icio_data(source="legacy") raises ValueError naming the bundled source and synthetic_icio_table.
  • The real-world tariff benchmark (and its re-export in game), the analysis stubs and result classes built on the named scenarios (among them EVDecompositionResult), the two ScenarioBatchResult table renderers for the named scenarios, plot.plot_country_impacts and plot.plot_tariff_escalation_curve, the "harberger" welfare metric (it now gives the "Unknown welfare metric" ValueError), and flexible.compute_homothetic_final_demand.
  • In mrio and condensed: the coarse tariff rules (coarse_tariff_rates, CoarseTariffResult), the sector concordance of the 4.x table and its code and name constants, the numbered model labels, condensed.measures.aggregation_gaps and the w_<code> alias.
  • puremacro.trade.extensions (scenario extensions built on the named scenarios) and puremacro.trade.tables (LaTeX tables of the named scenarios) do not return.

Changed relative to 4.8.1

  • No paper defaults. Dimensions are read from the table (any number of countries, sectors and final-use categories); the fiscal-recycling and capacity branches of solve_trade_equilibrium and solve_flexible_trade_equilibrium need their country and sector named and raise TypeError naming the missing argument; compute_unilateral_optimal_tariff and compute_welfare_payoff_matrix take their countries as required arguments, and solve_multilateral_nash_tariffs and evaluate_national_welfare raise TypeError without them; the Geary-Khamis numeraire is a country argument, and GearyKhamisResult's summary has no fixed-country PPP row. ICIOData needs its country and sector codes.
  • Renamed without aliases: the fold helpers are solve_fold_continuation and solve_fold_continuation_step (they take the stiff coordinate as an argument; the info key is idx_stiff), and CondensedMeasuresResult.headline() is key_measures().
  • solve_flexible_trade_equilibrium is rebuilt on the consistent accounting; at default settings its equations are those of solve_trade_equilibrium. method="quasi_condensed" on solve_trade_equilibrium raises NotImplementedError pointing to the flexible solver.
  • gpu: BatchedJacobianEvaluator(calib, tau_a, taufd_a, ...) evaluates the consistent-accounting residuals and no longer takes the national tariff-rate vectors; solve_homotopy_continuation has no national-rate targets.
  • get_country_code raises IndexError for a position outside the table instead of inventing a code.
  • solve_fold_continuation_step raises ValueError for an idx_stiff outside [0, nc); it used to warn and use country 0.
  • The condensed path of solve_flexible_trade_equilibrium keeps iterations within max_iter: Newton stops when its line search finds no decrease (metadata["newton_stalled"]). Before, a solve that could not converge ran Newton up to 21 times with the full max_iter each and could report more iterations than max_iter.
  • In that condensed path, the tariff homotopy runs only on the iteration budget the direct attempt leaves unused; with fewer iterations left than homotopy steps it is not tried.
  • When that condensed solve fails, iterations counts the homotopy stages as well as the direct attempt (metadata["tariff_homotopy_steps"] is 0 when a homotopy was tried and did not converge).
  • TradePolicySimulator.from_icio(year, *, tables=None, ...) calibrates to the table of year (from tables, or the bundled 2019 table) and raises ValueError listing the available years for any other year.

Fixed

  • CaliendoParroModel mapped the codes USA, CHN and DEU (and ROW) to positions 0-2 whatever the model's own country order, and manuf/serv to sectors 0-1. Countries and sectors are now found by their own codes (exact, then case-insensitive) or by position; an unknown code raises KeyError.
  • TradePolicySimulator.from_icio(year) ignored year and returned the nafta_china preset.
  • condensed.table.infer_merchandise_mask counted utilities (ISIC D and E) as merchandise, and accepted composite sector labels whose content is not documented; those now need an explicit merchandise_mask.

puremacro 5.1.0

Choose a tag to compare

@jalonso1979 jalonso1979 released this 15 Oct 19:30

First 5.x final release. puremacro.dp simulates panels of individual histories and reports exact inequality statistics, and it is now the stable way to declare dynamic-programming models; vfi.Model is deprecated and will be removed in 6.0. It also carries everything in the 5.0 pre-releases below: the research material is no longer part of the package (4.8.1 is the last release that contains it), puremacro.llm, and the review-defect fixes.

Added

  • DPSolution.simulate(n_agents, periods, seed=..., init=..., burn_in=..., record=...) simulates individual histories from a puremacro.dp solution (discrete VFI and EGM, infinite horizon and life cycle, with or without a discrete choice and taste shocks) and returns a dp.Panel. One seeded PCG64 generator draws, each period, the discrete-choice uniforms, then the shock uniforms, then the survival uniforms, so the same seed gives a bit-identical panel. Infinite-horizon panels start in the stationary distribution by default; life-cycle panels start with newborns and apply the declared survival (panel.alive). n_agents * periods above 5e8 raises ValueError with the memory it would need. Continuous-time (HJB) and Krusell-Smith solutions raise NotImplementedError.
  • dp.Panel, a container for simulated histories: to_frame, to_microframe, cross_section, path, and mean, quantile, gini, top_share and summary pooled over periods or grouped by="t" or by="age". Dead agents keep their rows and are left out of every statistic.
  • Exact statistics on the solved distribution: DPSolution.gini, top_share, quantile and lorenz. The Gini is the standard weighted Gini on the Lorenz curve, 1 - sum_i w_i (L_i + L_{i-1}), the same function as MicroFrame.gini (negative values allowed, NaN when the total is not positive). In a life cycle, age=j uses one cohort and age="all" returns one value per age. Krusell-Smith solutions raise ValueError (they have no stationary distribution).
  • DPSolution.population_shares() gives each age's population share in a life cycle with survival applied. sol.distribution still gives every cohort mass 1; the pooled statistics weight cohorts by these shares.
  • docs/dp.md (EN/ES) has two new sections, "Panels and inequality" and "Porting from vfi.Model". dp moves from Experimental to Stable in ARCHITECTURE.md.

Fixed

  • vfi.simulate_panel raises TypeError for a policy that is not an integer array of grid indices. It used to cast a float policy (for example a continuous EGM policy in asset levels) with astype(np.intp), which silently truncated it to wrong indices. Use DPSolution.simulate for such solutions.

Deprecated

  • vfi.Model issues a DeprecationWarning when constructed and will be removed in 6.0. Declare the model with puremacro.dp; docs/dp.md, "Porting from vfi.Model", shows the translation, and the dp version of notebook 21's model picks the same policy on every grid node.

Changed

  • The deprecated puremacro.lp.garch_utils module and puremacro.sigma.sigma_numpy.SigmaObject announced removal in 5.0.0; they now name 6.0.0, together with vfi.Model, so a minor release removes nothing.
  • Notebook 21 (EN/ES) declares its precautionary-savings model with puremacro.dp instead of vfi.Model. It now prints mean assets 6.5733 and a wealth Gini of 0.3494, the exact values of the stationary distribution. It used to print 6.6767 and 0.3924, which came from a 2000-household, 100-period simulation; that Gini was computed after subtracting the minimum asset holding, so it was not the standard Gini. The notebook adds a simulated panel whose cross-sectional Gini converges to the exact value, and its text no longer claims vectorised Bellman updates or pluggable Numba/MLX/CuPy backends.