model.xml and LONGBASE database, April 2026model.xml and LONGBASE database, April 2026score_reform refuses model="frbus", as no PolicyEngine reform bridge existsDiagnose the latest model-data vintage and forecast GDP and inflation - with uncertainty ranges. Coefficients are estimated through 2023Q2; + with uncertainty ranges. Coefficients are estimated through 2025Q1; the model does not score reforms.
read the model → diff --git a/integration/src/policyengine_macro/capabilities.py b/integration/src/policyengine_macro/capabilities.py index 0d2abde..b5e9f33 100644 --- a/integration/src/policyengine_macro/capabilities.py +++ b/integration/src/policyengine_macro/capabilities.py @@ -76,9 +76,9 @@ "runtime": "minutes per estimation and identification run", "uncertainty": "posterior 68% and 90% intervals", "status": "validated replication for selected outputs", - "estimation_sample": "1992Q1-2023Q2", + "estimation_sample": "1992Q1-2025Q1", "data_edge": "2026Q1", - "data_vintage": "2026Q1 conditioning data; estimation ends 2023Q2", + "data_vintage": "2026Q1 conditioning data; estimation ends 2025Q1", }, "frb-us": { "display_name": "Federal Reserve FRB-US implementation", diff --git a/integration/src/policyengine_macro/core.py b/integration/src/policyengine_macro/core.py index 6b42b0d..d4888dd 100644 --- a/integration/src/policyengine_macro/core.py +++ b/integration/src/policyengine_macro/core.py @@ -725,17 +725,24 @@ def _covid_dummies(index) -> np.ndarray: # Estimation window for the hosted SVAR (start of Great-Moderation sample; # end held fixed so hosted results only move on a deliberate refresh). The # provenance strings are derived from the actual df_est index, never retyped. +# +# 2026-07-24 refresh: bumped 2023Q2 -> 2025Q1 (126 -> 133 observations) to match +# the sample boe-var-model's production replication now runs on. Held one +# quarter short of the 2026Q1 data edge, matching the repo, so the forecast +# origin sits outside the estimation sample. _SVAR_EST_START = "1992Q1" -_SVAR_EST_END = "2023Q2" - -# Default posterior draws for hosted calls. Measured on the 2026Q1 vintage -# (Apple Silicon, single process): draws=500 -> 42 accepted, ESS 15.5, ~23s; -# 1000 -> 76 accepted, ESS 26.5, ~44s; 2000 -> 165 accepted, ESS 63.6, -# ~117s. Acceptance is ~8% and ESS ~3% of draws, so 2000 is the smallest -# default that clears the 100-accepted-draws reliability threshold at -# roughly two minutes of first-call runtime (results are cached in-process). -# Full ESS >= 100 needs ~3500 draws (~3.5 min); a `warnings` entry flags the -# residual ESS shortfall honestly instead of hiding it. +_SVAR_EST_END = "2025Q1" + +# Default posterior draws for hosted calls. Re-measured on the 1992Q1-2025Q1 +# estimation sample (seed 0), since the longer sample tightens the posterior +# and lowers the sign-restriction acceptance rate: draws=500 -> 35 accepted, +# ESS 13.6; 1000 -> 69, ESS 31.5; 2000 -> 135, ESS 65.3; 3500 -> 233, +# ESS 101.4. Acceptance is ~6.8% and ESS ~3.3% of draws (was ~8% and ~3% on +# the old 2023Q2 sample), so 2000 remains the smallest default that clears +# the 100-accepted-draws reliability threshold -- 1000 would not (69). +# Full ESS >= 100 still needs ~3500 draws; a `warnings` entry flags the +# residual ESS shortfall honestly instead of hiding it. First-call runtime is +# a couple of minutes end to end and results are cached in-process. _SVAR_DEFAULT_DRAWS = 2000 @@ -2705,7 +2712,7 @@ def svar_summary() -> dict: "provenance": _provenance( model_id="boe-svar", distribution="boe-var-model", - data_vintage="estimation through 2023Q2; conditioning data through 2026Q1", + data_vintage="estimation through 2025Q1; conditioning data through 2026Q1", baseline="committed replication and forecast artefacts", ), } diff --git a/integration/src/policyengine_macro/mcp_server.py b/integration/src/policyengine_macro/mcp_server.py index b5001ac..553e005 100644 --- a/integration/src/policyengine_macro/mcp_server.py +++ b/integration/src/policyengine_macro/mcp_server.py @@ -290,7 +290,7 @@ def frbus_summary() -> dict: def forecast_uk(horizons: int = 12, draws: int = 2000) -> dict: """Forecast UK YoY GDP growth and CPI inflation with the UK SVAR model. - Estimates a Bayesian VAR (1992Q1-2023Q2 sample, sign-identified structural + Estimates a Bayesian VAR (1992Q1-2025Q1 sample, sign-identified structural shocks) and simulates the predictive distribution from the latest data quarter. Returns, per future quarter, the median and 68%/90% bands for YoY GDP growth and YoY CPI inflation, both in percent. The response includes a @@ -299,11 +299,11 @@ def forecast_uk(horizons: int = 12, draws: int = 2000) -> dict: Args: horizons: Forecast horizon in quarters (default 12 = 3 years). - draws: Posterior draws (default 2000: ~165 accepted draws, ESS ~64, - about two minutes of runtime on first call; ~3500 draws reaches - ESS >= 100 in ~3.5 minutes; 500 responds in ~25s but yields ESS - ~15 and a warning). Results are cached in-process, so repeated - calls with the same (horizons, draws) are instant. + draws: Posterior draws (default 2000: ~135 accepted draws, ESS ~65, + a couple of minutes on first call; ~3500 draws reaches ESS >= 100; + 500 is faster but yields ~35 accepted, ESS ~14, and a warning). + Results are cached in-process, so repeated calls with the same + (horizons, draws) are instant. """ return core.svar_forecast(horizons=horizons, draws=draws) diff --git a/integration/tests/test_site_contract.py b/integration/tests/test_site_contract.py index 5072d82..aebaf3a 100644 --- a/integration/tests/test_site_contract.py +++ b/integration/tests/test_site_contract.py @@ -14,10 +14,25 @@ def _read(path: str) -> str: def test_svar_site_uses_runtime_estimation_endpoint(): + """The published estimation endpoint must match the one the server actually + estimates on. + + Derived from ``_SVAR_EST_END`` rather than hardcoded: the point of this + guard is to catch the site drifting away from the runtime, and a literal + makes it fail on the refresh it is supposed to permit while still missing + drift in the other direction.""" + from policyengine_macro.core import _SVAR_EST_END + pages = _read("svar/index.html") + _read("docs/index.html") - assert "estimated through 2023Q2" in pages - assert "estimation sample to 2025Q1" not in pages - assert "Estimation to 2025Q1" not in pages + assert f"estimated through {_SVAR_EST_END}" in pages + + # No other quarter may be presented as the estimation endpoint. + stale = { + q + for q in re.findall(r"estimat\w*\s+(?:through|to)\s+(\d{4}Q[1-4])", pages) + if q != _SVAR_EST_END + } + assert not stale, f"site quotes stale estimation endpoint(s): {sorted(stale)}" def test_og_is_not_labelled_as_us_model(): diff --git a/models/index.html b/models/index.html index c08262a..ff33b42 100644 --- a/models/index.html +++ b/models/index.html @@ -52,7 +52,7 @@Each guide covers inputs, outputs, method, validation, limitations, and how to run the model.
Selected near-term scenarios. Trace GDP, consumption, and investment against the official forecast. Validated for selected shocks; borrowing is not yet returned.
open model → -UK conditions and forecasts. Identify shocks and forecast GDP and inflation with uncertainty ranges. Estimated through 2023Q2; conditioned through 2026Q1.
open model → +UK conditions and forecasts. Identify shocks and forecast GDP and inflation with uncertainty ranges. Estimated through 2025Q1; conditioned through 2026Q1.
open model →Selected US shocks. Run raw monetary and fiscal levers under VAR expectations. No reform bridge or model-consistent expectations.
open model →Long-run incentives and capital. Run local steady-state reform analysis; transition paths are available from the package. Calibrated, not outcome-validated.
open model →Households and distribution. Calculate taxes, benefits, net income, and population impacts. Population results depend on survey data; no macro feedback.
open model → diff --git a/svar/index.html b/svar/index.html index b8bffa4..42b2807 100644 --- a/svar/index.html +++ b/svar/index.html @@ -137,7 +137,7 @@A VAR in eight quarterly variables — three global, five UK — estimated - over the inflation-targeting era — 1992Q1–2023Q2 in + over the inflation-targeting era — 1992Q1–2025Q1 in both the paper replication and the hosted adapter. The hosted forecast conditions on observations through 2026Q1, but does not re-estimate its coefficients on those later observations. It is estimated in @@ -178,7 +178,7 @@