Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,15 @@ heading here; `tests/test_repo_hygiene.py` asserts that they do.

## [Unreleased]

### BoE SVAR re-pin to `17f5d82`

- **Pin.** BoE SVAR `c827ab86` → `17f5d824` (boe-var-model #34, world CPI seasonal adjustment; #35, November 2026 MPR comparator and 2026Q3 scoring). After a redeploy the hosted `forecast_uk` and `latest_shocks` run this pin.
- **Replication figures** (`results/summary.md` at `17f5d82`): accepted draws 765 → 731, ESS 368.0 → 333.3; global share of UK GDP variance 38.7 → 37.6%, of UK CPI 41.5 → 52.0% (paper about 50%); domestic share of UK CPI 36.8 → 24.8% (paper about 40%), now the open gap.
- **Evaluation figures** (`rolling_evaluation.json`, `coverage_evaluation.json` at `17f5d82`): smallest adjusted q against drift 0.39 → 0.36; no-change wins after adjustment 3 of 64 → 1 of 64; UK CPI 90% band coverage 78% → 76% at one quarter, 54% → 56% at eight.
- **Shock reading, 2026Q2**: world supply "negative, 70%" → no call (55% positive); 430 of 6,000 draws accepted.
- **Band factors** follow the pin: eight of 32 moved, seven by one step of measured coverage (one origin in 50) and one by two (`forecasts/band_calibration.json`). Archived rounds are untouched.
- **Not changed**: the archived 12 Nov 2026 round shown as the current forecast (it predates #34), the working-paper PDF (the #24 vintage, now labelled) and the compare page.

### Model re-pin and site review (#278, #279)

- **#278: model pins.** OBR `2965e82d` → `0e44ec87` (main after obr-macroeconomic-model #74 and #75), BoE SVAR `c13a9c91` → `c827ab86`, FRB/US `846eeb5c` → `162cfe6a`, US HANK `551a40d3` → `58353e5f`. After a redeploy the hosted tools run these pins. The FRB/US and US HANK compare rows with stored paths re-run identically; the OBR rows were re-run at `0e44ec8`.
Expand Down
2 changes: 1 addition & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ verbatim — that registry is authoritative and this table is a copy of it.
|---|-------|--------|------|
| 01 | **`pe-microsim`** — PolicyEngine tax-benefit microsimulation | exact for implemented rules (household arithmetic); population estimates add survey error | [PolicyEngine/policyengine.py](https://github.com/PolicyEngine/policyengine.py) |
| 02 | **`obr-macro`** — OBR macroeconometric emulator | anchored to the October 2026 EFO, not validated as a forecaster; predictive and policy checks weak | [PolicyEngine/obr-macroeconomic-model](https://github.com/PolicyEngine/obr-macroeconomic-model) |
| 03 | **`boe-svar`** — Bank of England structural VAR replication | partial replication of Brignone and Piffer (2025): GDP decomposition consistent with the paper; CPI global-shock share about 8.5pp short; no forecast skill against a drifting random walk | [PolicyEngine/boe-var-model](https://github.com/PolicyEngine/boe-var-model) |
| 03 | **`boe-svar`** — Bank of England structural VAR replication | partial replication of Brignone and Piffer (2025): global share of GDP variance consistent with the paper; global share of CPI variance close on the mean, with a wide band; domestic CPI share about 15pp short; no forecast skill against a drifting random walk | [PolicyEngine/boe-var-model](https://github.com/PolicyEngine/boe-var-model) |
| 04 | **`frb-us`** — Federal Reserve FRB-US implementation | replicates the Fed's pyfrbus (1.0.0 and 1.1.1) on one 100bp funds-rate shock, within a 1e-6 tolerance; a solver check, not a test of the economics | [PolicyEngine/us-frb-model](https://github.com/PolicyEngine/us-frb-model) |
| 05 | **`us-hank`** — US two-asset HANK (Auclert–Bardóczy–Rognlie–Straub 2021) | partial replication of Auclert et al. (2021): six of seven solved parameters match directly and the seventh (vphi) after a wage-markup convention conversion; steady-state targets are inputs; shock signs are tested; one published magnitude (impact consumption after a monetary shock, units inferred) is matched within about 2%, the others are not checked; not a forecaster; distributional outputs are steady-state statistics and an MPC-proportional split of the aggregate | [PolicyEngine/us-hank-model](https://github.com/PolicyEngine/us-hank-model) |
| 06 | **`psl-og`** — OG-UK overlapping generations model | calibrated; no outcome benchmark; research prototype | [PSLmodels/OG-UK](https://github.com/PSLmodels/OG-UK) |
Expand Down
9 changes: 9 additions & 0 deletions assets/make_current_outlooks.py
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,15 @@ def _shock_run_note(data: dict) -> str:
if shocks.get("draws") == data.get("draws") and shocks.get("accepted_draws") == data.get("accepted"):
return ("A separate run from the shock probabilities with the same draw settings, "
f"so both show {data['draws']:,} draws and {data['accepted']} accepted.")
# The shock file can be newer than the archived round (it was after the
# 17f5d82 re-pin, which seasonally adjusted world CPI): say which run is
# which from the two files' own provenance rather than implying one model.
prov = shocks.get("provenance", {})
run_date, rev = str(prov.get("run_at", ""))[:10], str(prov.get("source_revision", ""))[:7]
if run_date and rev and run_date > str(data.get("generated", "")):
return (f"The shock probabilities are a later run ({run_date}, boe-var-model {rev}, "
f"{shocks.get('accepted_draws')} of {shocks.get('draws'):,} accepted), "
"so the draw counts differ and this forecast predates that model version.")
return "A separate run from the shock probabilities, so the draw counts differ."

def render(page: Path, name: str, content: str) -> str:
Expand Down
2 changes: 1 addition & 1 deletion economy/topics.py
Original file line number Diff line number Diff line change
Expand Up @@ -1193,7 +1193,7 @@ def inflation_model() -> str:
peak_q, peak_v = max(later, key=lambda r: r[1])
tail = [v for q, v in later if q > peak_q][-4:]
settle = f", then settles near {sum(tail) / len(tail):.1f}%" if tail else ""
# The plateau is the model, not a call: boe-var-model README at c827ab8
# The plateau is the model, not a call: boe-var-model README at 17f5d82
# ("no inflation anchor"); /svar/validation#no-anchor carries the detail.
anchor = (': the model has a unit root and no inflation anchor, so it extends '
'recent inflation rather than returning to 2% '
Expand Down
2 changes: 1 addition & 1 deletion economy/topics/growth/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -178,7 +178,7 @@ <h2>What the models see</h2>
<details class="about tg-limits">
<summary>Its limits</summary>
<ul>
<li>Status: partial replication of Brignone and Piffer (2025): GDP decomposition consistent with the paper; CPI global-shock share about 8.5pp short; no forecast skill against a drifting random walk.</li>
<li>Status: partial replication of Brignone and Piffer (2025): global share of GDP variance consistent with the paper; global share of CPI variance close on the mean, with a wide band; domestic CPI share about 15pp short; no forecast skill against a drifting random walk.</li>
<li>It cannot answer statutory policy reform effects.</li>
<li>Uncertainty is <a class="gloss-link" href="/glossary#prior-posterior">posterior</a> 68% and 90% intervals, which under-cover in back-testing (<a href="/svar/validation#rolling">coverage study</a>).</li>
<li>Estimated on 1992Q1-2025Q1 but conditioned on 2026Q2 data, so the estimation sample stops well short of the quarters being forecast.</li>
Expand Down
2 changes: 1 addition & 1 deletion economy/topics/inflation/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -177,7 +177,7 @@ <h2>What the models see</h2>
<details class="about tg-limits">
<summary>Its limits</summary>
<ul>
<li>Status: partial replication of Brignone and Piffer (2025): GDP decomposition consistent with the paper; CPI global-shock share about 8.5pp short; no forecast skill against a drifting random walk.</li>
<li>Status: partial replication of Brignone and Piffer (2025): global share of GDP variance consistent with the paper; global share of CPI variance close on the mean, with a wide band; domestic CPI share about 15pp short; no forecast skill against a drifting random walk.</li>
<li>It cannot answer statutory policy reform effects.</li>
<li>It forecasts headline CPI only. Core CPI is shown above as an outturn; no model in this suite forecasts it.</li>
<li>Uncertainty is <a class="gloss-link" href="/glossary#prior-posterior">posterior</a> 68% and 90% intervals; the sample ends 2025Q1 while the conditioning data run to 2026Q2.</li>
Expand Down
21 changes: 11 additions & 10 deletions forecasts/band_calibration.json
Original file line number Diff line number Diff line change
Expand Up @@ -3,8 +3,8 @@
"source": {
"repo": "PolicyEngine/boe-var-model",
"file": "results/coverage_evaluation.json",
"revision": "c13a9c912e106ad6de02809f99897d8ef79ea69d",
"code_version": "47e8ab54d92cee11bcfeed15a1e9e0d584b85d9b088fe2b9dc324643eff7bc92",
"revision": "17f5d8247e72e93e1d4bd34731666fea9d655ba1",
"code_version": "b12c1495bcd1f88aae564d79a352dc4718d7d79f498b4d66f9d89fc4442c6656",
"origins": 50,
"draws_per_origin": 1000,
"horizons_evaluated": 8,
Expand All @@ -18,13 +18,14 @@
"cpi": "cpisa"
},
"caveat": "Coverage is measured on four-quarter differences of the model's log series (pseudo-out-of-sample, overlapping origins), the same year-on-year basis as the published fans. The shock factor is the Cholesky factor (boe-var-model #26/#30), so the measured coverage is the same on every machine. If Bank Rate is ever fanned, use the level factors (band_scale_factors.production_spec), since it is published as a level.",
"model_vintage": "Factors follow the pinned model: measured at boe-var-model 17f5d82, after the world CPI seasonal adjustment (boe-var-model #34). Rounds archived up to 12 Nov 2026 were produced before that adjustment and are drawn with the same factors. Against the previous set (c13a9c9) eight of 32 factors moved, seven by one step of measured coverage (one origin in 50) and CPI 68% h1 by two: GDP 68% h1 0.8464 to 0.8829 and h3 1.1328 to 1.1816, GDP 90% h7 1.3999 to 1.4603, CPI 68% h1 1.1816 to 1.2878, h2 1.3460 to 1.2878 and h4 1.7062 to 1.6228, CPI 90% h1 1.3411 to 1.3999 and h8 2.2262 to 2.1301.",
"unemployment_satellite": "Left raw. The svar-unemployment (Okun) satellite is not among the series in the coverage evaluation, so no measured factor applies to it; borrowing the boe-svar factors would assert a calibration that was never tested. Its fan is labelled as raw model quantiles.",
"k": {
"gdp": {
"68": {
"h1": 0.8464,
"h1": 0.8829,
"h2": 0.9205,
"h3": 1.1328,
"h3": 1.1816,
"h4": 1.0864,
"h5": 1.0864,
"h6": 1.1328,
Expand All @@ -38,30 +39,30 @@
"h4": 1.3411,
"h5": 1.3999,
"h6": 1.3411,
"h7": 1.3999,
"h7": 1.4603,
"h8": 1.4603
}
},
"cpi": {
"68": {
"h1": 1.1816,
"h2": 1.346,
"h1": 1.2878,
"h2": 1.2878,
"h3": 1.408,
"h4": 1.7062,
"h4": 1.6228,
"h5": 1.797,
"h6": 1.797,
"h7": 1.797,
"h8": 1.6228
},
"90": {
"h1": 1.3411,
"h1": 1.3999,
"h2": 1.587,
"h3": 1.7239,
"h4": 1.9544,
"h5": 2.0397,
"h6": 2.0397,
"h7": 2.0397,
"h8": 2.2262
"h8": 2.1301
}
}
},
Expand Down
Loading