Bank of England structural VAR

BoE SVAR

The data, estimation, identification, sampling and the unemployment satellite.

Eight variables, six named shocks

From raw quarterly data to named structural shocks, following Bank of England Macro Technical Paper No. 3 (Brignone & Piffer, 2025).

1

The data: eight quarterly variables

Eight quarterly series in two blocks, global and UK.

Eight quarterly variables in two blocks global block (3) world GDP world CPI (UK-trade-weighted) real oil price in sterling UK block (5) Bank Rate · sterling exchange-rate index CPI · CPI energy real GDP 1992Q1 2025Q1 2026Q2 estimation sample, levels, 4 lags · Covid dummies 2020Q1–2021Q2 conditioning only time →
click to expand
The eight variablesSample and specification · Conditioning versus estimation
The eight variables
Global block
World GDP · World CPI (both UK-trade-weighted) · real oil price in sterling
UK block
Bank Rate · sterling exchange-rate index · CPI · CPI energy · real GDP
Sample
1992Q1–2025Q1, in levels, p=4 lags
Pandemic
exogenous Covid dummies, 2020Q1–2021Q2

Sample and specification

Estimation covers the inflation-targeting era, 1992Q1–2025Q1, in both the paper replication and the hosted adapter — in levels, with p=4 lags and exogenous Covid dummies for 2020Q1–2021Q2.

Conditioning versus estimation

The hosted forecast conditions on observations through 2026Q2 without re-estimating coefficients on them.

2

Bayesian estimation

Bayesian, with a Minnesota prior in the framework of Giannone, Lenza & Primiceri (2015).

Priors + data → a posterior, not point estimates priors Minnesota (normal-inverse-Wishart) sum-of-coefficients · dummy-initial-obs data eight variables, 1992Q1–2025Q1 levels, 4 lags posterior over (Π, Σ) IRFs decompositions fan charts all with bands
click to expand
PriorsThe posterior
Priors
Reduced form
yt=c+Π1yt−1+⋯+Πpyt−p+ut,p=4yt=c+Π1yt−1+⋯+Πpyt−p+ut,p=4
Main prior
Minnesota (normal-inverse-Wishart) on (Π,Σ)
Additional priors
sum-of-coefficients (μ = 1) · dummy-initial-observation (θ = 1); tightness λ = 0.2
Framework
Giannone, Lenza & Primiceri (2015)

The posterior

The output is a posterior over the reduced-form coefficients Π and error covariance Σ; every downstream object — impulse responses, decompositions, fan charts — is a posterior distribution, not a point estimate, and every one is reported with credible bands.

3

Identification: naming the shocks

Identification follows Arias–Rubio-Ramírez–Waggoner (2018); sign restrictions name six structural shocks: world demand, world energy, world supply, UK demand, UK supply, and UK monetary policy.

Restrictions turn residuals into named shocks posterior draws unnamed residuals zero restrictions small open economy: UK shocks cannot move world on impact sign restrictions impact signs name the shocks (Arias et al. 2018) six named shocks world demand · world energy world supply · UK demand UK supply · UK monetary policy + 2 unidentified(residual volatility) importance weights correct sampler non-uniformity · Chan–Matthes–Yu (2025) permutation search cuts rejections
click to expand
Restrictions and correctionsZero and sign restrictions · importance weights · permutation search
Zero and sign restrictions
Zero restrictions
small open economy: UK shocks cannot move world variables on impact
Sign restrictions
impact signs name the shocks (Arias et al. 2018)
Six named shocks
world demand · world energy · world supply · UK demand · UK supply · UK monetary policy
Residual
two unidentified shocks absorb remaining volatility
Corrections
importance weights correct the sampler's non-uniformity · the Chan–Matthes–Yu (2025) permutation search cuts rejections
4

Sampling: draws and acceptance

Accept/reject sampling on posterior (Π,Σ) draws.

Production run: draws narrow at each stage 10,000 posterior (Π, Σ) draws pass the sign restrictions 765 accepted (7.6%) importance-weighted ESS 368.0
click to expand
Accept/reject samplingProduction run · hosted default · forecast runs

Few draws pass the sign restrictions, so medians and bands are noisier than the paper's. The production run uses data through 2026Q2 and seed 20260812 (results/summary.md, after boe-var-model #24).

Accept/reject sampling
Production run
10,000 draws → 765 accepted (7.6%)
Effective sample size
368.0 (importance-weighted)
Hosted default
2,000 draws → 159 accepted, ESS 48.7
Separate forecast run
455 of 6,000 accepted, ESS 181.4 (overview shocks and forecast)

Hosted and forecast runs

  • The hosted forecast conditions on data through 2026Q2 and keeps the 2025Q1 coefficient-estimation endpoint.
  • Its default 2,000-draw run (159 accepted, ESS 48.7) is below the ESS of 100 at which the response warns of weak inference.
  • The current-outlook chart on the overview is the archived 12 Nov 2026 round: a separate forecast run, 455 of 6,000 accepted (ESS 181.4), not the production run above.
  • Forecast runs leave the dummy-initial-observation prior off, which moves medians by 0.02pp or less.
5

The svar-unemployment satellite

Unemployment is not one of the eight VAR variables. A separate satellite maps the VAR's UK GDP growth forecast into a path for the ONS unemployment rate (MGSX), leaving the replication core untouched.

The regressionHow it is fitted · how the bands work · how far it is validated
The regression
The mapping (dynamic)
Δut=α+βgt+γΔut−1+εt

How it is fitted

  • OLS of the quarterly change in unemployment on year-on-year GDP growth and its own lag, over the VAR's 1992Q1–2025Q1 sample.
  • The furlough quarters (2020Q1–2021Q2) are dummied out: GDP fell about 20% on a year earlier while unemployment barely moved.
  • Including them cuts the growth coefficient about threefold (β −0.047 → −0.016).

How the bands work

  • Archived 28 Jul 2026 round: unemployment falls as GDP growth rises, so the GDP median and bands pass through with the bounds swapped. These bands carry GDP-forecast uncertainty only, so they are too narrow.
  • The model now (boe-var-model #32): simulates along every GDP path with residual and coefficient draws, adding the satellite's own residual and estimation uncertainty. No round with these bands is archived yet.
  • Measured coverage of both: validation tab.

How far it is validated

  • The published path is capped at four quarters, the range that was tested.
  • 73-origin rolling test (origins 2005Q1 to 2023Q1): excluding the Covid quarters the satellite beats no-change at horizons 1–4, not significantly; on all 73 targets it is level at h=1 and worse from h=2.
Notation 7 symbols
SymbolMeaningStep
ytvector of the eight VAR variables in quarter t2
c, Π1,…,Πpconstant and lag-coefficient matrices2
pnumber of lags2
ut, Σreduced-form errors and their covariance2
Δutquarterly change in the unemployment rate (satellite)5
gtyear-on-year UK GDP growth (satellite)5
α, β, γ, εtsatellite regression coefficients and residual5

Sources

The paper, the method papers and the code.

Where each source lives
whatwhere
The source paper Macro Technical Paper No. 3 — Brignone & Piffer, Bank of England, July 2025.
Companion paper Brignone & Piffer, structural forecast analysis — the Section 5 forecast-revision method, later BoE Staff Working Paper No. 1,165 (January 2026).
Our working paper papers/boe-svar — the full replication write-up, quarterly forecast scorecard, and enumerated deviations.
This replication github.com/PolicyEngine/boe-var-model — code, prepared data, figures, and the annotated RESOURCES.md.
Method papers Arias, Rubio-Ramírez & Waggoner (2018) — zero/sign identification · Chan, Matthes & Yu (2025) — permutation-search speedup · Giannone, Lenza & Primiceri (2015) — the prior framework.
Reference toolboxes bsvarSIGNs (R, the exact Arias et al. machinery) · ECB BEAR · Cesa-Bianchi VAR-Toolbox — used in the independent econometrics review, which caught and fixed a genuine bias in the zero-restriction importance weights.