Models

Seven models, each with a trust label. Pick a question to narrow them down.

The seven models side by side

The seven models compared by country, question, reform scoring and run surface
ModelCountryAnswersScores reformsRuns
PolicyEngine microsimulationUKUSHousehold and population tax-benefit effectsYesHosted
OBR emulatorUKNear-term UK fiscal and macro scenariosYesHosted
BoE SVARUKWhat drove the UK economy; short-run forecastsNoHosted
FRB/USUSUS monetary and fiscal shocks, by quarterNoHostedraw shocks only
US HANKUSHow US shocks spread across household wealthNoHostedstylized shocks only
OG-UKUKLong-run work, saving and capitalYesnone publishedLocal
DEFINE-UKUKUK climate-policy scenario deltasNoLocal
Method, data and runtimePer model: method, vintage, joins, scope, runtime

PolicyEngine microsimulation pe-microsim

Method
Static tax-benefit microsimulation: direct rule evaluation, no behavioural response
Data vintage
UK & US statute; enhanced FRS 2023–24 microdata for UK population runs
Microsim connection
Is the microsimulation: every bridge starts or ends here
Reform scoring
Yes: a PolicyEngine parameter + a new value, applied directly to households
Out of scope
GDP, inflation, interest rates, or macro feedback
Run surface
Hosted
Typical runtime
Sub-second per household; minutes for population runs

OBR emulator obr-macro

Method
Structural macroeconometric emulator: Gauss–Seidel over 290 solved equations per quarter (of 372 in the published listing)
Data vintage
OBR Economic and Fiscal Outlook, October 2026
Microsim connection
Static costing in, macro feedback out: score_reform via HHDI_ADDFACTOR
Reform scoring
Yes: shocks to exogenous model variables; static-costing bridge from PolicyEngine reforms
Out of scope
Arbitrary reform incidence; borrowing (the model moves it, the adapter does not return it); capital spending (CGIPS does not reach GDP by default; the opt-in route is an identity passthrough) and corporation-tax magnitudes (TCPRO converges slowly)
Run surface
Hosted (CLI, MCP, Python API)
Typical runtime
About 1–5 minutes per raw shock; over 10 minutes for a translated reform (default five-year window)

BoE SVAR boe-svar

Method
Bayesian structural VAR: posterior sampling with zero and sign restrictions
Data vintage
ONS, BoE and FRED; estimated through 2025Q1, conditioned through 2026Q2 (12 Nov round)
Microsim connection
CPI path → benefit-uprating incidence (svar_inflation_incidence, experimental)
Reform scoring
No: the baseline/conditioning member: it reads the economy reforms are scored against
Out of scope
The causal effect of a specific statutory reform
Run surface
Hosted (CLI, MCP, Python API)
Typical runtime
Minutes per full estimation + identification run

FRB/US frb-us

Method
Large-scale macroeconometric model: 284 equations solved under VAR expectations
Data vintage
The Fed's model.xml and LONGBASE database, October 2026
Microsim connection
Wage/hours path → employment-income incidence (frbus_shock_incidence, experimental)
Reform scoring
No: shock experiments only; no reform bridge, so score_reform refuses model="frbus"
Out of scope
PolicyEngine reform effects, anticipated-policy paths, or MCE scenarios
Run surface
Hosted (CLI, MCP, Python API): raw shocks only
Typical runtime
About 3 seconds cold; under a second warm

US HANK us-hank

Method
Two-asset HANK (Auclert et al. 2021): first-order sequence-space impulse responses
Data vintage
The published Econometrica 2021 calibration; no data vintage
Microsim connection
Wage/labor IRFs → employment-income incidence (hank_shock_incidence, experimental)
Reform scoring
No: stylized shocks only; the labor tax is endogenous, so score_reform refuses model="hank"
Out of scope
Forecasts, detailed tax reforms, or nonlinear/state-dependent dynamics
Run surface
Hosted (CLI, MCP, Python API): stylized shocks only
Typical runtime
About 10 seconds cold on a laptop (cached); instant warm

OG-UK psl-og

Method
Dynamic general equilibrium (OLG): steady-state root-find and transition-path iteration
Data vintage
ONS, OBR and BoE aggregates as calibration targets; OG-UK 0.3.2
Microsim connection
Two-way: statute in via PE tax functions; long-run changes back into a second microsim run (og+microsim)
Reform scoring
Yes: a PolicyEngine policy translated into estimated tax functions
Out of scope
A short-run forecast or independently validated reform effect
Run surface
Local only
Typical runtime
Tens of minutes per score (two steady-state solves; not formally timed); transition paths far longer

DEFINE-UK define-uk

Method
Ecological stock-flow consistent (E-SFC): upstream R code run unmodified at a pinned commit
Data vintage
DEFINE-UK 1.1, April 2026, at pinned commit 846081a
Microsim connection
Scenario income deltas → employment-income incidence (define_scenario_incidence, experimental; hosted via define_payload)
Reform scoring
No: no reform bridge; score_reform will not accept it
Out of scope
Levels, forecasts, reform scoring
Run surface
Local CLI (pe-macro define-scenario); hosted MCP tools return run instructions
Typical runtime
Instant from the cached pinned run; a fresh upstream run takes tens of minutes in R 4.3
Same question, every modelSix questions put to every macro model, next to a benchmark

Where each model fails

None is validated as a forecaster. Each carries one of five status classes.

The seven models support five classes of check.

Full verdictsWhat each model is checked against, and the whole verdict
  • PolicyEngine microsimulationExact for implemented rulesChecked against: Implemented UK and US legislation, rule by ruleHousehold maths exact; population adds survey error
  • OBR emulatorAnchored to the October 2026 EFO, not validated as a forecasterChecked against: October 2026 EFO; HMRC ready reckoner; ONS outturnsTracks the EFO by construction (0.24% GDP MAPE); weak free-running; predictive, policy and uncertainty checks rated weak; every quarter of the reform solve converges under the nominal-GDP pin; CPI does not respond to rate or oil shocks
  • BoE SVARPartial replication of Brignone & Piffer (2025)Checked against: Brignone & Piffer (2025); ONS outturnsGDP decomposition consistent with the paper; CPI global-shock share about 8.5pp short; no forecast skill against a drifting random walk once the 64 tests are adjusted together; bands under-cover across 50 origins
  • FRB/USReplicates the Fed's pyfrbus on a 100bp shock and three further scenariosChecked against: The Fed's pyfrbus on an identical shock; the Board’s published 100bp run; published multiplier rangesOne 100bp funds-rate shock matches pyfrbus within a 1e−6 tolerance of both Fed releases, the same order as the release-to-release gap; of the economic checks, only the 100bp response sizes pass (within 10% of the Board’s published run, on its 2020Q1 design) and the rest are partial to failing; baseline tracking is an identity; no predictive claim
  • US HANKPartial replication of Auclert et al. (2021)Checked against: Auclert et al. (2021), via the authors' toolkitSix of seven solved parameters of the paper match directly and the seventh, vphi, after a wage-markup convention conversion; Y, K and r are inputs, not results; one IRF magnitude matched within about 2%, the rest unvalidated
  • OG-UKCalibrated; no outcome benchmarkChecked against: ONS/OBR aggregates as targets, not a validation setTargets met by construction; the nearest external benchmark (OBR Working Paper 22, Table 5.1) has not yet been run
  • DEFINE-UKPartial replication of the manual: macro block and scenario design-gate passChecked against: Manual Table 4; ONS outturns, OBR EFO, DESNZ emissions; no scenario figures are publishedMacro block matches manual Table 4, but 2025 growth misses by −0.30pp, clearing a gate set after the run by 0.01pp; emissions 23.3% below the manual in 2040; no published scenario results to check; baseline far from outturns; deltas only, never levels
Other institutions on the same questionsBenchmarks, how comparable each is, and the figures

Imposed matches are anchors, not validation. The stronger evidence is out-of-sample forecasts, independent costings, and agreement with another implementation.

QuestionCompared withHow comparableThe figures
Long-run effect of a 1pp basic-rate rise (psl-og) OBR Working Paper 22, Table 5.1 Future benchmark: same scenario; ours is not yet produced (the committed run is a transition path to 2030, not the long-run steady state WP 22 reports) psl-og validation
Yield of a 1p rise in the basic rate, 2026–27 (obr-macro + pe) HMRC ready reckoner (June 2025) Adjusted: below HMRC in the one matched year (2026–27); HMRC is post-behavioural on administrative data, ours static and survey-based. A benchmark, not a validation pe-microsim validation
Government-purchases multiplier (frb-us) CBO ranges (converted to a calendar-year ratio), Ramey (JEP 2019) and the 2010 FRB-US lines in Coenen et al. Adjusted: ours sit inside CBO’s converted ranges, in their lower part, and inside Ramey’s range, which pools estimates across horizons and monetary regimes; they sit below the 2010 FRB-US lines frb-us validation
Income-tax multiplier (frb-us, US) OBR’s published UK assumption for income tax and NICs Loose: different country and model class frb-us validation
GDP response to a 1pp rise in Bank Rate (obr-macro, boe-svar) Bank of England, Quarterly Bulletin 1999 Q2: a 1pp rise held for a year lowers output by 0.2% to 0.35% after about a year, and inflation by 0.2 to 0.4pp about a year after that Loose: a 1999 estimate, and the experiments differ in rate path; obr-macro sits below the range, boe-svar above it obr-macro validation; boe-svar validation
Forecast error, 2024Q3 onward (boe-svar) OBR FER (July 2025); BoE MPR (Aug 2024) Adjusted: different statistics, matched vintages; over 50 origins, no skill against a drifting random walk boe-svar validation

Macro models, with household detail

Three ways the microsimulation adds households to a macro answer. Pick one for the detail.

Score a reform: microsimulation, then the OBR emulator

What would a reform cost, and what would it do to the economy?

Reform Microsimulation OBR emulator Scored result
pe-macro score --country uk --reform '{"gov.hmrc.income_tax.rates.uk[0].rate":0.21}' --model obr
About this join

Worked example: 1p on the basic rate, costed by the microsimulation and fed through the emulator; the figures live on the OBR Overview. The static costing runs below HMRC’s ready reckoner in the one matched year, 2026–27: a benchmark, not a validation. UK only (why a US reform gets no macro feedback).

Read the caveats before the headline: the costing inherits survey-microdata uncertainty, and the emulator’s spending multiplier is fixed by construction (limits), so treat macro feedback as sign and order of magnitude. Setup on the Use page.

the join
chain
pe-microsim → obr-macro
enters
HHDI_ADDFACTOR, household disposable income
call
score_reform(model="obr")
country
UK only