Glossary
59 terms this site uses, in plain language.
No term matches. Ask us to add one.
Trust and checks
- Status vocabulary
Each model carries one of five status classes, defined below. No status calls a model validated.
How each model is graded- Exact for implemented rules
The rules are code, checked rule by rule, so a household answer can be verified by hand. Exact for the rules as implemented, which does not show that every area of legislation is covered. PolicyEngine microsimulation is in this class.
PolicyEngine microsimulation validation- Replicates the Fed’s solver
Our code reproduces a published model’s output within a stated error, on the checks named and no others. For FRB/US the check is solver against solver, not a test of the economics. FRB/US is in this class.
FRB/US validation- Partial replication
Some published outputs are reproduced, or are statistically consistent within a stated error; others miss or cannot be checked, and the status names both. The BoE SVAR, US HANK and DEFINE-UK are in this class; for DEFINE-UK the scenarios are only design-gated.
BoE SVAR validation · US HANK validation · DEFINE-UK validation
- Anchored, not validated
The model is set up to reproduce a published forecast by construction, so matching it is not evidence; its free-running and policy checks are reported separately. The OBR emulator is in this class.
OBR emulator validation- Calibrated, no outcome benchmark
Targets are met by construction and no external outcome has yet been run against the model, so there is no ground truth to test against. OG-UK is in this class.
OG-UK validation- Published anchor
A published figure a model is set up to match, such as an imposed steady-state target. An imposed match is an anchor, not a replication or a validation.
How each model is graded- By construction
A match that holds because the model was set up to produce it, for example by fixing adjustments so it reproduces a published path. It shows the copy is faithful, not that the model forecasts well.
OBR emulator validation- Anchored and free-running
Anchored: the model's adjustment terms are held so its baseline follows a published forecast. Free-running: those terms are released and the model's own equations drive the path. The OBR emulator is reported both ways.
OBR emulator validation- Deltas only
Results are read as differences from the model's own baseline, never as levels, because that baseline sits far from the outturns.
DEFINE-UK validation- Add-factor
An adjustment term added to a model equation. Held at the values that make the model reproduce a published forecast, add-factors anchor it; a reform can also enter as one, as the household-income add-factor
Score a reformHHDI_ADDFACTORdoes in the OBR emulator.- Design-gated
Checked that each scenario switches on exactly the policy settings its published definition lists, because no published results exist to check the numbers against. It catches a wrongly built scenario, not a wrong magnitude.
DEFINE-UK validation- Paper-anchored
Checked against a few figures quoted in the authors’ earlier papers rather than a full published results table. For DEFINE-UK these are two FMM-2023 figures: the peak GDP effect of green public investment and baseline 2030 emissions.
DEFINE-UK validation- Frozen edge
A forecast run from data cut off at a fixed past quarter, then scored against the outturns that followed. One frozen origin gives a handful of scores; a rolling evaluation over many origins is the stronger test.
BoE SVAR frozen origin- Noise floor
The smallest difference that numerical round-off alone can produce. Two implementations agreeing at the noise floor differ no more than the arithmetic does.
FRB/US validation- Release-to-release gap
The maximum absolute difference between the Fed’s pyfrbus 1.1.1 and 1.0.0 on the same shock, with the same model and baseline (1.4e−8 on the October 2026 LONGBASE). A yardstick for solver noise, not a bound or pass criterion: the pass criterion is the 1e−6 gate.
FRB/US validation- Benchmark, not a validation
A comparison with another institution's figure that was produced on a different basis, so agreement or disagreement is informative but does not prove either right. The site uses the word in two other senses, each named where it appears: a naive forecast to beat (the drifting random walk), and a published range a result is set beside (a plausibility band, not a test passed).
The HMRC costing comparison
Forecasts
- Outturn
The published official figure for a period, which a forecast is scored against once it exists.
The scored period- Forecast round
One dated set of forecasts, committed to Git so it cannot be revised after the fact; every round since the first was committed before its outturn existed.
The archive of every round- Horizon (h1, h2)
How many quarters after the round's data edge a forecast quarter falls: h1 is the first quarter the model had no data for, h2 the second.
The scored record- Backcast
A forecast for a quarter that had already ended when the round was archived, usually before its official figure was published. Easier than a true forecast, so it is tagged on the record; a round archived after the figure was published is also tagged Retroactive.
The scored record- Data vintage
A dated snapshot of the data as it was published on a given day. Snapshots are never edited, so any score can be reproduced from the data the model saw.
The data vintages- Credible band
The range the model puts a given probability on, such as 68% or 90%, from its posterior distribution. A band that contains every outturn can still be too wide.
BoE SVAR validation- Drifting random walk
A naive benchmark forecast: the last value carried forward plus the average past change. A model has forecast skill only if it beats benchmarks like this one.
BoE SVAR against benchmarks- Forecast skill
Forecasting better than a naive benchmark, out of sample, after allowing for the number of tests run. Matching a published forecast is not skill.
BoE SVAR against benchmarks
Model types
- Microsimulation
Applying the tax and benefit rules, written as code, to individual households, one at a time or across a weighted survey of the population.
PolicyEngine microsimulation- Static costing
The cost of a reform with no change in behaviour and no feedback through the wider economy.
Score a reform- Macro feedback
The second-round effect of a reform once output, prices and incomes respond; the OBR emulator adds it to a UK static costing.
Score a reform- Incidence
How a change is shared across households: who gains and who loses. The incidence tools lay a macro path onto households; they are experimental and are not reform scores.
A macro shock across households- SVAR
Structural vector autoregression: a statistical model of several series together, with named shocks identified by sign and zero restrictions.
BoE SVAR- HANK
Heterogeneous-agent New Keynesian model: households differ in wealth and income, so a shock can be traced through the whole distribution.
US HANK- OLG
Overlapping-generations model: age cohorts choose work and saving over their lifetimes, and the economy is solved for its long-run steady state or the path to it.
OG-UK- Steady state
The long-run position the model economy settles at once every adjustment has played out.
OG-UK- Stock-flow consistent
A model in which every flow of money is recorded as someone's payment and someone else's receipt, and stocks of assets and debts follow from those flows. DEFINE-UK is an ecological version that adds the climate.
DEFINE-UK- EFO
The OBR's Economic and Fiscal Outlook, the official UK forecast the OBR emulator is anchored to.
OBR emulator
Estimation and scoring
- Prior and posterior
In Bayesian estimation the prior states what is assumed about the parameters before seeing the data; the posterior combines that prior with the data. Posterior forecast bands are drawn from the posterior.
BoE SVAR methodology- Minnesota prior
A standard prior for vector autoregressions that pulls each series towards behaving like a random walk and shrinks the effects of longer lags and of other variables towards zero. It steadies estimates from short samples.
BoE SVAR methodology- Sign and zero restrictions
Rules that name a shock by how it moves the variables: a sign restriction says a variable must rise or fall, a zero restriction says it must not respond, on impact or in the long run. Draws that break the rules are discarded.
BoE SVAR methodology- Effective sample size (ESS)
When draws carry unequal weights, the number of equally weighted draws that would hold the same information. It is smaller than the raw count when a few draws carry most of the weight.
BoE SVAR methodology- Impulse response (IRF)
The path of each variable after a single shock, measured quarter by quarter as the difference from the path without the shock.
BoE SVAR- Variance decomposition (FEVD)
Forecast error variance decomposition: the share of a variable’s forecast uncertainty at a given horizon that each identified shock accounts for.
BoE SVAR validation- Historical decomposition
A split of the past movements of a series into the contribution of each identified shock, plus the part the starting conditions alone imply.
BoE SVAR- Band coverage
The share of outturns that fell inside a forecast band. Over many periods a 68% band should hold about 68% of outturns: far more means the band is too wide, far fewer means it is too narrow.
BoE SVAR validation- RMSE
Root mean squared error: the typical size of a forecast’s misses, in the units of the variable, with large misses counting more. An RMSE ratio below one means a forecast missed by less than the benchmark it is set against.
BoE SVAR validation- Percentage point (pp) and pp-quarter
A percentage point is the difference between two percentages: a rate going from 4% to 5% is up 1 pp. A pp-quarter measures the size of a rate path: the annualised rate change added up over the quarters it lasts.
Models side by side
Solving and simulating
- Exogenous and endogenous
Exogenous variables are set from outside the model, as inputs or shocks; endogenous variables are solved for by the model. Exogenising a variable fixes it at a chosen path, so its own equation no longer sets it.
OBR emulator methodology- Closure
The choice of which variables a model solves for and which it takes as given, for example whether taxes or spending adjust to meet a fiscal rule. Swapping the closure changes what a scenario holds fixed.
The OBR emulator uses three names. The published closure solves the OBR’s listing as written, which leaves nominal GDP undetermined. The nominal-GDP pin, the default, restores the OBR’s own line 690 so every quarter converges. The investment closure lets business investment respond to the cost of capital, and is what a corporation-tax shock uses.
OBR emulator methodology- Gauss–Seidel and Newton solvers
Two ways to solve a large set of equations each quarter. Gauss–Seidel works through the equations in turn, updating each variable with the latest values, until the changes fall below a tolerance; Newton solves them together using their derivatives.
OBR emulator methodology- Unit root and explosive root
Properties of how a shock fades. With a unit root the effect never dies out, so the series does not return to where it started; with an explosive root the effect grows over time instead of fading.
BoE SVAR validation- Output gap
How far actual output is above or below potential output, the level the economy can sustain without inflation rising or falling, as a percentage of potential.
FRB/US methodology- Taylor rule
A rule that sets the policy interest rate from how far inflation is from target and from the output gap. An inertial rule moves only part of the way each quarter, so the rate adjusts gradually.
FRB/US methodology- VAR expectations and model-consistent expectations (MCE)
Two ways a model forms expectations of the future. Under VAR expectations households and firms forecast with a small statistical model of the economy; under model-consistent expectations they foresee the path the full model itself produces.
FRB/US methodology- rstar
The name of the monetary shock in the US HANK code: a shift in the intercept of the central bank’s interest-rate rule. The policy rate does not move one for one with it, because the rule also responds to the inflation the shock causes. It is not an estimate of the neutral interest rate.
US HANK- Marginal propensity to consume (MPC)
The share of an extra pound or dollar of income that a household spends rather than saves, over a stated period such as a quarter or a year.
US HANK- Hand-to-mouth households
Households that hold little liquid wealth, so a change in income is expected to pass almost directly into spending. US HANK counts them by liquid wealth: most of its hand-to-mouth households hold illiquid wealth and spend little of a rebate.
US HANK- Sequence-space Jacobian
A fast way to solve heterogeneous-agent models: how the economy responds to a shock in each future period is computed once as a matrix, and any shock path is then a matrix product. At first order it is accurate for small shocks.
US HANK methodology- Euler equation
The condition that a household cannot gain by moving consumption between one period and the next. It ties the path of consumption to the interest rate and to how patient households are.
OG-UK methodology- Okun relation
An estimated link between output growth and the change in the unemployment rate, used to turn a GDP forecast into an unemployment forecast.
The jobs topic
Using the models
- MCP
Model Context Protocol: a standard way for an AI client such as Claude or ChatGPT to call tools on a server. The hosted server needs no account or API key.
Connect a client- Job
A slow call run in the background:
Access, limits and cold startsstart_jobreturns an id, andget_job_resultis polled until the result is ready.- Cold start
The first call to a container that has just started, which pays a one-off loading or estimation cost; later calls are fast.
Access, limits and cold starts
Missing a term? Tell us.