Fed FRB/US, rebuilt

FRB/US

The shocks against the Fed’s own solver, the multipliers, and the limits.

The baseline agreement is an identity; the shocks are the evidence

What was checked, against what, and the result
CheckAgainstResultVerdict
Solver, 100bp funds-rate shockFed’s pyfrbus 1.0.0 and 1.1.1, all 284 variables × 20 quarters; gate 1e−6; the Fed’s two releases differ by 1.4e−81.4e−8 and 1.8e−8pass
Solver, three further scenariosFed’s pyfrbus 1.0.0: federal purchases, personal tax rate, non-inertial Taylor rule; gate 1e−69.5e−9, 2.0e−8, 1.9e−8pass
100bp shock, size of responsesBoard’s published run (Laforte 2018) on its 2020Q1 design; gate ±10%. One design, not an exact replication; the unemployment peak is larger from later start dates0.96–1.04× the Board’spass
Purchases multiplier, yr 1, Taylor ruleCBO converted range 0.45–1.82 (lower third); FRB-US in Coenen et al. ≈0.95 (0.2 below)0.75partial
Transfers multiplier, yr 1CBO derived range 0.36–1.52 (typical Fed); FRB-US in Coenen et al. ≈0.48; below both (model economics, not implementation)0.28fail
Baseline, first projected quarterFRED outturns, 2026Q3; one quarter, and the Board’s baseline, not oursGDP growth −1.23pp; core PCE +0.80ppweak
Inflation effect of the 2022–23 hikesFRBSF Economic Letter 2025-03, about 3pp; mean 2022–24 (model economics, not implementation)−0.12ppfail
Baseline re-solve (tracking invariant)LONGBASE; holds for any input (9.0e−9 on scrambled data), so not evidence7.3e−12by construction
Baseline caveatMatching is not being right

Matching the Fed’s baseline does not make the baseline right: against the 2026Q3 outturn LONGBASE misses GDP growth and core PCE inflation (limits).

Evidence by dimension

Evidence by dimension: the registry’s categorical judgement, and what would change it
DimensionLevelWhat would change it
Implementation fidelitystrongLike-for-like gates on further official demos and closures
Predictive validationnot assessedVintage-preserving pseudo-out-of-sample evaluation
Identificationnot applicableNone: an equation model, not an identified VAR
Policy counterfactualsmoderateModel-consistent expectations in the hosted tools
UncertaintyweakReviewed residuals and published coverage
Vintage reproducibilitymoderateSeveral historical model and data pairs tested
How it is gated, and the uncertainty engine

Uncertainty engine. Seeded stochastic simulation jointly resamples the official residual vector, preserving contemporaneous dependence; model and LONGBASE packages carry SHA-256 provenance gates. Published probability results still need a reviewed simulation design, and model-consistent expectations are not yet available through the hosted tools.

  • A dedicated CI job re-runs the Board's pyfrbus from source and gates against that fresh solution, not a committed CSV.
  • A weekly run repeats the suite on Python 3.11 and 3.12.
  • The multipliers are re-gated weekly; the published comparators and this table are refreshed by hand.
  • Full tables: VALIDATION.md and the working paper (the paper is on the April 2026 vintage).
About this implementationIndependent build, not endorsed

An independent implementation of the Federal Reserve Board's published model equations and data. It is not produced, maintained or endorsed by the Federal Reserve; do not present its results as official Federal Reserve estimates.

The solver match, the multipliers, a worked shock

1

Against the Fed’s own solver

The same shock in both implementations, at the reference’s own noise floor.

Maximum absolute residual, log scale; further left is closer

FRB/US: residuals against the Fed’s pyfrbus, log scale Dot chart on a base-10 logarithmic axis of maximum absolute residuals; further left is closer. The rows fall into three kinds. ours vs LONGBASE — tracking identity, not evidence: 7.3e-12 (tracking identity); pyfrbus 1.1.1 vs LONGBASE — tracking identity: 1.9e-08 (tracking identity); ours vs pyfrbus 1.0.0 (shock): 1.4e-08 (ours vs the Fed); ours vs pyfrbus 1.1.1 (shock): 1.8e-08 (ours vs the Fed); pyfrbus 1.1.1 vs 1.0.0 — the Fed’s own releases: 1.4e-08 (the Fed vs itself). The two tracking rows are an identity — init_trac defines the add-factors as minus the residuals at the input data, so re-solving reproduces the input for any input at all — and are shown for completeness, not as evidence. The framing comparison is the last row: the Federal Reserve's own two pyfrbus releases disagree with each other by about as much as this implementation disagrees with either (the same order of magnitude), so our agreement sits at the scale of the reference implementation's own numerical noise; the pass criterion is the 1e-6 gate.
  • ours vs the Fed’s pyfrbus, under shock
  • the Fed’s two releases vs each other (dashed rule)
  • tracking identity — round-off, not fidelity
The top two rows are the tracking identity, not a ranking: they compare two implementations' round-off, not their fidelity to the Fed's economics.

Source: FRB/US replication

Source details

Source: us-frb-model VALIDATION.md at 162cfe6.

About this steptable, 3 points
Validation results against the Federal Reserve Board's own artefacts
TestWhat it checksResult
Tracking invariant Re-solving the baseline reproduces LONGBASE, 284 endogenous variables over 2026Q1–2030Q4. Against a 1e−8 gate, but an identity, not a comparison. init_trac sets each add-factor to minus that equation's residual at the input data, so the solve reproduces whatever it was tracked to, for any input: the model repository's test_tracking_invariant_holds_for_arbitrary_data passes at 9.0e−9 on a scrambled baseline with every accounting identity destroyed. Read the row below, not this one, as evidence.
Cross-validation The same 100bp rffintay_aerr shock run in both implementations, all 284 variables × 20 quarters. On the October 2026 LONGBASE our maximum absolute difference is 1.4e−8 against pyfrbus 1.0.0 and 1.8e−8 against 1.1.1; the two Fed releases differ from each other by 1.4e−8, so ours are the same order of magnitude as the release-to-release gap, and all are inside the 1e−6 gate.
Three further scenarios A +1% federal-purchases shock (egfe_aerr), a +2pp personal tax-rate shock (trp_aerr) and a +100bp shock under the non-inertial Taylor rule, against pyfrbus 1.0.0, all 284 variables × 20 quarters. Maximum absolute differences 9.5e−9, 2.0e−8 and 1.9e−8, all inside the 1e−6 gate.

The Fed-vs-Fed comparison frames the others.

  • The Board's two releases ship a byte-identical model.xml and LONGBASE, and differ only in how Newton reuses its LU factorisation.
  • Yet they disagree by as much as this implementation disagrees with either, so agreement sits at the reference's own numerical noise, not at a chosen tolerance.
  • All three residuals concentrate in the same series: the wage level wpsn in absolute terms, and near-zero gap terms (dpadj) in relative terms.
2

Fiscal multipliers

Purchases multipliers sit inside CBO’s ranges and 0.2–0.3 below FRB-US in 2010: plausibility bands, not a test passed. Transfers sit below every matched range.

About this steptable, 6 points
FRB/US fiscal multipliers (1% of GDP, sustained) against published ranges
QuantityOursPublishedReading
Gov. purchases, yr 1, inertial Taylor rule0.75CBO, Fed responding typically: 0.45–1.82 (converted); FRB-US in Coenen et al. Fig. 25, no accommodation: ≈0.95; Ramey (JEP 2019): "a surprisingly narrow range of 0.6 to 1"inside the CBO range, lower third; about 0.2 below the 2010 FRB-US line
Gov. purchases, yr 2, inertial Taylor rule0.66CBO, Fed responding typically: 0.24–1.21 (converted)inside the CBO range
Gov. purchases, yr 1, pegged funds rate0.79CBO, Fed response limited: 0.50–2.09 (converted)inside, near the low end
Gov. purchases, yr 2, pegged funds rate0.93CBO, Fed response limited: 0.50–2.50 (converted); FRB-US in Coenen et al. Fig. 25, 2-yr accommodation: ≈1.25inside the CBO range, near the low end; about 0.3 below the 2010 FRB-US line
Personal tax cut, yr 10.23CBO, Fed not tightening: lower- and middle-income 0.3–1.5; higher-income 0.1–0.6below the first row, inside the second; regimes differ (ours is under the Taylor rule)
Personal tax cut, yr 20.32CBO, Fed not tightening: two-year cut for lower- and middle-income people 0.3–1.5; one-year cut for higher-income people 0.1–0.6. CBO publishes no central valuebetween the two rows; regimes differ (ours is under the Taylor rule)
Transfers, yr 10.28CBO, Fed responding typically (derived): 0.36–1.52; CBO Table 3, Fed not tightening: 0.4–2.1; FRB-US in Coenen et al. Fig. 28: ≈0.48below all three (model economics, not implementation). Depends on the regime: 0.28 under the inertial Taylor rule with the surplus rule on, 0.32 with the funds rate pegged and the fiscal rule off. FRB/US gives transfer recipients the same MPC as earners (hand-to-mouth share 0.31)
  • CBO ranges marked “converted” come from Whalen and Reichling, CBO Working Paper 2015-02, Table 2: the quarterly response to a one-quarter $1 impulse.
  • We cumulate it into this page’s calendar-year ratio for a sustained shock. The conversion is ours, not CBO’s.
  • The tax-cut ranges are from CBO (Feb 2015, Table 3), which assumes the Fed did not tighten.
  • Coenen et al. (AEJ: Macroeconomics 2012) values are read from their charts (±0.1) and cover the 2009–10 FRB-US, whose expectations mode they do not state.
  • Ramey’s survey range pools estimates across horizons and monetary regimes.
  • The yr-1 purchases value depends on the shock: 0.75 for a +1%-of-GDP rise, about 0.83 for small shocks, 0.93 for a 1% cut. The cut/rise gap is an artefact of the model file adding growth in logs, not economics; CBO assumes symmetry.
  • Pegged purchases stay below 1 partly because the demo surplus rule raises the personal tax rate; with it off, the pegged yr-2 value is 1.04.
  • Sources and the conversion: us-frb-model docs/benchmarks.md.
3

A worked shock, against the Board

On the Board’s 2020Q1 design, all three published responses to a 100bp tightening are within 10% of the Board’s run.

More detailtable, notes
100bp funds-rate shock: the Board’s published run (Laforte 2018), ours on the Board’s design (2020Q1 start) and ours on the hosted default (2026Q1 start); q1 is the shock quarter
ResponseBoardOurs, 2020Q1Ours ÷ BoardOurs, 2026Q1
Core inflation, trough (pp)−0.0345, q8−0.0330, q90.96×−0.034, q8
Output gap, trough (pp)−0.486, q8−0.471, q80.97×−0.50, q7
Unemployment, peak (pp)+0.219, q8+0.228, q91.04×+0.28, q9
Real GDP, trough (%)not published––−0.56, q8
Funds rate (pp)+1.00, then −0.27––+1.00, then −0.265
  • The Board’s design. Laforte shocked in 2020Q1 with the Board’s ping settings: no zero lower bound, no thresholds and a fixed equilibrium real rate. The 2020Q1 column repeats that; the model repository gates each response within 10% of the note.
  • Not an exact replication. 2020Q1 is realised history in the October 2026 LONGBASE; the Board shocked its 2018 projection, which is archived only as an EViews database.
  • Across data vintages. The April 2026 LONGBASE gives 0.96×, 0.98× and 1.01×, so unemployment sits at 1.01–1.04× the Board’s.
  • Verdict re-banded on 18 Nov 2026. This row read “partial” when it was scored on the 2026Q1 run, where unemployment is 1.26× the Board’s. It reads “pass” now that it is scored on the Board’s own 2020Q1 design, against the model repository’s ±10% gate; the gate itself did not change.
  • The unemployment peak depends on the start date. It is 1.04× the Board’s from 2020Q1, 1.26× from 2026Q1 and up to 1.38× from 2035Q1. Core PCE and the output gap stay within 10% from every start date tested (2020Q1 to 2040Q1).
  • The hosted default. pe-macro frbus-shock --var rffintay_aerr --shock 1.0 starts in 2026Q1 and is the run in the last column (GDP trough in 2027Q4, unemployment peak in 2028Q1). The Overview example starts in 2026Q3, the first projected quarter, so its troughs fall two quarters later (GDP −0.55% in 2028Q2).
  • Earlier model generations move inflation more. The Board’s previous model gives −0.062pp for core PCE, and the −0.08pp in the 2014 FEDS Note describes the pre-2018 model. The Board attributes the smaller response to re-estimating the wage–price block through 2017.

Board: Laforte (2018), FEDS Notes, doi:10.17016/2380-7172.2306, Figure 2, VAR expectations; the Board reports the output gap, not GDP. Ours: us-frb-model results/inflation_transmission.md (2020Q1 and the start-date sweep) and results/site_responses.md (2026Q1) at 162cfe6, October 2026 LONGBASE. The 2026Q1 run also matches the Board’s own pyfrbus to the cross-validation result above (1.4e−8).

Known limitations

The baseline misses core inflation

Against the first outturn since its data edge, 2026Q3, the October 2026 LONGBASE tracks unemployment within 0.1pp and the funds rate within 0.2pp, undershoots GDP growth by 1.23pp (2.1% vs a 3.4% annualized outturn) and overshoots core PCE inflation by 0.80pp (3.3% vs 2.5%). The April 2026 vintage missed 2026Q1 core PCE by −1.65pp (April record). A weekly outturn backtest reports this rather than gating it, because the baseline is the Board’s. LONGBASE is an illustrative tracking baseline, not an official Fed forecast.

VAR expectations only

The hosted tools run VAR expectations only, so forward guidance is out of scope here. The package itself now supports opt-in model-consistent expectations (Frbus(path, mce=...)), gated against the Board’s pyfrbus at under 1e-6, for use from Python.

No mcontrol

Trajectory targeting is out of scope. stochsim_bands() does produce seeded bootstrap bands, but the hosted frbus_shock returns a single deterministic path.

Hosted, but only as raw shocks

One lever, one shock size, a policy rule and a headline set of responses — not the full 284-variable model. Anything more means installing the package.

No PolicyEngine reform bridge

Reforms enter as shocks, not statute: score_reform refuses model="frbus", so US reform scoring stops at PolicyEngine microsimulation's static answer.

A small inflation response

A 100bp shock moves core PCE inflation by −0.034pp at the trough, as in the Board’s own simulation of the current model (worked shock). Almost all of it runs through the VAR-expectations equations. A scenario whose conclusion depends on inflation should say so (us-frb-model #15).

Little disinflation from the 2022–23 hikes

FRB/US (VAR), as the Board publishes it, attributes about 0.1pp of the 2022–24 fall in core PCE inflation to the hikes (mean −0.12pp against never tightening, −0.11pp against following the rule from 2024). Empirical high-frequency estimates put it at about 3pp (FRBSF Economic Letter 2025-03). This follows from the small inflation response above (episode results, us-frb-model #14).

Vintage

New Board vintages of model.xml and LONGBASE are adopted by hand, not automatically; the one in use, October 2026, is on the model card. Its history runs through 2026Q2 while the simulation window still starts in 2026Q1.