Fed FRB/US, rebuilt

FRB/US

How the Board’s model file is parsed, solved and shocked.

Parse, add-factor, solve by Newton

From the Board's published artefacts to an impulse response. Full derivations are in the working paper.

1

Parse the Board's model file

The Federal Reserve Board publishes FRB/US whole: the equation system as model.xml, the LONGBASE historical and projection database, the documentation, and a reference Python solver, pyfrbus — all in the public domain.

One source of truth: the Board's model.xml model.xml variables, equations, coefficients, endo/exo public domain parse normalise rewrite EViews lags, substitute coefficients, flip to residual form 284 equations behavioural + identity, each with aerr + trac VAR ("standard") set nothing re-typed or re-specified — the XML declares the model's semantics
click to expand
ParseOne source of truth · Residual form
Parse
System
284 endogenous eqs (behavioural + identities), from model.xml
Residual form
Fi(x,data)+aerri+traci=0

One source of truth

The parse stage reads model.xml directly.

  • It reads variables, equations, per-equation coefficients and the endogenous / exogenous classification for the VAR-expectations ("standard") equation set, the only set this pipeline reads.
  • Nothing in the file is re-typed or re-specified: the XML file is the single source of truth for the model's semantics.
  • That yields 284 endogenous variables and their equations, with behavioural and identity blocks classified exactly as the Board's file declares them.

Residual form

Each equation is normalised — EViews lag and difference notation rewritten, coefficient references substituted by their published values — flipped to residual form Fi(⋅)=0, and augmented with its additive shock term (_aerr) and a tracking residual (_trac), so each equation solves Fi+aerri+traci=0.

2

Load LONGBASE, then add-factor

The load stage brings LONGBASE.TXT, the October 2026 vintage, into a pandas frame on a quarterly PeriodIndex. Every simulation is read as a deviation from this baseline.

Add-factoring pins the baseline LONGBASE.TXT October 2026 vintage history + projection quarterly PeriodIndex init_trac set each equation's tracking residual, per quarter solve = LONGBASE reproduced to identity by construction shocks are then imposed on top of the add-factored baseline
click to expand
LONGBASE conditioningTracking residuals (init_trac)
LONGBASE conditioning
Tracking residual
traci=−Fi(baseline) per equation, per quarter
Invariant
solve(baseline+trac)=LONGBASEsolve(baseline+trac)=LONGBASE exactly

Tracking residuals (init_trac)

init_trac makes the solve reproduce the database identically, for any input.

  • It sets each equation's add-factor to minus that equation's residual at the input data.
  • It is the Fed's own device, and the direct analogue of the OBR emulator's anchoring.
  • That is why a baseline match is not evidence.
  • Shocks are imposed on top of the add-factored baseline, where the solver does real work.
3

Solve by damped Newton

Under VAR expectations every expectations term depends only on current and lagged observables, so the model is simultaneous within each quarter but recursive across quarters.

Simultaneous within the quarter, recursive across quarters one quarter: damped Newton analytic sparse Jacobian (sympy) sparse LU solve (SuperLU) damp: halve the step to stay in-domain check convergence, else iterate Newton iteration done t t+1 t+2 each quarter starts from the previous solution
click to expand
Newton iterationAnalytic sparse Jacobian · One block, not many

The solver therefore works period by period, starting each quarter from the previous quarter's solution.

Newton iteration
Step
J(x(k))Δx=−F(x(k)), sparse LU (SuperLU)
Damping
halve Δx until F,J evaluate cleanly
Convergence
∥Δx∥≤10−8, then ∥F∥≤5×10−4

Analytic sparse Jacobian

Each quarter, a damped Newton iteration solves the full simultaneous system.

  • It uses an analytic sparse Jacobian for every step, differentiated symbolically with sympy and factorised with scipy.sparse.linalg.splu.
  • Damping backtracks the step only to stay inside the domain of definition: no logs of non-positive arguments and no NaNs.
  • The solver's default tolerances are xtol=10−8, rtol=5×10−4.

One block, not many

  • The reference implementation decomposes each period into recursive and simultaneous blocks; this one solves the whole period at once.
  • Both approaches solve the same F(x)=0.
  • So the two agree to solver tolerance, not to some negotiated threshold.
  • See how far to trust it.
4

Policy rules

Monetary policy is modular: exogenous 0/1 switches in the model file select which rule sets the funds rate.

0/1 switches select which rule sets the funds rate inertial_taylor default, LONGBASE rule shock: rffintay_aerr taylor same bracket, no inertia shock: rfftay_aerr fixed_funds_rate baseline path held no monetary offset funds rate rff — one rule active per run each rule reads its own add-error; the wrong one is rejected, not zeroed
click to expand
Policy rulesA modular funds rate · The rule choice is the exercise
Policy rules
Inertial Taylor
rt=0.85rt−1+0.15[rt∗+π¯t+0.5(π¯t−πtarg)+1.0x~t]+εtrt=0.85rt−1+0.15[rt∗+π¯t+0.5(π¯t−πtarg)+1.0x~t]+εt
Effective rate
rfft=max(rffrulet, rffmint)

A modular funds rate

Three are exposed through frbus_shock: inertial_taylor (the default, the LONGBASE rule and the one the validation numbers use), taylor, and fixed_funds_rate, which holds the funds rate on its baseline path so there is no endogenous monetary offset.

The rule choice is the exercise

The choice is usually the economic point of the exercise: the same government-purchases shock gives a larger output response with the funds rate fixed, and a larger price-level response without the monetary offset. The multipliers under each rule, against published ranges, are on Validation.

Each rule reads its own error term

Each policy rule reads its own add-error. The shock term εt in the inertial-Taylor rule above is rffintay_aerr; taylor applies the same reaction bracket without the inertia term and reads rfftay_aerr instead.

5

Shock and compare

A shock is an addition to an equation's error term (rffintay_aerr for the policy rule, for instance).

Impulse response = shocked path minus baseline path baseline solve add-factored LONGBASE shocked solve aerr += shock, same window difference sim − baseline 100bp from 2026Q1 GDP −0.56% lur +0.28pp picxfe −0.034pp shock units are per-lever — pp for the policy-rate error, log points of growth for spending
click to expand
Impulse responsesWhat comes out · What stays out of scope

Solve shocked and baseline paths over the same window and read the difference. Units differ by lever, each rule reads its own error term, and a shock held for several quarters stacks: see the three traps.

Impulse responses
Shock
aerr+=shock on the add-factored baseline
Response
IRFt=simt−baselinet over the same window

What comes out

The difference is an impulse response for every one of the 284 variables. The headline 100bp tightening, and how it compares with the Board's documented simulation properties, is the worked shock on Validation.

What stays out of scope

The hosted pipeline runs the VAR-expectations model only.

  • The package supports opt-in model-consistent expectations from Python, but no hosted tool exposes them, so forward guidance is out of scope here.
  • The reference implementation's mcontrol trajectory-targeting routine is not implemented.
  • No bridge links a PolicyEngine reform to model variables; see the known limits.
Notation 11 symbols
SymbolMeaningStep
Fi, xresidual form of endogenous equation i; the endogenous variables1
aerriadditive shock term (add-error) of equation i1
tracitracking residual of equation i1
LONGBASEthe Board's historical and projection database2
J, Δx, kJacobian of the system, Newton step, iteration3
xtol, rtolstep and residual convergence tolerances3
rt, rt∗policy rate from the inertial Taylor rule; equilibrium real rate4
π¯t, πtarg, x~tinflation, inflation target, output gap in the rule4
εtshock term of the policy rule4
rff, rffrule, rffmineffective funds rate, rule rate, floor4
IRFtimpulse response: simulated minus baseline path5