frbus — a modern Python implementation of FRB/US

A from-scratch, modern (numpy/scipy/sympy, Python 3.10+) implementation of the Federal Reserve Board's FRB/US macroeconomic model, with VAR (backward-looking) expectations.

Provenance

The model itself — the equation system (model.xml), the data vintage (LONGBASE.TXT), the reference implementation (pyfrbus), and the model documentation — is published by the Federal Reserve Board in the public domain (see vendor/pyfrbus_package/LICENSE). The raw materials are kept unmodified under vendor/ (see vendor/README.md). This package is an independent reimplementation that follows pyfrbus semantics exactly and is validated against it (see VALIDATION.md).

Install

uv venv && uv pip install -e ".[dev]"

Quickstart

import pandas as pd
from frbus import Frbus, load_data

data = load_data("vendor/data_only_package/LONGBASE.TXT")
model = Frbus("vendor/pyfrbus_package/models/model.xml")

start, end = pd.Period("2026Q1"), pd.Period("2030Q4")

# Fiscal-policy configuration used by the Fed's demos
data.loc[start:end, "dfpdbt"] = 0
data.loc[start:end, "dfpsrp"] = 1

# Add-factor the model so it reproduces the baseline exactly
with_adds = model.init_trac(start, end, data)

# 100bp monetary policy shock
with_adds.loc[start, "rffintay_aerr"] += 1
sim = model.solve(start, end, with_adds)

print((sim.loc[start:end, "xgdp"] / with_adds.loc[start:end, "xgdp"] - 1) * 100)

A runnable version is in examples/monetary_policy_shock.py.

What is implemented

• Parsing of model.xml: variables, equations, per-equation coefficients, endogenous/exogenous classification (VAR "standard" equations).
• LONGBASE.TXT loading into a pandas DataFrame with a quarterly PeriodIndex.
• Per-period damped Newton solver on the full simultaneous system with an analytic sparse Jacobian (sympy differentiation, scipy sparse LU).
• Frbus(path), .init_trac(start, end, data), .solve(start, end, data), .exogenize([...]) — mirroring the essentials of pyfrbus's Frbus class.
• Not implemented: MCE (rational-expectations) equation variants — Frbus(path, mce=...) raises NotImplementedError. mcontrol and stochastic simulations are also out of scope.

Validation

See VALIDATION.md for full tables. Summary, in order of how much each one actually tells you:

• Cross-validation (the real evidence). Four shock scenarios — monetary, fiscal egfe, tax trp, and a non-inertial Taylor variant — agree with the Fed's own pyfrbus across all 284 endogenous variables and 20 quarters to 6.0e-9 abs / 4.9e-8 rel. For scale, the Board's own pyfrbus 1.0.0 and 1.1.1 releases differ from each other by 1.3e-8, so this sits at the reference implementation's own noise floor.
• Fiscal multipliers lie inside published ranges (purchases 0.72 in year 1 under the inertial Taylor rule, 0.90 in year 2 with the funds rate pegged), and the full rule × horizon grid is ordered as theory requires.
• Tracking invariant. After init_trac, solving the baseline reproduces LONGBASE to 5.6e-17. This is an identity, not a validation result — init_trac defines the add-factors as minus the residuals at the input data, so it holds for any input; the repo gates a scrambled-baseline version of the same test to keep that honest. Do not cite it as agreement with the Fed's data.

Nothing here is a forecast evaluation. LONGBASE is the Board's illustrative baseline, not a Fed forecast, and no pseudo-out-of-sample accuracy of any kind is established for this model.

Development

pytest           # tests, including the tracking-invariant gate
ruff check src tests examples