Factor-Augmented Vector Autoregression (FAVAR)

favar is a Python package for estimating Factor-Augmented Vector Autoregressive models. It is designed for empirical macroeconomic research, monetary policy analysis, forecasting, and impulse-response analysis with large information panels.

The package implements the two-step FAVAR procedure of Bernanke, Boivin, and Eliasz (2005). It extracts latent factors from a large panel, removes the contemporaneous policy component from the estimated factors, estimates an augmented VAR system, and projects impulse responses back to any observable series in the information panel.

Contents

• Installation
• Key Features
• Quick Start
• Model Overview
• Data Requirements
• Data Preparation and Transformations
• Slow-Moving and Fast-Moving Variables
• Estimation Procedure
• Forecasting
• Impulse Response Functions
• Residual Autocorrelation Diagnostics
• Example Summary Output
• Examples and Notebook
• Public API
• Project Status and Releases
• Citation
• References

Installation

Install the latest release from PyPI:

pip install favar

Key Features

• Two-step FAVAR estimation following Bernanke, Boivin, and Eliasz (2005).
• Principal-component factor extraction from large information panels.
• Slow-moving variable adjustment for recursive monetary policy identification.
• Forecasts for observed variables with confidence intervals.
• Orthogonalized impulse response functions for the augmented system.
• Panel-projected impulse responses for any selected series in $X_t$.
• Lag-order selection table with AIC, BIC, FPE, and HQIC.
• Residual autocorrelation diagnostics with $2 / \sqrt{T}$ bounds.
• Clean pandas-based interface.

Quick Start

from favar import FAVAR

model = FAVAR(
    X=X,
    Y=Y,
    policy_var="policy_rate",
    k_factors=3,
    slow_columns=slow_columns,
    standardize=True,
)

order_selection = model.select_order(maxlags=12)
print(order_selection.summary())

results = model.fit(lags=4)
print(results.summary())

forecast = results.forecast(steps=12, confidence_level=0.95)
irf_y = results.impulse_response(periods=48, impulse_size=0.25)
irf_x = results.panel_impulse_response(
    periods=48,
    columns=["ip_growth", "inflation_core"],
    impulse_size=0.25,
)

Model Overview

A FAVAR combines a large information panel $X_t$ with a smaller set of observed variables $Y_t$ that enter the VAR directly.

• $X_t$ is a large panel of economic indicators with dimension $T \times N$.
• $Y_t$ is a smaller set of observed variables with dimension $T \times M$.
• $F_t$ is a low-dimensional vector of latent factors extracted from $X_t$.
• $R_t$ is the policy instrument, supplied through policy_var.

The estimated system is:

$$Z_t = c + A_1 Z_{t-1} + \cdots + A_p Z_{t-p} + u_t.$$

$$X_t = d + \Lambda F_t + \Gamma Y_t + e_t.$$

where:

$$ Z_t = \begin{bmatrix} F_t \\ Y_t \end{bmatrix}. $$

The policy variable is ordered last in $Y_t$ for recursive identification.

Data Requirements

Prepare two pandas DataFrame objects:

X  # large information panel, shape T x N
Y  # observed VAR variables, shape T x M

Both inputs must:

• have a compatible time index;
• be observed at the same frequency, such as monthly or quarterly;
• contain only numeric columns;
• contain no missing values after transformations;
• be aligned over the same sample period;
• include policy_var as a column of Y.

Example layout:

X
            ip_growth  employment_growth  inflation_core  credit_spread
2000-01         0.31               0.12            0.22           1.21
2000-02         0.28               0.10            0.19           1.18

Y
            output_growth  inflation  policy_rate
2000-01             0.31       0.22         5.75
2000-02             0.28       0.19         5.80

Data Preparation and Transformations

The package standardizes X internally when standardize=True, but it does not decide the economic transformation of each raw series. Transformations should be chosen before estimation.

Common transformations:

Log growth

For real quantities such as production, employment, credit, or monetary aggregates:

$$ \Delta \log(x_t) = 100 \left[\log(x_t) - \log(x_{t-1})\right]. $$

Inflation

For price indexes:

$$ \pi_t = 100 \left[\log(P_t) - \log(P_{t-1})\right]. $$

Interest rates and spreads

Interest rates, spreads, and percentages are often used in levels:

$$ r_t = R_t. $$

They can also be differenced when the empirical design calls for changes:

$$ \Delta r_t = r_t - r_{t-1}. $$

Practical preprocessing checklist

Before fitting the model:

• seasonally adjust series when appropriate;
• apply logs before differencing strictly positive level series;
• avoid mixing levels and growth rates without an economic reason;
• document outlier treatment and sample restrictions;
• call dropna() after transformations;
• verify that X.index.equals(Y.index) is True.

Example:

import numpy as np
import pandas as pd

raw = pd.read_csv("macro_panel.csv", parse_dates=["date"]).set_index("date")

X = pd.DataFrame(index=raw.index)
X["ip_growth"] = 100 * np.log(raw["industrial_production"]).diff()
X["employment_growth"] = 100 * np.log(raw["employment"]).diff()
X["inflation_core"] = 100 * np.log(raw["core_price_index"]).diff()
X["credit_spread"] = raw["credit_spread"]

Y = pd.DataFrame(index=raw.index)
Y["output_growth"] = X["ip_growth"]
Y["inflation"] = X["inflation_core"]
Y["policy_rate"] = raw["policy_rate"]

data = pd.concat([X, Y], axis=1).dropna()
X = data[X.columns]
Y = data[Y.columns]

Slow-Moving and Fast-Moving Variables

For monetary policy applications, the information panel is divided into:

• slow-moving variables: variables assumed not to react contemporaneously to the policy shock within the period, such as output, employment, consumption, and some prices;
• fast-moving variables: variables allowed to react within the period, such as interest rates, spreads, asset prices, and financial indicators.

Pass the slow-moving columns through slow_columns:

slow_columns = [
    "ip_growth",
    "employment_growth",
    "inflation_core",
]

If slow_columns=None, all columns in X are treated as slow-moving. This is allowed, but explicit classification is recommended for monetary policy work.

Estimation Procedure

Let $X^s$ denote the standardized information panel:

$$ X^s_{tj} = \frac{X_{tj} - \bar{X}_j}{s_j}. $$

The implemented two-step estimator proceeds as follows.

1. Principal components from the full panel

Estimate $K$ principal components from $X^s$:

$$ \widehat{C}_t = \widehat{C}(F_t, Y_t). $$

These components estimate the common space spanned by both latent factors and observed variables.

2. Principal components from slow-moving variables

Estimate $K$ principal components from the slow-moving subset of the panel:

$$\widehat{C}^{\star}_t = \widehat{C}^{\star}(F_t).$$

These components are used to isolate the latent factor space from the contemporaneous policy instrument.

3. Remove the contemporaneous policy component

Regress the full-panel principal components on a constant, the policy instrument, and the slow-moving principal components:

$$\widehat{C}_t = a + b_R R_t + B_S \widehat{C}^{\star}_t + v_t.$$

The cleaned factor estimate is:

$$ \widehat{F}_t = \widehat{C}_t - \widehat{b}_R R_t. $$

4. Estimate the augmented VAR

Stack the cleaned factors and observed variables:

$$ \widehat{Z}_t = \begin{bmatrix} \widehat{F}_t \\ Y_t \end{bmatrix}. $$

Estimate:

$$\widehat{Z}_t = c + A_1 \widehat{Z}_{t-1} + \cdots + A_p \widehat{Z}_{t-p} + u_t.$$

5. Estimate the measurement equation

Estimate the relationship between the standardized information panel and the augmented state:

$$ X^s_t = d + \Theta \widehat{Z}_t + e_t. $$

This measurement equation allows responses from the FAVAR system to be mapped back to each series in $X$:

$$\mathrm{IRF}_{X}(h) = \mathrm{IRF}_{Z}(h)\widehat{\Theta}.$$

When scale="original", projected panel responses are multiplied by the stored standard deviation of each original X column.

Basic Usage

from favar import FAVAR

model = FAVAR(
    X=X,
    Y=Y,
    policy_var="policy_rate",
    k_factors=3,
    slow_columns=slow_columns,
    standardize=True,
)

results = model.fit(lags=13)
print(results.summary())

Main arguments:

• X: large information panel.
• Y: observed variables included directly in the FAVAR system.
• policy_var: policy instrument column in Y.
• k_factors: number of latent factors.
• slow_columns: slow-moving columns from X.
• standardize: whether to standardize X before factor extraction.
• lags: fixed lag order for the augmented VAR.

Lag order can also be selected by an information criterion:

order_selection = model.select_order(maxlags=12)
print(order_selection.summary())

results = model.fit(select_order="aic", maxlags=12)

Accepted criteria are "aic", "bic", "hqic", and "fpe".

Compact order-selection example:

FAVAR Lag Order Selection (* highlights the minimums)
==================================
    AIC     BIC     FPE      HQIC
----------------------------------
0  -1.262  -1.208   0.2832  -1.240
1 -3.985* -3.769* 0.01859* -3.897*
2  -3.979  -3.601  0.01870  -3.826
3  -3.927  -3.386  0.01971  -3.708
4  -3.881  -3.178  0.02066  -3.596
----------------------------------

Forecasting

Use forecast() to forecast the observed variables in Y:

forecast = results.forecast(steps=12, confidence_level=0.95)
print(forecast.head())

For each variable in Y, the output includes:

• point forecast;
• lower confidence bound;
• upper confidence bound.

Example column names:

policy_rate  policy_rate_lower  policy_rate_upper

Impulse Response Functions

Use impulse_response() for responses of the augmented FAVAR system.

irf_system = results.impulse_response(
    periods=48,
    shock="policy_rate",
    impulse_size=0.25,
    include_factors=False,
)
print(irf_system.head())

impulse_size=0.25 rescales the shock so that the impact response of the policy variable is 0.25. If the policy rate is measured in percentage points, this corresponds to 25 basis points.

Use panel_impulse_response() to project responses back to selected series in the information panel:

irf_panel = results.panel_impulse_response(
    periods=48,
    shock="policy_rate",
    columns=["ip_growth", "inflation_core", "credit_spread"],
    scale="original",
    impulse_size=0.25,
)
print(irf_panel.head())

Use:

• scale="original" for projected responses in the transformed units supplied by the user;
• scale="std" for responses in standardized panel units.

Residual Autocorrelation Diagnostics

Use plot_acorr() to inspect residual autocorrelations and cross-correlations of the augmented FAVAR system:

fig = results.plot_acorr(nlags=10)

The figure contains one panel for each pair of variables in the augmented system. The dashed bands are $2 / \sqrt{T}$ bounds.

Example Summary Output

results.summary() returns a text summary with overall fit statistics, equation-by-equation coefficients, residual correlations, and FAVAR-specific metadata.

Compact example:

  Summary of FAVAR Regression Results
======================================
Model:                           FAVAR
Estimator:                Two-step PCA
VAR method:                        OLS
Date:               Sun, 28, Jun, 2026
Time:                         21:14:27
--------------------------------------------------------------------
No. of Equations:               3    BIC:                  -3.59224
Nobs:                         178    HQIC:                 -3.81539
Log likelihood:        -383.59539    FPE:                   0.01892
AIC:                     -3.96762    Det(Omega_mle):        0.01685
--------------------------------------------------------------------
FAVAR Model Information
====================================================================
No. of factors:                                                    2
No. of X variables:                                               40
No. of observed Y variables:                                       1
No. of slow-moving variables:                                     20
Policy variable:                                                 FFR
Policy position:                                                   3
Lag order:                                                         2
Standardized X:                                                 True
PC variance shares:                                     0.579, 0.248
--------------------------------------------------------------------
Identification: recursive policy shock with the policy variable ordered last.
Results for equation F1
============================================================================
                coefficient       std. error         t-stat          prob
----------------------------------------------------------------------------
const              0.024340         0.040805       0.596496         0.551
L1.F1              0.699803         0.078694       8.892735         0.000
L1.F2              0.055353         0.068085       0.813003         0.416
L1.FFR            -0.001386         0.028840      -0.048049         0.962
...

Correlation matrix of residuals
           F1        F2       FFR
F1   1.000000 -0.266766 -0.145655
F2  -0.266766  1.000000  0.674814
FFR -0.145655  0.674814  1.000000

Examples and Notebook

Run the self-contained script:

python examples/synthetic_demo.py

Open the walkthrough notebook:

notebooks/favar_synthetic_walkthrough.ipynb

The notebook demonstrates:

1. package import and installation check;
2. synthetic macroeconomic data generation;
3. preprocessing and transformations;
4. construction of X, Y, and slow_columns;
5. FAVAR estimation;
6. forecasts with confidence intervals;
7. impulse response functions;
8. panel-projected impulse responses.

Public API

from favar import FAVAR

Main methods:

• FAVAR(...).fit(...): estimate the model.
• results.summary(): print the estimation summary.
• model.select_order(maxlags=12): compare lag orders by information criteria.
• results.forecast(steps, confidence_level=0.95): forecast observed variables.
• results.impulse_response(periods, impulse_size=None): compute system IRFs.
• results.panel_impulse_response(periods, columns=None): compute IRFs projected to the information panel.
• results.plot_acorr(nlags=10): plot residual autocorrelations and cross-correlations.
• results.is_stable(): check dynamic stability of the augmented VAR.

Final Checklist Before Estimation

• X and Y have the same time index.
• No missing values remain after transformations.
• All columns in X and Y are numeric.
• policy_var is a column of Y.
• Every item in slow_columns is a column of X.
• k_factors <= min(T, N).
• The lag order is feasible for the available sample size.
• Transformations are economically justified and documented.

Project Status and Releases

This GitHub repository is the official source repository for the favar PyPI package. Public package releases are available at pypi.org/project/favar, and release history is tracked in CHANGELOG.md.

Development changes should be proposed through GitHub commits and pull requests. See CONTRIBUTING.md for the local development workflow and RELEASING.md for the PyPI release process.

Citation

If you use favar in academic work, please cite:

Siqueira, J. S. (2026). Factor-Augmented Vector Autoregression (FAVAR)
[Python package]. Zenodo. https://doi.org/10.5281/zenodo.21017290

Author ORCID: 0009-0009-8918-8946

DOI: 10.5281/zenodo.21017290

BibTeX:

@software{siqueira_favar_2026,
  author = {Siqueira, J. S.},
  title = {Factor-Augmented Vector Autoregression (FAVAR)},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.21017290},
  url = {https://doi.org/10.5281/zenodo.21017290},
  orcid = {0009-0009-8918-8946}
}

References

Bernanke, B. S., Boivin, J., & Eliasz, P. (2005). Measuring the Effects of Monetary Policy: A Factor-Augmented Vector Autoregressive (FAVAR) Approach. Quarterly Journal of Economics.

Lutkepohl, H. (2005). New Introduction to Multiple Time Series Analysis. Springer.

Seabold, S., & Perktold, J. (2010). statsmodels: Econometric and Statistical Modeling with Python. Proceedings of the 9th Python in Science Conference.

License

MIT.