alexandria-python 3.0.0

a software for Bayesian vector autoregressions and other Bayesian time-series applications

Alexandria

Alexandria is a Python package for Bayesian time-series econometrics applications. This is the second official release of the software, which introduces Bayesian vector autorgressions.

This is version 3.0, which includes Bayesian regression, Bayesian vector autoregression, Bayesian VEC/VARMA models, and Bayesian nowcasting.

Alexandria offers a range of Bayesian linear regression models:

• maximum likelihood / OLS regression (non-Bayesian)
• simple Bayesian regression
• hierarchical (natural conjugate) Bayesian regression
• independent Bayesian regression with Gibbs sampling
• heteroscedastic Bayesian regression
• autocorrelated Bayesian regression

Alexandria also offers a large number of Bayesian vector autoregression models and applications:

• maximum likelihood (OLS) VAR
• Litterman Minnesota prior
• normal-Wishart prior
• independent prior with Gibbs sampling
• dummy observation prior
• large Bayeisian VAR prior
• Bayesian oxy-SVAR

prior customization:

• constrained coefficients
• dummy extensions (sums-of-coefficients, initial observation,long-run prior)
• stationary priors
• hyperparameter optimization from marginal likelihood

structural identification:

• Cholesky
• triangular factorization
• restrictions: sign and zero restrictions on IRFs, narrative on shocks and historical decomposition

applications:

• forecasts
• impulse response function
• forecast error variance decomposition
• historical decomposition
• conditional forecasts (agnostic and sctructural approaches, allowing for hard and soft conditions)

Alexandria also includes Bayesian VEC and VARMA models, along with many applications:

• Bayesian VEC: uninformative, horseshoe and selection priors; general and reduced-rank approaches
• Bayesian VARMA: Minnesota prior on autoegressive and lag coefficients; residuals estimated from Bayesian state-space modelling
• structural identification and applications are the same as the Bayesian VAR models

The current version introduces Bayesian nowcasting models, along with many applications:

• Mixed Frequency Bayesian VAR: allows for mixed frequency, missing observations, and frequency decomposition; all BVAR applications are available with the MF-BVAR.
• Bayesian Dynamic Factor Model: extract structural factors from high and low frequency features; nowcasts in real time; also permits structural IRFs, variance and historical decomposition.
• Bayesian MIDA regression: 3 priors (Minnesota, Horseshoe, Bayesian lasso) and 3 representations (unrestricted MIDAS, Almon, Fourier).

Alexandria is user-friendly and can be used from a simple Graphical User Inteface (GUI). More experienced users can also run the models directly from the Python console by using the model classes and methods.

===============================

Installing Alexandria

Alexandria can be installed from pip:

pip install alexandria-python

A local installation can also obtain by copy-pasting the folder containing the toolbox programmes. The folder can be downloaded from the project website or Github repo:
https://alexandria-toolbox.github.io
https://github.com/alexandria-toolbox

===============================

Getting started

Simple Python example:

# imports
from alexandria import NormalWishartBayesianVar
from alexandria import DataSets
from alexandria import Graphics
import numpy as np

# load ISLM dataset
ds = DataSets()
islm_data = ds.load_islm()[:,:4]

# create and train Bayesian VAR with default settings
var = NormalWishartBayesianVar(endogenous = islm_data)
var.estimate()

# estimate forecasts for the next 4 periods, 60% credibility level
forecast_estimates = var.forecast(4, 0.6)

# create graphics of predictions
gp = Graphics(var)
gp.forecast_graphics(show=True, save=False)

===============================

Documentation

Complete manuals and user guides can be found on the project website and Github repo:
https://alexandria-toolbox.github.io
https://github.com/alexandria-toolbox

===============================

Contact

[e-mail]