Bayesian Vector Autoregressions for forecasting

User manual: see the documentation.

Features

• Natural-conjugate setting (Normal-Inverse-Wishart) with Minnesota shrinkage.
• Sum-of-coefficients and single-unit-root dummy observation priors (implemented with dummy observations).
• Hyperparameter optimisation following GLP (2015) but without the MH step.
• Covid dummies.
• Conditional and unconditional forecasting.
• Soft constraints in conditional forecasting, with uncertainty around the constraints.
• Generalised impulse response functions.
• Conditional forecast counterfactual analysis.
• Hyperparameter optimisation via cross-validation; see this doc.
  • Out-of-sample predictive likelihood.
  • Features: optimise hyperparameter to maximise the predictive likelihood at a given horizon and for a subset of targeted series.
• Support for skewed constraints (aka upside and downside risks); see this doc.
• Explicitly accounts for nowcasting uncertainty when exploiting nowcasts; see this doc.
• Unit tests and simulation experiments, including:
  • Simulation comparing the estimated parameters with the true data generating process.
  • Simulation evaluating forecast unbiasedness.
  • Simulation checking the moments of the constrained forecast distribution.

Project Structure

├── src/                        # Source code
├── docs/                       # Documentation and notebooks
├── tests/                      # Automated tests
├── ...

Installation

pip install bvar

Selected documentation

• General framework

Selected notebooks

• Illustration with simulated data

Contributing

• See CONTRIBUTING.md for details on how to contribute to the code.
• Have a question or an idea? Please open an issue.

Main references

• General model specification, priors and marginal likelihood:
  • GLP (2015)
• Implementation (matrix formulation and sampling algorithm):
  • Chan (2020)
• Covid dummies
  • Cascaldi-Garcia (2022)
• Conditional forecasting
  • Waggoner and Zha (1999) - Hard constraints
  • Antolín-Díaz et al. (2021) - Soft constraints
• Generalised IRFs
  • Pesaran and Shin (1998).

Data Classification

Bank of England Data Classification: OFFICIAL BLUE