Ardia-Bluteau-Boudt-Catania-Trottier (2019) Markov-Switching GARCH Models in R: The MSGARCH Package

Summary

This paper documents MSGARCH, an R package (with an efficient C++ object-oriented backend) that implements Markov-switching GARCH models. It lets the user simulate from, and estimate — by both maximum likelihood and Bayesian Markov chain Monte Carlo — a very large class of single-regime and Markov-switching GARCH-type models, combining several conditional-variance specifications with several conditional distributions. It provides single- and multi-step-ahead forecasts of the complete conditional density, and risk-management tools for conditional volatility, value-at-risk (VaR), and expected shortfall (ES), with backtesting. The functionality is illustrated on exchange-rate and stock-market return data.

Key Claims

• A broad model class. MSGARCH pairs regime-specific conditional-variance dynamics — sGARCH, eGARCH, gjrGARCH, tGARCH — with a menu of innovation distributions (normal, Student-t, GED, and their skewed variants), for any number of regimes; single-regime GARCH is the one-state special case.
• Two estimation paradigms. Models can be fit by maximum likelihood (with numerically robust C++ likelihood evaluation) or by Bayesian adaptive MCMC, delivering full posterior uncertainty for parameters and derived quantities.
• Path-independent formulation. To keep the likelihood tractable, the package uses the regime-specific (path-independent) Markov-switching GARCH formulation, so estimation avoids the exponential path-dependence of the natural specification.
• Density forecasting and risk. Beyond point volatility, MSGARCH produces the full predictive density at multiple horizons and computes VaR and expected shortfall, with tools to backtest these risk measures.
• Reproducible, tested software. The package is designed for practitioners: consistent S3 methods (fit, simulate, predict, risk), documentation, and worked FX/equity examples.

Concepts Introduced or Extended

• Markov-Switching GARCH
• GARCH and BEKK-GARCH

Entities Mentioned

• David Ardia
• Kris Boudt

Quotes

"The package MSGARCH allows the user to perform simulations as well as maximum likelihood and Bayesian Markov chain Monte Carlo estimations of a very large class of Markov-switching GARCH-type models."

My Take

This is the software that turned Markov-switching GARCH from a methods-paper technique into something a practitioner can actually run. Its most consequential design choice is using the path-independent regime-specific formulation (à la Haas–Mittnik–Paolella) rather than the natural path-dependent model of Bauwens–Preminger–Rombouts — that is exactly what makes both the ML likelihood and the MCMC tractable at package scale, and it's the practical resolution of the path-dependence problem. Offering ML and Bayesian estimation in one interface, plus full predictive densities and VaR/ES backtesting, makes it a natural companion to the theory pages and a bridge to the applied risk-management literature. As a Journal-of-Statistical-Software artifact its value is reproducibility rather than a new result, but that is precisely what the methodology needed.