bsvars-package: Bayesian Estimation of Structural Vector Autoregressive Models

Description

Provides fast and efficient procedures for Bayesian analysis of Structural Vector Autoregressions. This package estimates a wide range of models, including homo-, heteroskedastic and non-normal specifications. Structural models can be identified by adjustable exclusion restrictions, time-varying volatility, or non-normality. They all include a flexible three-level equation-specific local-global hierarchical prior distribution for the estimated level of shrinkage for autoregressive and structural parameters. Additionally, the package facilitates predictive and structural analyses such as impulse responses, forecast error variance and historical decompositions, forecasting, verification of heteroskedasticity and hypotheses on autoregressive parameters, and analyses of structural shocks, volatilities, and fitted values. Beautiful plots, informative summary functions, and extensive documentation complement all this. The implemented techniques align closely with those presented in Lütkepohl, Shang, Uzeda, & Woźniak (2024) <doi:10.48550/arXiv.2404.11057>, Lütkepohl & Woźniak (2020) <doi:10.1016/j.jedc.2020.103862>, Song & Woźniak (2021) <doi:10.1093/acrefore/9780190625979.013.174>, and Woźniak & Droumaguet (2015) <doi:10.13140/RG.2.2.19492.55687>.

Arguments

Author

Tomasz Woźniak wozniak.tom@pm.me

Details

Models. All the SVAR models in this package are specified by two equations, including the reduced form equation: $$Y = AX + E$$ where $Y$ is an NxT matrix of dependent variables, $X$ is a KxT matrix of explanatory variables, $E$ is an NxT matrix of reduced form error terms, and $A$ is an NxK matrix of autoregressive slope coefficients and parameters on deterministic terms in $X$.

The structural equation is given by: $$BE = U$$ where $U$ is an NxT matrix of structural form error terms, and $B$ is an NxN matrix of contemporaneous relationships.

Finally, all of the models share the following assumptions regarding the structural shocks U, namely, joint conditional normality given the past observations collected in matrix X, and temporal and contemporaneous independence. The latter implies zero correlations and autocorrelations.

The various SVAR models estimated differ by the specification of structural shocks variances. The different models include:

homoskedastic model with unit variances
heteroskedastic model with stationary Markov switching in the variances
heteroskedastic model with non-centred Stochastic Volatility process for variances
heteroskedastic model with centred Stochastic Volatility process for variances
non-normal model with a finite mixture of normal components and component-specific variances
heteroskedastic model with sparse Markov switching in the variances where the number of heteroskedastic components is estimated
non-normal model with a sparse mixture of normal components and component-specific variances where the number of heteroskedastic components is estimated

Prior distributions. All the models feature a Minnesota prior for autoregressive parameters in matrix $A$ and a generalised-normal distribution for the structural matrix $B$. Both of these distributions feature a 3-level equation-specific local-global hierarchical prior that make the shrinkage estimation flexible improving the model fit and its forecasting performance.

Estimation algorithm. The models are estimated using frontier numerical methods making the Gibbs sampler fast and efficient. The sampler of the structural matrix follows Waggoner & Zha (2003), whereas that for autoregressive parameters follows Chan, Koop, Yu (2022). The specification of Markov switching heteroskedasticity is inspired by Song & Woźniak (2021), and that of Stochastic Volatility model by Kastner & Frühwirth-Schnatter (2014). The estimation algorithms for particular models are scrutinised in Lütkepohl, Shang, Uzeda, & Woźniak (2024) and Woźniak & Droumaguet (2024) and some other inferential and identification problems are considered in Lütkepohl & Woźniak (2020).

References

Chan, J.C.C., Koop, G, and Yu, X. (2024) Large Order-Invariant Bayesian VARs with Stochastic Volatility. Journal of Business & Economic Statistics, 42, tools:::Rd_expr_doi("10.1080/07350015.2023.2252039").

Kastner, G. and Frühwirth-Schnatter, S. (2014) Ancillarity-Sufficiency Interweaving Strategy (ASIS) for Boosting MCMC Estimation of Stochastic Volatility Models. Computational Statistics & Data Analysis, 76, 408--423, tools:::Rd_expr_doi("10.1016/j.csda.2013.01.002").

Lütkepohl, H., Shang, F., Uzeda, L., and Woźniak, T. (2024) Partial Identification of Heteroskedastic Structural VARs: Theory and Bayesian Inference. University of Melbourne Working Paper, 1--57, tools:::Rd_expr_doi("10.48550/arXiv.2404.11057").

Lütkepohl, H., and Woźniak, T., (2020) Bayesian Inference for Structural Vector Autoregressions Identified by Markov-Switching Heteroskedasticity. Journal of Economic Dynamics and Control 113, 103862, tools:::Rd_expr_doi("10.1016/j.jedc.2020.103862").

Song, Y., and Woźniak, T. (2021) Markov Switching Heteroskedasticity in Time Series Analysis. In: Oxford Research Encyclopedia of Economics and Finance. Oxford University Press, tools:::Rd_expr_doi("10.1093/acrefore/9780190625979.013.174").

Waggoner, D.F., and Zha, T., (2003) A Gibbs sampler for structural vector autoregressions. Journal of Economic Dynamics and Control, 28, 349--366, tools:::Rd_expr_doi("10.1016/S0165-1889(02)00168-9").

Woźniak, T., and Droumaguet, M., (2024) Bayesian Assessment of Identifying Restrictions for Heteroskedastic Structural VARs.

Examples

Run this code

# upload data
data(us_fiscal_lsuw)    # upload dependent variables
data(us_fiscal_ex)      # upload exogenous variables

# specify the model and set seed
set.seed(123)
specification  = specify_bsvar_sv$new(us_fiscal_lsuw, p = 4, exogenous = us_fiscal_ex)

# run the burn-in
burn_in        = estimate(specification, 10)

# estimate the model
posterior      = estimate(burn_in, 20)

# compute impulse responses 2 years ahead
irf           = compute_impulse_responses(posterior, horizon = 8)

# compute forecast error variance decomposition 2 years ahead
fevd           = compute_variance_decompositions(posterior, horizon = 8)

# workflow with the pipe |>
############################################################
set.seed(123)
us_fiscal_lsuw |>
  specify_bsvar_sv$new(p = 4, exogenous = us_fiscal_ex) |>
  estimate(S = 10) |> 
  estimate(S = 20) |> 
  compute_variance_decompositions(horizon = 8) -> fevds