--- title: "AR(1) stochastic volatility steady-state BVAR" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{AR(1) stochastic volatility steady-state BVAR} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- Here we estimate a steady-state BVAR model with AR(1) stochastic volatility, see `?bvar` for details. We will estimate the model on a quarterly US data set from Koop and Korobilis (2010) on the inflation rate $\Delta \pi_t$ (the annual percentage change in a chain-weighted GDP price index), the unemployment rate $u_t$ (seasonally adjusted civilian unemployment rate, all civilian workers aged 16 years or older) and the interest rate $r_t$ (yield on the three-month Treasury bill rate). The sample is 1953Q1-2006Q3 and we have the data vector $$ y_t = \begin{pmatrix} \Delta \pi_t \\ u_t \\ r_t \end{pmatrix} $$ First, let's load the package, then import and plot the data. ``` r library(SteadyStateBVAR) data("KoopKorobilis2010") yt <- KoopKorobilis2010 plot.ts(yt) ``` ![](figure/AR(1)-1-1.png) Let's create the bvar object which we will use throughout here. ``` r bvar_obj <- bvar(data = yt) ``` We choose 2 lags and only a constant as the deterministic variable. ``` r bvar_obj <- setup(bvar_obj, p=2, deterministic = "constant") ``` We set the overall tightness to $\lambda_1 = 0.20$, cross-equation tightness to $\lambda_2 = 0.50$ and the lag decay rate to $\lambda_3 = 1.00$. For the prior means on the first own lags, we set them to $0.6$ for $\Delta \pi_t$ and $0.9$ for $u_t$ and $r_t$. Note that the prior mean on the first own lag of inflation is set to $0.6$ instead of $0$ to reflect some degree of persistence in the series (even though it is a growth rate variable). ``` r lambda_1 <- 0.20 lambda_2 <- 0.50 lambda_3 <- 1.00 fol_pm=c(0.6, # delta pi 0.9, #u 0.9) #R ``` Now, for the steady-state coefficients we use some toy values (let us pretend that they are expert based). Remember that we only have a constant now, so $q=1$ and therefore $\Psi$ only has one column $\psi_1=\Psi$. Since $d_t = 1 \ \forall \ t$, we have $\Psi d_t = \mu_t$ which simplifies to $\Psi = \mu$ and as such we can directly interpret $\Psi$ as the unconditional mean, i.e. the steady-state. ``` r theta_Psi <- c( ppi(1.90, 2.10, interval=0.95)$mean, #Psi: delta pi ppi(3.80, 4.50, interval=0.95)$mean, #Psi: u ppi(2.60, 3.90, interval=0.95)$mean #Psi: r ) Omega_Psi <- diag( c( ppi(1.90, 2.10, interval=0.95)$var, #Psi: delta pi ppi(3.80, 4.50, interval=0.95)$var, #Psi: u ppi(2.60, 3.90, interval=0.95)$var #Psi: r ) ) ``` Now we need to specify our stochastic volatility priors. See `?priors` for more information about the prior specification. Below I take inspiration from Carriero, Clark, and Marcellino (2024), which uses the exact same AR(1) stochastic volatility specification, but for a conventional BVAR. ``` r k <- bvar_obj$setup$k n_free_params_A <- bvar_obj$setup$n_free_params_A sigma2 <- diag(bvar_obj$setup$Sigma_AR) SV_priors_AR1 <- list( theta_A = rep(0, n_free_params_A), Omega_A = diag(10, n_free_params_A), theta_gamma_0 = 0.1 * log(sigma2), Omega_gamma_0 = diag(2, k), theta_gamma_1 = rep(0.9, k), Omega_gamma_1 = diag(0.04, k), theta_log_lambda_1 = log(sigma2), Omega_log_lambda_1 = diag(2, k), V_Phi = (10 - k - 1) * 0.03 * diag(k), m_Phi = 10 ) ``` Here `sigma2` contains the residual variances from AR($p$) models (the same ones we used in the Minnesota prior). Let's put everything into the `priors()` function. ``` r bvar_obj <- priors(bvar_obj, lambda_1 = lambda_1, lambda_2 = lambda_2, lambda_3 = lambda_3, first_own_lag_prior_mean =fol_pm, theta_Psi = theta_Psi, Omega_Psi = Omega_Psi, SV = TRUE, SV_type = "AR1", SV_priors = SV_priors_AR1) ``` Now we can fit the model. Note that we can use arguments from `rstan::sampling()` such as `control` where we can tweak `max_treedepth` and `adapt_delta`. ``` r bvar_obj <- fit(bvar_obj, H = 40, d_pred = matrix(rep(1, 40)), iter = 4000, warmup = 1000, chains = 2, cores = 2, control = list(max_treedepth = 12, adapt_delta = 0.999)) ``` Now lets see the posterior means ``` r summary(bvar_obj, stat="mean", t = 215) #t = 215 for covariance matrix #> Posterior mean estimates #> ------------------------ #> #> #> beta #> -------------------------------------------------------------------------------- #> delta pi u r #> delta pi.l1 1.27 0.02 0.17 #> u.l1 -0.09 1.17 -0.16 #> r.l1 0.00 -0.01 1.04 #> delta pi.l2 -0.28 0.01 -0.11 #> u.l2 0.07 -0.23 0.17 #> r.l2 -0.01 0.02 -0.11 #> -------------------------------------------------------------------------------- #> #> #> Psi #> -------------------------------------------------------------------------------- #> [,1] #> delta pi 2.00 #> u 4.28 #> r 3.51 #> -------------------------------------------------------------------------------- #> #> #> Sigma_u,t (t = 215) #> -------------------------------------------------------------------------------- #> delta pi u r #> delta pi 0.07 -0.01 0.02 #> u -0.01 0.03 -0.02 #> r 0.02 -0.02 0.17 #> -------------------------------------------------------------------------------- #> #> #> A #> -------------------------------------------------------------------------------- #> delta pi u r #> delta pi 1.00 0.00 0 #> u 0.13 1.00 0 #> r -0.24 0.45 1 #> -------------------------------------------------------------------------------- #> #> #> gamma_0 #> -------------------------------------------------------------------------------- #> delta pi u r #> -0.14 -0.15 -0.09 #> -------------------------------------------------------------------------------- #> #> #> gamma_1 #> -------------------------------------------------------------------------------- #> delta pi u r #> 0.95 0.95 0.94 #> -------------------------------------------------------------------------------- #> #> #> Phi #> -------------------------------------------------------------------------------- #> delta pi u r #> delta pi 0.06 0.04 0.06 #> u 0.04 0.07 0.07 #> r 0.06 0.07 0.14 #> -------------------------------------------------------------------------------- ``` You can always look at the `stanfit` object `bvar_obj$fit$stan` directly if you want. Note that the `z`'s below are not parameters per se, they are simply used in a reparameterization trick to sample the log volatilities more efficiently. ``` r print(bvar_obj$fit$stan) #> Inference for Stan model: steady_state_bvar_AR1_stochastic_volatility. #> 2 chains, each with iter=4000; warmup=1000; thin=1; #> post-warmup draws per chain=3000, total post-warmup draws=6000. #> #> mean se_mean sd 2.5% 25% 50% 75% 97.5% n_eff Rhat #> beta[1,1] 1.27 0.00 0.06 1.16 1.23 1.27 1.31 1.38 5160 1 #> beta[1,2] 0.02 0.00 0.04 -0.05 0.00 0.02 0.05 0.09 5901 1 #> beta[1,3] 0.17 0.00 0.08 0.00 0.11 0.17 0.22 0.33 6117 1 #> beta[2,1] -0.09 0.00 0.04 -0.16 -0.11 -0.09 -0.07 -0.02 5951 1 #> beta[2,2] 1.17 0.00 0.05 1.06 1.13 1.17 1.21 1.27 5008 1 #> beta[2,3] -0.16 0.00 0.08 -0.30 -0.21 -0.16 -0.11 0.00 5132 1 #> beta[3,1] 0.00 0.00 0.02 -0.03 -0.01 0.00 0.01 0.04 5985 1 #> beta[3,2] -0.01 0.00 0.02 -0.04 -0.02 -0.01 0.00 0.02 6203 1 #> beta[3,3] 1.04 0.00 0.06 0.93 1.00 1.04 1.08 1.16 5446 1 #> beta[4,1] -0.28 0.00 0.06 -0.39 -0.32 -0.28 -0.24 -0.16 5177 1 #> beta[4,2] 0.01 0.00 0.04 -0.07 -0.02 0.01 0.03 0.08 6059 1 #> beta[4,3] -0.11 0.00 0.08 -0.27 -0.17 -0.11 -0.05 0.05 6217 1 #> beta[5,1] 0.07 0.00 0.03 0.00 0.05 0.07 0.09 0.14 6188 1 #> beta[5,2] -0.23 0.00 0.05 -0.33 -0.26 -0.23 -0.19 -0.13 4843 1 #> beta[5,3] 0.17 0.00 0.07 0.03 0.12 0.17 0.22 0.31 5318 1 #> beta[6,1] -0.01 0.00 0.02 -0.04 -0.02 -0.01 0.01 0.03 6261 1 #> beta[6,2] 0.02 0.00 0.02 -0.01 0.01 0.02 0.04 0.06 6238 1 #> beta[6,3] -0.11 0.00 0.06 -0.22 -0.14 -0.11 -0.07 0.00 5555 1 #> Psi[1,1] 2.00 0.00 0.05 1.90 1.96 2.00 2.03 2.09 12709 1 #> Psi[2,1] 4.28 0.00 0.18 3.93 4.16 4.28 4.40 4.62 10680 1 #> Psi[3,1] 3.51 0.00 0.33 2.83 3.29 3.52 3.74 4.16 10833 1 #> z[1,1] 0.01 0.00 0.41 -0.76 -0.28 -0.01 0.27 0.84 7481 1 #> z[1,2] 1.25 0.00 0.43 0.43 0.96 1.23 1.54 2.15 7952 1 #> z[1,3] -1.02 0.01 0.53 -2.01 -1.36 -1.04 -0.68 0.06 7495 1 #> z[2,1] 0.00 0.01 0.94 -1.85 -0.63 -0.01 0.63 1.82 11636 1 #> z[2,2] 0.24 0.01 0.97 -1.70 -0.42 0.25 0.89 2.11 10038 1 #> z[2,3] -0.08 0.01 0.98 -1.98 -0.76 -0.08 0.58 1.85 12678 1 #> z[3,1] 0.12 0.01 0.96 -1.75 -0.52 0.13 0.77 2.01 11131 1 #> z[3,2] 0.26 0.01 1.01 -1.75 -0.42 0.28 0.95 2.22 11982 1 #> z[3,3] -0.11 0.01 1.00 -2.09 -0.79 -0.10 0.57 1.83 14312 1 #> z[4,1] -0.38 0.01 0.95 -2.24 -1.04 -0.38 0.29 1.42 11655 1 #> z[4,2] -0.22 0.01 0.97 -2.13 -0.86 -0.22 0.43 1.67 13269 1 #> z[4,3] -0.24 0.01 0.99 -2.18 -0.92 -0.22 0.43 1.71 13366 1 #> z[5,1] -0.25 0.01 0.95 -2.11 -0.90 -0.26 0.39 1.60 11303 1 #> z[5,2] -0.24 0.01 0.99 -2.19 -0.91 -0.24 0.44 1.68 11485 1 #> z[5,3] -0.23 0.01 0.99 -2.15 -0.89 -0.22 0.44 1.65 12966 1 #> z[6,1] -0.15 0.01 0.96 -2.00 -0.81 -0.15 0.49 1.74 14683 1 #> z[6,2] -0.11 0.01 0.97 -1.97 -0.76 -0.11 0.52 1.84 13910 1 #> z[6,3] -0.19 0.01 0.98 -2.15 -0.84 -0.19 0.47 1.74 11176 1 #> z[7,1] 0.08 0.01 0.97 -1.80 -0.57 0.08 0.74 2.00 13398 1 #> z[7,2] 0.03 0.01 0.97 -1.85 -0.62 0.01 0.67 1.92 12425 1 #> z[7,3] -0.14 0.01 0.98 -2.05 -0.80 -0.16 0.53 1.86 13863 1 #> z[8,1] 0.27 0.01 0.92 -1.58 -0.33 0.28 0.89 2.05 11512 1 #> z[8,2] 0.04 0.01 0.99 -1.92 -0.64 0.04 0.73 1.96 12908 1 #> z[8,3] -0.08 0.01 0.99 -2.05 -0.74 -0.08 0.57 1.85 13953 1 #> z[9,1] 0.43 0.01 0.93 -1.38 -0.21 0.43 1.06 2.26 13284 1 #> z[9,2] 0.23 0.01 0.98 -1.67 -0.43 0.22 0.87 2.15 12659 1 #> z[9,3] 0.01 0.01 0.96 -1.80 -0.63 0.02 0.66 1.92 11587 1 #> z[10,1] -0.15 0.01 0.92 -1.96 -0.77 -0.15 0.47 1.69 11896 1 #> z[10,2] 0.13 0.01 0.96 -1.78 -0.53 0.13 0.78 2.03 12824 1 #> z[10,3] -0.14 0.01 0.99 -2.03 -0.80 -0.15 0.52 1.77 12533 1 #> z[11,1] -0.21 0.01 0.93 -2.06 -0.83 -0.21 0.42 1.59 13594 1 #> z[11,2] 0.10 0.01 0.96 -1.75 -0.54 0.09 0.73 2.05 11469 1 #> z[11,3] -0.21 0.01 0.97 -2.17 -0.86 -0.20 0.45 1.66 11544 1 #> z[12,1] -0.32 0.01 0.96 -2.19 -0.97 -0.32 0.32 1.59 13902 1 #> z[12,2] 0.21 0.01 0.96 -1.71 -0.43 0.22 0.85 2.08 13973 1 #> z[12,3] -0.19 0.01 1.00 -2.17 -0.85 -0.18 0.49 1.80 11531 1 #> z[13,1] -0.11 0.01 0.92 -1.94 -0.75 -0.10 0.51 1.69 12618 1 #> z[13,2] 0.37 0.01 0.98 -1.53 -0.25 0.37 1.02 2.33 11793 1 #> z[13,3] -0.10 0.01 0.97 -1.97 -0.77 -0.10 0.55 1.84 11747 1 #> z[14,1] -0.19 0.01 0.97 -2.05 -0.87 -0.18 0.48 1.66 12023 1 #> z[14,2] 0.40 0.01 0.98 -1.50 -0.26 0.40 1.05 2.32 11069 1 #> z[14,3] -0.14 0.01 0.97 -2.06 -0.80 -0.13 0.49 1.81 12887 1 #> z[15,1] -0.14 0.01 0.95 -2.04 -0.80 -0.15 0.50 1.73 12416 1 #> z[15,2] 0.35 0.01 0.97 -1.55 -0.29 0.36 1.00 2.28 13583 1 #> z[15,3] -0.06 0.01 1.00 -2.03 -0.73 -0.05 0.62 1.89 13108 1 #> z[16,1] 0.06 0.01 0.96 -1.83 -0.58 0.05 0.71 1.91 13162 1 #> z[16,2] 0.48 0.01 0.96 -1.43 -0.16 0.48 1.11 2.37 12278 1 #> z[16,3] 0.01 0.01 0.98 -1.86 -0.66 0.02 0.69 1.91 12646 1 #> z[17,1] 0.11 0.01 0.95 -1.78 -0.52 0.11 0.76 1.95 11898 1 #> z[17,2] 0.55 0.01 0.97 -1.34 -0.12 0.54 1.21 2.45 12485 1 #> z[17,3] 0.07 0.01 0.97 -1.82 -0.58 0.07 0.73 1.95 13700 1 #> z[18,1] 0.17 0.01 0.94 -1.67 -0.46 0.16 0.80 2.00 9939 1 #> z[18,2] 0.72 0.01 1.00 -1.30 0.05 0.73 1.37 2.69 10622 1 #> z[18,3] 0.16 0.01 0.99 -1.77 -0.50 0.16 0.83 2.10 12037 1 #> z[19,1] 0.37 0.01 0.96 -1.58 -0.26 0.39 1.03 2.22 9302 1 #> z[19,2] 0.84 0.01 0.97 -1.07 0.20 0.85 1.49 2.70 10167 1 #> z[19,3] 0.16 0.01 0.99 -1.78 -0.51 0.15 0.84 2.08 12696 1 #> z[20,1] 0.13 0.01 0.95 -1.73 -0.50 0.13 0.77 1.98 13397 1 #> z[20,2] 0.51 0.01 0.96 -1.36 -0.14 0.51 1.17 2.41 10028 1 #> z[20,3] 0.13 0.01 0.96 -1.73 -0.51 0.12 0.77 2.02 12549 1 #> z[21,1] -0.07 0.01 0.95 -1.94 -0.71 -0.06 0.58 1.78 11022 1 #> z[21,2] 0.09 0.01 0.98 -1.84 -0.57 0.07 0.77 2.02 13856 1 #> z[21,3] 0.17 0.01 0.97 -1.73 -0.50 0.17 0.84 2.05 15289 1 #> z[22,1] 0.16 0.01 0.94 -1.67 -0.46 0.15 0.79 2.01 11092 1 #> z[22,2] 0.26 0.01 0.97 -1.61 -0.39 0.27 0.93 2.15 11579 1 #> z[22,3] 0.28 0.01 0.99 -1.67 -0.38 0.28 0.94 2.22 12314 1 #> z[23,1] -0.24 0.01 0.96 -2.15 -0.87 -0.24 0.39 1.62 13287 1 #> z[23,2] -0.03 0.01 1.00 -1.96 -0.71 -0.04 0.64 1.92 13724 1 #> z[23,3] -0.11 0.01 0.99 -2.05 -0.76 -0.11 0.55 1.83 13234 1 #> z[24,1] -0.22 0.01 0.95 -2.09 -0.86 -0.22 0.41 1.68 13098 1 #> z[24,2] 0.11 0.01 0.95 -1.75 -0.52 0.10 0.75 1.98 11951 1 #> z[24,3] -0.04 0.01 0.97 -1.95 -0.70 -0.03 0.61 1.87 11868 1 #> z[25,1] -0.11 0.01 0.94 -1.93 -0.73 -0.12 0.52 1.74 12240 1 #> z[25,2] 0.14 0.01 0.97 -1.76 -0.51 0.14 0.79 2.03 13125 1 #> z[25,3] 0.06 0.01 0.97 -1.83 -0.59 0.07 0.69 1.99 11364 1 #> z[26,1] 0.21 0.01 0.95 -1.67 -0.44 0.21 0.86 2.06 11865 1 #> z[26,2] 0.31 0.01 0.95 -1.59 -0.30 0.30 0.94 2.19 12368 1 #> z[26,3] 0.16 0.01 0.98 -1.74 -0.50 0.16 0.82 2.10 13009 1 #> z[27,1] -0.17 0.01 0.92 -1.92 -0.80 -0.17 0.46 1.62 10960 1 #> [ reached 'max' / getOption("max.print") -- omitted 3783 rows ] #> #> Samples were drawn using NUTS(diag_e) at Tue Jul 28 05:20:52 2026. #> For each parameter, n_eff is a crude measure of effective sample size, #> and Rhat is the potential scale reduction factor on split chains (at #> convergence, Rhat=1). ``` We can forecast ``` r forecast(bvar_obj, pi = 0.68, show_all = TRUE) ``` ![](figure/AR(1)-2-1.png)![](figure/AR(1)-2-2.png)![](figure/AR(1)-2-3.png) Let us plot the log volatility estimates and predictions ``` r stochastic_volatility_plot(bvar_obj, ci = 0.95, vol = "log_lambda") ``` ![](figure/AR(1)-3-1.png)![](figure/AR(1)-3-2.png)![](figure/AR(1)-3-3.png) Let us plot the estimates and predictions of the implied innovation standard deviations ``` r stochastic_volatility_plot(bvar_obj, vol = "sd") ``` ![](figure/AR(1)-4-1.png)![](figure/AR(1)-4-2.png)![](figure/AR(1)-4-3.png) We can also produce orthogonalized IRFs ``` r IRF(bvar_obj, method = "OIRF", t=215, ci=0.68) #latest t ``` ![](figure/AR(1)-5 -1.png) ## References Carriero, A., Clark, T. E., and Marcellino, M. (2024). Capturing macro-economic tail risks with Bayesian vector autoregressions. *Journal of Money, Credit and Banking*, 56(5), pp. 1099–1127. Koop, G. and Korobilis, D. (2010). Bayesian multivariate time series methods for empirical macroeconomics. *Foundations and Trends in Econometrics*, 3(4), pp. 267–358. Villani, M. (2009). Steady-state priors for vector autoregressions. *Journal of Applied Econometrics*, 24(4), pp. 630–650.