[,1] [,2] [,3] [,4]
[1,] 1 NA NA NA
[2,] NA NA NA NA
[3,] -1 NA NA NA
[4,] 1 NA NA NA
Di Cook Award Presentation
central banks of Switzerland, Italy, Canada, Basil, and Mexico
postgraduate students
anonymous internet users
VAR model: \[\begin{align*} y_t &= A_1 y_{t-1} + \ldots + A_p y_{t-p} + A_dd_t + \varepsilon_t\\ y_t &= Ax_t + \varepsilon_t \end{align*}\]
Structural VAR (SVAR) identifies well-isolated structural shocks: \[u_t=B\varepsilon_t,\quad \text{var}(u_t)=I\]
SVAR is a simultaneous equations model: \[\begin{align*} y_t &= Ax_t + B^{-1}u_t\\ By_t &= BAx_t + u_t \end{align*}\]
Commonly used in empirical macroeconomics
Can’t identify \(A\) and \(B\) due to simultaneity, SIGN restriction helps!1
An Australian example2, consider 4 variables: cash rate, real GDP, trimmed-mean CPI, trade-weighted exchange rate
To identify positive monetary policy shock, may restrict \(B^{-1}\) s.t. \[\begin{align*} y_t &= \dots + B^{-1}u_t\\ \begin{bmatrix} \text{CASH}_t\\\text{GDP}_t\\\text{CPI}_t\\\text{TWI}_t \end{bmatrix} &= \dots+ \begin{bmatrix} +&*&*&*\\*&*&*&*\\-&*&*&*\\+&*&*&* \end{bmatrix} \begin{bmatrix} u_{1t}^{\text{MP}}\\u_{2t}\\u_{3t}\\u_{4t} \end{bmatrix} \end{align*}\]
\(B^{-1}\) are the impulse responses at horizon 0, can extend to futher horizons
The package alo supports restrictrions on \(B\) i.e. contemporaneous relations \[\begin{align*} By_t&=\dots\\ \begin{bmatrix} b_{11}&b_{12}&b_{13}&b_{14}\\ b_{21}&b_{22}&b_{23}&b_{24}\\ b_{31}&b_{32}&b_{33}&b_{34}\\ b_{41}&b_{42}&b_{43}&b_{44} \end{bmatrix} \begin{bmatrix} \text{CASH}_t \\\text{GDP}_t \\\text{CPI}_t \\\text{TWI}_t \end{bmatrix} &=\dots \end{align*}\]
Where the first row can be interpreted as policy reaction function \[b_{11}\text{CASH}_t=-b_{12}\text{GDP}_t-b_{13}\text{CPI}_t-b_{14}\text{TWI}_t+\dots\]
Economic theory says \[\begin{align*} \begin{bmatrix} +&-&-&+\\ *&*&*&*\\*&*&*&*\\*&*&*&* \end{bmatrix} \begin{bmatrix} \text{CASH}_t \\\text{GDP}_t \\\text{CPI}_t \\\text{TWI}_t \end{bmatrix} &=\dots \end{align*}\]
[,1] [,2] [,3] [,4]
[1,] 1 -1 -1 1
[2,] NA NA NA NA
[3,] NA NA NA NA
[4,] NA NA NA NA
# specify the model
spec <- specify_bsvarSIGN$new(
Y,
p = 4,
exogenous = Z,
sign_irf = sign_irf,
sign_structural = sign_structural
)
# sample posterior draws
post <- estimate(spec, S = 5000, show_progress = FALSE)
# compute impulse response functions
irf <- compute_impulse_responses(post, horizon = 20)
Recall, \[\varepsilon_t=B^{-1}u_t,\quad \text{var}(u_t)=I\]
Let \(\Sigma=\text{var}(\varepsilon_t)\) and \(Q\) be some orthogonal matrix, can decompose \[B^{-1}=\text{chol}(\Sigma)Q\]
Sample \(Q\) uniformly by QR decomposition of \(X\), where \(X_{ij}\overset{i.i.d}{\sim}N(0,1)\)1 \[X=QR\]
Until the SIGN restrictions for \(B\) and \(B^{-1}\) are satisfied
**************************************************|
bsvars: Bayesian Structural Vector Autoregressions|
**************************************************|
Posterior summary of impulse responses |
**************************************************|
mean sd 5% quantile 95% quantile
0 -0.1837484 0.1438489 -0.4295756 0.0254521230
1 -0.1919969 0.1462399 -0.4442807 0.0241581974
2 -0.2125791 0.1469114 -0.4621927 0.0112834321
3 -0.2323939 0.1460791 -0.4757095 0.0008290802
4 -0.2528213 0.1488622 -0.4977457 -0.0104613974
5 -0.2637240 0.1525702 -0.5153036 -0.0134905369
6 -0.2664673 0.1563720 -0.5256955 -0.0122216476
7 -0.2622209 0.1593557 -0.5315106 -0.0038934003
8 -0.2547855 0.1615475 -0.5298977 0.0063592121
9 -0.2457313 0.1631078 -0.5232646 0.0128340909
10 -0.2362044 0.1642777 -0.5143287 0.0254801247
11 -0.2267619 0.1652260 -0.5048653 0.0374801324
12 -0.2177028 0.1660716 -0.4972879 0.0423198619
13 -0.2091278 0.1668833 -0.4949965 0.0501458378
14 -0.2010556 0.1677055 -0.4830306 0.0564884001
15 -0.1934627 0.1685626 -0.4773320 0.0621637593
16 -0.1863216 0.1694680 -0.4703457 0.0667087572
17 -0.1796096 0.1704267 -0.4640941 0.0726520546
18 -0.1733106 0.1714390 -0.4596467 0.0781103445
19 -0.1674123 0.1725033 -0.4538969 0.0854539864
20 -0.1619036 0.1736181 -0.4535250 0.0896760294
Hyperparameter estimation via adaptive Metropolis1
Zero and sign restrictions2
Narrative restrictions3
All restrictions at once4
adamwang15[at]gmail.com
adamwang15
adamwang15
adamwang15.bsky.social