1Why did Sims (1980) advocate VAR models?
1. Traditional macroeconometric models imposed many untested exclusion restrictions.
2. VARs allow each variable to depend on the lags of every variable in the system, imposing fewer a priori restrictions.
2What is the Lucas critique of reduced-form econometric models?
1. Estimated coefficients may depend on the prevailing policy regime.
2. A policy change can alter agents' behaviour and therefore the coefficients, making historical relationships unreliable for policy evaluation.
3What is a structural VAR($p$)?
1. $\mathbf{A}_0\mathbf{Y}_t=\sum_{j=1}^{p}\mathbf{A}_j\mathbf{Y}_{t-j}+\boldsymbol{\epsilon}_t$.
2. $\mathbf{Y}_t$ is an $n\times1$ vector of endogenous variables.
3. $\boldsymbol{\epsilon}_t$ contains serially uncorrelated structural shocks.
4How is the reduced-form VAR($p$) derived?
1. Assume $\mathbf{A}_0$ is invertible.
2. $\mathbf{Y}_t=\sum_{j=1}^{p}\mathbf{A}_0^{-1}\mathbf{A}_j\mathbf{Y}_{t-j}+\mathbf{A}_0^{-1}\boldsymbol{\epsilon}_t$.
3. Define $\mathbf{C}_j=\mathbf{A}_0^{-1}\mathbf{A}_j$ and $\mathbf{u}_t=\mathbf{A}_0^{-1}\boldsymbol{\epsilon}_t$.
4. $\mathbf{Y}_t=\sum_{j=1}^{p}\mathbf{C}_j\mathbf{Y}_{t-j}+\mathbf{u}_t$.
5What can be estimated directly from a reduced-form VAR?
1. The coefficient matrices $\mathbf{C}_1,\ldots,\mathbf{C}_p$.
2. $\boldsymbol{\Sigma}_u=\operatorname{Cov}(\mathbf{u}_t)=\mathbf{A}_0^{-1}\boldsymbol{\Sigma}_{\epsilon}(\mathbf{A}_0^{-1})^{\top}$.
3. Recovering structural parameters requires additional identifying restrictions.
6How should variables be transformed before estimating a VAR?
1. Test for unit roots using ADF tests.
2. Stationary variables can be modelled in levels.
3. $I(1)$ variables without cointegration can be modelled in first differences.
4. Cointegrated $I(1)$ variables should be modelled using a VECM.
7What is the stability condition for a VAR($p$)?
1. The roots of $\det(\mathbf{I}_n-\mathbf{C}_1z-\cdots-\mathbf{C}_pz^p)=0$ lie outside the unit circle.
2. Equivalently, the companion matrix has eigenvalues with modulus smaller than $1$.
8How is a reduced-form VAR estimated and its lag order selected?
1. Estimate each equation separately by OLS, since the equations contain identical regressors.
2. Select the lag order using information criteria or nested lag-order tests.
3. Check residuals for remaining serial correlation.
9How are nested lag orders tested in a VAR?
1. To compare VAR(3) against VAR(2), estimate VAR(3) and test $H_0:\mathbf{C}_3=\mathbf{0}$.
2. To compare VAR(4) against VAR(3), estimate VAR(4) and test $H_0:\mathbf{C}_4=\mathbf{0}$.
3. Each test imposes $n^2$ restrictions, examined using an appropriate joint $F$-test or asymptotic Wald test with $\chi^2_{n^2}$ distribution.
10How are the coefficients of a reduced-form VAR interpreted?
1. Own-variable lags capture persistence.
2. Cross-variable lags capture conditional predictive relationships.
3. These coefficients do not establish structural causality.
11How is Granger causality tested in a bivariate VAR($p$)?
1. Test whether all $p$ lagged coefficients of $X_t$ in the $Y_t$ equation equal zero.
2. Repeat for lagged $Y_t$ in the $X_t$ equation.
3. Rejecting both nulls implies two-way Granger causality, not one-way causality.
12How are forecasts produced from a VAR(1)?
1. Suppose $\mathbf{Y}_t=\mathbf{C}\mathbf{Y}_{t-1}+\mathbf{u}_t$.
2. $\hat{\mathbf{Y}}_{T+1|T}=\mathbf{C}\mathbf{Y}_T$.
3. $\hat{\mathbf{Y}}_{T+h|T}=\mathbf{C}^h\mathbf{Y}_T$.
4. With an intercept $\mathbf{c}$, iterate $\hat{\mathbf{Y}}_{T+h|T}=\mathbf{c}+\mathbf{C}\hat{\mathbf{Y}}_{T+h-1|T}$.
13How are level forecasts recovered from forecasts of first differences?
1. $X_{T+h}=X_T+\sum_{j=1}^{h}\Delta X_{T+j}$.
2. $\hat X_{T+h|T}=X_T+\sum_{j=1}^{h}\widehat{\Delta X}_{T+j|T}$.
14What is an impulse response function (IRF)?
1. It measures the effect of a shock at time $t$ on a variable at time $t+h$, holding the history and subsequent shocks fixed.
2. $IRF(h)=\mathbb{E}[Y_{t+h}|\epsilon_t=\delta]-\mathbb{E}[Y_{t+h}|\epsilon_t=0]$, under identical other conditions.
15What is the IRF of an AR(1) process?
1. $Y_t=\rho Y_{t-1}+\epsilon_t$.
2. For a unit shock, $IRF(h)=\rho^h$.
16What is the IRF of an MA(1) process?
1. $Y_t=\epsilon_t+\theta\epsilon_{t-1}$.
2. For a unit shock, $IRF(0)=1$, $IRF(1)=\theta$, and $IRF(h)=0$ for $h\geq2$.
17How do the IRFs of an $I(1)$ process differ between first differences and levels?
1. If $\Delta Y_t=c(L)\epsilon_t$, then $IRF_{\Delta Y}(h)=c_h$.
2. $IRF_Y(h)=\sum_{j=0}^{h}c_j$.
3. The permanent effect is $\sum_{j=0}^{\infty}c_j$, when the sum converges.
18How is the moving-average representation of a stationary VAR(1) derived?
1. Start with $\mathbf{Y}_t=\mathbf{C}\mathbf{Y}_{t-1}+\mathbf{u}_t$.
2. Repeated substitution gives $\mathbf{Y}_t=\mathbf{u}_t+\mathbf{C}\mathbf{u}_{t-1}+\mathbf{C}^2\mathbf{u}_{t-2}+\cdots$.
3. $\mathbf{Y}_t=\sum_{h=0}^{\infty}\mathbf{C}^h\mathbf{u}_{t-h}$.
4. The reduced-form IRF matrix at horizon $h$ is $\mathbf{C}^h$.
19How are reduced-form IRFs calculated for a VAR($p$)?
1. Write $\mathbf{Y}_t=\sum_{h=0}^{\infty}\boldsymbol{\Psi}_h\mathbf{u}_{t-h}$.
2. $\boldsymbol{\Psi}_0=\mathbf{I}_n$.
3. For $h\geq1$, $\boldsymbol{\Psi}_h=\sum_{j=1}^{\min(p,h)}\mathbf{C}_j\boldsymbol{\Psi}_{h-j}$.
4. $[\boldsymbol{\Psi}_h]_{ij}$ gives the response of variable $i$ to reduced-form shock $j$ at horizon $h$.
20Why are reduced-form IRFs not necessarily structural IRFs?
1. Reduced-form shocks may be contemporaneously correlated.
2. They can therefore be mixtures of different underlying structural shocks.
3. Recovering structural IRFs requires identifying the structural shocks.
21How are structural IRFs obtained in a VAR(1)?
1. $\mathbf{u}_t=\mathbf{A}_0^{-1}\boldsymbol{\epsilon}_t$.
2. $\mathbf{Y}_t=\sum_{h=0}^{\infty}\mathbf{C}^h\mathbf{A}_0^{-1}\boldsymbol{\epsilon}_{t-h}$.
3. The structural IRF matrix for a unit shock is $\boldsymbol{\Theta}_h=\mathbf{C}^h\mathbf{A}_0^{-1}$.
4. For a shock of size $\delta$, multiply the corresponding response by $\delta$.
22How many restrictions are needed to identify a structural VAR?
1. The structural model contains $n^2$ more parameters than the reduced form.
2. Normalising the diagonal of $\mathbf{A}_0$ gives $n$ restrictions.
3. Assuming uncorrelated structural shocks gives $\frac{n(n-1)}{2}$ restrictions.
4. A further $\frac{n(n-1)}{2}$ restrictions are required.
23How are contemporaneous structural effects derived in a bivariate VAR?
1. Let $\mathbf{A}_0=\begin{pmatrix}1&a_{12}\\a_{21}&1\end{pmatrix}$.
2. $\mathbf{A}_0^{-1}=\frac{1}{1-a_{12}a_{21}}\begin{pmatrix}1&-a_{12}\\-a_{21}&1\end{pmatrix}$.
3. The effect of shock $1$ on variable $2$ at horizon $0$ is $-\frac{a_{21}}{1-a_{12}a_{21}}$.
4. The effect of shock $2$ on variable $1$ at horizon $0$ is $-\frac{a_{12}}{1-a_{12}a_{21}}$.
24How do zero contemporaneous effects translate into restrictions on $\mathbf{A}_0$?
1. If shock $1$ cannot affect variable $2$ contemporaneously, impose $a_{21}=0$.
2. If shock $2$ cannot affect variable $1$ contemporaneously, impose $a_{12}=0$.
25How does Sims' causal ordering identify a bivariate SVAR?
1. Assume structural shocks are uncorrelated.
2. Impose $a_{21}=0$, giving $\mathbf{A}_0=\begin{pmatrix}1&a_{12}\\0&1\end{pmatrix}$.
3. Structural shock $1$ cannot affect variable $2$ contemporaneously, but may affect it with a lag.
4. With normalisation and shock orthogonality, the model is just-identified.
26How is $a_{12}$ recovered under Sims' causal ordering?
1. With $a_{21}=0$, $u_{1t}=\epsilon_{1t}-a_{12}\epsilon_{2t}$ and $u_{2t}=\epsilon_{2t}$.
2. By shock orthogonality, $\sigma_{u,12}=-a_{12}\sigma_{\epsilon,2}^2=-a_{12}\sigma_{u,22}$.
3. $a_{12}=-\frac{\sigma_{u,12}}{\sigma_{u,22}}$.
27How are the remaining structural parameters recovered under Sims' causal ordering?
1. $\sigma_{\epsilon,2}^2=\sigma_{u,22}$.
2. $\sigma_{\epsilon,1}^2=\sigma_{u,11}-\frac{\sigma_{u,12}^2}{\sigma_{u,22}}$.
3. Recover the structural lag matrices using $\mathbf{A}_j=\mathbf{A}_0\mathbf{C}_j$.
28What is the principal limitation of Sims' causal ordering?
1. It relies on an economic assumption about contemporaneous causal effects.
2. Results may depend on variable ordering and sampling frequency.
3. A just-identifying restriction cannot be tested against the reduced-form data alone.
29How does Blanchard–Quah identification differ from Sims' approach?
1. Sims restricts contemporaneous responses, while Blanchard–Quah restricts long-run responses.
2. For a VAR in $\Delta X_t$, a structural shock with no permanent effect on $X_t$ satisfies $\sum_{h=0}^{\infty}[\boldsymbol{\Theta}_h]_{X,j}=0$.
3. This can identify a transitory demand shock separately from a permanent supply shock.
4. A just-identifying long-run restriction cannot itself be tested from the reduced-form data alone.
30Why is a VAR in first differences inappropriate for cointegrated variables?
1. Differencing removes information about the long-run equilibrium relationship.
2. A differenced VAR omits the stationary cointegrating residual.
3. A VECM includes this residual alongside short-run changes.
31How is a VAR(1) rewritten as a VECM?
1. Start with $\mathbf{Y}_t=\mathbf{A}\mathbf{Y}_{t-1}+\boldsymbol{\epsilon}_t$.
2. Subtract $\mathbf{Y}_{t-1}$: $\Delta\mathbf{Y}_t=(\mathbf{A}-\mathbf{I}_n)\mathbf{Y}_{t-1}+\boldsymbol{\epsilon}_t$.
3. Define $\boldsymbol{\Pi}=\mathbf{A}-\mathbf{I}_n$.
4. $\Delta\mathbf{Y}_t=\boldsymbol{\Pi}\mathbf{Y}_{t-1}+\boldsymbol{\epsilon}_t$.
32How is a general VAR($p$) rewritten as a VECM?
1. For VAR(2), $\Delta\mathbf{Y}_t=(\mathbf{A}_1+\mathbf{A}_2-\mathbf{I}_n)\mathbf{Y}_{t-1}-\mathbf{A}_2\Delta\mathbf{Y}_{t-1}+\boldsymbol{\epsilon}_t$.
2. Generally, $\boldsymbol{\Pi}=\sum_{j=1}^{p}\mathbf{A}_j-\mathbf{I}_n$.
3. $\Delta\mathbf{Y}_t=\mathbf{c}+\boldsymbol{\Pi}\mathbf{Y}_{t-1}+\sum_{j=1}^{p-1}\boldsymbol{\Gamma}_j\Delta\mathbf{Y}_{t-j}+\boldsymbol{\epsilon}_t$.
33What does the rank of $\boldsymbol{\Pi}$ imply in a VECM with $n$ variables?
1. $\operatorname{rank}(\boldsymbol{\Pi})=0$: no cointegrating relationships.
2. $0<\operatorname{rank}(\boldsymbol{\Pi})=r<n$: $r$ cointegrating relationships and $n-r$ common stochastic trends.
3. $\operatorname{rank}(\boldsymbol{\Pi})=n$: the variables are stationary under the standard stable-system assumptions.
34How is $\boldsymbol{\Pi}$ decomposed when the variables are cointegrated?
1. $\boldsymbol{\Pi}=\boldsymbol{\alpha}\boldsymbol{\beta}^{\top}$, where $\boldsymbol{\alpha}$ and $\boldsymbol{\beta}$ are $n\times r$.
2. $\boldsymbol{\beta}^{\top}\mathbf{Y}_t$ contains the $r$ stationary cointegrating relationships.
3. $\boldsymbol{\alpha}$ contains the adjustment coefficients governing responses to disequilibrium.
35Why must $\mathbf{A}-\mathbf{I}_n$ be singular in a VAR(1) of $I(1)$ variables?
1. Suppose $\mathbf{A}-\mathbf{I}_n$ were invertible.
2. Then $\mathbf{Y}_{t-1}=(\mathbf{A}-\mathbf{I}_n)^{-1}(\Delta\mathbf{Y}_t-\boldsymbol{\epsilon}_t)$.
3. The RHS is $I(0)$, contradicting the presence of $I(1)$ variables in $\mathbf{Y}_t$.
4. Therefore, $\mathbf{A}-\mathbf{I}_n$ is singular.
36How can stationarity of the cointegrating relationships be verified in a VAR(1)?
1. $\mathbf{Y}_t=(\mathbf{I}_n+\boldsymbol{\alpha}\boldsymbol{\beta}^{\top})\mathbf{Y}_{t-1}+\boldsymbol{\epsilon}_t$.
2. Premultiply by $\boldsymbol{\beta}^{\top}$ to obtain $\boldsymbol{\beta}^{\top}\mathbf{Y}_t=(\mathbf{I}_r+\boldsymbol{\beta}^{\top}\boldsymbol{\alpha})\boldsymbol{\beta}^{\top}\mathbf{Y}_{t-1}+\boldsymbol{\beta}^{\top}\boldsymbol{\epsilon}_t$.
3. Check that $\mathbf{I}_r+\boldsymbol{\beta}^{\top}\boldsymbol{\alpha}$ has eigenvalues with modulus smaller than $1$.
37What are the preliminary steps in Johansen's cointegration procedure?
1. Establish the integration orders of the variables.
2. Select the VAR lag order and deterministic specification.
3. Estimate the VECM and test the rank of $\boldsymbol{\Pi}$ using Johansen's procedure.
38What are the two Johansen tests for cointegration?
1. Maximum-eigenvalue test: $H_0:\operatorname{rank}(\boldsymbol{\Pi})=r$ against $H_1:\operatorname{rank}(\boldsymbol{\Pi})=r+1$.
2. Trace test: $H_0:\operatorname{rank}(\boldsymbol{\Pi})\leq r$ against $H_1:\operatorname{rank}(\boldsymbol{\Pi})>r$.
3. Test sequentially from $r=0$ to determine the rank, using non-standard critical values appropriate to the deterministic specification.
39Why are multiple cointegrating vectors not uniquely identified?
1. Suppose $\boldsymbol{\Pi}=\boldsymbol{\alpha}\boldsymbol{\beta}^{\top}$.
2. For any invertible $r\times r$ matrix $\mathbf{R}$, define $\tilde{\boldsymbol{\beta}}=\boldsymbol{\beta}\mathbf{R}$ and $\tilde{\boldsymbol{\alpha}}=\boldsymbol{\alpha}(\mathbf{R}^{\top})^{-1}$.
3. Then $\tilde{\boldsymbol{\alpha}}\tilde{\boldsymbol{\beta}}^{\top}=\boldsymbol{\Pi}$.
4. Normalisation and further restrictions are needed to identify particular cointegrating vectors.
40How are two common price-spread cointegrating relationships represented for three $I(1)$ prices?
1. Suppose $Y_{1t}-Y_{2t}\sim I(0)$ and $Y_{1t}-Y_{3t}\sim I(0)$.
2. $\boldsymbol{\beta}=\begin{pmatrix}1&1\\-1&0\\0&-1\end{pmatrix}$.
3. $Y_{2t}-Y_{3t}=(Y_{1t}-Y_{3t})-(Y_{1t}-Y_{2t})\sim I(0)$.
4. There are two linearly independent cointegrating vectors.
41How can economic restrictions on cointegrating vectors be tested?
1. If the complete relationship is known, directly test the proposed combination for stationarity.
2. Within Johansen's framework, impose the proposed restrictions on $\boldsymbol{\beta}$.
3. Test overidentifying restrictions using an appropriate likelihood-ratio test.
42How are the adjustment coefficients $\boldsymbol{\alpha}$ interpreted?
1. $\alpha_{ij}$ measures the response of $\Delta Y_{it}$ to disequilibrium in cointegrating relationship $j$.
2. Its sign determines the direction of adjustment, conditional on the normalisation of $\boldsymbol{\beta}$.
3. If an entire row of $\boldsymbol{\alpha}$ is zero, that variable does not adjust to the long-run disequilibria.
43How is a common stochastic trend identified in a cointegrated VAR(1)?
1. Find $\boldsymbol{\alpha}_{\perp}\neq\mathbf{0}$ such that $\boldsymbol{\alpha}_{\perp}^{\top}\boldsymbol{\alpha}=\mathbf{0}$.
2. Premultiply $\Delta\mathbf{Y}_t=\boldsymbol{\alpha}\boldsymbol{\beta}^{\top}\mathbf{Y}_{t-1}+\boldsymbol{\epsilon}_t$ by $\boldsymbol{\alpha}_{\perp}^{\top}$.
3. $\boldsymbol{\alpha}_{\perp}^{\top}\Delta\mathbf{Y}_t=\boldsymbol{\alpha}_{\perp}^{\top}\boldsymbol{\epsilon}_t$.
4. Therefore, $\boldsymbol{\alpha}_{\perp}^{\top}\mathbf{Y}_t$ follows a random walk when the innovations are iid and mean-zero.
44How can three cointegrated variables be decomposed into a common trend and stationary deviations?
1. Define $D_{12,t}=Y_{1t}-Y_{2t}$, $D_{23,t}=Y_{2t}-Y_{3t}$ and $\tau_t=(Y_{1t}+Y_{2t}+Y_{3t})/3$.
2. $\mathbf{Y}_t=\mathbf{1}_3\tau_t+\frac{1}{3}\begin{pmatrix}2D_{12,t}+D_{23,t}\\-D_{12,t}+D_{23,t}\\-D_{12,t}-2D_{23,t}\end{pmatrix}$.
3. If both spreads are $I(0)$ and $\tau_t$ is $I(1)$, the variables share the common stochastic trend $\tau_t$.
45How are forecasts produced from a VECM?
1. Forecast $\Delta\mathbf{Y}_{T+h}$ recursively using lagged differences and the error-correction terms.
2. Recover level forecasts by cumulating the predicted changes.
3. The error-correction mechanism incorporates the estimated long-run equilibrium relationships into forecasts.
46How can a VECM test the expectations hypothesis of the term structure?
1. Under stationary term premia, yields at different maturities should share stochastic trends and have stationary spreads.
2. Test the cointegration rank using Johansen's method.
3. Test the theoretically implied restrictions on the cointegrating vectors.
4. Examine the estimated adjustment dynamics for consistency with the hypothesised long-run relationships.
47What does a deterministic trend in a bond-yield cointegrating relationship imply?
1. The long-run equilibrium spread varies systematically over time rather than remaining constant.
2. This is inconsistent with a constant equilibrium term premium, though it may reflect a changing premium or model misspecification.
3. Forecasts must incorporate the estimated deterministic trend.