You are on page 1of 11

Lecture 6

VAR Models
Dr. Junsoo Lee
EC 413

(Handbook Ch 5)

but they share the same independent variables. for each equation. x1t = c1 + φ11x1t-1 + φ12x2t-1 + u1t x2t = c2 + φ21x1t-1 + φ22x2t-1 + u2t Point: (1) x1t is explained by past values of x1t and x2t. ARMA → Vector ARMA (VARMA) can be possibly considered. Thus. x2t. simpler and easy to apply. Ex 2) VAR(2) model with 2 variables (p = 2. we ignore the MA components for simplicity.. Also. Model Xt = C + Φ1 Xt-1 + Φ2 Xt-2 + … + Φp Xt-p + ut where Xt is a vector. OLS can be used. (2) The dependent variables are different (x1t and x2t). g = 2) .. OLS cannot be used. Not popular: Reason: Non-linear estimation. The coefficients are different. Xt = (x1t. g = 2) In a vector form.2 What is the VAR Model? VAR = Vector AR Model (With more than one equations) AR → Vector AR (VAR) Very popular. Ex 1) VAR(1) model with 2 variables (p = 1.. x2t is explained by past values of x1t and x2t. Or. xgt)′ and Φi is a matrix.

013856801 0.14290266 3. g = 3) Example) VAR(4) with two variables [var_ex1.63343646 3. INF{3} -0.10613 0. GM1{3} 0.092121415 0.766937632 0.83934 0.932844 T x R**2 114. INF{2} -0.84617656 4.111008752 0. SEASONS{-2} 0.283368966 0. Constant -0.52403 0.91566988 8.030613277 1.87083 0.3602 0.009769159 5.353845176 0.043911306 0. INF{1} -0.853282 Variable Coeff Std Error T-Stat Signif ******************************************************************************* 1.030017353 0.291340934 -0.00707626 .3 Ex 3) VAR(1) model with 3 variables (p = 1.19445 0. GM1{2} 0. SEASONS 0.017902786 0.265682719 0.090956764 1.89702418 5.74418 0.286885571 -1.866510 R Bar **2 0. GM1{4} 0.22046 0.437221751 0.38572460 7.003893796 0.045170635 0.006353871 4. SEASONS{-1} 0.02482 0.00001377 11.13034571 6.00218498 5.090328162 3.006175990 -4.030179705 0.009650233 4.00001615 12.74677 0.650536 Uncentered R**2 0.13711 0. INF{4} 0.55028 0.0122948 INF 2.096816800 2.22487836 2.279213885 2.47822 0.092067368 0.024857254 0.740 Variable Coeff Std Error T-Stat Signif ******************************************************************************* 1.1472 0.028658364 0. GM1{1} 0.00000003 F-Tests.00702412 9.45844 0.64753550 2.095336219 3. Dependent Variable GM1 (will be discussed later) Variable F-Statistic Signif GM1 3.025679738 0.00032375 6. INF{2} 0. GM1{2} 0.12972 0.0796779 Dependent Variable INF Quarterly Data From 1961:02 To 1991:04 Usable Observations 123 Degrees of Freedom 111 Centered R**2 0. GM1{3} -0.40308506 4.51038 0.97691 0.253709419 0.682045 R Bar **2 0.71155 0.030595316 -0.284362412 -0.prg] Two variables: g = 2 [GM1 = ∆log(M1t) and INF = ∆log(GDPDEFt)] Lags: p = 4 VAR/System .47552 0.00010450 10. GM1{4} 0.Estimation by Least Squares Dependent Variable GM1 Quarterly Data From 1961:02 To 1991:04 Usable Observations 123 Degrees of Freedom 111 Centered R**2 0. INF{1} 0.044054004 0.058389418 0.030226246 0. GM1{1} 0.

002016630 0. Dependent Variable INF (will be discussed later) Variable F-Statistic Signif GM1 0.0000000 System of Equations The VAR model is a system of equations with more than one dependent variables! (# of equations = # of dependent variables = # of endogenous variables) But.37270891 0. No endogeneity issue: No need to use advanced techniques of simultaneous equation models. However. but OLS is fine for VAR models.53575258 F-Tests. An alternative estimator. . Why is the VAR model popular? 1.002217940 0. In this case.69161 0.001735558 -0.71215 -0.002111483 0. The GLS utilizes the covariance of the error terms.001889836 -0. due to the fact that the same independent variables are used in each of the multivariate equations.89503 -0. the usual OLS is just fine in estimating VAR models.4 7.5489180 INF 56.7670 0. OLS = GLS. Allows for impulse response analysis and innovation accounting (later on). 9.003246433 2.256292126 -0. 10.62118 0. Its forecasting ability is super! Better than ARMA models.007487793 0.00774955 0.93582673 0. VAR model is also a system of equations. Allows for causality tests (later on). It is a system of equations that may require GLS.003206912 0. Easy to estimate. VAR models include only deterministic terms (lagged variables and exogenous variables: no endogenous variables) in the independent variables. No need to use GLS b. Estimation of VAR models OLS is just fine. 11. is usually necessary to estimate a system of equations.094497737 0. 3.08070 0. 4. 2. INF{3} INF{4} Constant SEASONS{-2} SEASONS{-1} SEASONS 0.49062484 0. called GLS (Generalized Least squared).1908 0.39957518 0.092786807 0.84564 -0. 12.002052371 0. a. 8.

We cannot use an OLS estimation for this. (b) Likelihood RATIO test LR = (T – c) [log|Ω2| . N is the total # of parameters in all equations = g2p + g (add g if each equation includes a constant term.5 Issues in VAR models 1. Ex) Back to VAR_ex1.) 2. (This issue is beyond our scope.prg LR test for 4 vs 8 lags . How could we determine the optimal number of lags (p)? Lag determinations (a) Information Criteria AIC = T log|Ω| + 2 N SBC = T log|Ω| + N * log (T) where Ω is the covariance matrix of residuals. We need a restriction β1 = 0 or β2 = 0 to avoid the so-called identification problem. One popular question: why not include contemporaneous terms? x1t = c1 + β 1 x2t + φ11x1t-1 + φ12x2t-1 + u1t x2t = c2 + β 2 x1t + φ21x1t-1 + φ22x2t-1 + u2t If we include these terms.05). g = # of equations) c = small sample corrections = # of parameters in an unrestricted model * # of equations + # of other exogenous variables → If the null is rejected (p-value > 0. then choose an unrestricted model (longer lags). it becomes a structural VAR model.log|Ω1|] ~ χmg2 Degree of freedom = mg2 (m=# of lag difference.) → Choose a model with a lower AIC value.

j=1. we have VAR models and this is an Vector MA(∞) representation form. Here.… (future periods) = dXi.t+s/dujt = Change of Xi over time at period t+s (s-period ahead) when there is a one unit shock (or one standard deviation shock) in the j-th variable. we choose the unrestricted model with 8 lags.0030 0.6 Log Determinants are 44. Impulse Response Analysis Impulse response = dynamic multiplier at the period t+s.3.066340 with Significance Level 0.2. X2t = GM1 ) Plot of Responses of INF 0. Ex) Back to the example (X1t = INF.0015 0. = Reaction of the i-th series occurring at the s-period ahead to one unit shock (or one standard deviation shock) of the j-th series Note: This is an MA(∞) representation form of AR models.009578 -5.0020 0.0005 0.0005 0 2 4 6 8 10 INF 12 14 16 GM1 This is the plot of dX1.666962 Chi-Squared(16) = 5365.2 18 20 22 .00000000 Since the null is rejected. s = 1.0010 0.0035 0.t+s/dujt.0000 -0.0025 0.

012 0.000 1.866 9.219 92.206 10.073 88.011734426 0.011294651 0.011528274 0.703 10.011678509 0.004 0.2 Decomposition of Error Variances “Proportion of the s-step ahead forecast error variance of the i-th series accounted for the shock of the j-th series.008 0.004 0 2 4 6 8 10 12 INF 14 16 18 20 22 GM1 This is the plot of dX2.529 89.766 88.” Ex) Back to the example Decomposition of Variance for Series GM1 Step 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Std Error 0.011756587 0.002 0.974 88.461 95.010 0.927 11.651 88.429 7.002 -0.7 Plot of Responses of GM1 0.571 92.000 98.006 0.011676560 0.539 4.011622571 0.011771760 0.781 7.000 -0.870 89.349 11.471 10.334 88.797 89. j=1.297 89.794 89.134 90.011029989 0.797 .666 11.011817442 GM1 100.203 INF 0.234 11.600 92.123 10.203 10.130 10.877 89.026 11.400 7.010801080 0.010531271 0.010629928 0.011803582 0.t+s/dujt.011703163 0.011411462 0.011588752 0.011780951 0.

093 12.000 12.192 96. then put X first. the ordering of VAR models matters.003484629 0.322 4.239 8.648 6.257 91.005496268 0.005807294 0.613 94.962 91.280 6.697 11.004952995 0.777 4.866 8. A natural question is.005969105 0.005991985 0.755 87.345 96.005751989 0.005296908 0.100 88.038 8.569 96.353 94.000 87.678 95.005911234 0. put GM1 first.004069207 0.003721402 0.005709597 0.843 INF 99.203 12.006016572 0.406 92.526 8.770 99.701 97.797 87.332 7.900 12. put INF first.720 93.005940779 0. how do we know? .005841251 0.907 87.990 7.010 92.005646276 0.134 91.387 5.761 91.431 3.647 5.869 96.005071735 0.733 6.668 92.808 3.” Does GM1 affect INF but not vice versa? If so.011857878 0.352 93.230 0.356 91.005577657 0.005381514 0.743 8. “If X affects Y.004466115 0.157 Important Note on the Impulse response Analysis and Variance Decomposition Due to an orthogonalization method to calculate the impulse response functions. Put exogenous variables first.609 91.005880154 0.006036233 0.131 3.011852587 0.004720198 0.011828197 0.011862732 88.011841943 0.303 Decomposition of Variance for Series INF Step 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Std Error 0.245 12.006056490 GM1 0.011834937 0.655 3.8 19 20 21 22 23 24 0. but Y does not affect X.474 91.594 7.223 95. Does INF affect GM1 but not vice versa? If so.644 8.391 8.299 2.267 93.

. To see if X1t causes X2t.9 Granger Causality Tests “X2t Granger-causes X1t if past values of X2t help in predicting X1t in the presence of past values of X1t. Thus. If the null is not rejected. = αp = 0 (no causality from X1t to X2t) Do F-test. = βp = 0 (no causality from INF to GM1) The null is NOT rejected at the 5% level. we conclude that . If the null is not rejected.+ α px1t-p + β1x2t-1 + .” x1t = c1 + α1x1t-1 + .+ αpx1t-p + β 1x2t-1 + ..0796779 GM1t = c1 + α1GM1t-1 + . = βp = 0 (no causality from X2t to X1t) Do F-test.+ αpGM1t-p + β 1INFt-1 + ..3602 0.Estimation by Least Squares Dependent Variable GM1 = x1t F-Tests. . we conclude that . (*) Thus.. then X2t does not granger-cause X1t. we can use: x2t = c1 + α 1x1t-1 + . + β pINF2t-p + u1t H0: β1 = .. then X1t does not granger-cause X2t. Ex) Back to VAR estimation 1st equation: VAR/System . The null is REJECTED at the 10% level. Dependent Variable GM1 (will be discussed later) Variable F-Statistic Signif GM1 3... + βpx2t-p + u1t H0: α1 = ... + β px2t-p + u1t H0: β1 = .0122948 INF 2.1472 0.

Thus.5489180 56. Omitted variables affect the results. we conclude that . Thus.7670 INF 0... A-theoretical model a. 4. INF Note: This is an ad-hoc model without allowing for enough # of endogenous variables or careful lag selections.0000000 INFt = c1 + α 1GM1t-1 + . we need to put first. + βpINF2t-p + u1t H0: α1 = . Critically depends on the lag determination. INF Granger-causes GM1 GM1 does NOT Granger-cause INF. = αp = 0 (no causality from GM1 to INF) The null is NOT rejected at the 10% level. 3. we do not put much value on this exercise..10 2nd equation: Dependent Variable INF F-Tests. Limitations of VAR models 1. Not with a good economic theory b. Too many parameters in the model (loss of degree of freedom) 5. Thus. . Forecasting oriented model 2.1908 0.+ α pGM1t-p + β1INFt-1 + . At the 10% level. Ordering matters for the impulse analysis and variance decomposition. Dependent Variable INF Variable F-Statistic GM1 (will be discussed later) Signif 0.

. OLS is used in estimating VAR models while VAR models can be considered as a system of equations. Define the error variance decomposition. Discuss why the contemporaneous terms do not appear in the standard VAR model. Also Provide specific forms of impulse response function of a VAR(1) model. 8. 3. Why? 2. What are the advantages of using VAR models? 4. Discuss the difference between the VAR and VARMA models. Provide specific forms of an AR(1) model. 7. 9.11 Review Questions on VAR models 1. Define the impulse response function. What are the limitations of using VAR models? 5. 6. Does the ordering of variables matter in estimating VAR models? Why or why not? 10. Does the ordering of the variables matter in the impulse response function analysis? Does ordering matter in the error variance decomposition? Why or why not? 11. Define the Granger-causality. Write down the equations of the VAR(2) model with 3 endogenous variables.