Vector Autoregression (VAR) Models

February 19, 2026

1 / 59
VAR Exercises

1 Stationarity Testing

2 Lag Length Selection

3 Granger Causality

4 OLS Estimation and Interpretation

5 From VAR to VMA Representation

6 Impulse Response Functions

7 Bonus: Structural VAR

8 Wrap-Up and Next Steps

2 / 59
Exercise 1: Stationarity Testing - Question

Exercise 1.1: Theoretical Understanding

      
y1t 0.6 0.2 y1,t−1 u
Consider a bivariate VAR(1) model: = + 1t
y2t 0.1 0.5 y2,t−1 u2t
1 Is this VAR(1) stationary? Verify using the stationarity condition.

2 Compute the eigenvalues of the coefficient matrix.

3 What is the companion form of this system?

3 / 59
Exercise 1.1: Hint

Hint
Key concepts to remember:
A VAR(1) is stationary if all eigenvalues of A1 lie inside the unit circle
Equivalently: |λi | < 1 for all i
For VAR(1): companion form is the same as the coefficient matrix
To find eigenvalues, solve: det(A1 − λI) = 0
 
a b
Calculation tip: For a 2 × 2 matrix :
c d
√
(a+d)± (a+d)2 −4(ad−bc)
λ1,2 = 2

4 / 59
Exercise 1.1: Solution (Part 1)

Solution
Step 1: Stationarity condition
For a VAR(1), stationarity requires all eigenvalues of A1 to satisfy |λi | < 1.

Step 2: Compute
 eigenvalues
0.6 0.2
Let A1 =
0.1 0.5
Characteristic
  det(A1 − λI) = 0
equation:
0.6 − λ 0.2
det =0
0.1 0.5 − λ
(0.6 − λ)(0.5 − λ) − (0.2)(0.1) = 0 λ2 − 1.1λ + 0.28 = 0

5 / 59
Exercise 1.1: Solution (Part 2)

Solution
√ √
Using the quadratic formula: λ1,2 = 1.1± 1.21−1.12
2 = 1.1± 0.09
2 = 1.1±0.3
2
λ1 = 1.1+0.3
2 = 0.7, λ2 = 1.1−0.3
2 = 0.4

Step 3: Check stationarity

|λ1 | = 0.7 < 1 X
|λ2 | = 0.4 < 1 X
Conclusion: The VAR(1) is stationary since both eigenvalues lie inside
the unit circle.
Step 4: Companion form
For VAR(1), the companion form is identical to the original system
(already in VAR(1) form).

6 / 59
 Exercise 1.2: Stationarity in R - Question

Exercise 1.2: Practical Implementation

You have estimated a VAR(2) model on three variables (GDP growth,

inflation, interest rate) using R:
1 library ( vars )
2 data <- read . csv ( " macro _ data . csv " )
3 var _ model <- VAR ( data , p =2 , type = " const " )

Tasks:
1 Write R code to check the stationarity of this VAR(2) model
2 Interpret the output: what do you look for?
3 What should you do if the model is non-stationary?

7 / 59
 Exercise 1.2: Solution
R code for stationarity testing:
1 # Method 1: Using roots () function
2 stability _ roots <- roots ( var _ model )
3 print ( stability _ roots )
4 # Check condition : all roots should be < 1
5 all ( stability _ roots < 1) # Should return TRUE
6 # Method 2: Compute eigenvalues of companion matrix
7 companion _ matrix <- companionMatrix ( var _ model )
8 eigenvalues <- eigen ( companion _ matrix ) $ values
9 modulus <- Mod ( eigenvalues )
0 print ( modulus )
1 # Check : all moduli should be < 1
2 all ( modulus < 1) # Should return TRUE
3 # Visual check
4 plot ( stability _ roots ,
5 main = " Inverse Roots of AR Characteristic Polynomial "
,
6 xlab = " Real Part " , ylab = " Imaginary Part " )
7 symbols (0 , 0 , circles =1 , add = TRUE , inches = FALSE ) 8 / 59
Exercise 1.2: Interpretation
Solution
What to look for:
1 Roots ¡ 1: If roots(var model) returns values all < 1, the system
is stationary
2 Eigenvalues modulus ¡ 1: Equivalently, check |λi | < 1
3 Visual check: In the plot, all points should lie inside the unit circle

If non-stationary:
Check for unit roots in individual series (ADF test)
Consider differencing the data
Test for cointegration (if variables share common trends)
Estimate a Vector Error Correction Model (VECM)

Warning: Standard VAR asymptotics require stationarity. Non-stationary

VARs can lead to spurious results!
9 / 59
Exercise 2: Lag Length Selection - Question
Exercise 2.1: Information Criteria

You estimated VAR models with different lag lengths on a dataset with
K = 3 variables and T = 200 observations. The results are:

Lag (p) AIC BIC HQ FPE

0 -5.234 -5.180 -5.212 0.0054
1 -6.782 -6.674 -6.738 0.0011
2 -7.125 -6.963 -7.059 0.0008
3 -7.089 -6.873 -7.001 0.0008
4 -7.012 -6.742 -6.902 0.0009

Questions:
1 Which lag length does each criterion select?

2 Why do AIC and BIC sometimes disagree?

3 Which criterion would you recommend and why?

10 / 59
Exercise 2.1: Hint

Hint
Key principles:
Selection rule: Choose p that minimizes the criterion (most
negative for log-likelihood based criteria)
2K 2 p
AIC: Akaike Information Criterion AIC(p) = ln |Σ̃u (p)| + T
BIC: Bayesian/Schwarz Information Criterion
2
BIC(p) = ln |Σ̃u (p)| + K pTln T
Key difference: BIC penalizes additional parameters more heavily
(especially for large T )

Practical advice:
BIC tends to select more parsimonious models (fewer lags)
AIC can be liberal (may overfit)

11 / 59
Exercise 2.1: Solution
Solution
Step 1: Identify optimal lags for each criterion
Looking at the table (bold values indicate minimum):
AIC: Selects p = 2 (most negative: -7.125)
BIC: Selects p = 3 (-6.873)
HQ: Selects p = 2 (-7.059)
FPE: Selects p = 2 (0.0008)

Step 2: Why do they disagree?

BIC has a stronger penalty for additional parameters
K 2 p ln 200
With T = 200: BIC penalty = 200 ≈ 0.0265 × K 2 p
2K 2 p 2
AIC penalty = 200 = 0.01 × K p
BIC penalty is approximately 2.65 times larger
BIC favors more parsimonious models
12 / 59
Exercise 2.1: Recommendation

Solution
Step 3: Which criterion to use?

Recommendation: Use p = 2 (majority vote: AIC, HQ, FPE)

Reasoning:
1 Three out of four criteria agree on p = 2
2 BIC’s selection of p = 3 is marginal (only 0.016 difference from p = 2)
3 Parsimony principle: prefer simpler model when performance is similar
4 p = 2 is more interpretable (quarterly data → half-year dynamics)

Additional checks:
Perform residual diagnostics (serial correlation test)
If p = 2 residuals show autocorrelation → consider p = 3
Check if results are robust across lag specifications
13 / 59
Exercise 2.2: Lag Selection in R - Question

Exercise 2.2: Practical Implementation

Write R code to:

1 Load the sample dataset Canada from the vars package
2 Select optimal lag length using multiple criteria
3 Estimate the VAR with the selected lag
4 Verify that residuals are not serially correlated

14 / 59
 Exercise 2.2: Solution

1 library ( vars )
2 # Load data
3 data ( Canada )
4 # Step 1: Lag selection ( test p = 1 to 10)
5 lag _ select <- VARselect ( Canada , lag . max =10 ,
6 type = " const " )
7 print ( lag _ select $ selection )
8 # AIC ( n ) HQ ( n ) SC ( n ) FPE ( n )
9 # 3 2 1 3
0 # Step 2: Estimate VAR with BIC - selected lag ( p =1)
1 var _ bic <- VAR ( Canada , p =1 , type = " const " )
2 # Step 3: Serial correlation test
3 serial _ test <- serial . test ( var _ bic , lags . pt =10 ,
4 type = " PT . asymptotic " )
5 print ( serial _ test )
6 # Null : No serial correlation up to lag 10

15 / 59
 Exercise 2.2: Interpretation

Interpreting VARselect output:

1 # AIC ( n ) HQ ( n ) SC ( n ) FPE ( n )
2 # 3 2 1 3

AIC and FPE select p = 3 (more flexible)

HQ selects p = 2 (middle ground)
SC (= BIC) selects p = 1 (most parsimonious)
Serial correlation test interpretation:
If p-value > 0.05: No evidence of serial correlation X
If p-value < 0.05: Residuals are serially correlated → increase lag
length
Decision: Start with BIC (p = 1), check diagnostics. If serial correlation
detected, move to p = 2 or p = 3.

16 / 59
Exercise 3: Granger Causality - Question

Exercise 3.1: Theoretical Understanding

Consider a bivariate VAR(2) with GDP growth (gt ) and interest rate (rt ):
gt = 0.3 + 0.5gt−1 + 0.2gt−2 − 0.1rt−1 − 0.05rt−2 + u1t
rt = 1.2 + 0.1gt−1 + 0.05gt−2 + 0.7rt−1 + 0.15rt−2 + u2t
Questions:
1 Does the interest rate Granger-cause GDP growth?
2 Does GDP growth Granger-cause the interest rate?
3 What are the null hypotheses for each test?
4 How would you test these hypotheses?

17 / 59
Exercise 3.1: Hint

Hint
Granger causality definition:
X Granger-causes Y if past values of X help predict Y beyond what past
values of Y alone can predict.

Testing procedure:
1 Write the equation for the “caused” variable
2 Identify coefficients on lags of the “causing” variable
3 Set up null hypothesis: H0 : all those coefficients = 0
4 Use F-test or Wald test

For our example:

“Does r GC g ?”: Look at rt−1 , rt−2 coefficients in gt equation
“Does g GC r ?”: Look at gt−1 , gt−2 coefficients in rt equation

18 / 59
Exercise 3.1: Solution (Part 1)

Solution
Question 1: Does interest rate Granger-cause GDP growth?
Look at the gt equation:
gt = 0.3 + 0.5gt−1 + 0.2gt−2 − 0.1rt−1 − 0.05rt−2 + u1t
Coefficients on rt−1 and rt−2 : −0.1 and −0.05
These are non-zero
Null hypothesis: H0 : βr ,1 = βr ,2 = 0
R −RSSU )/q
Test statistic: F = (RSS
RSSU /(T −k) ∼ F (q, T − k) where q = 2
(number of restrictions), RSSR = restricted sum of squares, RSSU =
unrestricted sum of squares

Conclusion: If we reject H0 , then r Granger-causes g .

Based on the non-zero coefficients, we would expect to reject H0 →
interest rate does Granger-cause GDP growth.

19 / 59
Exercise 3.1: Solution (Part 2)
Solution
Question 2: Does GDP growth Granger-cause interest rate?
Look at the rt equation:
rt = 1.2 + 0.1gt−1 + 0.05gt−2 + 0.7rt−1 + 0.15rt−2 + u2t
Coefficients on gt−1 and gt−2 : 0.1 and 0.05
These are non-zero
Null hypothesis: H0 : γg ,1 = γg ,2 = 0
Use same F-test procedure

Conclusion: Based on non-zero coefficients, we would expect GDP

growth does Granger-cause interest rate.

Key insight: This VAR exhibits bidirectional Granger causality → GDP

and interest rate have feedback effects.
Economic interpretation: Monetary policy (interest rate) affects output,
and output conditions influence policy decisions.
20 / 59
Exercise 3.2: Granger Causality in R - Question

Exercise 3.2: Testing in Practice

Using the Canada dataset:

1 Test whether employment (e) Granger-causes productivity (prod)
2 Test the reverse: does prod Granger-cause e?
3 Interpret the results

21 / 59
 Exercise 3.2: Solution

1 library ( vars )
2 data ( Canada )
3 # Estimate VAR (2)
4 var _ model <- VAR ( Canada , p =2 , type = " const " )
5 # Test : Does employment GC productivity ?
6 gc _ e _ to _ prod <- causality ( var _ model ,
7 cause = " e " )
8 print ( gc _ e _ to _ prod $ Granger )
9 # Test : Does productivity GC employment ?
0 gc _ prod _ to _ e <- causality ( var _ model ,
1 cause = " prod " )
2 print ( gc _ prod _ to _ e $ Granger )
3 # Interpret
4 # H0 : [ cause ] does NOT Granger - cause [ others ]
5 # If p - value < 0.05: Reject H0
6 # = > [ cause ] DOES Granger - cause [ others ]

22 / 59
Exercise 3.2: Interpretation Guide
How to interpret causality() output:
Example output:
Granger causality H0: e do not Granger-cause prod
F-Test:
statistic = 4.521, p-value = 0.012
Interpretation:
p-value = 0.012 < 0.05
Reject H0
Conclusion: Employment does Granger-cause productivity
Important caveats:
1 Granger causality 6= true causality (it’s predictive association)

2 Depends on lag length choice

3 Sensitive to omitted variables

4 Requires stationarity

23 / 59
Exercise 4: OLS Estimation - Question

Exercise 4.1: Understanding VAR Coefficients

You
 estimated
  a bivariate
 VAR(1)
  and obtained:
ŷ1t 0.5 0.7 −0.3 y1,t−1
= +
ŷ2t 1.2 0.2 0.6 y2,t−1
Standard errors: SE (â11 ) = 0.08, SE (â12 ) = 0.12, SE (â21 ) = 0.09,
SE (â22 ) = 0.07

Questions:
1 Test whether a11 is significantly different from zero (α = 0.05)
2 Test whether a12 is significantly different from zero
3 Interpret the coefficient a12 = −0.3
4 What does a21 = 0.2 tell you about the relationship?

24 / 59
Exercise 4.1: Hint

Hint
Hypothesis testing for coefficients:
H0 : aij = 0 (no effect)
H1 : aij 6= 0 (significant effect)
âij
Test statistic: t = SE (âij )
Critical value (5% level, two-tailed): |t| > 1.96 (for large T )
Alternatively: p-value < 0.05

Interpretation guide:
aij measures effect of yj,t−1 on yi,t
Positive coefficient: positive relationship
Negative coefficient: negative relationship
Magnitude indicates strength of effect

25 / 59
Exercise 4.1: Solution (Part 1)

Solution
Question 1: Test a11 = 0.7
t11 = SEâ(â1111 ) = 0.08
0.7
= 8.75
|t| = 8.75 > 1.96 (critical value at 5%)
Reject H0
a11 is highly significant
Interpretation: y1,t−1 has strong predictive power for y1t (strong
persistence)

Question 2: Test a12 = −0.3

t12 = SEâ(â1212 ) = −0.3
0.12 = −2.5
|t| = 2.5 > 1.96
Reject H0
a12 is significant at 5% level
26 / 59
Exercise 4.1: Solution (Part 2)
Solution
Question 3: Interpret a12 = −0.3
Coefficient a12 measures the effect of y2,t−1 on y1,t .
Sign: Negative → inverse relationship
Magnitude: A 1-unit increase in y2,t−1 leads to a 0.3-unit decrease
in y1,t (holding y1,t−1 constant)
Significance: t = −2.5 → statistically significant
Economic example: If y1 = GDP growth, y2 = interest rate:
Higher interest rate in previous period → lower GDP growth today
Consistent with monetary policy transmission mechanism
Question 4: Interpret a21 = 0.2
Measures effect of y1,t−1 on y2,t
Positive coefficient: y1 positively affects y2 with one-period lag
Test: t = 0.2/0.09 = 2.22 > 1.96 → significant
27 / 59
 Exercise 4.2: VAR Estimation Output in R

Exercise 4.2: Reading R Output

1 summary ( var _ model )

2 # Output for equation 1: GDP growth
3 # Estimate Std . Error t value Pr ( >| t |)
4 # const 0.523 0.145 3.60 0.0004 ***
5 # gdp . l1 0.351 0.089 3.94 0.0001 ***
6 # inf . l1 -0.087 0.042 -2.07 0.0402 *
7 # rate . l1 -0.124 0.056 -2.21 0.0284 *

Questions:
1 Which coefficients are significant at 5%?
2 Interpret the coefficient on rate.l1
3 What is the R-squared interpretation in VAR?

28 / 59
Exercise 4.2: Solution
Solution
Question 1: Significant coefficients
const: p = 0.0004 < 0.05 → Significant ***
gdp.l1: p = 0.0001 < 0.05 → Significant ***
inf.l1: p = 0.0402 < 0.05 → Significant *
rate.l1: p = 0.0284 < 0.05 → Significant *
All coefficients are significant at 5% level! Question 2: Interpret
rate.l1 coefficient
Coefficient: −0.124
Interpretation: A 1 percentage point increase in the interest rate in
period t − 1 leads to a 0.124 percentage point decrease in GDP
growth in period t
Economic meaning: Contractionary monetary policy (higher rates)
→ lower economic growth
Statistical significance confirms this effect is not due to chance 29 / 59
Exercise 4.2: R-squared in VAR
Solution
Question 3: R-squared interpretation
In VAR models, each equation has its own R-squared:
R-squared for equation i: Proportion of variance in yit explained by
all lagged variables in the VAR
Range: 0 to 1
Higher R2 → better fit (but watch for overfitting!)
Typical values:
Growth rates, returns: R2 often 0.05–0.30 (low predictability)
Persistent series (e.g., levels): R2 often 0.80–0.99
High R2 doesn’t mean causality!
VAR-specific note:
Focus more on joint significance (F-tests) than individual R2
System-level diagnostics matter more than equation-level fit
30 / 59
Exercise 5: VMA Representation - Question

Exercise 5.1: Deriving VMA

Consider a simple
 VAR(1):
 yt = A1 yt−1 + ut
0.5 0
where A1 = (diagonal for simplicity)
0 0.3
Questions:
P∞
1 Write the VMA representation: yt = i=0 Φi ut−i
2 Compute the first three VMA coefficient matrices: Φ0 , Φ1 , Φ2
3 Show that Φi → 0 as i → ∞ (explain why)
What is the cumulative impulse response: ∞
P
i=0 Φi ?
4

31 / 59
Exercise 5.1: Hint

Hint
VAR to VMA transformation:
Starting from: yt = A1 yt−1 + ut
Recursive substitution:

yt = ut + A1 (A1 yt−2 + ut−1 )

= ut + A1 ut−1 + A21 yt−2
= ut + A1 ut−1 + A21 ut−2 + A31 yt−3
..
.

If stationary: yt = ∞ i
P
i=0 A1 ut−i

Thus: Φi = Ai1 (for VAR(1))

32 / 59
Exercise 5.1: Solution (Part 1)

Solution
Step 1: VMA representation P
From recursive substitution: yt = ∞
i=0 Φi ut−i
i
where Φi = A1 for VAR(1)

Step 2: Compute  VMA  matrices

1 0
Φ0 = A01 = I =
0 1
 
1 0.5 0
Φ1 = A1 =
0 0.3
 2  
0.5 0 0.25 0
Φ2 = A21 = =
0 0.3 0 0.09

33 / 59
Exercise 5.1: Solution (Part 2)

Solution
Step 3: Show Φi → 0
0.5i
 
0
For a diagonal matrix: Ai1 =
0 0.3i
As i → ∞: 0.5i → 0 and 0.3i → 0
Because both eigenvalues (0.5 and 0.3) are inside the unit circle
General result: If VAR is stationary (|λj | < 1), then Φi → 0
exponentially fast

Step
P∞ 4: Cumulative
P∞ i impulse response
−1
Φ
i=0 i = i=0 A1 = (I − A1 ) −1  
−1 1 − 0.5 0 2 0
For our example: (I − A1 ) = =
0 1 − 0.3 0 1.43
Interpretation: Long-run multiplier — cumulative effect of a shock on
the level of each variable.
34 / 59
Exercise 5.2: Computing VMA in R - Question

Exercise 5.2: Practical Implementation

Using the vars package:

1 Estimate a VAR(2) on the Canada dataset
2 Compute the VMA coefficient matrices Φi for i = 0, 1, . . . , 10
3 Plot how the coefficients decay over time
4 Interpret the results

35 / 59
 Exercise 5.2: Solution
1 library ( vars )
2 data ( Canada )
3 # Estimate VAR (2)
4 var _ model <- VAR ( Canada , p =2 , type = " const " )
5 # Compute VMA matrices ( Phi coefficients )
6 # Phi ( i ) represents effect of shock at t - i on y at t
7 phi _ matrices <- Phi ( var _ model , nstep =10)
8 # phi _ matrices is a 3 D array :
9 # [ variable , shock , horizon ]
0 # Example : Effect of shock to ’e ’ on ’ prod ’ at horizon5
1 effect _ h5 <- phi _ matrices [ " prod " , " e " , 5]
2 print ( effect _ h5 )
3 # Visualize decay
4 horizons <- 0:10
5 effects <- phi _ matrices [ " e " , " e " , ]
6 plot ( horizons , effects , type = " b " ,
7 main = " VMA Coefficient Decay " ,
8 xlab = " Horizon " , ylab = " Phi _ i [1 ,1] " )
9 abline ( h =0 , lty =2 , col = " red " )
36 / 59
Exercise 5.2: Interpretation
Solution
Interpreting VMA coefficients:
The Φi matrices tell us:
[Φi ]jk : Effect of a 1-unit shock to variable k at time t on variable j at
time t + i
Φ0 = I: Contemporaneous impact (shock affects itself immediately)
Φ1 , Φ2 , . . .: Dynamic propagation of shocks over time
What to look for in the plot:
1 Initial response: Φ0 = 1 (by construction)
2 Persistence: How fast do coefficients decay?
Fast decay → shocks dissipate quickly
Slow decay → long-lasting effects
3 Sign patterns:
Always positive → no overshooting
Sign changes → oscillatory response
4 Convergence to zero: Confirms stationarity 37 / 59
Exercise 6: Impulse Response Functions - Question

Exercise 6.1: Understanding IRFs

Consider the following statements about Impulse Response Functions:

A. IRFs measure the effect of a one-time shock to one variable on all
variables in the system
B. Reduced-form IRFs are unique and do not depend on variable ordering
C. Cholesky decomposition assumes a recursive causal structure
D. The area under an IRF curve gives the cumulative effect of the shock
E. IRFs from a stationary VAR should eventually return to zero

Questions:
1 Which statements are TRUE? Which are FALSE?
2 For each false statement, explain why and correct it

38 / 59
Exercise 6.1: Solution
Solution
Evaluating each statement:

X A is TRUE: IRFs trace out dynamic response of all variables to a

shock in one variable over time
× B is FALSE:
Reduced-form shocks ut are correlated: E [ut u0t ] = Σu (non-diagonal)
Cannot isolate effect of “one shock” without identification
Cholesky IRFs do depend on variable ordering
Correction: “Structural IRFs require identification; Cholesky IRFs
depend on ordering”
X C is TRUE: Cholesky assumes variable 1 affects all others
contemporaneously, variable 2 affects 3,. . . ,K contemporaneously,
etc. (recursive/triangular structure)
X D is TRUE: ∞
P
h=0 IRFh = cumulative (long-run) effect
X E is TRUE: Stationarity → shocks dissipate over time → IRF → 0
39 / 59
Exercise 6.2: IRF Identification Problem

Exercise 6.2: Why Do We Need Structural Identification?

 
1.0 0.6
You estimated a bivariate VAR and obtained: Σu =
0.6 0.8
Questions:
1 What does the off-diagonal element (0.6) tell you?
2 Why can’t you directly interpret a 1-unit shock to u1t ?
3 How does Cholesky decomposition solve this?
4 Compute the Cholesky factor (lower triangular) P such that Σu = PP0

40 / 59
Exercise 6.2: Hint

Hint
Cholesky decomposition procedure:
σ12 σ12

For Σu =
σ σ22
 12 
p 0
Find P = 11 such that Σu = PP0
p21 p22
Steps:
q
1 p11 = σ12
σ12
2 p21 = p11
q
3 p22 = σ22 − p21
2

Structural shocks: ut = Pεt

where εt are orthogonal structural shocks with E [εt ε0t ] = I

41 / 59
Exercise 6.2: Solution (Part 1)

Solution
Question 1: Interpretation of σ12 = 0.6
Correlation: ρ = √ 0.6 = 0.6
= 0.671
1.0×0.8 0.894
High positive correlation between reduced-form shocks
Shocks tend to move together
Problem: Can’t isolate effect of “just u1 ” or “just u2 ”

Question 2: Why can’t we interpret u1t directly?

If u1t increases by 1, u2t typically also changes (due to correlation)
We don’t observe “pure” shocks
Need to orthogonalize: find uncorrelated structural shocks εt

42 / 59
Exercise 6.2: Solution (Part 2)
Solution
Question 3: How Cholesky helps
Cholesky decomposes Σu = PP0 with P lower triangular
Define structural shocks: ut = Pεt
By construction: E [εt ε0t ] = I (orthogonal!)
Now can interpret IRFs as responses to εjt (pure shocks)
 
1.0 0.6
Question 4: Compute Cholesky factor Given: Σu =
0.6 0.8
√
p11 = 1.0 = 1.0
0.6
p21 = = 0.6
1.0
p √ √
p22 = 0.8 − 0.62 = 0.8 − 0.36 = 0.44 = 0.663
 
1.0 0
P=
0.6 0.663
43 / 59
Exercise 6.2: Verification

Solution
Verify: PP0= Σu
 
1.0 0 1.0 0.6
0.6 0.663 0 0.663
 
1.0 × 1.0 + 0 × 0 1.0 × 0.6 + 0 × 0.663
=
0.6 × 1.0 + 0.663 × 0 0.6 × 0.6 + 0.663 × 0.663
   
1.0 0.6 1.0 0.6
= = X
0.6 0.36 + 0.44 0.6 0.8
Interpretation of P:
Row 1: u1t = 1.0 · ε1t → y1 responds only to first structural shock
Row 2: u2t = 0.6 · ε1t + 0.663 · ε2t → y2 responds to both shocks
Recursive structure: first variable causally prior

44 / 59
Exercise 6.3: Computing IRFs in R - Question

Exercise 6.3: Practical IRF Analysis

Tasks:
1 Estimate a VAR(2) on the Canada dataset
2 Compute orthogonalized IRFs (Cholesky)
3 Plot the response of prod (productivity) to a shock in e
(employment)
4 Plot IRFs for 20 periods ahead
5 Add confidence bands (bootstrap)
6 Interpret the results

45 / 59
 Exercise 6.3: Solution (Part 1)
1 library ( vars )
2 data ( Canada )
3
4 # Step 1: Estimate VAR (2)
5 var _ model <- VAR ( Canada , p =2 , type = " const " )
6
7 # Step 2: Compute orthogonalized IRFs
8 # ortho = TRUE uses Cholesky decomposition
9 # n . ahead =20: 20 periods ahead
0 # boot = TRUE : bootstrap confidence intervals
1 # runs =500: number of bootstrap replications
2 irf _ results <- irf ( var _ model ,
3 impulse = " e " ,
4 response = " prod " ,
5 n . ahead =20 ,
6 ortho = TRUE ,
7 boot = TRUE ,
8 runs =500 ,
9 ci =0.95)
46 / 59
 Exercise 6.3: Solution (Part 2)

1 # Step 3: Plot IRF

2 plot ( irf _ results ,
3 main = " Response of Productivity to Employment Shock " ,
4 ylab = " Response " ,
5 xlab = " Periods Ahead " )
6 # Access numerical values
7 print ( irf _ results $ irf $ e ) # Point estimates
8 print ( irf _ results $ Lower $ e ) # Lower CI
9 print ( irf _ results $ Upper $ e ) # Upper CI

47 / 59
 Exercise 6.3: Solution (Part 3)

1 # Custom plot
2 horizons <- 0:20
3 point _ est <- irf _ results $ irf $ e
4 lower <- irf _ results $ Lower $ e
5 upper <- irf _ results $ Upper $ e
6 plot ( horizons , point _ est , type = " l " ,
lwd =2 ,
7 ylim = range ( c ( lower , upper ) ) ,
8 main = " IRF : Employment -> Productivity "
,
9 xlab = " Horizon " , ylab = " Response " )
0 polygon ( c ( horizons , rev ( horizons ) ) ,
1 c ( lower , rev ( upper ) ) ,
2 col = rgb (0 ,0 ,1 ,0.2) , border = NA )
3 abline ( h =0 , lty =2 , col = " red " )

48 / 59
Exercise 6.3: Interpretation Guide
Solution
How to interpret IRF plots: Key features to identify:
1 Impact effect (horizon 0):

Contemporaneous response
Cholesky: impact depends on ordering
2 Peak response:
When is the maximum effect?
How large is it?
3 Persistence:
How long until IRF returns to zero?
Half-life: horizon when response drops to 50% of peak
4 Shape:
Monotonic decay vs hump-shaped
Overshooting (goes negative after being positive)
5 Statistical significance:
Does confidence band include zero?
If yes at some horizon: effect not significant 49 / 59
Exercise 6.3: Economic Interpretation Example
Solution
Example interpretation: Suppose the IRF shows:
Impact: 0.02 (horizon 0)
Peak: 0.08 at horizon 3
Returns to zero: around horizon 12
Confidence bands: exclude zero for horizons 0–8
Economic story: “A positive 1-standard-deviation shock to employment
leads to:
An immediate increase in productivity of 0.02%
The peak effect of 0.08% occurs after 3 quarters
The effect gradually dissipates, becoming statistically insignificant
after 2 years
This pattern suggests that increased employment boosts productivity
through learning-by-doing and scale effects, but the effect is
temporary” 50 / 59
Bonus Exercise: SVAR Identification

Bonus Question: Alternative Identification Schemes

Consider a monetary policy VAR with three variables:

yt : Output gap
πt : Inflation
rt : Interest rate (policy instrument)

Questions:
1 Under Cholesky ordering (y , π, r ), what does the identification
assume about monetary policy?
2 Is this assumption realistic? Why or why not?
3 Propose an alternative identification strategy
4 What external instrument could you use for monetary policy shocks?

51 / 59
Bonus Exercise: Solution (Part 1)
Solution
Question 1: Cholesky (y , π, r ) assumption Recursive structure:

yt = f (yt−1 , πt−1 , rt−1 ) + εyt

πt = g (yt , πt−1 , rt−1 ) + επt
rt = h(yt , πt , rt−1 ) + εrt

Implications:
Output shocks affect inflation and interest rate contemporaneously
Inflation shocks affect interest rate contemporaneously (but not
output)
Interest rate shocks affect nothing contemporaneously
Economic interpretation:
Central bank observes yt and πt and sets rt accordingly (within
period)
Monetary policy has no contemporaneous effect on output/inflation
52 / 59
Bonus Exercise: Solution (Part 2)

Solution
Question 2: Is this realistic?
Potential problems:
Information lag: Central bank may not observe yt , πt in real time
(data released with delay)
Decision lag: Monetary policy committee meets periodically (not
continuously)
Financial markets: Interest rate changes affect asset prices,
exchange rates immediately
Expectations: Forward-looking agents react to policy announcements

53 / 59
Bonus Exercise: Solution (Part 3)

Solution
Question 3: Alternative identification
Option 1: Reverse ordering (r , π, y )
Assumes policy affects inflation and output contemporaneously
But then: output can’t affect policy within period (unrealistic!)
Option 2: Sign restrictions
Contractionary policy shock: r ↑, y ↓, π ↓ (at least for some horizons)
Doesn’t impose exact timing of effects

54 / 59
Bonus Exercise: Solution (Part 4)
Solution
Question 4: External instruments (Proxy SVAR)
Idea: Use an instrument Zt correlated with true monetary policy shock
but uncorrelated with other shocks
Candidate instruments:
1 High-frequency identification (Gertler & Karadi, 2015):

Interest rate futures change in narrow window around FOMC

announcement
Captures “surprise” component of policy
2 Narrative approach (Romer & Romer, 2004):
Read FOMC minutes, identify intentions
Code policy changes as expansionary/contractionary
3 Policy rule residuals:
Estimate Taylor rule, use residuals as shocks
Problem: relies on correct specification
Advantage: Identification doesn’t rely on timing assumptions!
55 / 59
Summary
:
1 Stationarity Testing
Eigenvalue computation
Practical testing in R
2 Lag Length Selection
Information criteria (AIC, BIC, HQ)
Reconciling conflicting recommendations
3 Granger Causality
Hypothesis testing
Interpretation vs true causality
4 OLS Estimation
Coefficient interpretation
Statistical significance
5 VMA Representation
Deriving VMA from VAR
Computing Φ matrices
6 Impulse Response Functions
Identification problem
Cholesky decomposition
IRF interpretation 56 / 59
Key Takeaways

1 Stationarity is crucial
Always check before proceeding
Non-stationary VARs can give spurious results
2 Model selection matters
Use multiple criteria
Check residual diagnostics
3 Granger causality 6= true causality
It’s about predictive power, not mechanisms
4 Identification is key for IRFs
Reduced-form shocks are correlated
Need structure to isolate effects
Choice of identification affects conclusions
5 Always provide economic interpretation
Statistical results must connect to economic theory
IRFs tell stories about dynamic adjustment

57 / 59
Practice Recommendations
To master VAR analysis:
1 Replicate classic papers:

Sims (1980): Macroeconomics and Reality

Christiano, Eichenbaum, Evans (1999): Monetary policy shocks
Blanchard & Perotti (2002): Fiscal policy
2 Work with real data:
FRED (Federal Reserve Economic Data)
ECB Statistical Data Warehouse
IMF International Financial Statistics
3 Compare identification schemes:
How do results change with different orderings?
Try sign restrictions, external instruments
4 Read the vars package documentation:
Vignette: comprehensive examples
Function help: technical details
5 Explore extensions:
Time-varying VAR (TVP-VAR)
Bayesian VAR (BVAR)
Factor-augmented VAR (FAVAR)
58 / 59
Additional Resources

Textbooks:
Lütkepohl (2005): New Introduction to Multiple Time Series Analysis
Hamilton (1994): Time Series Analysis
Kilian & Lütkepohl (2017): Structural Vector Autoregressive Analysis

Software:
R: vars, svars, BVAR
Python: [Link]
MATLAB: built-in VAR functions, Uhlig’s toolkit

Datasets:
vars::Canada (practice dataset)
FRED-QD (quarterly macroeconomic database)
Stock & Watson macroeconomic data
59 / 59