Questions for Practice

Question 1
Suppose you are analyzing the interaction between Export Growth (%) and
the Exchange Rate (₹ per USD) in a trade-dependent economy.
Policymakers believe that higher exports strengthen the domestic currency
(lower exchange rate), while a depreciating exchange rate can boost
exports by making them more competitive.
Recent data (last 4 quarters):
Export Growth
Quarter Exchange Rate (₹/USD)
(%)
t=1 3.0 75.0
t=2 3.5 74.5
t=3 4.0 74.2
t=4 3.8 74.0

Estimated VAR(1) model:

EX P t=1.0+ 0.6 EX Pt −1−0.4 EX R t−1

EX R t=70.0−0.3 EX P t−1 +0.7 EX R t−1

Questions
1. Using the data from quarter 4 (EXP=3.8, EXR=74.0), compute the
predicted values of export growth and exchange rate for quarters 5, 6, and
7. Comment briefly on the economic dynamics.

2. Suppose in quarter 5, the exchange rate experiences a sudden shock

depreciation (+2 units above forecast). Trace the effect of this shock on
export growth over the next 3 quarters, keeping other dynamics constant.

Solution
Given
Model:
EX P t=1.0+ 0.6 EX Pt −1−0.4 EX R t−1

EX R t=70.0−0.3 EX P t−1 +0.7 EX R t−1

Last observed (quarter 4): EX P 4=3.8 , EX R 4=74.0 .

Part 1 Baseline forecasts (no shock) for quarters 5, 6, 7

Quarter 5
EX P 5=1.0+0.6 ( 3.8 )−0.4 ( 74.0 ) =1.0+2.28−29.6=−26.3200

EX R 5=70.0−0.3 ( 3.8 ) +0.7 ( 74.0 )=70.0−1.14+ 51.8=120.6600

Quarter 6
EX P 6=1.0+0.6 (−26.32 )−0.4 ( 120.66 )=1.0−15.792−48.264=−63.0560

EX R 6=70.0−0.3 (−26.32 ) +0.7 ( 120.66 ) =70.0+7.896+84.462=162.3580

Quarter 7

EX P 7=1.0+0.6 (−63.056 )−0.4 ( 162.358 )=1.0−37.8336−64.9432=−101.7768

EX R 7=70.0−0.3 (−63.056 ) +0.7 ( 162.358 )=70.0+18.9168+113.6506=202.5674

Baseline forecasts (rounded):

Export growth : EX P5=−26.3200 , EX P6=−63.0560 , EX P 7=−101.7768 .

Exchange rate : EX R 5=120.6600 , EX R6 =162.3580 , EX R7=202.5674 .

The system explodes quickly , exchange rate depreciates sharply (large

rise in ₹/USD) and export growth turns deeply negative.
Economically this suggests the estimated coefficients imply depreciation
reduces export growth (the −0.4 term): perhaps the economy is import-
intensive (higher EXR raises export costs), or the model is misspecified.

Part 2 — Shock experiment: exchange-rate depreciation of +2 units at

quarter 5
Assume at quarter 5 the exchange rate is 2 units above forecast:

shock
EX R 5 =120.6600+2.00=122.6600 .

Keep EX P 5 at its baseline value (−26.3200) for the shock timing, then
propagate forward. We trace the effect of this EXR shock on export growth
for the next three quarters (i.e. , EX P6 , EX P7, EX P 8 under shock).

Quarter 6 (after shock)

shock
EX P 6 =1.0+0.6 (−26.32 ) −0.4 ( 122.66 )
¿ 1.0−15.792−49.064=−63.8560

shock
EX R 6 =70.0−0.3 (−26.32 ) +0.7 ( 122.66 )
¿ 70.0+7.896+ 85.862=163.7580

Quarter 7
shock
EX P 7 =1.0+0.6 (−63.8560 )−0.4 ( 163.7580 )
¿ 1.0−38.3136−65.5032=−102.8168
shock
EX R 7 =70.0−0.3 (−63.8560 ) +0.7 ( 163.7580 )
¿ 70.0+19.1568+114.6306=203.7874

# Quarter 8
 shock
EX P 8 =1.0+0.6 (−102.8168 )−0.4 ( 203.7874 )
¿ 1.0−61.6901−81.5149=−142.2050
shock
EX R 8 =70.0−0.3 (−102.8168 ) +0.7 ( 203.7874 )
¿ 70.0+30.8440+142.6512=243.4952

Shock-path for export growth (rounded):

shock
EX P 6 =−63.8560
shock
EX P 7 =−102.8168
shock
EX P 8 =−142.2050

Comparison with baseline:

EXP6: baseline −63.0560 → shock −63.8560 (worse by ≈ −0.8000)

EXP7: baseline −101.7768 → shock −102.8168 (worse by ≈ −1.0400)
EXP8: baseline −141.0930 → shock −142.2050 (worse by ≈ −1.1120)

So the +2 depreciation shock makes export growth slightly more negative

in subsequent quarters, and the adverse effect accumulates in the explosive
trajectory.
Interpretation
1. Sign of the EXR coefficient in the EXP equation (−0.4) implies: a higher
numerical exchange rate (larger ₹/USD = weaker rupee) reduces export
growth in this estimated system. That is economically plausible if exporters
depend heavily on imported inputs or if depreciation raises domestic costs
more than it helps competitiveness.

2. Shock impact: A depreciation shock causes export growth to worsen

further (more negative) in the model because of that negative coefficient.
The effect is transmitted immediately and amplified over time because the
system’s lag structure is explosive.
3. Model stability warning: The forecasts explode quickly (EXR rising to
very large values, EXP becoming very negative). This indicates the
estimated VAR is unstable (roots outside the unit circle). Before relying on
multi-step forecasts or shock analysis, you should:
# Check stability (companion matrix eigenvalues).
# Re-specify (include more lags, deterministic terms, or transform
variables) or use cointegration/VECM if appropriate.
# Re-examine data scaling and estimation — the large constant in the
EXR equation (70) and interaction signs are producing explosive
dynamics.

Question 2
You are studying the relationship between Exchange Rate (USD/INR) and
Foreign Capital Inflows (in billion) using a VAR model. Before estimation,
your professor asks you to check whether both series are stationary.
The Augmented Dickey-Fuller (ADF) test is applied on the two variables
with the following results (at 5% significance):

Variable Test Statistic Critical Value p-value Result

Exchange Rate –2.10 –2.95 0.25 ?
Capital Inflows –3.50 –2.95 0.01 ?

Questions:
1. Explain the steps of the ADF test in this context.
2. State whether each variable is stationary or not based on the above table.
3. Why is the ADF test a necessary pre-step before estimating a VAR model?

Solution:
1. Steps in the ADF test:
# Null hypothesis: Variable has a unit root (non-stationary).
# Estimate the regression of the variable on its lag(s) and test if the
coefficient on lagged level = 0.
# Compare test statistic with critical value (or use p-value).
# Reject null if test statistic < critical value (or p < 0.05).

2. Interpretation:
# Exchange Rate: Test statistic = –2.10 > –2.95; p=0.25 > 0.05 → Fail to
reject null → non-stationary.
# Capital Inflows: Test statistic = –3.50 < –2.95; p=0.01 < 0.05 → Reject
null → stationary.

3. Why ADF test before VAR?

# VAR requires stationary series, otherwise estimated relationships
may be spurious.
# If variables are non-stationary, shocks do not die out and variance
grows over time, violating VAR assumptions.
# Hence, the ADF test helps decide whether to difference the data or
use VAR in levels/with cointegration (VECM).

Multivariate Volatility Models

In financial markets, assets rarely move in isolation — stocks, bonds, or
currencies often rise and fall together. This makes it important to study not
just how risky each asset is on its own, but also how their risks interact.
For example, the risk of a portfolio depends on both the individual
volatilities and the way these volatilities co-move, which is captured by
their covariances.
This is crucial for portfolio management, hedging strategies, and even
pricing of derivatives, where changing relationships between assets
directly affect decisions.
During crises, we also see shocks spreading from one market or country to
another, and multivariate volatility models help us measure and
understand these spillover effects.
In essence, these models allow us to capture the joint behavior of financial
assets, providing a fuller and more realistic picture of risk than studying
each asset separately.

Multivariate volatility models are designed to capture not only the volatility
of individual financial assets but also the way their volatilities and risks
move together over time.

The most widely used models include the VECH model (which directly
models all elements of the variance–covariance matrix), the BEKK model (a
more parsimonious version that ensures the covariance matrix remains
positive definite), and the DCC model (Dynamic Conditional Correlation),
which focuses on time-varying correlations across assets.

These models are crucial in finance because portfolio risk, hedging

strategies, and contagion analysis depend on understanding both
volatilities and co-movements. By applying them, analysts can better
measure spillovers between markets, design effective hedge ratios, and
study how correlations change, especially during periods of financial stress

Estimating BEKK(1,1)
Estimating a BEKK(1,1) model (Baba, Engle, Kraft, and Kroner, 1990)
follows a well-structured process

1. Recall the BEKK(1,1) Model Structure

For a k-dimensional return vector r t with innovation ε t :

ε t ∼ ( 0 , H t ) , r t =μt + ε t

Conditional covariance matrix:

' ' ' '
H t =C C+ A ε t−1 ε t−1 A + B H t −1 B

H t =k × k conditional variance–covariance matrix.

C = lower triangular matrix (C ' C ensures H t is positive definite).

A = captures ARCH effects (impact of lagged shocks).
B = captures GARCH effects (persistence of past volatility).

2. Prepare the Data

# Collect multivariate time series (e.g., daily returns on two stock
indices).
# Ensure stationarity (e.g., via ADF test on returns).
# Standardize or de-mean if necessary.

3. Specify the Mean Equation

# For each series, write a mean model (e.g., constant mean or VAR(p) if
needed).
# Residuals from the mean equation become ε t .

4. Initialize Parameters
 # Choose starting values for C, A, B.
# Often C is initialized as a Cholesky factor of the unconditional
covariance of residuals.
# A and B can start with small diagonal values (e.g., 0.1).

5. Construct the Likelihood Function

For each t, compute H t recursively using the BEKK equation.

The log-likelihood for a Gaussian BEKK(1,1) is:
T
−Tk 1
l ( θ )= ln ( 2 π )− ∑ ( ln |H t|+ ε 't H −1
t εt )
2 2 t=1

where θ=\{ C , A , B \} .

Here:

 T = number of time periods (sample size).

 k = dimension of the system, i.e. number of variables in the
VAR/BEKK system (say stock returns of 2 assets → k=2).
 H t = conditional variance–covariance matrix of size k×k
 ε t=k × 1 vector of innovations (errors) at time t.
 k: Refers to the number of series being modeled. For
example:

k=2: modeling return series of two assets.

k=3: modeling return series of three assets.

 T: Refers to the time dimension, i.e. how many observations

you have (sample size). For example:

 T=500: daily data for 2 years.

 T=120: monthly data for 10 years.

6. Estimate Parameters
 # Use numerical optimization (e.g., BFGS, Newton-Raphson, Quasi-
Newton) to maximize l ( θ ) .
# Software packages: R (ccgarch, rmgarch), Python (arch, mgarch),
MATLAB (MFE toolbox), or EViews.

7. Check Model Diagnostics

# Verify positivity and stationarity (eigenvalues of A ⊗ A+ B ⊗ B< 1¿ .
# Check residuals for remaining ARCH effects (multivariate LM test).
# Ensure H t is positive definite throughout the sample.

8. Interpret Results
# Diagonal elements of A: own-market shock effects.
# Off-diagonal elements of A: cross-market shock spillovers.
# Diagonal elements of B: own volatility persistence.
# Off-diagonal elements of B: volatility transmission across series.

9. Applications
# Volatility spillover analysis (e.g., S&P 500 to Nikkei).
# Portfolio risk management (time-varying covariance).
# Impulse response in volatility (shock transmission).

Important topics
Which model (IGARCH, EGARCH, or GARCH-M) would you choose
and why?
Steps in Arima Model building and estimation, stationarity testing in
ARIMA
Cointegration in time series
Steps in VAR Model and Estimation {discussed , refer to previous
notes}
Difference in BEKK vs DCC, ARCH vs GARCH, VECH vs BEKK
Main time series assumptions like Stationarity, Cointegration,
ADF test steps (demonstrated in class already)
Volatility Clustering (discussed and shown numerically)
Stationarity (discussed)
Granger Causality (discussed)