Multivariate Models (VAR and Cointegration)

Here is an explanation of Multivariate Models, focusing on Vector Autoregression (VAR) and

Cointegration.

1. Motivation: Why go Multivariate?

In single-equation models (structural models), variables are classified as endogenous
(determined within the system, e.g., Quantity Demanded) or exogenous (determined outside,
e.g., Weather),. However, in finance and economics, variables often influence each other
simultaneously (e.g., House Prices ↔ Quantity of Houses).
• Simultaneity Bias: If you estimate a structural equation where a regressor is actually
endogenous (correlated with the error term) using OLS, the estimates will be biased and
inconsistent,.
• The Solution: Multivariate models allow for feedback loops where variables affect each other
over time without strictly imposing a theoretical structure of "cause and effect" a priori.

2. Vector Autoregression (VAR)

A VAR is a systems regression model where there is more than one dependent variable,. It can
be viewed as a hybrid between univariate time series models (like ARMA) and simultaneous
equations models.
• Structure: In a VAR, all variables are treated as endogenous,. Each variable in the system
depends on its own past values (lags) and the past values of all other variables in the system,.
◦ Example (Bivariate VAR with 1 lag):
y1t=β10+β11y1t−1+α11y2t−1+u1t
y2t=β20+β21y2t−1+α21y1t−1+u2t.
• Estimation: Provided there are no contemporaneous terms (current values) on the right-hand
side, a standard VAR can be estimated using OLS separately for each equation,.
• Advantages:
◦ Flexible and easy to estimate.
◦ Avoids the "incredible" identifying restrictions required in structural models to decide which
variables are exogenous.
◦ Often produces better forecasts than complex structural models,.
• Disadvantages:
◦ A-theoretical: It uses little financial theory to structure the model, making theoretical
interpretation difficult,.
◦ Parameter Bloat: The number of parameters increases quadratically with the number of
variables. For g variables and k lags, you must estimate g+kg2 parameters, which consumes
degrees of freedom,.

Tools for Analyzing VARs

Because VAR coefficients are hard to interpret directly, researchers use three key tools:
1. Granger Causality Tests: A test to see if past values of variable X contain statistically
significant information to predict variable Y,. Note: This indicates predictive correlation, not
necessarily physical causation.
2. Impulse Responses: These trace out the responsiveness of the dependent variables in the
VAR to a unit shock (innovation) in the error term of a specific variable,. It shows the magnitude
and persistence of a shock over time.
3. Variance Decompositions: This determines how much of the forecast error variance of a
given variable is explained by shocks to itself versus shocks to the other variables in the system,.

--------------------------------------------------------------------------------

3. Cointegration
Cointegration addresses long-run relationships between non-stationary variables.
• The Problem: Many financial time series (like stock prices or exchange rates) are non-
stationary (I(1)). Regressing one non-stationary series on another can lead to spurious
regressions (high R2, significant t-stats, but no real relationship).
• The Concept: Variables are cointegrated if they are individually non-stationary (I(1)) but a
linear combination of them is stationary (I(0)).
◦ Intuition: While the series wander individually, some economic force (like arbitrage or
market efficiency) binds them together in the long run, preventing them from drifting too far
apart.
• Examples: Spot and futures prices, equity prices and dividends.
Testing for Cointegration
1. Engle-Granger (2-Step) Method:
◦ Step 1: Regress yt on xt and save the residuals.
◦ Step 2: Test the residuals for stationarity (using ADF). If residuals are stationary (I(0)), the
variables are cointegrated,.
◦ Limitation: It can only detect one cointegrating relationship, even if more exist.
2. Johansen Technique:
◦ A systems-based approach using VARs that allows testing for multiple cointegrating
relationships.
◦ It tests the rank of the matrix Π using eigenvalues (λ). Two test statistics are used: the
Trace test (λtrace) and the Maximum Eigenvalue test (λmax).
--------------------------------------------------------------------------------
4. Error Correction Models (ECM)
If variables are cointegrated, they should be modeled using a Vector Error Correction Model
(VECM) (or ECM for single equations),.
• Granger Representation Theorem: If variables are cointegrated, there must exist a valid
Error Correction Model.
• Mechanism: The VECM allows for short-run dynamics (via first differences, Δyt, which are
stationary) while preserving the long-run equilibrium relationship (via the error correction
term, usually the lagged residuals from the cointegrating equation).
◦ The error correction term corrects for disequilibrium: if variables drift apart last period, this
term pulls them back together in the current period

Granger Causality Test

1. Concept and Definition
The Granger Causality Test is a statistical hypothesis test used to determine whether one time
series is useful in forecasting another.
• The Core Question: It seeks to answer questions such as, "Do changes in variable y1 cause
changes in variable y2?",.
• Predictive Power: A variable y1 is said to "Granger-cause" y2 if the past values (lags) of y1
contain statistically significant information that helps predict the current value of y2, beyond the
information contained in the past values of y2 itself.
• "Misnomer" Warning: The term "causality" is somewhat of a misnomer. It does not imply
that movements in one variable physically cause movements in another. Instead, it implies a
chronological ordering of movements or a correlation between the current value of one variable
and the past values of others,.

2. Context and Methodology

Granger causality tests are typically conducted within the framework of Vector Autoregression
(VAR) models. In a VAR system, all variables are treated as endogenous (dependent on each
other's past values).
To test for causality in a bivariate VAR system (with variables y1 and y2), researchers examine
the significance of the coefficients on the lagged variables:
• If we want to test if y1 Granger-causes y2, we look at the equation where y2 is the dependent
variable.
• We test if the coefficients on the lags of y1 are jointly significantly different from zero,.

3. The Hypotheses
The test uses an F-test to examine joint restrictions on the parameter matrices,.
• Null Hypothesis (H0): Lags of y1 do not explain current y2. (There is no Granger causality
from y1 to y2).
◦ Mathematical Implication: All coefficients on the lags of y1 in the y2 equation are zero
(e.g., β21=0,γ21=0,δ21=0),.
• Alternative Hypothesis (H1): Lags of y1 do explain current y2. (y1 Granger-causes y2).

4. Possible Outcomes
When analyzing the relationship between two variables, there are three main outcomes:
1. Unidirectional Causality: Occurs if lags of y1 are significant in the y2 equation, but lags of
y2 are not significant in the y1 equation (or vice-versa). We say y1 Granger-causes y2.
2. Bi-directional Causality (Feedback): Occurs if lags of y1 are significant in the y2 equation
AND lags of y2 are significant in the y1 equation. This implies the variables influence each
other.
3. Independence: Occurs if neither set of lags is statistically significant in the equation for the
other variable.
5. Empirical Application
In the provided Sample Exam, Question 24 confirms that if the null hypothesis is rejected, it
means that "Past values of X provide statistically significant information for predicting Y",.
In practice (e.g., using EViews), these tests are often referred to as Block Exogeneity Tests or
Block Significance Tests, where the software tests whether the parameters for a given lag of all
variables in a particular equation can be restricted to zero

Volatility Modeling (ARCH/GARCH)

Here is an explanation of Volatility Modeling, focusing on ARCH and GARCH.

1. Motivation: Why Model Volatility?

Standard regression models (like OLS) assume homoscedasticity, meaning the variance of the
error term is constant over time. However, financial time series (stock returns, exchange rates)
typically exhibit "Stylized Facts" that violate this assumption:
• Volatility Clustering: Large changes in prices tend to be followed by large changes, and small
changes by small changes. Volatility comes in "bursts",.
• Leptokurtosis: The distribution of returns has "fat tails" (extreme events happen more
frequently than a normal distribution predicts),.
• Leverage Effects: Price falls (bad news) often increase volatility more than price rises (good
news) of the same magnitude.
To capture these features, we use models that allow the variance (σ2) to change over time
(conditional heteroscedasticity).

2. The ARCH Model

Proposed by Engle (1982), the Autoregressive Conditional Heteroskedasticity (ARCH) model
assumes that the current variance depends on the squared errors from previous periods.
• The Equation (ARCH(1)): σt2=α0+α1ut−12 Here, the conditional variance (σt2) is
determined by a constant and the "news" or shock from the previous period (ut−12).
• Limitations:
◦ It often requires many lags (q) to model the data accurately, making the model non-
parsimonious.
◦ The parameters (α) must be positive to ensure variance is non-negative, which can be
violated in estimation.
3. The GARCH Model
Developed by Bollerslev (1986), the Generalized ARCH (GARCH) model is the most widely
used tool for volatility modeling because it is more parsimonious than ARCH.
• The Equation (GARCH(1,1)): σt2=α0+α1ut−12+βσt−12 The current variance depends on:
1. α0: A long-term average level.
2. ut−12 (ARCH term): The lag of the squared residual (recent news/shocks).
3. σt−12 (GARCH term): The lag of the variance itself (persistence).
• Interpretation:
◦ A high β indicates that volatility is highly persistent (shocks die out slowly).
 ◦ Stationarity Condition: For the variance to be stable, we require α1+β<1. If α1+β=1, it
is an Integrated GARCH (IGARCH) model, where shocks persist infinitely.
◦ Parsimony: A GARCH(1,1) model is mathematically equivalent to an ARCH model with
infinite lags, making it very efficient.

4. Estimation Method
You cannot use OLS to estimate GARCH models because OLS minimizes the residual sum of
squares based on the mean, not the variance.
• Maximum Likelihood (ML): GARCH models are estimated using Maximum Likelihood,
which finds the parameter values that are "most likely" to have produced the observed data,.

5. Extensions: Asymmetric Models

Standard GARCH models are symmetric: because they square the error term (ut−12), positive
shocks and negative shocks have the same effect on volatility. To handle leverage effects (where
bad news creates more volatility), asymmetric models are used:
• GJR-GARCH: Adds a dummy variable that activates only when the shock is negative (ut−1
<0). If the coefficient on this dummy is positive, it confirms that bad news increases volatility
more than good news,.
• EGARCH (Exponential GARCH): Models the natural log of the variance (lnσt2). This
ensures the variance is always positive without imposing restrictions on the coefficients and
allows for asymmetric responses to shocks