Finalytics Experiential Learning Assignment

By: Group-10 Division B (Fintech)

Simona Sanjeebita- 23020942028
Aakarshika Shukla- 23020942003
Vipul Mishra- 23020942012
Shreyas Prabhu- 23020942047
Q.1 Describe several methods for estimating simultaneous equation models with suitable
example
Simultaneous equation models (SEMs) consist of multiple interrelated equations where
endogenous variables in one equation can appear as explanatory variables in others.
Estimating these models requires specialized techniques due to the potential endogeneity
and simultaneity biases. Common methods include:

Indirect Least Squares (ILS): This method involves estimating the reduced form of the
system (where each endogenous variable is expressed solely in terms of exogenous
variables) using Ordinary Least Squares (OLS). The structural parameters are then derived
from these reduced-form estimates.

Two-Stage Least Squares (2SLS): In the first stage, the endogenous explanatory variables are
regressed on all exogenous variables in the system to obtain predicted values. In the second
stage, the original structural equation is estimated using these predicted values in place of
the endogenous variables.

Three-Stage Least Squares (3SLS): An extension of 2SLS, this method accounts for potential
correlations between the error terms of different equations. It combines 2SLS with
seemingly unrelated regressions (SUR) to provide more efficient estimates.

Example: Consider a simple supply and demand model where both price (P) and quantity
(Q) are endogenous:
Here,
Demand: Qd = α – βP + ud
Supply: Qs = γ + δP + us

Here, Qd and Qs represent the quantity demanded and supplied, respectively, ud

And us are error terms. Since price and quantity are determined simultaneously, estimating
these equations requires methods like 2SLS to address the endogeneity of P and Q.

Q.2 Explain relative advantages and disadvantages of VAR modelling with suitable
examples.
Vector Autoregressive (VAR) models are among the few tools that try to capture the linear
dependencies among a collection of time series.
Advantages:

1. Simplicity: These models treat every single variable as endogenous and appraise
each variable as a function of its own lags and the lags of other variables; therefore,
it simplifies the modelling.
 2. Flexibility: It captures complex dynamic behaviours without a priori restrictions on
the relations.
3. Forecasting: Best-performance in forecasting than univariate models by utilizing the
information from multiple time series.
Disadvantages:
1. Overparameterization: Adding too many lags or variables can result in overfitting
and reduce the predictive power of the model.
2. Interpretation Problems: Without structural identification, it may be quite difficult to
specify an interpretation for the relations between variables.
3. Assumption of Stationarity: A VAR model assumes that the time series must be
stationary. Unless the time series is transformed correctly, it will produce spurious
results.
Example: It would be possible to apply a VAR model to the relationship between interest
rates, inflation, and GDP growth rates; each would be modelled as a function of its own past
observations and those of the other variables.

Q.3 Determine whether an equation from a system is identified.

Identification in SEMs is defined as the condition under which it is possible to obtain a
unique estimate of the structural parameters based on the information at hand. An equation
is said to be identified when it is possible to identify from the available data unique values
for all its parameters.

Order Condition for Identification:

For an equation to be identified, the number of excluded exogenous variables (present in
one equation but not in the equation of interest) must be at least the number of
endogenous variables minus one.

Just identified: The number of excluded exogenous variables equals the number of
endogenous variables minus one.
Overidentified: There are more excluded exogenous variables than the number of
endogenous variables minus one.
Under identified: There are fewer excluded exogenous variables than the number of
endogenous variables minus one.
By way of example, an equation is just identified in a system with three endogenous
variables (Y1, Y2, Y3) and many exogenous variables (X1, X2, X3) if two of the exogenous
variables are excluded and there are two other endogenous variables in the system.
Q4: Estimate optimal lag lengths, impulse responses, and variance decompositions.
Selecting the appropriate lag length for VAR modelling is important for estimating the
dynamics of the system without overfitting.
Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and Hannan-
Quinn criterion (HQC): Criteria commonly used to determine the optimal number of lags,
maximizing the trade-off between model fit and parsimony.
Impulse Response Functions (IRFs): The IRFs trace the effect of a one-time shock of one of
the variables on the values of all variables in the system for the current and future periods at
a given point in time. They are useful in unfolding the dynamic interactions accompanying
each other with regard to an impulse.
The results of the VAR model done on time series data of stocks and volume data of Bajaj
FinServ ltd., are as follows:
 Optimal lag length: 1 (based on AIC)
 Model: VAR
 Method: OLS

 Date: Friday, January 31, 2025

 Time: 12:10:52
 Number of Equations: 2.00000
 Nobs: 99.0000
 Log likelihood: -1097.27

 AIC: 16.5317
 BIC: 16.5842
 HQIC: 16.5529
 FPE: 1.51230e+07
 Det(Omega_mle): 1.48220e+07

Results for equation Stock Price_diff:

Results for equation Trade Volume_diff:

Correlation matrix of residuals:

Variance Decompositions: This technique decomposes the forecast error variance of each
variable into proportions that can be traced back to shocks of each variable in the system,
giving one some perspective on the relative importance of each shock.
Example: After estimating a VAR model with GDP growth, inflation, and interest rates, one
can use AIC to select the lag length of two. Then, one may proceed to conduct IRFs of
macroeconomic variables to see how over time a shock to interest rates affects GDP growth
and inflation, with variance decomposition showing that shocks to interest rates account for
the considerable share of the GDP growth forecast error variance.

Q.5 Conduct Granger causality tests

The Granger Causality Test is a way to ascertain whether one time series can predict another.
To put it another way, it tests if the past values of any one time series contain significant
information in predicting the future value of any other series.
A few outcomes that may result from the test results and their respective interpretations are
highlighted below:
1. Existence of Granger Causality (Cause and Effect/Relationship): A p-value less than the
significance level (typically taken as 0.05) suggests that the Granger causality exists. This
means past values of one variable contain relevant information for the prediction of the
future values of the other variable.
Example: If testing if variable X Granger-causes variable Y, then with a p-value equal to less
than 0.05; we will conclude that X Granger-causes Y.

2. There is No Granger Causality: A p-value greater than significance predefined rather

means we cannot reject the null; past values of the first variable do not provide any
predictive power for the second variable.

Example: When testing between X and Y, if the resulting p-value is >0.05, we conclude that
there is no Granger causality between X and Y.

3. Bidirectional Granger Causality:

If both variables Granger-cause one another (both p-values are significant), we say there is
bidirectional Granger causality; this means they have a feedback loop running between each
other.
Example: If both X Granger-causes Y and Y Granger-causes X it implies a reciprocal
relationship between them.
4. Unidirectional Granger Causality: If only one among the two variables Granger-causes the
other, we say there is unidirectional Granger causality; this suggests that the causal influence
runs only in one direction.
Example: When X Granger-causes Y while the reverse does not hold certainly, we conclude
unidirectional causality from X to Y.
Ganger Causality Test on time series data of Bajaj FinServ ltd.,
Granger causality tests is conducted to determine if changes in Trade Volume can predict
changes in Stock Price, and vice versa. Then estimation of impulse responses and variance
decompositions to understand the dynamic relationships between these two variables is
done.
The results show that there is no significant causal relationship between Trade Volume_diff
and Stock Price_diff in either direction.

The impulse response and variance decomposition analyses did not return valid results. This
is because the optimal lag length for the VAR model is 0, indicating that there is no dynamic
relationship between the variables. Therefore, even though the VAR model with a lag length
of 1, it does not change the fact that there is no dynamic relationship between the variables.