Reference:

[Link]

4a) Elasticity is defined as (dy/y)/(dx/x) = {(dy/dx)*(x/y)} .We’ll calculate elasticities for all the
stated regression.

(a) y = 2 + 0.8x

Differentiating w.r.t. x on both sides

dy/dx = 0 + 0.8 = 0.8
thus , elasticity = 0.8 * (x/y) = 0.8 *(x/(2+0.8x))

(b) ln(y)= 0.1 +0.4x

Differentiating w.r.t. x on both sides

(1/y)*(dy/dx) = 0 + 0.4 => dy/dx = 0.4y
thus, elasticity = (0.4y)*(x/y) = 0.4x

(c) ln(y) = 0.1 + 0.25ln(x)

Differentiating w.r.t. x on both sides

(1/y)*(dy/dx) = 0 + 0.25*(1/x) => dy/dx = 0.25*(y/x)
thus, elasticity = 0.25*(y/x)*(x/y) = 0.25

(d) y = 0.15+1.2ln(x)

Differentiating w.r.t. x on both sides

dy/dx = 0 + 1.2*(1/x) => dy/dx = 1.2/x
thus, elasticity = (1.2/x)*(x/y) = 1.2/y = 1.2/(0.15 + 1.2ln(x))

4b) to find out the elasticity of y w.r.t. x we must use regression with constant elasticity and that we
can see by above solution is
Regression (c) ln(y) = 0.1 + 0.25ln(x) with constant elasticity of 0.25 which is independent of x.
where elasticity of 0.25 signifies that if we make changes in x by 1% then y will be changed by
0.25% and this interpretation doesn’t depend on the specific x or y values.

5a)

Issue of stationarity for time series model:

A stationary time series means that its mean and auto covariance are independent of absolute time
location of data observations. For example, Cov(Xt,Xs) must depend only on the difference t-s.
The estimation of ACF and PACF will be accurate and stable without the worry of which section of
data we choose to estimate. Hence, it is critical to obtain a stationary time series before doing mod-
elling [Link] instance in an ARMA model, the time series is required to be stationary before
modelling. There are also coefficient restrictions in both the Box-Jenkins method and GARCH
model to ensure the modelled process is stationary.

How to test for presence of a unit root:

ACF and PACF Interpretation : For a stationary time series ACF and PACF drops to zero quickly.
This indicates short-term dependence, typical of stationary processes like AR(p) or MA(q) with low
orders. But it only provides rough guess for unit root. We have more formal tests like Dickey fuller
test, Augmented dickey fuller test and KPSS test and for checking presence of unit root.
Which have discussed Dickey fuller and Augmented Dickey fuller test below.

1) Dickey Fuller Test

The Dickey Fuller test determines if unit roots exist in a time series. It is only valid for the AR(1)
[Link] the following AR(1) process:

xt=α*x(t−1)+vt, where vt is white noise

we can rewrite the above equation as follows:

xt−x(t−1)=αx(t−1)−x(t−1)+υt

⇒∇xt=(α−1)x(t−1)+υt

⇒∇xt=γx(t−1)+υt, where γ=(α−1)

Hence, the basic Dickey Fuller test can be set up as follows:

H0: γ=0 H1: γ≠0

The above Dickey Fuller test is equivalent to test if α=1. However, under Ho, the time series is not
stationary and the usual normal distribution tests do not apply. The test statistic we will use is
known as a unit root test statistic or Dickey Fuller (DF) statistic. If we set the significance level at
5% and the DF statistic > 5% critical value, we fail to reject [Link] means there is a unit root
for xt.
2.)Augmented Dickey Fuller test (ADF): The augmented Dickey Fuller test (ADF) is the most
popular unit root test to use in research. Compared to the Dickey Fuller test, the ADF test includes
the lag difference terms of the time series in the equation to control the serial correlation in the time
series of interest. We can write the three versions of ADF equations and tests as follows:

2.1) ADF test without Drift and Trend:

∇xt= γ*x(t-1) +∑pi=2βi∇xt−i+1+υt

(τ1):H0:γ=0

2.2) ADF Test with Drift and without Trend

∇xt=γxt−1+α0+∑pi=2βi∇xt−i+1+υt
(ϕ1):H0:γ=0,α0=0
(τ2):H0:γ=0

2.3) ADF Test with Drift and with Trend:

∇xt=γxt−1+α0+α2t+∑pi=2βi∇xt−i+1+υt
(ϕ2):H0:γ=0,α0=0,α2=0

(ϕ3):H0:γ=0,α0=0
(τ3):H0:γ=0

Arrange the equations

The versions of the ADF test are similar to the Dickey Fuller test, except that now we need to find
the optimal lag order p for the equation. The software we use will automatically search for the opti-
mal p based on the information criteria (AIC or BIC) we select for the test.

What to do if unit root is found:

Once we conclude that a time series is non-stationary due to presence of unit root, we can take the
first difference of the time series and then check the new time series again with a unit root test. It is
an iterative process until the time series has no unit root and is stationary. Also, if there is a time
trend, we can implement the regression method to remove the trend. Once we are sure the time se-
ries is stationary, we can move on to build an ARMA model or conduct the next analysis.

5b) We have downloaded data for Microsoft share from Kaggle.

We have used the dataset for adjusted closed share price of Microsoft from 2016-22. First we plot-
ted the dataset, it shows exponential growth in share price. Then we have performed 3 versions of
ADF test and KPSS test.

is -1.94. Hence, we cannot reject 𝐻o. The conclusion is that there is a presence of unit root for
1. ADF Test without Drift and Trend: calculated test statistic is 2.655, while the 5% critical value

Microsoft share price time series.

value is -2.86. Hence, we cannot reject 𝐻o. The conclusion is that there is a presence of unit
2. ADF Test with Drift and without Trend: calculated test statistic is 0.619, while the 5% critical

root for Microsoft share price time series.

value is -3.41. Hence, we cannot reject 𝐻o. The conclusion is that there is a unit root in Micro-
3. ADF Test with Drift and with Trend: calculated test statistic is -2.189, while the 5% critical

soft's stock price.

We see that we can not reject Ho at a 5% critical value for all three models. This means that the
time trend should not be included in the model.

KPSS test : Results for KPSS test shows that p-value is 0.000 which is smaller than 0.05 thus we
can reject the null hypothesis which is that time series is weakly stationary.

Conclusion: above time series is not trend stationary and has a unit root.

5c)
Assume xt follows an AR(1) process:
xt=α0+α1xt−1+vt

Where vt is a stationary ARMA process with mean = 0

We are concerned about unit root because it leads to non stationarity and usually natural time series
in economic and finance is non stationary while if time series is with root of 1.5 then for AR(1)
process phi coefficient will be 1/1.5 = 0.667 and that leads to stationary time series and that is easy
to analyse.

While if we consider time series with phi coefficient of 1.5 than that is not important in field of eco-
nomic and finance as it will explode.

If |α1| > 1 , then xt is an explosive process. If xt follows an explosive process, it means that as time
goes by, x would either keep growing larger and larger or reducing more and more without bounds.
Explosive processes are not seen in economics and finance, that’s why we are not concerned about
a coefficient α1 which has absolute value greater than 1 (for instance1.5).

We are presenting simulated time series for all the 3 cases by assuming AR(1) process.
1. With unit root ( α1= 1)
2. With root value 1.5 (α1= 0.667)
3. With root value 0.667(α1 = 1.5)

6a) considering below linear regression model

Y(t) = α + βX(t) + ε(t) t=1, 2, .., 20
we want to check for structural break of the β parameter at t=10 due to suspicion.
We need to define a dummy variable to test it in single regression.

define D(t) dummy variable which takes value 0 for t<=10 and takes value 1 for 11=<t<=20
And consider regression:
Y(t) = α + β1X(t) +(β2 - β1)(D(t)*X(t)) + ε(t) t=1, 2, .., 20

Define β3 = β2 - β1
Y(t) = α + β1X(t) + β3(D(t)*X(t)) + ε(t)

Now simplifies above equation for t<=10 implies D(t) = 0

and regression equation becomes Y(t) = α + β1X(t) + ε(t) t=1, 2, .., 10

And for 10<t<=20 it implies D(t) = 1

thus regression equation becomes
Y(t) = α + β1X(t) +(β2 - β1)(1*X(t)) + ε(t)

Y(t) = α + β2X(t) + ε(t) for t = 11,12,….20 thus by defining dummy variable D(t) single regression
Y(t) = α + β1X(t) +(β2 - β1)(D(t)*X(t)) + ε(t) t=1, 2, .., 20 is valid for whole analysis.

now we’ll define null hypothesis as (β2 - β1)= β3 = 0

alternate hypothesis (β2 - β1 = β3 != 0

If β3 coefficient is statistically significant then we have an evidence of structural break in the β pa-
rameter at t=10.