Assignment

Section 1: Linear Regression Analysis

By interpreting the coefficients of the estimation of the linear regression model, we explore that
the coefficients of all the variables expect SMB, have a positive effect on the Microsoft’s risk
premium. Moreover, the only significant variable in the model is the risk-free rate, an increase of
1% of the value of the explanatory variable risk-free rate has a positive impact of the Microsoft’s
risk premium with an increase of 1,306243%. An increase of 1% in SMB variable impacts the
Microsoft’s risk premium with a decrease of 0,0016571%, an increase of 1 point in HML
variable impacts the Microsoft’s risk premium with 0,0024777%, the linear regression model has
a value of the constant of 0,0013498.

To test the significance of the slope coefficients or the explanatory variables in the model, we
use the 95% confidence interval approach. The first slope with a value of 1,306243 with a 95%
confidence interval that does not contains a 0 in its width, though we claim that the slope of the
risk-free rate is significant at a 95% level of confidence, the second slope of the SMB variable
with a value of -0,0016571, we explore that it’s 95% confidence interval does contain a 0 in its
width then we conclude that this slope is insignificant in the model, the HML slope has a value
of 0,0024777, the 95% confidence interval does contain a 0 in its width, though, we claim that
this slope is also insignificant in the linear regression model. Last the constant is also
insignificant in the model because it’s 95% confidence interval contains the value 0 in its width.
We can also explore this observation by the p-value approach, we see that the p-value of the risk-
free rate is equal to 0.000 less than 0.05 then in a 95% level of confidence we reject the null
hypothesis that the slope of this variable is equal to 0. Hence the slope of the risk-free rate is a
significant variable in the model explaining the Microsoft risk premium by the explanatory
variables risk-free rate SMB and HMLg. The other variables HML and SMB and the constant
has P values higher than 0.05 then at a 95% level of confidence we reject the null hypothesis and
assume that those slopes of these variables are insignificant.

If we wish to test the hypothesis that the variables SMB and HML have a statistically significant
joint effect on Microsoft’s risk premium, the Null hypothesis that we set up is that jointly SMB
and HML has no significant effect on the output variable, formally we set up the Null hypothesis
β SMB= β HML=0 .

We want to compare the estimated models in Table 1 and Table 2, the first model using 3
explanatory variables which are HML, SMB and risk-free rate, but we explore that it has only
one significant variable which is the only variable tested in the second model, it is the risk-free
rate which is significant in explaining the Microsoft’s risk premium, also we explore that the
width of the 95% confidence interval of the slope of the risk-free rate in the model 1 is larger
than the model 2, though we claim that the slop of the risk-free rate is estimated in the second
model with more precision at a 95% level of confidence. Last, we compare the R-squared, the
model 2 has a bigger value of R-squared. Thus we claim that the model in Table2 fits the data
better than the model in Table 1. Also we can assume that the adding of the variables HML and
SMB from the model 1 to model 2 impact the skewness and the precision of the estimations in
explaining the Microsoft’s risk premium with the explanatory variable risk-free rate.

Section 2: Time Series Analysis

With reference to ACF / PACF & dickey fuller test, one can make the comparison amongst two-
time series - close price & return. On the comparison of ACF in the context of two-time series,
one can make the conclusion as ACF related to close prices indicate that value of time series has
a positive correlation with respect to its first lagged value. Furthermore, acf in the context of
returns shows that series has positive correlation & negative in context of past lagged value.
PACF of two-time series corroborates initial observation. So on removal of effect of first lag,
one observes that PACF related to close price specifies that time series has correlation just with
first lagged value, i.e. lag 1. But, PACF in the context of return series shows that series values
have the correlations which depend on all previous lagged values.

Prior to the success of Dickey Fuller for confirming the observations one would claim that one
can easily fit ARMA model to return time series. Exploration of Dickey Fuller test shows
stationary time series. Dickey Fuller test related to close price time series shows T statistics of
test = -0,675 and does not exceed critical value of test. So we cannot reject Null hypothesis
meaning that time series is stationary. Also it is observed p-value of test equals 0,8530. This also
does not exceed Alpha value of 0,05. So it can be said that time series of close prices cannot be
a stationary time series. Also, p-value of Lag 1 in the time series equals 0,50 so one can say that
lag cannot be significant in context of the time series. Interpretation of Dickey Fuller test of
return time series shows that Test statistics equals -33,326. This does not exceed critical value of
the test at level 1%, 5% & 10%. Thus, p-value of test equals 0.000. so one can say that time
series of the return is stationary time series. Also, p-value of lag 1 in the model equals 0. It is
also lesser than all significant levels. So one can say that lag 1 has significance in time series
model of the return.

Comparing the conclusions, one can say that ARMA model fits as time series of return. ARMA
model can also be fitted to returns series. As one looks at table of information criteria for various
ARMA (p, q) specifications of the S&P500 return, one can confirm as to which specification is
best specification. In the context of rule of order selection having AIC criterion, one can say that
best order of ARMA model enables minimization of the values of information criterion AIC. It
can also be said that specification ARMA (3,3) indicates best order model. This is so as its
information criterion equals o -6807,541. It also indicates that value which minimizes AIC
information Criterion. So one can select order ARMA (3,3) of S&P500 returns.

Creating volatility model lets the making of assumption related to mean to find out whether time
series can be stationary. Also, seeing the related figure one can always make an assumption that
mean of S&P500 returns will not change with time. So one can easily make an assumption that,
it will be stationary process. On exploring the histogram of return, one can also make an
assumption that return exactly follows centered reduced normal probability. It is really a
justifiable assumption of normality of return variables. Exploration of descriptive statistics reveal
that computation on mean of range, IQ range and standard deviation respectively is 0.0004487,
0.0076794 & 0.008083. Their respective maxima indicates the values of 0.026795, 0.0337594 &
0.0484032. These calculations present quite low dispersion as it approaches minima, and these
values respectively are: -0.0418425, -0.0402114 & -00365808.

Use of Arch LM test is done for determining as to which appropriate order would be the arch
specification in volatility specification model. Normally, these are significant arch effect in
lagged 5. Exploration of chi square value indicates it as 208,784 which is significant and reveals
that there is an arch effect for lags 5. It can also be assumed that GARCH type model fits data of
S&P500 returns time series. Then, first Garch (1,1) model estimated equation can be represented
as underneath:

y t =μt +u t

Where ut =σ t . ε t & σ t ²❑=ω +α 1 u ²t −1+ β 1 σ t −1 ²

ε t i . i. d ( 0,1 )

If μt =0,0008277 & ω=4,25∗10−6 α 1=0,2136617∧β1 =0,7323723 ,

we can write as: y t =0. 0008277❑+ut

ut =σ t . ε t H ere σ t ²❑=4. 25∗10−6 +0. 2136617u 2t−1+ 0.7323723 σ t−1 ²

From 7/1/2016 - 30/12/2019 the GARCH (1,1) model gets conditional volatility. For constraints
one can write: α 1+ β 1<1 (= 0,946034 <1), in the context of stationarity.

The GARCH (1,1) process that satisfies these constraints is stable. But, when constraints do not
satisfy, one can say that process is not stable. It is explosive and so using GARCH model is not
justified.

Similarly, applying a GARCH model to the stock price time series, one can utilize the returns
instead of prices. The returns are rather more stationary but prices are not stationary. One may
also make an assumption that the sum of α+β is a parameter that measures as to how rapidly the
disruptions in variance disperse over the time. When these adds up to more than one, then these
disruptions or shocks do not vanish with the time. Lower quality of sample would have a
tendency of generating numbers and result in such situation which cannot be relied on.

As for the 2nd Garch (1,1) model is concerned, it is evaluated form the equation as mentioned
underneath:
ϑ2t =ω+ α 1 ( I ε t−1 I −γ ε t −1 ) + β 1 ϑ2t −1

Here, ω=3,59∗10−6 α 1=−0,2514419∧β1 =0,7773843

So, we have y t =μt +u t

ut =ϑ❑t . ε t

μt =0,0004799

ϑ2t =3,59∗10−6−0,2514419 ( I ε t −1 I −γ ε t−1 ) +0,7773843 ϑ2t −1

For attaining stationarity, below constraints need to be placed on coefficients as follows:

ω> 0 ; α 1 >0∧β 1> 0

In this case, constraints do not completely satisfy as α 1=−0. 2514419<0

As for the 3rd E Garch (1,1) model is concerned it is evaluated as is estimated from the below
mentioned equations as follows:

ε t=σ t z t and lnσ t ²❑= ω+ α 1 ε ²t −1+ β1 lnσ t−1 ²❑ here,

ω=−0. 5686262 α 1=0. 9419173∧β 1=0.1898031∧so

ε t=σ t z t & lnσ t ²❑= −0,5686262+0. 9419173 ε 2t −1 +0. 1898031 lnσ t−1 ²❑

Since volatility is not found so one should have good proxy related to it and so when the
conditional mean equals 0, so we will have squared returns that will provision for the unbiased
estimators of veritable underlying volatility processes. So the squared returns would be quite
noisy measures. Now suppose that r t =ln St −ln St −1 Here St represents stock price. Thus, r t
represents continuous compounded returns for underlying assets. Suppose, I t−1 is sigma field that
contains pertinent info until time t-1. So, we have E(r ²t | I t−1) = σ ² t

This value indicates that examined volatility of return gets a complete failure to accomplish the
precision over the forecast of 5 points. The research is undertaken from data sourced from
S&P500 return. Apparently not so good result is due to the reason that main parts of volatility
occurs as the new is released. It gets very difficult for simple GARCH (1,1) model to calculate
the pertinent coefficients. The enhancement of this model can be achieved by adding dummy
variables to date such as new, like new medical discovery or test result.

Test of normality in residual is comprehensively rejected. This indicates that distribution of