You are on page 1of 15

ACFI 225 Coursework Assignment Joseph Palin 2015 69849

ACFI225: Econometrics for Finance

Joseph Palin
201569849
hsjpalin@liverpool.ac.uk

REPORT ON FORD
STOCKFord
A report analysing RETURN
stock
return using regression analysis.

2,241 Words
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Table of Contents
Introduction1
Converting Stock Prices into Log returns and Excess Returns 1
Benefits of Using Log Returns 1
Treatment of Independent Variables 1
Summary Statistics of Required Variables 2
Running the Regression 3
Assumptions of Classical Linear Regression Model 4
Detecting Heteroskedasticity Graphically 4
Checking for heteroskedasticity using White’s test 5
Re-estimating Regression Coefficients with Adjusted Standard Errors 5
Interpretation of Durbin-Watson Statistic 6
Use of Breusch-Godfrey Test 6
Using Newey-West Procedure on Regression 6
Distribution of Log Return of FORD 7
Analysis to Determine the Presence of Multicollinearity 7
Conclusion 8
Reference List 9
Appendices 10
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Introduction

This report seeks to interpret the results of a linear regression conducted on the stock returns of
FORD as influenced by the Industrial Production Index (INDPRO), the measure of narrow money
supply (M1) i.e. currency, demand deposits, and other liquid deposits (Investopedia, 2021), as well
as the Consumer Price Index (CPI).

Converting Stock Prices into Log returns and Excess Returns

In order to use US treasury bill data as the risk-free rate in our CAPM model from which to calculate
excess returns it needed to be converted into monthly returns as so:

us3tbill = us3tbill/12

(See appendix 1 for converted data)

Then, in order to use the stock price data for FORD in the linear regression, it needed to be
converted into log returns (creating a new variable). This was achieved in EViews using the following
formula:

rford=100*log(ford/ford(-1))

This process was repeated for the S&P 500 series. Which represented market return in the model.

The rford variable was then converted into excess return in the following way:

erford= rford-us3tbill

Benefits of Using Log Returns

The benefit of using log return is that the returns are symmetric i.e. log returns of equal magnitude
and opposite signs will cancel each other out. For example, an investment of £100 that yields a
logarithmic return of 50% followed by -50% will return to a value of £100. The same situation with
simple returns results in a value of £75 (Lucas, 2021).

Taking log returns can help to overcome heteroskedasticity, can help to make the data be distributed
in a way which is closer to a normal distribution, as well as making multiplicative relationships
between variables into linear additive ones (Brooks, 2019).

Treatment of Independent Variables

My three independent variables were then to be converted into three series of first differences. I
named these my change series. They have been calculated using the following formulae:

m1change= m1-m1(-1)

cpichange= cpi-cpi(-1)

Page | 1
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

indprochange= indpro-indpro(-1)

Summary Statistics of Required Variables

Figure 1 Summary Statistics Table

Fig. 1. shows the summary statistics for the 5


variables. The most notable information that Erford Histogram
has been gleaned, is the negative mean of
excess returns of Ford compared with the
excess returns of the S&P 500 which is
positive.

The excess returns of Ford are very slightly


positively skewed meaning there are more
extreme positive returns in the data set than
negative.

The kurtosis is very high. Much higher than


that of a normal distribution at 17.84. This is
Figure 2 ERFORD HISTOGRAM
shown in fig. 2 by the sharp peak in the
middle and long thin tails caused by the large
outliers in the data set.

Figure 3 Table Showing Correlation of Variables with ERFORD

As shown fig. 3 the strongest correlation between variables is that which exists between ERFORD
and ERS&P500 at 0.579. None of the other correlations really bare mentioning, other than perhaps
that ERFORD is slightly negatively correlated with the Industrial Production Index.

Plotting the ERFORD and ERS&P500 (fig 4.) it is


possible to see how much more volatile FORD
returns are. It is also possible to see that the
S&P500 and Ford seem to move together.
Which is what one would expect to see based
on the correlation statistics (fig.3).

Page | 2
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Figure 4 Plotthe
Running showing ERFORD and ERS&P500
regression:

The regression was completed in eviews using the formula:

erford c ersandp500 indprochange m1change cpichange

Figure 5 Regression Results

This formula yielded the results in Fig. 5. To check whether these coefficients are statistically
significant it is necessary to do a hypothesis test using the t-statistic generated and compare it with
the critical T value. Due to the degrees of freedom falling between 101-1000, the correct critical
value to use is that of 1000 degrees of freedom (see appendix 2). This value is 1.962.
5%
Coefficient t-Statistic </> Critical Value
ERS&P500 1.931 10.08 > 1.962
CPICHANGE -0.808 -0.842 < 1.962
INDPROCHANGE -2.990 -2.575 > 1.962
M1CHANGE -0.005 -0.213 < 1.962
C -0.466 -0.496 < 1.962
As shown in Figure
Fig. 66the
Table showing coefficient
absolute values ofvalues and t-stats. compared
the T-statistics for onlywith
twocritical T-value.
variables are greater than the
critical value at the 5% significance level. These are ERS&P500 and INDPROCHANGE. Therefore, the
null hypothesis that the actual coefficient values for these variables are zero must be rejected.

Looking at the P-Values for CPICHANGE and M1CHANGE, the level of significance at which the null
hypotheses must be rejected are 40% and 83% respectively. This size of test would clearly be open
to Type I error (false positive).

Page | 3
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Looking at the F-Statistic (26.53) for the regression Fig. 5 it is possible to reject the null hypothesis
that the Betas of the independent variables are jointly equal to 0. This is because the F-statistic of
the regression is higher than the critical F-value for a set with 4 independent variables, and 192
degrees of freedom which is 1.94486.

Taking the Beta of ERSAND500, and INDPRO out of the equation it is possible to test whether Betas
of the other 2 independent variables are jointly different from 0 using a restricted test. The result of
the Wald test showed that coefficients for CPICHANGE and M1CHANGE were not jointly different
from 0 (See appendix 2) for results.

The standard error for the regression shown in fig. 5, shows the average distance that the observed
values fall from the regression line. It is measured in the units of the dependent variable. So, in this
case the standard error is in the units of ERFORD which is percentage return.

R2 of 36.1% shows that the independent variables do not jointly explain the dependent variables
very well. The majority of variation in ERFORD (63.9%) is not explained by the independent variables
at all. To improve the relevance of the model, other independent variables must be included.

The standard errors shown in Fig. 5 show that coefficient value for INDRPOCHANGE is very likely to
be different across a different sample. It is therefore unlikely to be representative of the true value.
ERSANDP500 has a relatively low standard error and so the coefficient can be expected to be the
same over many different samples.

Assumptions of the Classical Linear Regression Model

The first assumption according to Brooks (Brooks, 2019) is that the average value of errors is zero. In
such a situation where this is not the case, it is possible to achieve a negative R 2. By including a
constant term α this will never be the case.

The second assumption of the classical linear regression model is the assumption of
homoskedasticity. That is to say that the variance of the errors is assumed to be constant. Where
this is not the case the issue of heteroskedasticity will occur. More analysis of the data for the
presence of heteroskedasticity follows below.

Assumption three of the CLRM is that the covariance of error terms over time is zero. If this is not
the case the errors are said to be ‘autocorrelated’. Essentially, if this is the case it means that there
are important variables that have been omitted from the model but are persistent through time and
are still affecting the dependent variable (Lambert, 2013).

The fourth assumption of the classical linear regression model is that the regressors are non-
stochastic. That is to say that it cannot be predicted even under a particular set of circumstances.

The fifth assumption of CLRM is that the disturbance terms are normally distributed. It is not obvious
what should be done in a case where the error terms are not normally distributed. According to
Brooks For sufficiently large sample sizes, violation of the normality assumption is inconsequential
(Brooks, 2019).

Detecting Heteroskedasticity Graphically

In order to maintain the second assumption of the CLRM, it is necessary to look for
heteroskedasticity in the residuals. It is possible to do this graphically. According to C. Robero, after
plotting the residuals, if there is an evident pattern in the plot then heteroskedasticity is present

Page | 4
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

(Robero, 2016). Looking at the plot (see appendix 6) There is no obvious pattern in the data and
based on this alone heteroskedasticity is
unlikely.

Checking for heteroskedasticity using


White’s test.

However, looking for heteroskedasticity


graphically is not enough to prove its
presence one way or the other. One test
which can be carried out to formally check
for heteroskedasticity is White’s test. This
has been carried out in Eviews and the results can be seen in Fig. 8.

Looking at the F-statistic for this auxiliary regression and the accompanying Prob. F, it is clear that
the null hypothesis of homoskedasticity Figure 8 Results of White Test
should be rejected as heteroskedasticity is
very likely present. The implication of this is that the standard errors in the initial regression are not
reliable and need to be adjusted for the model to have any utility.

Re-estimating Regression Coefficients with Adjusted Standard Errors

Fig. 9 Shows the re-estimated regression results using adjusted standard errors. Naturally the
coefficient values are the same as in the original regression equation.

The T-statistics are different however and hypothesis testing must once again be carried out.

Fig. 10 shows the effects of using heteroskedasticity-robust standard errors. When comparing the
absolute value of the adjusted T-stats with the critical value of 1.962 (as in the original analysis) only
one change occurs. ERSANDP500 is still statistically significant. However, INDPROCHANGE is now
insignificant as it is below the critical T-value of 1.962 (at 1.56). Using adjusted standard errors, it is
possible to say that in order to make predictions about FORD stock price one only needs to know its
Beta with the market.

Unajusted T-stat Adjusted T-stat


ERS&P500 10.08 4.81
CPICHANGE -0.84 -0.52
INDPROCHANGE -2.58 -1.56
M1CHANGE -0.21 -0.25
C -0.50 -0.39

Unajusted SE Adjusted SE
ERS&P500 0.19 0.40
CPICHANGE 0.96 1.56
INDPROCHANGE 1.16 1.91
M1CHANGE 0.02 0.02
C 0.94 1.20

Fig. 10 Effect of Using heteroskedasticity-


Fig. 9 Regression Results with Adjusted Standard robust standard errors.
Errors Interpretation of Durbin-Watson Statistic

Page | 5
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Fig. 11 Excerpt from Chris Brooks Book (Brooks, 2019)

The Durbin Watson Statistic from the regression is 2.507 (Fig. 9). Using Durbin-Watson Statistic
Tables (See appendix 4) the upper and lower critical values highlighted for 4 regressors with 193
observations are 1.68 and 1.79.

0 1.68 1.79 2.21 2.32 4


Fig. 12 Updated Durbin Watson Table

Fig. 12 Shows the Chris Brook’s chart with critical values applied (see appendix 4). Clearly the Durbin
Watson Statistic falls in the negative autocorrelation range based on my calculations and
understanding. This means that there is first order autocorrelation present in the data.

Use of Breusch-Godfrey Test

This test analyses whether or not there is auto correlation up to the 10 th lag of the residuals.

Fig. 13 Breusch-Godfrey Test Result

Simply by looking at the Prob. F statistic, it is clear to see that the null hypothesis can be rejected
even up to the 1% significance level. This means that there is auto correlation up to the 10 th lag of
the residuals. This test agrees with the results of the Durbin Watson test carried out above.

Using Newey-West Procedure on Regression

Running Newey-West procedure on regression to adjust for autocorrelation and heteroskedasticity


(See appendix 5).

Page | 6
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Huber White Newey-West Fig 14. Shows the affect of adjusting for
Adjusted T-stat Adjusted T-stat both heteroskedasticity and autocorrelation
ERS&P500 4.81 7.26 in the regression analysis. Still using a
critical T-Value of 1.962 (as above) it is
CPICHANGE -0.52 -0.99
possible to reject the null hypothesis that
INDPROCHANGE -1.56 -1.98
the true value of the coefficient for INDPRO
M1CHANGE -0.25 -0.26
is = 0. Meaning it is statistically significant.
C -0.39 -0.64
R2 is unchanged between both procedures
and the T-value for ERSANDP500 is
Huber-White Newey-West
statistically significant in both.
Adjusted SE Adjusted SE
ERS&P500 0.40 0.27 Standard errors are reduced for almost all
CPICHANGE 1.56 0.82 regressors with the Newey-West
INDPROCHANGE 1.91 1.51 adjustment for autocorrelation. This
M1CHANGE 0.02 0.02 confirms the above assertions that
C 1.20 0.73 autocorrelation is present in the data.
Fig. 14 Comparing Std. E & T-Stat for difference procedures

Distribution of Log Return of FORD

Ford returns are not normally distributed. If the Jarque-Bera statistic were close to
0 it would indicate that the returns are normally distributed. The JB statistic in this
case is very far from zero and the null hypothesis of normality is strongly rejected
at every level (Fig. 15).

However, ordinary least squares method of linear regression is not dependent on


normally distributed data, so this is not a cause for concern.

Fig. 15 Summary Stats of


Analysis to Determine the Presence of Multicollinearity
Ford Log Return

By laying out the correlations between variables in a table such as Fig. 16 it is possible to assess
whether or not multicollinearity exists between variables. Based on the below table it seems fair to
suggest that no multicollinearity exists between variables as correlations are low on the whole.

It is also possible to test for


multicollinearity using variance inflation
factors. When it comes to interpreting
variance inflation factors, factors of
more than 4 or 5 are considered Fig. 16 Table showing correlations of
moderately high (Bock, 2018). independent variables
If multicollinearity were present in the
model, it would be very difficult to
determine which of the independent

Page | 7
Fig. 17 Table showing variance inflation factors.
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

variables is causing any change in the independent variable as they both act on it in a similar way at
the same time.

Analysing the centred VIF Fig. 17, It is very unlikely that multicollinearity among the independent
variables is present in the model.

Conclusion

To conclude, the results of this linear regression show that only one of the independent variables
explained the returns of FORD exhaustively and this variable is the market Beta. However, it is
possible that INDPRO change explains some of variation in ERFORD however it is likely to be a very
small amount indeed.

In order to improve the model, more relevant independent variables should be included. I do not
believe that all of the independent variables are BLUE as several fall at the first hurdle of being the
BEST explanatory variables and better explanatory variables surely exist.

Reference List:

Investopedia. (2021) M1 Money Supply: How It Works and How to Calculate It. Available at
https://www.investopedia.com/terms/m/m1.asp#:~:text=M1%20is%20a%20narrow
%20measure,financial%20assets%2C%20such%20as%20bonds. (Accessed 15/12/2022)

Page | 8
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Lucas, L. (2021) Why Use Logarithmic Returns In Time Series Modelling. Available at
https://lucaslouca.com/Why-Use-Logarithmic-Returns-In-Time-Series-Modelling/
#:~:text=Logarithmic%20returns%20are%20useful%20for,will%20cancel%20each%20other%20out.
(Accessed 15/12/2022).

Brooks, C. (2019) Introductory Econometrics for Finance. 4th edn. Cambridge: Cambridge University
Press.

Lambert, B. (2013) Autocorrealation and introduction (2013) YouTube. YouTube. Available at:
https://www.youtube.com/watch?v=jt5nl2VEpwg (Accessed: December 16, 2022).

Robero, C. (2016) Methods of Detecting and Resolving Heteroskedasticity. Available at


https://rstudio-pubs-static.s3.amazonaws.com/187387_3ca34c107405427db0e0f01252b3fbdb.html
#:~:text=Residual%20Plots,plot%2C%20then%20heteroskedasticity%20is%20present. (Accessed
17/12/2022).

Statology. (2019) Durbin-Watson Table. Available at https://www.statology.org/durbin-watson-


table/ (Accessed 17/12/2022)

Ttable.org. (2019) T Table. Available at T Table - T Score Table - T Critical Value (Accessed
17/12/2022)

Bock, T. (2019) What are Variance Inflation Factors (VIFs)?. Available at


https://www.displayr.com/variance-inflation-factors-vifs/ (Accessed 17/12/2022).

Appendices
Appendix 1: Data before and after dividing US3TBILL by 12

Page | 9
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Figure 2 US3TBILL After

Appendix 2. Results of Wald Test on Variables CPICHANGE & M1CHANGE

Figure 2 Wald Test Result

Page | 10
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Appendix 3. T-table

Figure 3 T-tables available at www.ttable.org

Page | 11
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Appendix 4 Durbin Watson Table

Figure 4 Durbin Watson Table with Highlighted Critical Values (1.68) and (1.79) (Statology, 2019)

Page | 12
ACFI 225 Coursework Assignment Joseph Palin 2015 69849

Appendix 5 Newey-West Adjusted Regression

Appendix 6 Testing for heteroskedasticity graphically

Page | 13

You might also like