LIST OF MEMBERS

Student Name Student ID Contribution

Huỳnh Mỹ Hạnh 31231022116 100%

Đoàn Mỹ Huyền 31231027879 100%

Trần Nguyễn Mai Linh 31231021888 100%

Nguyễn Hoàng Minh Thư 31231023060 100%

Session 1:
2.7.
1. Introduction
As part of our group discussion for the Financial Econometrics course, we evaluated three research
ideas against the criteria outlined in Box 2.1 (Saunders et al., Research Methods for Business
Students). The checklist assesses whether a research topic meets three key dimensions:
Appropriateness, Capability, and Fulfilment. Each idea was discussed collectively before reaching a
group conclusion.

The three research ideas assessed are:

- Idea 1: The impact of leverage (debt-to-equity ratio) on firm profitability of listed companies
in Vietnam - a panel data approach.
- Idea 2: Testing the Efficient Market Hypothesis (EMH) in the Vietnamese stock market using
GARCH models and event study methodology.
- Idea 3: The effect of investor sentiment on stock return volatility - a quantile regression
analysis of the VN-Index.
2. Research Idea 1: Leverage and Firm Profitability
"The impact of leverage (debt-to-equity ratio) on firm profitability of listed companies in Vietnam: A
panel data regression approach"

Criterion Evaluation

Appropriateness This topic has a clear theoretical foundation grounded in established

corporate finance theories (Trade-off Theory and Pecking Order
Theory). The research question - how does leverage affect ROA and
ROE among Vietnamese-listed firms? is specific, measurable, and
directly linked to theory. The topic meets institutional standards for a
financial econometrics project and can provide fresh insights if recent
post-COVID data (2020-2023) is used. Regardless of whether the
relationship is positive or negative, the findings will be of equal
academic value, satisfying the symmetry requirement.

Capability The required econometric techniques, panel data regression with

Fixed Effects and Random Effects models, and the Hausman
specification test, are well within the scope of our course. Data can be
sourced from publicly available platforms such as Cafef, Vietstock, or
FiinPro. The scope is manageable within the project timeline of 3-4
months.

Fulfilment The topic is of genuine interest to group members with career

aspirations in corporate finance, investment banking, and financial
analysis. It provides practical skills in panel data modelling that are
directly applicable to professional roles in the Vietnamese financial
sector.

Conclusion Good research topic - recommended as the primary option.

3. Research Idea 2: Testing the Efficient Market Hypothesis

"Testing the Efficient Market Hypothesis (EMH) in the Vietnamese stock market using GARCH
models and event study methodology"

Criterion Evaluation

Appropriateness The EMH is one of the most widely studied theories in financial
economics. The proposed topic is well-grounded in theory and can
generate novel contributions by focusing on the Vietnamese market,
an emerging market with unique structural characteristics. Using data
from the post-COVID period or comparing pre- and post-pandemic
market efficiency would add a fresh dimension. The research question
is clearly stated, and the findings (whether the market is efficient or
not) will be equally informative.

Capability This topic requires more advanced econometric techniques, including

GARCH/EGARCH modelling, ARIMA for time-series analysis, unit
root tests (ADF, PP), and event study methodology. These are
technically demanding but achievable if at least one group member
has prior experience with time-series econometrics. Stock price data is
readily available from Cafef and Yahoo Finance. The main risk is the
time required to implement and interpret the models correctly.

Fulfilment This topic is highly motivating for group members interested in

investment analysis, quantitative finance, and risk management. It
develops skills in time-series econometrics that are highly valued in
asset management and trading roles. The challenge of the topic adds
to its appeal.

Conclusion Good research topic - viable if the group has strong time-series
econometrics skills.

4. Research Idea 3: Investor Sentiment and Stock Return Volatility

"The effect of investor sentiment on stock return volatility: A quantile regression analysis of VN-
Index"

Criterion Evaluation

Appropriateness This topic sits at the intersection of behavioural finance and financial
econometrics, an increasingly prominent research area. The idea of
using quantile regression to analyse how investor sentiment affects
volatility across different market conditions is original and
theoretically sound, drawing on behavioural finance theory (e.g., De
Long et al., 1990; Baker and Wurgler, 2006). It is capable of
producing genuinely novel findings for the Vietnamese market
context.

Capability The main challenge lies in data availability. Constructing a reliable

investor sentiment proxy requires either access to proprietary survey
data, indirect proxies such as Google Trends or the Consumer
Confidence Index, or text-based sentiment scores from financial
news. These are not straightforward to obtain. Quantile regression
itself is feasible given our training, but the data collection process
could significantly constrain the timeline. This is the primary risk
factor for this topic.

Fulfilment The topic is intellectually stimulating and stands out from more
traditional financial econometrics topics. Group members interested
in behavioural finance and fund management find this topic highly
appealing. However, the practical difficulties with data sourcing
somewhat reduce its feasibility within the project constraints.

Conclusion Potential research topic - only advisable if data access can be

confirmed in advance.

5. Comparative Summary
The table below summarises the group's assessment of all three ideas across the three Box 2.1
dimensions:

Research Idea Appropriaten Capability Fulfilment Overall

ess

Idea 1: Leverage & High High High Recommended

Profitability

Idea 2: EMH Testing High Medium High Recommended (if

(GARCH) skilled in time-
series)

Idea 3: Investor High Low-Medium High Consider carefully

Sentiment (data access
required)

6. Group Discussion and Conclusion

During our group discussion, we considered each idea in terms of data accessibility, technical
requirements, and alignment with our collective skills and interests. The following key points were
raised:
- Idea 1 was unanimously regarded as the most feasible option. The availability of financial
statement data from Vietnamese-listed companies and the manageable complexity of panel
data regression make it the most suitable choice within our timeline.
- Idea 2 was considered a strong candidate by group members with experience in time-series
modelling. We concluded that it is viable provided the group commits sufficient time to
mastering GARCH estimation and event study design.
- Idea 3, while intellectually appealing, poses significant challenges in terms of obtaining a
valid and reliable sentiment measure for the Vietnamese market. We agreed it would only be
pursued if a credible data source could be secured in advance.
Based on the Box 2.1 checklist and our group discussion, Idea 1 is our primary recommendation as a
good research topic. Idea 2 is a strong alternative for a group with the appropriate technical
background. Idea 3 remains a possibility but requires further investigation into data availability before
committing.

2.8.
In the field of financial economics, an article titled “EFFECT OF SELECT MACROECONOMIC
VARIABLES ON FINANCIAL PERFORMANCE OF TRADERS IN EASTLEIGH MARKET NAIROBI
COUNTY, KENYA” was selected as it is based on primary data collected through questionnaires
administered to traders. The study employs quantitative techniques such as regression analysis and
ANOVA to examine the relationships between variables.
Assuming that the research question and objectives are not explicitly stated, they can be inferred from
the variables examined and the empirical analysis presented. The study focuses on financial
performance as the dependent variable, while exchange rate, inflation, and gross domestic product
(GDP) are treated as the main independent variables. The analysis reports regression results and
ANOVA tests for each variable, indicating that the study evaluates the individual effects of these
macroeconomic factors.
Based on this, the underlying research question can be inferred as: To what extent do exchange rate,
inflation, and GDP affect the financial performance of traders in Eastleigh Market, Nairobi
County, Kenya?
Similarly, it can be inferred that the study aims to:
- Assess the effect of exchange rate on traders’ financial performance
- Examine the impact of inflation on financial outcomes
- Evaluate the relationship between GDP and financial performance
These inferred objectives are consistent with the analytical approach used, where each independent
variable is tested against the dependent variable to determine its statistical significance and direction
of influence.
In conclusion, even in the absence of explicitly stated research questions and objectives, they can be
logically derived from the study’s variables and empirical structure.
Reference: [Link]
%[Link]?sequence=1&isAllowed=y

Session 3:

End-of-chapter exercises from C.3 Chris Brooks

1.

a. Why does OLS estimation involve taking vertical deviations of the points to the line rather
than horizontal distances?

b. Why are the vertical distances squared before being added together?

c. Why are the squares of the vertical distances taken rather than the absolute values?

Answer

a. Why does OLS estimation involve taking vertical deviations of the points to the line
rather than horizontal distances?

The reason OLS (Ordinary Least Squares) estimation focuses on vertical deviations of the points to
the line instead of horizontal distances is because the objective of OLS is to minimize the sum of
squared vertical deviations, which are also referred to as residuals or errors.

In the process of OLS, our goal is to identify the line that provides the closest fit to the given data
points. This line is determined by the equation:

y=β0+β1x
where y represents the dependent variable, x represents the independent variable, β0 represents the
intercept, and β1 represents the slope. The main objective of OLS estimation is to determine the
values of β0 and β1 that minimize the sum of squared residuals.

The residuals represent the vertical distance between observed data points and the fitted line. By
minimizing the sum of squared residuals, the line with the smallest overall vertical deviations is
found. This approach is useful in economic and statistical applications, as it minimizes the overall
discrepancy between observed and predicted values.

b. Why are the vertical distances squared before being added together?

Due to the possibility of some points lying above or below a line, the sum of squared distances is
minimized rather than the sum nearest to zero. As long as the total of the distances of the points above
and below the line is equal, this guarantees that the distances cancel out and practically any line may
be fitted to the data. This removes the requirement for a special solution for the estimated coefficients
because the total would always equal zero for any fitted line passing through the observation mean.

c. Why are the squares of the vertical distances taken rather than the absolute values?

Squaring residuals simplifies differentiation, derivative computations, and estimating by allowing for
the derivation of analytical formulas for statistical quantities like variance and mean.

However, the absolute value function's non-differentiability at zero complicates analytical and
optimization solutions, making estimating more difficult. While the problem of positive and negative
residuals canceling out can be resolved by using their sum, the absolute value function is more
challenging to optimize compared with the square. Analytical formulas for mean and variance are
easier to obtain when squared residuals are used.

2. Explain, with the use of equations, the difference between the sample regression function and
the population regression function.

Answer

The connection between the variables is represented by the population regression function (PRF),
which is a description of the model assumed to be producing the actual data. The data generating
process (DGP) is another name for the population regression function.
Based on a sample of data, the SRF shows the estimated connection between the independent variable
(X) and the dependent variable (Y). It is computed with sample data and is employed to draw
conclusions on the population relationship.

The equation for the SRF can be expressed as:

The fitted value from the model and a residual term are the two halves of the observed value of y that
are separated by this equation. The PRF's likely values are deduced using the SRF. In other words,
estimates A and B are created using the sample data.

The bivariate model's PRF, which represents the actual values of and , may be written as alpha and
beta

Taking into account all potential observations, the PRF characterizes the ideal or underlying
relationship between Y and X in the population. The population parameters that establish the link are
represented by the actual coefficients.

3. What is an estimator? Is the OLS estimator superior to all other estimators? Why or why
not?

An estimator is a statistical method or technique that is used to obtain available data in order to
estimate an unknown parameter or value of interest. It is a mathematical function that gives the
parameter under estimate a value or a range of values.

To estimate the unknown parameters in a linear model, the Ordinary Least Squares (OLS) estimator is
frequently used in linear regression. Under some circumstances, OLS can have favorable qualities like
unbiasedness and efficiency, but it is not always better than alternative estimators. The particulars of
the data and the underlying statistical model influence the estimator selection. In some circumstances,
other estimators, like robust or maximum likelihood estimators, might be more appropriate. This is
particularly true when working with non-linear models or when the assumptions of OLS are broken.
As a result, the context and particular needs of the statistical study determine whether an estimator is
superior.

4. What five assumptions are usually made about the unobservable error terms in the classical
linear regression model (CLRM)? Briefly explain the meaning of each. Why are these
assumptions made?

Answer

Technical Notation Interpretation

(1) E(ut) = 0 The errors have zero mean

(2) var(ut) = σ^2 < ∞ The variance of the errors is constant and finite
over all values of x

(3) cov(ui, uj) = 0 The errors are linearly independent of one

another

(4) cov(ut, xt) = 0 There is no relationship between the error and

corresponding x variate

(5) u(t) ~ N(0, σ^2) i.e., that ut is normally distributed.

As long as assumption (1) holds, assumption (4) can be equivalently written E(xt ut) = 0. Both
formulations imply that the regressor is orthogonal to (i.e., unrelated to) the error term. An alternative
assumption to (4), which is slightly stronger, is that the x(t) are non-stochastic or fixed in repeated
samples. This means that there is no sampling variation in x(t), and that its value is determined
outside the model.

A fifth assumption is required to make valid inferences about the population parameters from the
sample parameters estimated using a finite amount of data, namely that the disturbances follow a
normal distribution.

If assumptions (1)–(4) hold, then the estimators alpha hat and betahat determined by OLS will have a
number of desirable properties, and are known as best linear unbiased estimators (BLUE) with
minimum variances among CLRM. Under assumptions (1)–(4) listed above, the OLS estimator can be
shown to have the desirable properties that it is consistent, unbiased and efficient. This is proved by
the Gauss-Markov theorem, so in the case of violating one of these four assumptions, it would affect
the accuracy and cause fluctuations between samples.

To test coefficient hypothesis, or draw statistical conclusions about population parameters from
sample data, we have to assume the fifth thing: the disturbances are regularly distributed. Under the
premise that the other assumptions remain true, test statistics are expected to follow a t-distribution.

5. Which of the following models can be estimated (following a suitable rearrangement if

necessary) using ordinary least squares (OLS), where X, y, Z are variables and α, β, γ are
parameters to be estimated? (Hint: the models need to be linear in the parameters.)

yt = α + βxt + ut (3.39)

yt = e αxβ t e ut (3.40)

yt = α + βγ xt + ut (3.41)

ln(yt) = α + β ln(xt) + ut (3.42)

yt = α + βxt zt + ut (3.43)
 Answer

The 3.39 is the basic and simple model of CLRM so we can use it - the OLS

The 3.40 can be linearised by taking logarithms of both sides and by rearranging to form a parameter
as a linear estimator because the parameters now are in the exponential position.

The 3.41 can be used to be estimated by the OLS; however, the value of the two coefficients
multiplied together will be obtained rather than the values of each since we can see that the formulae
is presented nearly similar to the simple linear regression model.

The 3.42 is accepted for OLS since this is another form of linear but in logarit term. Models of this
type, which are linear in the logarithms, have the intriguing characteristic that the coefficients alpha
and beta can be understood as elasticities.

The 3.43 is a quite dissimilar formulae when compared to the original one. However, when we denote
x(t) multiple with z(t) into e(t) for example, we can transform the given formula to the well-known
one that we have learned. As shown below here is the substitute formula and you can solve for the
value of x(t) and z(t) in the end:

6. The capital asset pricing model (CAPM) can be written as

E(Ri) = Rf + βi[E(Rm) − Rf ] (3.44)

using the standard notation.

The first step in using the CAPM is to estimate the stock’s beta using the market model. The
market model can be written as

Rit = αi + βi Rmt + uit (3.45)

where Rit is the excess return for security i at time t, Rmt is the excess return on a proxy for the
market portfolio at time t, and ut is an iid random disturbance term. The cofficient beta in this
case is also the CAPM beta for security i .

Suppose that you had estimated (3.45) and found that the estimated value of beta for a stock, βˆ
was 1.147. The standard error associated with this coefficient SE(βˆ) is estimated to be 0.0548.

A city analyst has told you that this security closely follows the market, but that it is no more
risky, on average, than the market. This can be tested by the null hypotheses that the value of
beta is one. The model is estimated over sixty-two daily observations. Test this hypothesis
against a one-sided alternative that the security is more risky than the market, at the 5% level.
Write down the null and alternative hypothesis. What do you conclude? Are the analyst’s
claims empirically verified?

Answer
In contrast to the one-sided alternative that suggests beta is bigger than one, the null hypothesis states
that the true (but unknown) value of beta is equal to one:

H0 : β = 1

H1 : β > 1

The test statistic is provided by:

^ β∗¿ 1.147−1
Test stat ¿ β❑ − = =2.682 ¿
SE ( ^
β 0) 0.0548

T is the sample size, and in this case, T-2 = 60. We wish to compare this value with a value from the
t-table with T-2 degrees of freedom. Given that the test is one-sided, we want a number with 5% in
each tail. The t-table's crucial t-value is 1.671.

We can reject the null hypothesis because the test statistic's value is within the rejection zone. This
security is substantially riskier than the market as a whole, as evidenced by statistically strong
evidence that its beta value is larger than one.

9. Are hypotheses tested concerning the actual values of the coefficients (i.e. β) or their
estimated values (i.e. βˆ) and why?

Answer

Y = Bo + B1x + t

H0: B1 = C

Ha: B1 different with C

t = (B1^ - B1)/ (SE (B1^) with tn-2

For testing we are using test estimated value B1^ but we are concerned with the actual value of
parameters B1. Thus, tested hypotheses concern the actual value of coefficients B1 not their estimated
values, because the goal is to make inferences about the true population parameters values.

End-of-chapter 2 from Wooldridge's 7th edition

4.

Answer

( i ) bwght = 119.77 - 0.154 * cigs

● When cigs = 0 ⇔ bwght = 119.77 - 0.514 * 0 = 119.77 ( oz )

● When cigs = 20 ⇔ bwght = 119.77 - 0.514 * 20 = 109,49 ( oz )
● The difference between bwght when cigs = 0 vs cigs = 20 is

119.77-109.49 = 10.28 (oz)

=> The effect of cigarettes on infant birth weight is negative. On average, one more cigarette reduces
the infant birth weight for 0,514 oz.

( ii ) No, this simple regression does not necessarily establish a causal relationship between a child’s
birth weight and the mother’s smoking habits. This is because various other factors, such as the
mother’s overall health and the quality of prenatal care, can influence birth weight and may also be
linked to smoking during pregnancy. Additionally, factors like caffeine consumption, which can
impact birth weight, might also be associated with cigarette smoking.

( iii ) To predict a birth weight of 125 oz, cigs would be :

bwght = 119.77 - 0.514 * cigs

⇔ 125 = 119.77 - 0.514 * cigs

⇔ cigs = ( 125 - 119.77 ) / 0.514

⇔ cigs = ~ -10.18 or ~ -10 cigs

=> It’s impossible, it happens when we use the linear regression model but the true relationship is
non-linear. Moreover, when we look at the simple regression model and the fitted equation, we’ll find
that the highest birth weight which we can estimate from the equation is 119.77 oz when the cigs = 0.
However, nearly 700 births in the sample had a birth weight exceeding 119.77.
( iv ) Out of 1,388 women, 1,176, or approximately 84.7%, did not smoke during pregnancy.

5.

Answer

(i) The intercept estimates the amount of consumption when income is zero. There is a
negative sign on it representing that if family income is zero, the estimated consumption would be
negative.

(ii) inc = $30,000 we have:

Therefore, the predicted consumption amounts to $25,465.16 when the family income is $30,000.

(iii) With income on the x-axis and consumption on the y-axis.

The slope – Marginal propensity to consume (MPC)

Calculate Average Propensity to Consume (APC)

The graph begins at an annual income level of $1,000.

Then we will have the following graph:

 ● MPC is constant because it’s defined as the change in consumption per unit of income,
independent of the level of income.
● APC varies with income, the APC curve is asymptotic, meaning t gets closer to 0.853 but
never reaches it, no matter how large income becomes

6.

Answer

(i) Coefficient 0.312 of log(dist):

● The price of houses depends on the distance between the house and the incinerator.
● The coefficient 0.312 indicates that, on average, a 1% increase in the distance from the
incinerator is associated with a 0.312% increase in housing price (assuming everything else
remains constant).

The sign of estimates:

● The positive sign in the regression equation above suggests that houses farther from the
incinerator tend to be more expensive, which aligns with expectations since proximity to an
incinerator is typically considered undesirable.

(ii) Simple regression may not provide an unbiased estimate of the true elasticity. If the incinerator's
location is selected in areas farther from higher-priced neighborhoods, then log(dist) may be
positively associated with housing quality. This relationship could introduce bias into the estimate.

(iii) Several other factors likely influence housing prices, including:

● Size and condition of the house.

● Neighborhood characteristics (e.g., school quality, crime rate).
● Accessibility to amenities (e.g., public transport, parks, shopping centers).
● Environmental factors (e.g., air quality, noise levels).

These factors may have an association with distance from the incinerator. For instance, if lower-
income areas are more likely to host incinerators, this could confound the relationship between
distance and housing prices. Including these variables in a regression model would help control for
such confounding effects.
Session 4:
Chris Brooks’s Books Chapter 4

1. It can be demonstrated that the t-distribution is a specific case of the broader F-distribution.
Specifically, the square of a t-distribution with T-k degrees of freedom is equivalent to an F-
distribution with (1,T-k) degrees of freedom. However, it is important to note that for a test conducted
at the 5% significance level, the critical value for the F-distribution. For example when conducting at
5% level from the table:

T-k F critical value T critical value

10 4.96 2.29

20 4.35 2.09

40 4.08 2.02

120 3.92 1.98

∞ 3.84 1.96

2.

(a) H0 : β 3 = 2

We could use both the F-test and the T-test. Because it’s a single hypothesis and involves one
coefficient, we could use the T-test for simpler.

(b) H0 : β 3 + β 4 = 1

It involves more than one coefficient and 1 restriction, so use the F-test.

(c) H0 : β3 + β4 = 1 and β5 = 1

There are more than one hypothesis and 2 restrictions, so use the F-test.

(d) H0 : β2 =0 and β3 = 0 and β4 = 0 and β5 = 0

There are more than one hypothesis and 4 restrictions, so use the F-test.

(e) H0 : β2 β3= 1

This is a multiplicative restriction, so we cannot use both the F-test or the T-test.

3.

In the context of equation: y t =β 1+ β 2 x 2t + β3 x 3 t + β 4 x 4 t + β 5 x5 t +ut ,

null hypothesis (d) H 0 : β 2=0 and β 3=0 and β 4 =0 and β 5=0 constitutes ‘THE’ regression F-
statistic
This null hypothesis is always of interest in regression analysis because it tests the joint significance
of all explanatory variables (other than the intercept) in the model. This helps answer the fundamental
question: Do the independent variables in the model matter? If holds, then all the independent
variables do not contribute to explaining variations in , then we have a completely useless regression

4.

- The restricted residual sum of squares (RRSS) is always larger than or equal to the unrestricted
residual sum of squares (URSS).

- This happens because, in conducting regression, the model chooses parameter values that make the
RSS as small as possible. If we add restrictions to the model, limiting how the parameters can be
chosen, the model cannot work as effectively as before.

- As a result, the RRSS should be higher. If it isn’t, it would mean the OLS didn’t actually choose the
best parameter values originally, which goes against how the OLS method is supposed to work.

Wooldridge’s Book Chapter 3:

3.2

(i) Due to budget limitations, it's reasonable to expect that the more siblings a family has, the less
education each individual child receives. So we got an increase in the number of siblings resulting in a
reduction of one year in the predicted level of education:

1
1=0.094 × Δsibs → Δsibs= =10.6
0.094

(ii) Keeping the number of siblings and father's education constant, each additional year of a mother's
education corresponds to a 0.131-year increase in the predicted education of her child. Thus, if a
mother has four additional years of education, her son is expected to have approximately 0.524 more
years of education.

(iii) Since the number of siblings remains constant, but both mother’s and father’s education levels
differ, the coefficients for mother’s and father’s education must both be considered.

Predicted difference ∈education between individuals α∧β =0.131× 4 +0.210 × 4=1.364

This suggests that β₁ is less than 0 in the equation.

3.3

(i) If adults balance sleep and work, an increase in work results in less sleep, assuming all other
factors remain constant.
(ii) The signs of β₂ and β₃ are not immediately clear to me. It could be argued that more educated
individuals may want to experience more in life, leading them to sleep less, implying β₂ < 0. The link
between sleep and age is more complex than this model suggests, and economists may not be the best
suited to assess this.
(iii) Since total work time is measured in minutes, we need to convert five hours into minutes:
totwrk=5 × 60=300 minutes . Consequently, sleep is predicted to decrease by
0.148300=44.4 minutes. Over the course of a week, losing nearly 45 minutes of sleep is not a
significant change.

(iv) Higher education levels are associated with less predicted sleep time, but the impact is minimal.
Assuming a four-year difference between high school and college, a college graduate is expected to
sleep about 45 minutes less per week, all else being equal.
(v) It’s unsurprising that the three explanatory variables account for only about 11.3% of the
variation in sleep. Significant factors in the error term include overall health, marital status, and
whether the individual has children. Health, marital status, and the number and ages of children are
generally correlated with total work time (totwrk), as less healthy individuals typically work less.

Chapter 4:

4.6

(i)

^
price=−14.47+0.976 assess

(16.27) (0.049)

n=88 ; SSR=165,644.51 ; R2=0.820

^
β0 −0 −14.47−0
The t-test: t-statistics ¿ = =−0.889
SE ( ^
β 0) 16.27

The two-sided 10% critical value with 86 degree of freedom is 1.66

-1.6628 < -0.899 < 1.6628

→ Do not reject H 0

The 90% confidence interval is −14.47 ± 1.6628× 16.27=(−41.52 ,12.58)

→ Do not reject H 0

(ii)

H 0: β 0=0∧β 1=1

H 1: β 0 ≠ 0∧β 1 ≠ 1

SS R r−SS Rur (n−k ) 209,448.99−165,644.51 88−2

F= × = × =11.37
SS Rur m 165,644.51 2
 m = number of restrictions

n = number of observations

k = number of regressors in unrestricted regression including a constant in the unrestricted

regression (or the total number of parameters to be estimated).

The critical value is F (2.86)=3.10(5 %)∧4.85(1 %)

→ Reject H 0

(iii)

price=β 0 + β 1 assess+ β 2 lotsize+ β3 sqrft + β 4 bdrms+u

2
n=88 ; R =0.829

H 0 : β 2=0 , β3 =0∧β 4=1

H 1 : β2 ≠ 0∨β 3 ≠ 0∧β 4 ≠ 0

2 2
R ur −Rr (n−k ) 0.829−0.820 88−5
F= 2
× = × =1.456
1−R ur
m 1−0.829 3

The critical value is F (3.83)=2.72(5 %)∧4.04 (1 %)

→ Do not reject

→ Lotsize, sqrft, bdrms are jointly statistically insignificant

(iv)

If the variance of price changes with assess, lotsize, sqrft, or bdrms; the important assumption
2
underlying the classical linear regression model, var (ut )=σ <∞ (The variance of the errors is
constant and finite over all values of ) is violated → The F-test is no longer valid

4.9

(i) ^
sleep=3,368.25−0.148 totwrk +11.13educ +2.20 age

(112.28) (0.17) (5.88) (1.45)

2 ❑
n=706 , R =0.113❑

* Educ

H 0 : β educ=0
 H 1 : β educ ≠0

df =T −k =706−4=702

Standard critical value (df=∞ , in table G.2)= 1.96

^β −11.13
educ
t ^β = = =−1.893
educ ^
SE ( β educ ) 5.89

| |< 1.96 => Fail to reject H 0 at the 5% level

Since t β^ educ

* Age

df =T −k =706−4=702

Standard critical value (df=∞ , in table G.2)= 1.96

^β 2.20
age
t ^β = = =1.517
educ ^
SE ( β age) 1.45

| |< 1.96 => Fail to reject H 0 at the 5% level

Since t β^ educ

=> Neither educ or age is significant at the 5% level agains a two-sided alternative.

(ii)

Restricted model:

^
sleep=3,586.38−0.151 totwrk

(38.91) (0.017)

2 ❑
n=706 , R =0.113❑
2 ư
(Rur−R r ) /q (0.113−0.103)/2
Test statistic= = =3.957
(1−R )/(n−k −1) (1−0.113)/(706−3−1)
2
ur

The 5% critical value with 2 and 702 (table G.3b) df is 3.00

Since F=3.957>3.00 → Reject H0 → educ and age are jointly statistically significant at the 5% level.

(iii) No. The model's estimated tradeoff between working and sleeping is not significantly impacted
by factors like age and educ. The estimate changes slightly from -0.151 to -0.148

(iv) Heteroskedasticity violates the homoskedasticity assumption. The t-test and the F-test are no
longer valid. The t statistics and the F statistics do not have the t distribution and the F distribution
under the null hypothesis if CLM assumptions are not met.
Session 5:

7.

The Ramsey RESET test is designed to check whether the functional form of a regression model is
correctly specified. Specifically, it examines if the correlation between the independent and dependent
variables should be linear or if a non-linear form would be more suitable. This is done by including
powers of the predicted values from the initial regression into a second regression model. If the linear
model is appropriate, these additional terms will not be statistically significant.

If the RESET test indicates a problem, one simple approach is to transform all variables into
logarithms. This transformation converts a multiplicative relationship into an additive one, which may
better fit the data.

However, if the test still fails, it suggests that correlation between the independent and dependent
variables might not be linear. In such cases, we would need to either estimate a non-linear model
(though this may go beyond the course scope) or revisit the model by selecting different variables for
the regression.

8.

(a) While the normality assumption is not necessary for deriving the estimates of 𝛼 and 𝛽 or their
standard errors, it becomes crucial when testing hypotheses about the regression coefficients. The
assumption ensures that the test statistics (e.g., t or F) follow their respective distributions, enabling us
to draw valid conclusions from hypothesis tests.
(b) If residuals do not follow a normal distribution, one option is to use estimation and inference
methods that do not rely on normality. However, these methods tend to be complex and their
performance in various situations is not always well understood.
A practical approach is to examine the residuals for extreme outliers, which are often the cause of
normality violations. Such outliers, whether unusually large positive or negative values, can
significantly impact the skewness or kurtosis of the residuals. Identifying these outliers typically
involves plotting the residuals and analyzing their behavior.
If outliers are detected, their causes should be investigated. For example, they may correspond to
specific events like wars, government changes, or market disruptions. In such cases, dummy variables
can be included in the model to account for these events. This approach allows us to adjust for the
impact of outliers without completely removing valuable information from the dataset.
9.
a) The stability of a regression equation's coefficient estimates over time is known as parameter
structural stability. The coefficient estimates for some subsamples of the data would differ
from those for other subsamples if the regression is not structurally stable. Since we implicitly
assume that the regression parameters are constant over the whole sample period under
consideration when we estimate a regression, this is obviously not what we want to find.
b)
1981M1-1995M12
r t =0.0215+1.491 r mt RSS=0.189 ;T =180
1981M1-1987M10
r t =0.0163+1.308 r mt RSS=0.079 ;T =82
1987M11-1995M12
r t =0.0360+1.613 r mt RSS=0.082; T =98
c) Assuming that the coefficient estimates for the first and second halves of the sample are
α 1∧β 1, respectively, and α 2∧β 2 the alternative and null hypotheses are:
H 0 :α 1=α 2∧β 1=β 2
and H 1: α 1 ≠ α 2∧β 1 ≠ β 2❑
d) The test statistic is computed as:
Test stat =
RSS−(RSS1 + RSS 2) 0.189−(0.079+0.082)
∗(T −2k ) ∗180−4
RSS1 + RSS2 0.079+0.082
= =15.304
k 2
 This has (k,T-2k) degrees of freedom and is distributed according to a F distribution. At 5%,
F(2,176) = 3.05. The null hypothesis, according to which the coefficients in the two sub-
periods are equal, is obviously rejected.
10.
By definition, variables having associated parameters that are not significantly different from zero are
not, from a statistical perspective, helping to explain variations in the dependent variable about its
mean value. One could therefore argue that empirically, they serve no purpose in the fitted regression
model. But leaving such variables in the model will use up valuable degrees of freedom, implying that
the standard errors on all of the other parameters in the regression model, will be unnecessarily higher
as a result. If the number of degrees of freedom is relatively small, then saving a couple by deleting
two variables with insignificant parameters could be useful. On the other hand, if the number of
degrees of freedom is already very large, the impact of these additional irrelevant variables on the
others is likely to be inconsequential.

11.

Variables with parameters that are statistically insignificant (not significantly different from zero) do
not contribute to explaining the variation in the dependent variable. Including such variables in the
regression model serves no practical purpose and may reduce the model's efficiency.

Specifically, retaining these variables uses up degrees of freedom, which can lead to higher standard
errors for the other estimated parameters. This effect is more pronounced when the available degrees
of freedom are limited. In contrast, if the degrees of freedom are already ample, the impact of
including insignificant variables is likely to be negligible. Removing these variables can therefore
streamline the model and improve its precision, particularly in cases where sample size is a constraint.

12.

An outlier dummy variable will take the value one for one observation in the sample and zero for all
others. The Chow test involves splitting the sample into two parts. If we then try to run the regression
on both the sub-parts but the model contains such an outlier dummy, then the observations on that
dummy will be zero everywhere for one of the regressions. For that sub-sample, the outlier dummy
would show perfect multicollinearity with the intercept and therefore the model could not be
estimated.

13.
(a) "Measurement error," also referred to as the errors-in-variables problem, is the discrepancy
between the variable's true values and observed values.

(b) Measurement errors can occur in a variety of circumstances:

- Macroeconomic variables are almost always estimated quantities (GDP, inflation, and so on),
as is most information contained in company accounts
- Sometimes we cannot observe or obtain data on a variable we require and so we need to use a
proxy variable – for instance, many models include expected quantities (e.g., expected
inflation) but we cannot typically measure expectations.

(c) Measurement error becomes a problem only when it appears in the regression's independent
variables, where it violates the assumption that such variables are not stochastic and makes them so,
which leads to a loss of consistency in parameter estimation. However, since the disturbance term in
the regression model will account for it, measurement error in the explained variable is considerably
less significant. The disturbance term will therefore comprise both the typical disturbance term and an
additional noise source from the measurement error when the explained variable is measured with
error. In these conditions, the parameter estimates will not change.

(d) Since betas are estimated at the first stage by the standard approach to testing the CAPM, which
was pioneered by Fama and MacBeth (1973), rather than being directly observable, they will
undoubtedly contain measurement error. As a result, the impact of measurement error on CAPM tests
can be quite significant. Explanatory variables will then be included in a series of cross-sectional
regressions in the second stage. The effect of this has sometimes been termed attenuation bias. Tests
of the CAPM showed that the relationship between beta and returns was smaller than expected, and
this is precisely what would happen as a result of measurement error

- Possible solutions:
+ Various approaches to solving this issue have been proposed, the most common of
which is to use portfolio betas in place of individual betas
+ An alternative approach (Shanken,1992) is to modify the standard errors in the
second stage regression to adjust directly for the measurement errors.

End-of-chapter exercises from C.9 of the Wooldridge: 1,2,3

1.
There is functional form misspecification if β 6 ≠ 0 or β 7 ≠ 0, where these are the population
parameters on ceoten2 and comten2, respectively. Therefore, we test the joint significance of
these variables using the R-squared form of the F test: F = [(.375 − .353)/(1 − .375)][(177 – 8)/2]
2.97. With 2 and ∞ df, the 10% critical value is 2.30 while the 5% critical value is 3.00. Thus, the
p-value is slightly above .05, which is reasonable evidence of functional form misspecification.

2.
(i) The coefficient on voteA88 implies that if candidate A had one more percentage point of the vote
in 1988, she/he is predicted to have only .067 more percentage points in 1990. Or, 10 more percentage
points in 1988 implies .67 points, or less than one point, in 1990. The t statistic is only about 1.26, and
so the variable is insignificant at the 10% level against the positive one-sided alternative. (The critical
value is 1.282.) While this small effect initially seems surprising, it is much less so when we
remember that candidate A in 1990 is always the incumbent. Therefore, what we are finding is that,
conditional on being the incumbent, the percent of the vote received in 1988 does not have a strong
effect on the percent of the vote in 1990.
(ii) Naturally, the coefficients change, but not in important ways,
especially once statistical significance is taken into account. For
example, while the coefficient on log(expendA) goes from −.929 to −.839,
the coefficient is not statistically or practically significant anyway
(and its sign is not what we expect). The magnitudes of the coefficients in
both equations are quite similar, and there are certainly no sign
changes. This is not surprising given the insignificance of voteA88.
3.
(i) Eligibility for the federally funded school lunch program is very tightly linked to living in poverty.
Therefore, the percentage of students eligible for the lunch program is very similar to the percentage
of students living in poverty.
(ii) We can use our usual reasoning on omitting important variables from a regression equation. The
variables log(expend) and lnchprg are negatively correlated: school districts with poorer children
spend, on average, less on schools. Further, β3 < 0. From Table 3.2, omitting lnchprg (the proxy for
poverty) from the regression produces an upward biased estimator of β1 (ignoring the presence of
log(enroll) in the model). So when we control for the poverty rate, the effect of spending falls.
(iii) Once we control for lnchprg, the coefficient on log(enroll) becomes
negative and has a t of about –2.17, which is significant at the 5% level
against a two-sided alternative. The coefficient implies that −(1.26/100)
(%Δenroll) = −.0126(%Δenroll). Therefore, a 10% increase in enrollment
leads to a drop in math10 of .126 percentage points. Δmath10 ≈ 72
(iv) Both math10 and lnchprg are percentages. Therefore, a ten percentage point increase in lnchprg
leads to about a 3.23 percentage point fall in math10, a sizeable effect.
(v) In column (1), we are explaining very little of the variation in pass rates on the MEAP math test:
less than 3%. In column (2), we are explaining almost 19% (which still leaves much variation
unexplained). Clearly, most of the variation in math10 is explained by variation in lnchprg. This is a
common finding in studies of school performance: family income (or related factors, such as living in
poverty) are much more important in explaining student performance than are spending per student or
other school characteristics.

Session 6:
1.
The inclusion of a squared term for roe to capture a diminishing effect on log(salary) does not appear
to be necessary. When testing the null hypothesis that the coefficient on roe² equals zero, we are
essentially examining whether the relationship between roe and log(salary) is purely linear. The
estimated t-statistic for roe² is approximately -0.31, which is far below the critical values required for
statistical significance at conventional levels. Thus, there is no strong statistical evidence to confirm
that the effect of roe on log(salary) diminishes as roe increases.

Moreover, even from a practical perspective, the impact is minimal, when roe reaches a value of 25
(approximately one standard deviation above the mean), the marginal effect declines only slightly,
from 0.215 to 0.211. This very small change suggests that the diminishing return implied by the
squared term has little economic relevance within the observed data range. Overall, both statistical
and practical considerations indicate that a simpler linear specification in roe is adequate.

3.
^
β1
(i) The turning point can be computed as the ratio of the linear and squared coefficients, , which
2| ^
β|
2
simplifies to:
0.0003/(0.000000014) ≈ 21,428
(ii) Most likely. The t-statistic is approximately -1.89, which is
statistically significant under a one-sided test at the 5% level (critical
value ≈ -1.70 with df = 29). The corresponding p-value is about 0.036,
supporting rejection of the null hypothesis.
(iii) When sales is rescaled by dividing by 1,000 to form salesbil, the coefficient must be adjusted
accordingly. Specifically, the coefficient on sales is multiplied by 1,000:
(1,000)×0.00030=0.30.

The standard error is also scaled by the same factor. For the squared term, since salesbil² corresponds
to sales divided by 1,000,000, the coefficient is multiplied by 1,000,000:

(1,000,000)×0.0000000070=0.0070.

The standard error of this term is also increased by a factor of one million. The intercept remains
unchanged because rdintens is not rescaled, and the R² value is unaffected.

2
rdinten=2.613+ 0.30 ×salesbil−0.0070 × salesbi

(0.429) (0.14) (0.0037)

where n = 32
(iv) The specification in part (iii) is preferable because it avoids unnecessarily small decimal values,
making the equation easier to interpret. Although the numerical form changes due to rescaling, the
underlying relationship remains the same.

4.
(i) We have:
Δlog(wage) = β1Δeduc + β2Δeduc ⋅ pareduc
Δlog(wage) = Δeduc(β1 + β2pareduc)
Δlog(wage)/Δeduc = β1 + β2pareduc

Holding other factors constant, it is reasonable to expect that individuals with more highly educated
parents benefit more from an additional year of schooling. This implies a positive interaction effect,
so β₂ > 0. Conversely, if parental education has no reinforcing role, β₂ would be zero or negative.

(ii) We have:
 Δlog(wage)/Δeduc = β1 + β2pareduc = 0.0047 + 0.00078pareduc

● If pareduc = 32, Δlog(wage)/Δeduc = 0.072

● If pareduc = 24, Δlog(wage)/Δeduc = 0.066

These results suggest that individuals whose parents have a college-level education experience an
approximate return to schooling of 7.2%. In contrast, those whose parents only completed high school
have a return of about 6.6%. The difference of 0.6 percentage points indicates that parental education
modestly enhances the return to an additional year of schooling.

(iii) The regression model is: log(wage) = β0 + β1educ + β2educ ⋅ pareduc + β3pareduc +
β4exper + β5tenure + μ

We test:

H0: β2 = 0

H1: β2 ≠ 0

T-statistics = -1.33 => Fail to reject H0 at the 10% level

This suggests that once pareduc is included as a separate explanatory variable, the interaction term
does not meaningfully affect the estimated return to education.

7.

Among the three specifications, the second model is preferred because it yields a higher adjusted R².
It is also more parsimonious than the 3rd equation, but the same number as 1st. In addition, its structure
is more straightforward, making the results easier to interpret.

End-of-chapter computer exercises from Wooldridge's 7th edition: 2, 3, 4, 5

2.
2
(i) log (wage)=.128+.0904 educ +.0410 exper−.000714 exper
(.106) (.0075) (.0052) (.000116)
2 2
Where n=526 , R =300 , R =.296
(ii) The t-statistic for exper 2 is around -6.16, with p−value = 0.000. This indicates that exper 2 is
statistically significant at the 1% level.
 ∂^
log log wage ^
(iii) = β 2−2 ^
β 3 exper=.041−.00143 exper
∂ exper

∂^
log log wage
In the fifth year of experience, exper = 4, ≈ 3.53%.
∂ exper
∂^log log wage
Similarly, if exper = 19 in the twentieth year of experience, ≈ 1.39%
∂ exper
(iv) The turnaround point can be caculated:
∂^
log log wage ^
= β 2−2 ^
β 3 exper=.041−.00143 exper=0 This results in an estimated
∂ exper
value of approximately 28.7 years of experience. In the sample, 121 individuals have at
least this level of experience, accounting for roughly 23% of all observations

3.
(i) Assuming exper (and elements in u) remain constant, we have the equation:
Δ log(wage)=β1 Δeduc + β 3 (Δeduc)exper=¿+ β 3 exper ¿(Δeduc)
Alternatively, it can be rewritten as:
Δ log (wage)
=¿+ β 3 exper ¿
Δeduc

This reflects the approximate percentage change in wages associated with one additional year of
education.

(ii) The null hypothesis is H₀: β₃ = 0. If we expect a positive interaction between education and
experience, such that the productivity gains from education are larger for more experienced
individuals, then the appropriate alternative hypothesis is β₃ > 0.

(iii) The estimated regression equation:

Δ log (wage)=5.95+0.044∗educ−0.0215∗exper+0.0032∗(educ∗exper)
(0.24) (0.0174) (0.02) (0.0153)
Sample size: n = 935
Coefficient of determination: R² = 0.135, Adjusted R² = 0.132
The t-statistic for the interaction term is approximately 2.13, yielding a p-value below 0.02 against the
alternative hypothesis H₁: β₃ > 0. As a result, we reject the null hypothesis H₀: β₃ = 0 in favor of H₁:
β₃ > 0 at a 2% significance level.

(iv) The equation can be reformulated as follows:

Δ log(wage)=β1 +Θ1 Δeduc + β 2 exper + β 3 educ∗(educ−10)+u

We then regress log(wage) on educ, exper, and educ(exper - 10), with a focus on the coefficient of
educ. The estimated value of θ₁ is approximately 0.0761, with a standard error of 0.0066. The 95%
confidence interval for θ₁ is roughly [0.063, 0.089].
4.

The quadratic term is highly statistically significant (t-statistic) of around -3.87.

(ii) The value of class size at which the predicted SAT score reaches its maximum, denoted as hsize,
is obtained from the turning point of the quadratic function. Using the standard formula gives: hsize ≈
4.65

Since hsize is measured in hundreds, this corresponds to approximately 465 students. However, the R²
is very low, indicating that class size explains only a small fraction of the variation in SAT scores.

(iii) The sample is not representative of all high school seniors because it includes only students who
took the SAT. To make valid inferences about the entire population, we would need a random sample
of all students taking a common standardized test.

(iv) Using log(sat) - dependent variable:

log(sat) 6.896 + 0.0196 hsize - 0.00209 hsize²

(0.006) (0.0040) (0.00054)

n = 4,137, R2 = 0.0078.

From this specification, the implied optimal class size is about 4.69 (469 students), which is very
close to the estimate obtained from the level-level model.

5.
 (i) The model's estimated outcomes are
^ = -1.297 + 0.168 log(lotsize) + 0.700 log (sqrft) + 0.037 bdrms
lprice
(0.65) (.038) (.093) (.028)

n = 88, R2 = 0.643

(ii) With sqrft = 2,500, bdrms = 4, lotsize = 20,000:

^ = -1,297 + .168 ⋅ log(20,000) + .700 ⋅ log(2,500) + .037(4) = 5.99
lprice
At part (i), we have 0^ 2 = 0.034078134 in the result table.

^ ^ ) = exp(0.034078134/2) × exp(5.9928985) = $407.458

price = exp(0^ 2 /2) × exp(lprice
(iii) When estimating the model in levels, the R² is approximately 0.672. In contrast, the
correlation between the actual price and the predicted values from the log model is 0.859.
Squaring this correlation yields about 0.738, which exceeds the R² of the level
specification. This indicates that the log model provides a better fit and has stronger
predictive performance for price.
Session 7:

1.

In an Autoregressive (AR) model, the current value of a time series, y t ,is determined by its own past
p values along with a contemporaneous error term, ut . Conversely, a Moving Average (MA) model
represents ut as a function of the present and past q values of the error term, ut . AR models exhibit
long memory properties, meaning that historical observations continue to have a significant influence
on future values over an extended period. A shock to y t gradually diminishes over time, depending on
the autoregressive coefficients. In contrast, MA models have short memory, meaning only a few past
error terms influence the current value and a shock to y t affects only the next q periods before dying
out completely.

2.

ARMA models are especially useful for financial time series because they are flexible and relatively
easy to apply. They allow for reasonable forecasting without needing detailed knowledge of external
economic factors, which is often required in traditional econometric approaches. One major advantage
is that financial data is often available at high frequencies, while key explanatory variables like
macroeconomic indicators or accounting ratios are usually reported much less frequently. This makes
ARMA models a practical choice when external data is limited or unavailable.

Regarding the differences between AR, MA, and ARMA processes:

- AR (AutoRegressive) models assume that past values of a time series influence its future values.
Essentially, the current value is based on a combination of previous observations.

- MA (Moving Average) models focus on past forecasting errors rather than past values themselves.
The current value is determined by past random shocks, making it useful for capturing short-term
fluctuations.

- ARMA (AutoRegressive Moving Average) models combine both AR and MA components,

allowing them to account for both past values and past errors, making them more adaptable to
different types of data patterns.

This combination of flexibility and ease of use is what makes ARMA models particularly valuable for
analyzing financial data.

3.

y t = y t −1 +ut (1)

y t =0.5 y t −1 +ut (2)

y t =0.8 ut −1 +u t (3)

a)

Equation (1) is AR(1) model with ϕ=1 ; which makes it a random walk model.

Equation (2) is AR model with ϕ=0.5; meaning it is a stationary autoregressive process.

Equation (3) is MA(1) model with ϕ=0.8

b)

Model (1):

- The ACF will show high and persistent correlations, decreasing very slowly over time.
- Since the process is non-stationary, the ACF does not decay to zero in a predictable way.
- The PACF would exhibit a pronounced autocorrelation at the first lag and near-zero values at
higher lags.

Model (2)

- The ACF will decay exponentially as the lag increases.

- Since < 1, the process is stationary, meaning that past values have diminishing influence
over time.
- The PACF would exhibit a pronounced autocorrelation at the first lag and near-zero values at
higher lags.

Model (3)
 - The ACF will show a sharp cutoff after lag 1, meaning that there will be a significant
correlation at lag 1, but no correlation at higher lags.
- This is because an MA(1) process only depends on the most recent shock.
- The PACF declines geometrically rather than cutting off at a specific lag.

c)

Stock prices are often modeled as a random walk because they incorporate all available information
and are unpredictable (Efficient Market Hypothesis - EMH). According to Rational Expectations
Theory, people’s current expectations of the economy are, themselves, able to influence what the
future state of the economy will become, hence today’s price should only differ from yesterday’s by
the amount of unexpected news which influences dividend payments. This suggests that model (1)
best represents stock prices in practice.

However, if stock prices truly followed Model (1) (random walk), forecasting would be useless, as
tomorrow’s price is just today’s price plus an unpredictable shock. If the stock market truly adhered to
the process defined by equations (2) or (3), our model could generate statistically meaningful
forecasts of the time series.

d)

Model (1): A shock persists indefinitely because each shock permanently affects the level of the
series.

Model (2): Shocks dissipate over time because of the factor 0.5

Model (3): A shock affects the series for only two periods: directly in period tt and indirectly in period
t+1 through 0.8ut , after which it vanishes.

4.

a) Box and Jenkins (1976) proposed 3 main steps for constructing an ARMA model:

1. Identification: Use autocorrelation (ACF) and partial autocorrelation (PACF) plots to

determine the appropriate order of the model.
2. Estimation: Once the model structure is identified, estimate its parameters using methods like
least squares or maximum likelihood estimation.
3. Diagnostic Checking: Assess whether the model adequately fits the data by:
○ Overfitting Test: Fit a larger model and test if additional parameters can be
eliminated.
○ Residual Analysis: Check if the residuals are uncorrelated, for example, using the
Ljung-Box test.

If the model is adequate, it can be used for forecasting. Otherwise, adjustments should be made, and
the process should restart.

b) A major criticism of this methodology is the subjectivity in model identification. ACF and PACF
plots can be difficult to interpret accurately, especially with real-world data, making model selection
more of an educated guess than a precise scientific approach. Additionally, the diagnostic checking
step only indicates when a model is too small but does not effectively identify when a model is too
complex.

c. An alternative approach is to use information criteria, such as:

- Akaike Information Criterion (AIC):

2 2k
AIC=ln (σ )+
T

- Schwarz’s Bayesian Information Criterion (BIC):

k × ln(T )
BIC=ln(σ 2)+
T

Where:

● is the variance of the residuals.

● k is the number of parameters in the model.
● T is the number of observations in the dataset.

These criteria help balance model accuracy and complexity. As the number of parameters increases,
the penalty term in the formula also increases to prevent overfitting. The optimal model is the one
with the lowest AIC or BIC value.

5.

Express the model as a lag polynomial in y t

2
y t (1−0.803 L−0.682 L )=ut

The characteristic equation for stationary is:

2
(1−0.803 L−0.682 L )=0

1
To simplify, multiply through by setting z for the characteristic equation and rearranging
−0.682
2
z + 1.177 z−1.466=0

Using the standard formula for obtaining the roots of a quadratic equation,

−1.177 ± √ 1.177 2+ 4 ×1 ×1.466

z= =0.758∨1.934
2

For the process to be stationary, both roots must be greater than 1 in absolute value, we conclude that
the estimated model is not stationary in this case.
6.

Using formulas, we calculate AIC and SBIC for each model order.

2 2k
AIC=ln (σ )+
T

k × ln(T )
BIC=ln(σ 2)+
T

ARMA (p,q) model order log( ) AIC SBIC

(0,0) 0.932 0.942 0.944

(1,0) 0.864 0.884 0.887

(0,1) 0.902 0.922 0.925

(1,1) 0.836 0.866 0.870

(2,1) 0.801 0.841 0.847

(1,2) 0.821 0.861 0.867

(2,2) 0.789 0.839 0.846

(3,2) 0.773 0.833* 0.842*

(2,3) 0.782 0.842 0.851

(3,3) 0.764 0.834 0.844

Both the Schwarz Bayesian Information Criterion (SBIC) and the Akaike Information Criterion (AIC)
indicate that the optimal model specification is an ARMA(3,2) process.

7.

To evaluate the adequacy of the estimated models, we can apply the Ljung-Box test to the residuals to
detect any remaining linear dependence that the model may not have captured. Another approach is
out-of-sample forecasting, where a subset of the data is withheld during model estimation. For
instance, with 2000 observations, we might estimate the model using the first 1,800 and reserve the
remaining 200 for forecasting. The model that yields the most precise forecasts would be considered
the superior choice.

8.

A larger ARMA model does not necessarily mean a better model. While adding more lags may
improve the fit to historical data, financial time series often contain a lot of noise-random fluctuations
that are unlikely to repeat in the future. If a model captures this noise instead of meaningful patterns,
it will not generalize well for forecasting.

This is why the principle of parsimony is important. A good model should be as simple as possible
while still capturing the key features of the data. Using too many parameters reduces the degrees of
freedom, leading to larger standard errors and weaker hypothesis test results.

In summary, an overly complex model may fit past data well but can fail to make accurate
predictions. The best approach is to strike a balance between fit and simplicity.

9.

a)

To determine the appropriate time-series model, we assess the autocorrelation function (ACF) and
partial autocorrelation function (PACF) while using the significance threshold:

1 1
± 1.96 =±1.96 =± 0.196
√T √ 100
The ACF is significant at lags 1 and 4, while all other lags are insignificant.

The PACF at lag 1,2,3,4,5 is all significant, then gradually declines.

This pattern aligns with a moving average process of order 1 (MA(1)), since:

- The ACF shows only the first lag as significant, while the others are close to zero (except lag
4, which may be noise).
- The PACF has a slow, gradual decline, which is characteristic of an MA process rather than
an AR process.

b)

The Ljung-Box Q test checks formula

m 2
τk
Q∗¿ T (T +2) ∑ ❑
k=1 T −k

where:

● T = 100 (sample size),

● m = 3 (since we are testing the first three autocorrelations),
● are the sample autocorrelations.

The null hypothesis is H0: = 0 and = 0 and = 0.

The test statistic is calculated as

 [ ]
2 2 2
0.420 0.104 0.032
Q∗¿ 100× 102× + + =19.41
100−1 100−2 100−3

At the 5% and 1% significance levels, the critical values for a χ2\chi^2 distribution with 3 degrees of
freedom are 7.81 and 11.3, respectively. If the test statistic exceeds these thresholds, we reject the null
hypothesis.

10.

To forecast the series y t , we use the given ARMA(1,1) model:

y t 0.036+ 0.69 y t−1 +0.42 ut−1 +ut

Given that y t −1=3.4∧ut−1=−1.3 ,we can compute the forecast for t:

E( y t∨ y t −1 , ut −1)=0.036+0.69(3.4 )+ 0.42(−1.3)=0.036+ 2.346−0.546=1.836

Since ut is unknown, we replace y t with our forecast f t−1=1.836 :

f t +1=0.036+0.036 +0.69(1.836)=0.036+1.268=1.302

Similarly, for t+2:

f t +2=0.036+ 0.69(1.302)=0.036+0.898=0.935

Thus, our forecasts are:

- f t=1.836
- f t +1=1.302
- f t +2=0.935

b)

- y t =−0.032
- y t +1=0.96
- y t +2=0.203

1
MSE= ¿
3

We observe that most of the error comes from the first forecast, meaning that errors may be higher
when there is significant deviation in initial conditions.

c) A colleague suggests using an exponential smoothing model with a smoothing constant of 0.15 and
a smoothed value St−1=0.0305. Given that exponential smoothing inherently factors in past
observations, we assume that:

St =St +1=St +2=0.0305

d) To compare the two models, we calculate the MSE for the exponential smoothing forecasts:

1
MSE= ¿
3

Since the MSE for exponential smoothing (0.2998) is much lower than that of the ARMA model
(1.380), the exponential smoothing method provides more accurate forecasts in this case. However,
because we only tested three forecasts, further evaluation over a longer period would be needed to
make a conclusive judgment.

11.

a)

White Noise

● ACF: The autocorrelation function (ACF) is approximately zero for all lags (except lag 0,
which is 1).
● PACF: The partial autocorrelation function (PACF) is also approximately zero for all lags.

AR(2) Process

● ACF: The ACF follows a damped sinusoidal or exponential decay, depending on the roots of
the characteristic equation.
● PACF: The PACF shows significant spikes at lags 1 and 2 but cuts off after that.

MA(1) Process

● ACF: The ACF has a significant spike at lag 1 and cuts off abruptly afterward.
● PACF: The PACF declines geometrically toward zero.

ARMA(2,1) Process

● ACF: The ACF exhibits a mix of damped sinusoidal behavior (from the AR(2) component)
and a cut-off after a short lag due to the MA(1) component.
● PACF: The PACF has a mix of the characteristics of AR(2) and MA(1), showing a slow
decay rather than a sharp cut-off.

b)

For an MA(q) process to be invertible, the MA coefficients must have roots outside the unit circle.
Here, the characteristic equation for the MA part is: (1+0.42z) = 0 -> z = -2.38, which lies outside the
unit circle, and therefore the MA part of the model is invertible.

c)

To compute multi-step forecasts, we express the model in a way that facilities forecasting:

y t +1=0.036 +0.69 y t +0.42 ut +u t +1

 y t +2=0.036 +0.69 y t +1+ 0.42u t+1 +ut +2

y t +3=0.036 +0.69 y t +2+ 0.42u t+ 2+u t+3

y t +4 =0.036+0.69 y t +3 +0.42 ut +3 +ut +4

Since expectations of future error terms are zero, forecasts are computed as follows:

1. One-Step-Ahead Forecasts:

E( y t+ 1∨ y t , y t−1 ,...)=f t , 1=Et [ y t +1 ]=E t [ 0.036+0.69 y t + 0.42u t +ut +1 ] =0.036+0.69 y t +0.42 ut

2. Two-Step-Ahead Forecasts:

E( y t+ 2∨ y t , y t−1 ,...)=f t ,2=Et [ yt +2 ]= Et [ 0.036+ 0.69 y t +1 +0.42 ut +1 +ut +2 ]=0.036+ 0.69 f t ,1

3. Three-Step-Ahead Forecasts:

E( y t+ 3∨ y t , y t−1 ,...)=f t ,1=Et [ y t+3 ] =Et [ 0.036+ 0.69 y t+2 +0.42 ut +2 +ut +3 ]=0.036+0.69 f t ,2

4. One-Step-Ahead Forecasts:

E( y t+ 4∨ y t , y t −1 , ...)=f t ,1 =Et [ y t +4 ]=Et [ 0.036+ 0.69 y t+3 +0.42 ut +3 +ut + 4 ]=0.036+ 0.69 f t ,3

d)

Two Forecast Evaluation Criteria

Mean Squared Error (MSE)

· Measures the average squared difference between actual and forecasted values.

· Sensitive to large errors due to squaring.

Mean Absolute Error (MAE)

· Measures the average absolute difference between actual and forecasted values.

· Less sensitive to large deviations compared to MSE.

MSE emphasizes large forecast errors, while MAE treats all errors equally.

e)

Ordinary Least Squares (OLS) is not appropriate for ARMA models because the presence of a moving
average component makes the error terms dependent on past observations, violating OLS
assumptions, besides, the MA terms introduce a non-linear structure, making standard OLS estimation
inconsistent.
The most common method for estimating ARMA parameters is Maximum Likelihood Estimation
(MLE) or Conditional Least Squares (CLS).

Autoregressive (AR) models can be estimated using Ordinary Least Squares (OLS) or Maximum
Likelihood estimation. These models include only lagged values of the dependent variable as
regressors on the right-hand side (RHS) of the equation. These lagged values are observed data, so
OLS can treat them as standard regressors and estimate parameters efficiently.

MLE is used to estimate the parameters of MA and ARMA models. It works by selecting the
parameter values that maximize the likelihood of observing the given data [Link] requires
assuming a probability distribution for the error terms (typically normal distribution).

Unlike OLS, MLE estimates both the model parameters and the residuals simultaneously.

12.

a) Macroeconomic and financial data have distinct characteristics, particularly in terms of how
frequently they are recorded. Financial data, such as asset returns and exchange rates, can be
measured at very high frequencies, even down to the second or minute. In contrast, macroeconomic
indicators like GDP, inflation, and unemployment rates are typically recorded at lower frequencies,
such as quarterly or monthly. This difference makes it challenging to use macroeconomic variables in
models that require high-frequency data, such as those used for predicting short-term fluctuations in
financial markets.

Due to this discrepancy, different econometric approaches are often applied to each type of data.
Financial data, which tends to exhibit high volatility and rapid changes, is well-suited for time-series
models like ARMA or GARCH, which can effectively capture short-term trends and fluctuations. On
the other hand, macroeconomic data is typically analyzed using structural models, such as VAR
(Vector Autoregression) or cointegration techniques, which help understand long-term relationships
between variables.

Another key distinction is the forecasting approach. Financial models often focus on real-time
forecasting, where time-series models iteratively generate predictions based on past values. In
contrast, structural models require forecasts for underlying macroeconomic variables, making the
process more complex. This difference further supports the use of separate econometric tools tailored
to the characteristics of each data type.

b) To determine whether the autocorrelation and partial autocorrelation coefficients are statistically
significant at the 5% level, we can use a simple "rule of thumb." Specifically, a coefficient is
1
considered significant if its absolute value exceeds 1.96 ± , where T is the sample size. Given that
√T
T =500, the threshold for significance in this case is approximately ± 0.08.

According to this criterion, the autocorrelation coefficients at lags 1 and 5, as well as the partial
autocorrelation coefficients at lags 1, 2, and 3, surpass the threshold, indicating statistical significance.

To formally test whether the first five autocorrelation coefficients are jointly zero, we can apply the
Box-Pierce and Ljung-Box tests. The Box-Pierce test statistic is computed as:
 5
Q=T ∑ ❑ r 2k
k=1

Substituting the given values:

2 2 2 2 2
Q=500 ×(0.307 +(−0.013) +0.086 +0.031 +(−0.197) )=70.79

Similarly, the Ljung-Box test statistic is:

3 2
rk
Q∗¿ T (T +2) ∑ ❑
k=1 T −k

Using the given coefficients:

2 2 2 2
502 0.307 (−0.013) 0.086 (−0.197)
Q∗¿ 500× ×( + + + )=71.39
500 500−1 500−2 500−4 500−5

Both test statistics follow a χ 2 distribution

df =5

The critical values are 11.07 at 5% and 15.09 at 1%.

Since both test statistics(70.79 and 71.39) are far larger than these critical values => Reject the null
hypothesis that the first five autocorrelation coefficients are jointly equal to zero.

c) Suggested Model for the Series in Part (b): An MA(1) Approach

Based on the autocorrelation and partial autocorrelation patterns presented, an MA(1) model appears
to be a tentatively suitable choice for representing this series.

- The autocorrelation coefficients exhibit a distinct characteristic: significance primarily at lag 1,

followed by a rapid decline towards zero. Conversely, the partial autocorrelation coefficients
demonstrate a more gradual decay as the lag length increases. This contrasting behavior is indicative
of an MA(1) process.

- It is important to acknowledge the significant autocorrelation coefficient observed at lag 5. While

this deviates from the expected MA(1) pattern, such anomalies are not uncommon in real-world data.
Actual data-generating processes rarely perfectly align with theoretical ARMA models.

- The goal is not necessarily to identify the "true" underlying model (which is likely far more
complex), but rather to find a parsimonious and useful approximation that captures the salient features
of the data. In this context, an MA(1) model offers a potentially valuable framework for
understanding and potentially forecasting the series, despite the presence of the lag 5 anomaly.

d) Given two different approaches to modeling the USD/GBP exchange rate returns:

Model A (MA(1)): ^χ t =0.38+ 0.10u t−1

Model B (AR(2)): ^χ t =0.63+ 0.17 xt −1−0.09 x t−2

Using the given data:

x z =0.31

x z −1 =0.02

x z −2=−0.16

u z=−0.02

u z−1=0.13

u z−2=0.19

Forecasts for Model A (MA(1))

The MA(1) model only depends on the most recent error term. This means:

- At z+1: f z , 1=0.38+0.10 (−0.02)=0.378

- At z+2 and beyond: Since MA(1) only relies on the previous error term, and future errors are
unknown (assumed to be zero in forecasting), all further forecasts remain constant at:

f z , 2=f z , 3=f z , 4=0.38

Forecasts for Model B (AR(2))

For Model B, the AR(2) model relies on past values of x t :

- At z+1: f t +1=0.63+0.17 (0.31)−0.09(0.02)=0.681

- At z+2: f t +2=0.63+0.17 (0.681)−0.09( 0.31)=0.718

- At z+3: f t +3=0.63+0.17 (0.718)−0.09 (0.681)=0.690

- At z+4: f t +4=0.63+0.17(0.690)−0.09(0.716)=0.683

e) Box and Jenkins (1976) proposed two methods to check if an ARMA model is appropriate:
overfitting tests and residual diagnostics.

1. Overfitting Test – Expands the model by adding extra AR and MA terms. If the new terms are
statistically insignificant, the original model is sufficient. For example:

Model A (MA(1)) → Test with ARMA(1,2)

Model B (AR(2)) → Test with ARMA(3,1) If the added complexity doesn’t improve performance,
it’s unnecessary.

2. Residual Diagnostics – Checks if the residuals (errors) are random by analyzing their ACF
and PACF. If residuals show correlation, the model is missing important patterns and may be
inadequate.

Limitation: These methods mainly detect underfitting (model too simple) but won’t flag overfitting
(model too complex).

f)

Calculating MSE for Each Model

Given actual values:

● x z +1=0.62
● x z +2=0.19
● x z +3=0.32
● x z +4 =0.72

Model A (MA(1)) Predictions:

● f z +1=0.378
● f z +2=0.38
● f z +3=0.38
● f z +4 =0.38

MSE for Model A:

1
MSEA= ¿
4

Model B (AR(2)) Predictions:

● f z +1=0.681
● f z +2=0.718
● f z +3=0.690
● f z +4 =0.683

MSE for Model B:

1
MSEB= ¿
4
Since Model A (MSE=0.175) has a lower MSE than Model B (MSE=0.326), it provided more
accurate forecasts in this case. Thus, Researcher A's model (MA(1)) performed better.