EF tentafrågor Flashcards

1
Q

Explain the difference between PROBIT and LOGIT models

A

The probit model uses the standard normal CDF to map x’B into probability, while the LOGIT models uses the logit function to do this.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Nikki states that the Yieldspread has a higher economic impact in the LOGIT model than in the PROBIT model, as the corresponding coefficients are equal to 0.31 and 0.17 respectively.

c) Explain why Nikki is wrong. What should Nikki then do to assess the economic impact of YLDSPREAD on the bullbear market variable in both models? (2 pt)

A

Since Probit and Logit models are non-linear models while at the same time they have a different link function, you can not compare the estimated coefficients directly. (1 pt) [important: different LINK function]

To compare the economic impact, Nikki should consider the marginal effects of both models. (1 pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

How can we evaluate whether an estimated ARMA model is ok?

A

Several possibilities exist. Many \misspecication tests” aim
to test whether the residuals of the ARMA model satisfy
the white noise properties:

-Test of no residual autocorrelation

-Test of homoskedasticity (constant variance), often based
on autocorrelations of squared residuals

-Test of normality: Skewness = 0, Kurtosis = 3.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

What is the difference bwtween a CAP and a ROC curve?

A

See slide 38. The CAP curve plots the proportion of data (j/N) against the hit rate, while the ROC curve uses the false alarm rate 𝐹𝑗/𝐹̅ against the hit-rate (1 pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

You would like to investigate the Capital structure of 250 firms listed at the NYSE from 1970 – 2018.
As a dependent variable, you use the leverage (𝐿𝑖𝑡) whereas the independent variables consist of the logarithm of Size (ln𝑆𝑖𝑡), the logarithm of 1 + the age of the firm (ln(1+𝐴𝑔𝑒𝑖𝑡)), the Profits/sales ratio (𝑃𝑆𝑖𝑡), the asset tangibility (𝑇𝑎𝑛𝑖𝑡) and the R&D/Sales ratio (𝑅𝐷𝑆𝑖𝑡) with t in years. The 250 firms of the sample can be classified into 10 different industries.

b) What is the economic Interpretation of the coefficient corresponding to the ln Size variable? (1 pt)

A

If size of a firm at time t increases with 1%, then the leverage will change with 1/100 𝛽1 units (1 pt)

Note: I would like to see the relative change of x w.r.t. absolute effect of y!

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

You would like to investigate the Capital structure of 250 firms listed at the NYSE from 1970 – 2018.
As a dependent variable, you use the leverage (𝐿𝑖𝑡) whereas the independent variables consist of the logarithm of Size (ln𝑆𝑖𝑡), the logarithm of 1 + the age of the firm (ln(1+𝐴𝑔𝑒𝑖𝑡)), the Profits/sales ratio (𝑃𝑆𝑖𝑡), the asset tangibility (𝑇𝑎𝑛𝑖𝑡) and the R&D/Sales ratio (𝑅𝐷𝑆𝑖𝑡) with t in years. The 250 firms of the sample can be classified into 10 different industries.

a) Write down the pooled regression model for 𝐿𝑖𝑡 ? (1 pt)
b) What is the economic Interpretation of the coefficient corresponding to the ln Size variable? (1 pt)

A

𝐿𝑖𝑡=𝛼+ 𝛽1ln𝑆𝑖𝑡+𝛽2ln(1+𝐴𝑔𝑒𝑖𝑡)+ 𝛽3𝑃𝑆𝑖𝑡+𝛽4𝑇𝑎𝑛𝑖𝑡+ 𝛽5𝑅𝐷𝑆𝑖𝑡+ 𝜖𝑖𝑡

b)

If size of a firm at time t increases with 1%, then the leverage will change with 1/100 𝛽1 units (1 pt)

Note: I would like to see the relative change of x w.r.t. absolute effect of y!

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

You expect in the pooled regression model that for each firm the (in)dependent variable(s) do not change a lot through time.

d) Which standard OLS assumption will be violated? Also provide the EXACT solution to this problem? (2 pt)

A

The assumption that 𝐶𝑜𝑣(𝜖𝑖𝑡,𝜖𝑖,𝑡+1)≠0 (1 pt)

Exact solution is to use clustered standard errors, and cluster per firm. (1 pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

Suppose that the relationship between leverage and the independent variables changes drastically after 2008 (the Global Financial Crisis).

e) Write down the complete procedure to test this. That is: write the equation(s) to be estimated, the H(0) and H(A) of the test, and the corresponding test statistic. Be as complete as possible! (3 pt)

A

This is the Chow break test
Define 𝐷𝑡 as a dummy that equals 1 after 2008 and zero elsewhere.
The model now reads:

𝐿𝑖𝑡=𝛼+ 𝛽1ln𝑆𝑖𝑡+𝛽2ln(1+𝐴𝑔𝑒𝑖𝑡)+ 𝛽3𝑃𝑆𝑖𝑡+𝛽4𝑇𝑎𝑛𝑖𝑡+ 𝛽5𝑅𝐷𝑆𝑖𝑡+ 𝛽6ln𝑆𝑖𝑡𝐷𝑡+𝛽7ln(1+𝐴𝑔𝑒𝑖𝑡)𝐷𝑡+ 𝛽8𝑃𝑆𝑖𝑡𝐷𝑡+𝛽9𝑇𝑎𝑛𝑖𝑡𝐷𝑡+ 𝛽10𝑅𝐷𝑆𝑖𝑡𝐷𝑡+ 𝛽11𝐷𝑡+ 𝜖𝑖𝑡
Now perform an F-test
H0: 𝛽𝑖 =0 for i = 6,…, 11.
HA: at least one 𝛽𝑖 (i = 6,…11) is non-zero. (equation and H0: 2 pt)

Procedure:

1) estimate the pooled regression model of a), and store 𝑆𝑆𝑅0.
2) Estimate the model above, and store 𝑆𝑆𝑅1

3) The F-stat is now given by
F(𝑘1−𝑘0,𝑇−𝑘1)= 𝑇−𝑘0𝑘1−𝑘0(𝑆𝑆𝑅0−𝑆𝑆𝑅1) 𝑆𝑆𝑅1
With T = 250 * 49, 𝑘1 = 12 and 𝑘0=6 . (test stat 1pt)

Note: it is also ok if you omit 𝛽11𝐷𝑡.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

You regress leverage on the other variables (hence without any dummies) using a random effects specification, and obtain very different results from a fixed effects specification.

f) How do you interpret this result, and what will you do next? (2pt)

A

I conclude that the random effect might be correlated with the regressors and there might be an endogeneity concern. (1pt)

I could do a Hausman test to see whether this is actually the case, but for sure would trust the fixed effect result more than the random effects result. (1pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

a) What is the crucial difference between CONDITIONAL and UNCONDITIONAL volatility? (1 pt)

A

a) What is the crucial difference between CONDITIONAL and UNCONDITIONAL volatility? (1 pt)
Unconditional volatility means that you do not take into account any information, while conditional volatility means that you consider the volatility, given some information (usually we use information up to one period earlier in time). (1 pt)
Note: the CONDITIONING is important, NOT if it is fixed through time or not.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q
Question 4 (Volatility modeling; 10 pt)
Sanne would like to model the volatility of stock returns. She has daily stock returns (in percentages) of the Bank of America Inc. (BAC) from 2001 – 2014. In addition, she also has daily values of the VIX index. Figure 4.1 plots the daily returns.

Sanne hypothesizes that the VIX could be related to the volatility of BAC stock returns.
b) Explain why Sanne is possibly right? (1 pt)

A

The VIX is a forward looking indicator of the volatility of the SP500 Index. If the general stock market volatility goes up, the volatility of a big bank (BAC) will probably also go up. (1 pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

Andre does not believe in the effect of the VIX on the variance of BAC returns and estimates a simple GARCH(1,1) model, assuming a conditional Normal distribution for the returns. After estimating the parameters, he obtains the fitted variances 𝜎𝑡2̂.

d) Describe the full procedure how to test if the volatility model is correctly specified given the fitted volatilities? [Be explicit] (2 pt)

A

1) Compute the square of the standardized residuals 𝑢̂𝑡 with
𝑢̂𝑡=(𝑟𝑡−𝜇)/𝜎𝑡̂
(1 pt)

2) Check if there is any autocorrelation left in the squared values of 𝑢̂𝑡 with a Ljung Box test. (1 pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

e) Explain what is meant by an omitted variables bias?

A

It is a bias in the coefficients of the regressors included in the model due to a regressor that impacts BM but that is left out of the model.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

a) What does the 𝑅^2 of this regression tell you?

A

The % of explained variation of the dependent variable by the regression model

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

Unbiasedness of OLS

THEOREM (normality of OLS ^ ): If

A

Assumption 0a (correct specication)

Assumption 0b (no multi-collinearity)

Assumption 1 (zero mean errors)

Assumption 2 (homoskedasticity)

Assumption 3 (uncorrelated errors)

Assumption 4 (regressors not stochastic)

Assumption 5 (normality)

Violation 1: heteroskedasticity
Test: Engle LM test, LB on squared residuals (but also other possibilities)

Violation 2: autocorrelation
Test: Ljung-Box or Box-Pierce or others
Violation 2 (alt): non-normality, or even endogeneity (omitted variables)
Test (alt): Bera-Jarque, or adding additional risk factors

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

Looking at the results of your regression, you see that none of the coefficients has an absolute t-value
higher than 1.5.

c) Argue why you can/cannot conclude from this that the 5 risk factors jointly fail to adequately
describe the risk in “dedicated short bias” hedge fund returns.

A

You cannot conclude this. Testing for multiple restrictions at the same time requires an F-test
This is particularly important if there are multicollinearity issues. You may expect these to be absent here given the 5 risk factors.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

d) Using HAC standard errors, the significance of the individual coefficients drops further. Explain
how this drop in significance by the use of HAC standard errors may come about.

A

d)
If the regression residuals are large in absolute magnitude at the same time as the regressors are far from their mean. Alternatively, if errors are positively correlated over time.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
18
Q

You repeat your regression for the “event driven” hedge fund returns and obtain different
coefficients than for “dedicated short bias”.

e) Explain how you will test whether these differences are statistically significant
[hint: what alternative regression or what additional auxiliary regression is needed to do this]

A

e)
You can run a pooled regression for the two styles, store the SSR as SSR0. The sum of the SSR’s for the separate regressions is SSR1. Now you can make an F-test.
You can also make a joint regression model with a dummy interaction for the DSB style, and do an F-test. This amounts to doing a Chow break test.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
19
Q

Going back to the regression results for the “dedicated short bias”, you regress the “dedicated short bias” returns on a constant and on the market excess return 𝑀𝑀𝑡𝑡 only. You find a regression
coefficient for the market return of -0.51. You worry that you might have forgotten an explanatory
variable measuring the illiquidity climate of the market. You expect this liquidity variable to have a
direct negative impact both on the hedge fund return and the market return.

f) Explain intuitively why this omitted regressor may bias the coefficient estimate for 𝑀𝑀𝑡𝑡.
g) Argue what is the direction of the bias, i.e., why you expect the true coefficient to be more negative than −0.51, or why you expect it to be less negative (or even positive).

A

Part of the effect of the omitted illiquidity on HF return now runs through the included market return variable.

(mention: effect via market return variable,)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
20
Q

h) Argue why in the above regressions it would / would not be recommended to use robust standard errors rather than common standard errors given the data set (returns (panel) and risk factors (time series)) at hand.

A

Returns are often known to be heteroskedastic due to volatility clustering, so heteroscedasticity robust standard errors make sense to prevent flawed inference.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
21
Q

Suppose that you as a company has a portfolio that contains all 500 assets. You would like to compute the minimum required capital. You are given the EAD of each company, denoted by 𝐸𝐸𝐸𝐸𝐸𝐸 𝑖𝑖 . Assume a LGD of 50% and assume that the fitted probabilities of default from your Logit model are the true Basel based probabilities.

You would like to account for parameter uncertainty while computing the minimum required capital.
f) Provide an argument why it makes sense to account for this uncertainty? You may use the output from Table 1.1 (1 pt)

g) Describe the full procedure how to account for parameter uncertainty when calculating this minimum requirement ? (3 pt)

A

E.g: By neglecting the parameter uncertainty you don’t take into account that the precision of one certain estimated coefficient is much higher than the other. This will affect in the probability of default and in the end the distribution of the min req capital.

g) Describe the full procedure how to account for parameter uncertainty when calculating this minimum requirement ? (3 pt)

𝑀𝑀𝑀 𝐶𝐶𝑖𝑖= 𝑃𝑃𝐷𝐷𝑖𝑖×𝐸 𝐷𝐷𝑖𝑖×𝐿𝐿𝐿𝐿𝐿 𝑀𝑀𝑀 𝐶𝐶𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃=Σ𝑀𝑀𝑀 𝐶𝐶𝑖𝑖

The procedure reads

1) Estimate the logit model and store the parameter vector 𝛽𝛽̂ and the covariance matrix 𝑉 􀷠
2) Simulate 𝛽𝛽𝑗𝑗 from 𝑁𝑁(𝛽𝛽̂,𝑉 􀷠)
3) Compute 𝑃𝑃𝐷𝐷𝑖𝑖𝑗𝑗 for each company via 𝑃𝑃𝐷𝐷𝑖𝑖𝑗𝑗= exp (𝑥𝑥𝑖𝑖′𝛽𝛽𝑗𝑗)1+ exp (𝑥𝑥𝑖𝑖′𝛽𝛽𝑗𝑗)
4) Compute 𝑀𝑀𝑀 𝐶𝐶𝑖𝑖𝑗𝑗= 𝑃𝑃𝐷𝐷𝑖𝑖𝑗𝑗×𝐸 𝐷𝐷𝑖𝑖×𝐿𝐿𝐿𝐿𝐿
5) Compute 𝑀𝑀𝑀 𝐶𝐶𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑗𝑗=Σ 𝑀𝑀𝑀 𝐶𝐶𝑖𝑖𝑗𝑗
6) Repeat steps 2 until 5 N times (j = 1,… N)

In the end, we get N simulated MRC values of the portfolio.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
22
Q

The positive found relationship between ln(wage) and education can be a result of just correlation or a real causal effect.

c) Explain the difference between correlation and causality ? (1 pt)

A

Correlation is a statistical measure of dependency between two variables (without any reasoning/theory). (0.5 pt)

Causality means that y goes of because x changes due to a certain (economic) relationship between y and x. (0.5 pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
23
Q

The estimated coefficient of 𝛽𝛽1 could be biased due to endogeneity problems. For example, we do not include 𝐴 Ability𝑖 (since we can not measure this variable).

d) Explain why the above example could lead to a biased coefficient 𝛽𝛽1? (1 pt)

Rosy states that by using panel data, the aforementioned bias problem could vanish.
e) Is Rosy right? Argue WHY or WHY NOT? (1 pt)

A

Ability has an effect on education such that we omit a variable here. This could lead to an endogeneity problem. (correlation between X and the error term) (0.5 pt)
Since Ability is correlated with education, we indeed will get a biased beta, (0.5 pt)

Note: you should mention that ability is correlated with education!!!

Yes, because Ability𝑖 does not change over time! (0.5 pt)

Hence by using panel data with individual specific effects, this affect will be mopped up by individual specific intercept. (0.5 pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
24
Q

You are given panel data of 100 S&P 500 listed firms, which could be grouped into 10 different industries. You study a famous corporate finance subject whether the composition of the board influences the leverage of the firm. A theory from sociology says that men would like to take more risk than women at a board.
You run the following regression model:

EQ 3.1 𝐿𝐿𝐿𝐿𝑣𝑣𝑖𝑖𝑖 =𝛽𝛽0+𝛽𝛽1 𝐵𝐵𝐵 𝐷𝐷𝑖𝑖𝑖 + 𝑥𝑥𝑖𝑖𝑖 ′𝛽𝛽+ 𝜖𝜖𝑖𝑖𝑖
with t in years (1990 – 2017) and 𝐵𝐵𝐵 𝐷𝐷𝑖𝑖𝑖 the percentage of males at the board of company i at the end of year t, and 𝑥𝑥𝑖𝑖𝑖 ′ a row vector containing several control variables. Standard errors are computed using the normal OLS standard errors. You find a positive 𝛽𝛽1 of 0.50 with a standard error equal to 0.10.

a) Argue why the usual OLS standard errors are possibly wrong here? Also provide a possible solution? (2 pt)

A

OLS standard errors assume that there is no correlation between 𝜖𝜖𝑖𝑖,𝑡 𝑎 𝑎𝑎𝑎 𝜖𝜖𝑖𝑖,𝑡 +1. However, it could be that the Leverage values and/or BGD values of a certain company i does not vary that much over time, such that 𝜖𝜖𝑖𝑖,𝑡 and 𝜖𝜖𝑖𝑖,𝑡 +1 are correlated! (1 pt)

Hence one should use clustered standard errors, and then cluster on firm (1 pt)

ALSO OK: Heteroskedasticity (1 pt): however recall that robust standard errors (vce robust) does NOT solve the problem of correlation between 𝜖𝜖𝑖𝑖,𝑡 𝑎 𝑎𝑎𝑎 𝜖𝜖𝑖𝑖,𝑡 +1

Note: HAC/White is wrong here as we deal with panel data!

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
25
Q

You use the data until 2009 to estimate the parameters. The remaining data is used as out-of-sample observations. After generating 2-step ahead forecasts, you compute the associated forecast errors.

e) Do you expect that the MSPE of the AR(0) will be larger or lower than the MSPE of the ARMA(1,1) model? Explain your result! (1 pt)

A

Since there is a lot of autocorrelation, this means that the ARMA(1,1) will capture a part of it for sure such that the forecasts are much better than the forecasts of a model without NO autocorrelation (the AR-0). Hence the MSPE is expected to be lower of the ARMA(1,1) model. (1 pt)

Keyword: autocorrelation, that is not captured by AR(0)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
26
Q

You discover that there is another risk factor out there in the literature, namely stock return differentials between portfolios of low minus high liquidity stocks. You expect this factor to carry a positive risk premium. You expect the stock liquidity to be correlated with market capitalization (i.e., with size): smallcap stocks are less liquid on average.

g) Explain the direction of the omitted variables bias you expect when regressing 𝑟𝑟𝑖𝑖𝑖 on the four risk factors 𝑟𝑟𝑡𝑡𝑀𝑀,𝑆𝑆𝑆 𝑆𝑆𝑡𝑡,𝐻𝐻𝐻 𝐻𝐻𝑡𝑡,𝑈𝑈𝑈 𝑈𝑈𝑡𝑡, only, thus omitting the liquidity risk factor. (2pt)

A

You apparently expect a positive impact on the omitted risk factor, and a positive correlation between low liquidity and low market cap, so a positive correlation between SMB and the new risk factor. Therefore, SMB also captures part of the positive (1pt) impact of the omitted risk factor on the returns, and the SMB coefficient will be biased upwards (have a positive bias). (1pt)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
27
Q
Question 2 (6pt)
Your colleague uses a default indicator 𝑦𝑦𝑖𝑖 and regresses it directly on firm characteristics, and uses the fitted values to predict the probability of default.

a) Comment on the advantages and disadvantages of this approach. (1pt)

A

Advantage: easy interpretation (directly as prob.default), simple, and correct as long as the prediction is between zero and one.

Disadvantage: you cannot guarantee the predicted probabilities to be between zero and one.

28
Q

c) Explain why the difference between AIC of the logit and probit is the same as the difference in BIC between the two models. (1pt)

A

Both models have the same number of parameters. So only the difference in fit remains. This is the same for the two criteria.

29
Q

d) Mention (at least) two other ways to assess the adequacy and fit of a logit model? (1pt)

A

ROC curve, CAP curve, hit rates, pseudo-R2.

30
Q

Another colleague is interested in modeling the volatility of the S&P 500 Index returns. He has observations of 𝑅 𝑉𝑉𝑡𝑡, the realized variance which is defined as the sum of all squared 5 minute returns during a trading day. The data runs from 2001 until 2014.
You would like to estimate an ARMA(2,1) model on the RV series using Stata.

e) Argue why/why (NOT) this is feasible using OLS. If possible, also provide the Stata code to estimate the parameters (1 pt)

A

The lagged error term is not available, hence OLS can not be applied. (1 pt)

31
Q

Joyce has a panel data set of 700 companies observed over 10 years of annual data. She wants to relate
the CEO compensation level to firm characteristics such as size of the company (measured as log market
value), profitability of the company, and a social network variable relating the CEO to the board of other
companies.

To compute standard errors, Joyce can use HAC standard errors, White robust standard errors, or (time
and firm) clustered standard errors.

d) Which option should Joyce select and why?

A

HAC standard errors are for time series and stationary, non-constant time effects. Robust standard
errors are mainly for heteroskedasticity correction. Clustered standard errors combine robustness and
error components for time and firm effects. They are the most suitable for panel data.

32
Q

Assume a ban is introduced on algorithmic trading of mid‐cap and small‐cap stocks in an attempt to
increase market stability. To measure the effect of the ban, you measure key variables that reflect
market quality in your perception, such as trading volume, bid‐ask spread, and stock price volatility.
You do so before and after the introduction of the ban for a number of mid‐cap and small‐cap
stocks.

a) Provide a simple regression set‐up to test whether the ban affected any of your key variables in a
statistically significant way. Also provide your null‐hypothesis and the test you carry out. (1pt)
[note: you have multiple stocks, so both a cross‐section and a time index!]

b) Explain why the above approach and data set may be insufficient (or even problematic) to pin
down the causal effect of the implementation of the ban on market quality. (1pt)

A

We can use a simple t‐test for the hypothesis that the effect is absent, e.g., 𝑦𝑦𝑖𝑖𝑖𝑖= 𝜇𝜇𝑖𝑖+𝛾𝛾𝐷𝐷𝑡 +𝜖𝜖𝑖𝑖𝑖𝑖
where 𝐷𝐷𝑡 is a dummy that is 1 after the ban is effective, and 𝛾𝛾 measures the effect on the dependent
variable 𝑦𝑦𝑖𝑖𝑖𝑖, which measures spread, volume, volatility, or something else.
b)

It is a diff set‐up. There is no outside benchmarking. We cannot say whether the variables have not
gone down/up because of other reasons than the ban (other simultaneous events, changes in the
environment, etc.

33
Q

What is endogeneity problems and give Three important causes for endogeneity problems.

A

Meaning that covariance between Xi and ei is not equal to zero…which leads to biases in the parameters.

  1. Measurement errors
  2. Omitted variables
  3. Siumltaneity (not covered in this course)
34
Q

Difference between correlation and causality.

A

Causal: if you study 1 more year, then you het higher income (Direct relationship)

Correlation: individuals with higher income have chosen higher education (for whatever unobserved characterics/reasons)

35
Q

You build a regression model that relates next year’s firm performance (left‐hand side) to a number
of firm characteristics and to a variable indicating whether the family CEO has a family successor, or
an outside successor. Your friend claims that if the firm is expected to do badly next year, it is more
likely that an outside successor is going to be brought in.

e) Supposing your friend is right, explain why this would/would not create an endogeneity problem
for your regression.

f) In case of an endogeneity problem, you should find an instrumental variable. Explain the two conditions that should hold for such a variable in the context of the current setting.

A

If correct ,there would be a correlation between the residual and the explanatory variable, and thus
indeed an endogeneity problem.

f)
The instrumental variable should be related to whether there is a family successor, but at the same time not directly related to next year’s firm performance (only indirectly via the family successor
variable).

36
Q

Question 3

You are interested in the capital structure of non‐financial firms. You gather a panel data set of 1000
firms, each firm observed over the 10 year period 2005‐2014.
Your variables are Leverage (𝐿𝐿𝑖𝑖𝑖𝑖) as your dependent variable, growth opportunities (measured as
Tobin’s 𝑄𝑄𝑖𝑖𝑖𝑖), tax shields (𝑇𝑇𝑖𝑖𝑖𝑖), and collateral value of plant, inventories and equipment (𝐶𝐶𝑖𝑖𝑖𝑖).

a) Provide the specification of the pooled regression model using the above variables, where the
error term has an error components structure.

b) Provide the definition of the fixed effects and the random effects model, and discuss when you
would use one or the other.

A

a)
𝐿𝐿𝑖𝑖𝑖𝑖= 𝛽𝛽0+𝛽𝛽1𝑄𝑄𝑖𝑖𝑖𝑖+𝛽𝛽2𝑇𝑇𝑖𝑖𝑖𝑖+𝛽𝛽3𝐶𝐶𝑖𝑖𝑖𝑖+𝑒 𝑖𝑖𝑖𝑖 (no subscript on coefficients)
𝑒
𝑖𝑖𝑖𝑖=𝜇𝜇𝑖𝑖+𝜀𝜀𝑖𝑖𝑖𝑖. (also ok with time effects)

b)
Fixed effects: 𝜇𝜇𝑖𝑖 are fixed parameters, and 𝛽𝛽0 is dropped. Used if you have the entire universe of units, or if you are worried about inconsistency of the random effects in a random effects specification.
Random effects: see 1.a, with 𝜇𝜇𝑖𝑖 random. Use if you only observe a sample of units and want to claim results for the population; use if you want to include unit specific, time invariant regressors.

37
Q

d) Explain what are clustered standard errors and why you would use them.

A

d)
They are needed if you have either leftover time or cross unit correlation. This correlation among the error terms causes bias in the standard errors. The bias in the standard errors can be corrected by using clustered standard errors.

38
Q

For many of your firms, if you look at the Durbin‐Watson statistic based on the pooled regression
residuals, you find values close to 2. If you compute the cross‐sectional correlations of the pooled
regression residuals across all unit pairs (i,j) you ind significantly positive values for each of your
cross‐sections 2005 … 2014.

e) Argue how you would compute clustered standard errors: i.e., would you cluster over units or
over time?

A

e)
The DW of 2 for each unit suggests no firm specific effect. The significant correlation across units for each cross section suggests a time effect. So you would cluster over time.

39
Q

As an instrument, someone suggests to you the percentage of the female LinkedIn contacts of the
highest ranking male in the board.

d) Give your critical assessment of this instrument suggestion given the conditions an instrument should satisfy.

A
  1. needs to be correlated with the regressor; we can check this. Not unreasonable: if there are more
    female contacts, possibly also more likely that female contacts end up in the board.
  2. needs to have no direct impact on riskiness; harder to argue; males with many female contacts
    might have a different risk attitude for the same reasons that females might have a different risk
    attitude, and be therefore selected on the board of particular companies.
    Other lines of argument also possible (some argue against (1) being reasonable).
40
Q

e) Explain whether you would still have the same endogeneity issue if you would specify the
regression in changes, i.e., relating the increase in riskiness to the increase in female board participation from one year to the next.

A

The above problem would become less. All time‐invariant differences between riskness and/or
female board participation would be removed, and thus all these time‐invariant omitted variables
would no longer cause an omitted‐variables‐bias.

41
Q

Due to a new regulation, each board needs to have at least 2 female members.
f) Explain whether you use this new regulation in a diff‐in‐diff analysis to uncover the impact of gender composition on firm riskiness, or not?

A

f)

No. There is no treated and non‐treated group. All firms are treated.

42
Q

a) Using the same problem on estimating the determinants of risk aversion (𝑅𝑅𝐴𝐴𝑖𝑖,𝑡 ) as in Question 5 of the exercise set on Linear Regression, someone warns you that the dummies for the program type (𝑃𝑃𝑅𝑅𝑖𝑖𝑗𝑗) and last year’s wealth level of the student (𝑊𝑊𝑖𝑖,𝑡 −1) might be correlated. Explain whether or not this gives rise to an endogeneity concern in the regression?

A

No, this is a multicollinearity problem possibly, but not an endogeneity problem.

43
Q

b) Someone else warns you that you may also have forgotten a regressor, namely a dummy indicating whether the student’s great-grandparents were business people. She shows this variable is correlated with last year’s wealth level. Explain under what conditions you would now have an endogeneity issue? And when not.

A

Endogeneity issue when:
If the dummy has a direct effect on the student’s risk attitude, e.g., because a business people is more risk seeking for multiple generations.
No endogeneity issue when:
If the dummy has no direct impact on the students risk attitude, e.g., because the effect of reduced risk aversion of business people dies out after the next generation (or is not hereditary at all).

44
Q

c) Introduce a valid instrument for last year’s wealth and argue why it is a valid instrument.

A

Multiple answers possible, but you should have an argument why the instrument is correlated with the wealth level (relevant), and does not have a direct impact on risk averseness (exclusion restriction).
E.g., you could say IQ (correlated (arguably) with wealth), but (again arguably) not related to risk averseness.

45
Q

Question 1
Demi works as the data-scientist at the marketing department of Apple. Apple is interested in modeling and forecasting the daily number of sales of the iPhone 7. Demi has a daily times series of 1000 observations and aims to model sales using ARMA models. Assume that the time series is stationary.

a) What is THE necessary property of your time series such that using ARMA models makes sense?

A

a

The time series variable should contain significant autocorrelation.

46
Q

Sacha, a colleague of Demi, states that the table suggests to use the ARMA(1,1) model for ln(Sales), since the BIC and AIC values are the lowest across all AIC/BIC values listed in the table.

c) Explain why Sacha is / is not right ?

A

Anne is wrong, as the BIC/AIC values are only useful to compare between models if the dependent variable is the same!! Hence we cannot compare model (1) and (2) with model (3) and (4)

47
Q

a) Define the concept of (weak) stationarity. (1pt)

Tom types the following two STATA commands with subsequent output.

A

a
A time series 𝑦𝑡 is weakly stationary if 𝐸[𝑦𝑡]=𝜇 (not dependent on time) and 𝐸[𝑦𝑡 𝑦𝑡−𝑘]=𝛾𝑘 (not
dependent on time).

48
Q

What does endogeneity problems mean

A

Refers to situations where explanatory variable is related to the error term

49
Q
Consider Governance (G) regression for the top firms from 11 industries (N=11) over 10 years (T=10):
Leverage=a+bG+e

Do we really have N*T = 110 observations as precision?

A

No, dependent variable and regressor vary little over time, so more like 11 observations.
-> default standard errors are wrong.
Solution: use Clustered standard errors
Cluster over firms as there is probably a strong firm-specific component in the error term and the regressor.

50
Q

Suppose you want to see if the relationship between leverage and all independent variables is different after the 2008 GFC. How would you test this

A

Do a Chow break test: define dummys (D) that is equal to 1 after 2008.

Then perform an F test

51
Q

Parameters in a regression

A

equal to number of betas. beta=parameter

52
Q

When can we use LINEAR regression?

A

if the model is linear in the parameters.

It can still be non-linear in the variables.

Sometimes, you can transform a non-linear model to make it linear; this changes the interpretation of the coefficient

.Example:
transformation of a model: Cobb-Douglas production function
Yi = 0K1
i L2
i , take logs to obtain the linear (in the
parameters) specication yi = ~ 0 + 1ki + 2li , with ~ 0 a
constant (equal to ln 0), and ki = lnKi etc.
the coecient 2 is now an elasticity @ ln yi=@ ln xi rather
than the usual marginal eect @yi=@xi

53
Q

R squared measures:

A

The percentage of explained variation in the dependent variable by using the regression model.

Note: R squared never decrease if you add a variable

54
Q

Adjusted R squared measures:

A

Similar to normal R-squared it is the percentage of explained variation in the dependent variable by using the regression model. But in addtion it accounts for the number of variables.

The reason for this is that R^2 never decrease if you add a variable.

55
Q

Difference between AIC and BIC

They are tools for measuring model adequacy

A

BIC favors models with fewer parameters.

No clear guidance as to which model selection criterion to use, so check robustness of your results with respect to your choice.

56
Q

Three ways to deal with outliers:

A
  1. Dummy variable for each outlier observation
  2. Winsorizing
  3. Quantile regression or other forms of robust regression
57
Q

How to find outliers:

A
  1. Histograms of each variable
  2. Scatters of pairs of variables.
  3. Plot residuals
  4. DFBetas
58
Q

Which are the two desirable properties of the OLS estimator:

A

> Unbiasedness - estimated(fitted) betas should on average be a correct estimate of beta.

> Low variance - the estimated(fitted) should be a reliable estimator of beta. i.e it should have a low variance.

59
Q

for OLS estimator to be Unbiased it requires 4 assumptions to hold:

And for the inferential framework (to make predictions) to hold 3 additional:

A

Unbiasedness OLS if:

  1. Correct specification
  2. No multicollinearity
  3. Zero mean errors
  4. Regressors not stochastic

Inferential:

  1. Uncorrelated errors
  2. homoscedasticity
  3. normality

The OLS estimator is only exactly normally distributed if all 7 assumptions hold

60
Q

What does it mean that the OLS estimator is BLUE?

A

Best Linear Unbiased Estimator.

i.e. all other linear unbiased estimators have a ‘larger’ covariance matrix

61
Q

Can you omit the F-test (for example for b1=b2=0) if both the individual t-tests are highly insignificant?

A

NOO!!!!

This particularly holds if regressors are strongly correlated!!
This may easily happen in Corporate Finance if you use
several indicators for the same construct, e.g., company size
by number of employees (coecient 1) as well as sales
(coecient 2)

62
Q

Logit models (and many other models) are estimated by:

A

Maximum likelihood.

definition: For a given model and given set of parameters (betas) and a given data set, the log-likelihood function gives the logarithm of the probability of observing the data given the model and the parameters.

The more curved the likelihood function, the more precise the estimator.

63
Q

Why is it important to test a model out of sample after estimating it in-sample?

A
  • Prevents overfitting in-sample

- Checks robustness of model out-of-sample.

64
Q

a)Briefly describe the ROC curve as a means to check the adequacy of credit scoring models

A

It can be used to test of well a model is at predicting default. A Larger AUROC means that it is better a predicting default.

Therefore, larger AUROC -> better credit scoring model.

65
Q

Give 4 examples of financial application for the logit model

A

Default prediction

Pay dividend yes/no

Crisis prediction

Sell recommendation

Going to trade

66
Q

You want to explain why some firms pay a dividend, whereas others do not.
For this, you estimate a logit model (𝐷𝐷𝑖𝑖=1 if a dividend is paid, and zero otherwise).

a) Your friend argues that you could just regress 𝐷𝐷𝑖𝑖 on a set of explanatory variables. Explain why it
might be a better idea to use a logit model instead.

A

logit predicts a probability, and can never be outside [0,1]. Also, the suggested regression is unlikely
to meet the CLRM assumptions.

CLRM = Classical Linear regression model

67
Q

d) Using HAC standard errors, the significance of the individual coefficients drops further. Explain
how this drop in significance by the use of HAC standard errors may come about.

A

d)
If the regression residuals are large in absolute magnitude at the same time as the regressors are far from their mean. Alternatively, if errors are positively correlated over time.