Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit

Examination Paper for STAT0011

2020

Question 1 [8 marks]

A data analytics company wishes to estimate the proportion of voters that will vote for candidate A for an elective office in a particular country. Let ✓ denote the proportion of voters that will vote for candidate A. In a survey before an election, a sample was taken of 200 potential voters, of which 130 said they would vote for Candidate A. The goal is to estimate ✓ based on this sample information.

(a) Let Y denote the number of people who will vote for candidate A.  Show that

Beta(↵ , β) distribution is the conjugate prior’ for the likelihood function.          [2]

(b) Using the Beta(↵ , β) prior from part (a), show that the posterior distribution of ✓

given the data is Beta(↵ +130, β + 70). You do not need to evaluate any integrals or normalising constants.                                                                                         [2]

(c) The company decides to model their prior beliefs about ✓ using a Beta distribution with a mean of 0.6 and standard deviation 0.2. Write down the parameters ↵ and β of the Beta(↵ , β) distribution which represent this belief.                                  [2]

(d) Write down the posterior distribution for ✓ using the above survey results and the prior from part (c).                                                                                                  [2]

 

Question 2 [20 marks]

(a) Let Y Bernoulli(✓). It is of interest to estimate ✓ using the loss function L(✓ , ✓) = (✓ − ✓)2 . Let the estimator be as follows:

 =  +

Compute the risk function, and find values of ↵ and β for which ✓ is minimax. [8]

(b) Let X N(2, σ 2 ). It is of interest to estimate σ2 using the loss function L(σ2 , σ 2 ) = (σ2  − σ2 )2 . Let the estimator be as follows: σ 2  =  P (Xi − 2)2 . Compute the risk function.                                                            [7]

(c) A car manufacturer is faced with the following two actions: a1  - advertise new model in various media, or a2 - do not advertise. The e↵ectiveness of advertisement depends on the state of the economy: ✓ 1   is “bad”, ✓2   is “moderate”, and ✓3   is “good”.

The losses corresponding to each action ai , i = 1, 2, and type ✓j , j = 1, 2, 3, are represented by the following loss matrix :

 

1

2

 1

1

0

2

0.5

0.5

3

0

1

Find the minimax randomized action.                                      [5]

 

Question 3 [10 marks]

Let Yt  = ln(Pt )− ln(Pt1 ) denote the daily log-return on a nancial asset, where Pt  is the opening daily price at time t.

(a) The following model has been proposed to describe the behaviour of log-returns:

Yt  = ut

ut  = σt"t

" N(0, 1)

σt(2)  = ! + ↵ut(2)− 1 + γut(2)− 1 I[ut1 <0] + βσt(2)− 1

where

I[ut1 <0]  = 0(1)

if ut1  < 0

if ut1  ≥ 0

Derive the unconditional variance of Yt  clearly showing all your steps.                 [6]

(b)  [Type] Now the following model has been proposed for log-returns Yt :

Yt  = ut

ut  = σt"t

" N(0, 1)

σt(2)  = ! + σt(2)1 + (ut(2)1 − σt(2)1 )

where ! > 0 and 0 < ↵ < 1.  Write down the name of this model and explain in your own words using no more than one A4 page what are the implications of this

model on the conditional and unconditional variance of Yt .                                   [4]

 

Question 4 [14 marks]

(a)  Consider the following information on the hypothetical portfolio of £10,000 invested

in two assets.  The information on each daily asset return is provided in the table below. It is assumed that these returns are jointly normally distributed.

Asset 1   Asset 2

Mean

Standard deviation

Portfolio weights

0.01

0.09

0 7

0.05

0.2

0 3

Portfolio value            Correlation coecient

£10,000

0.1

Standard normal distribution table.

z          -2.326   -2.054   -1.881   -1.751   -1.645   -1.555

Φ(z)     0.01      0.02      0.03      0.04      0.05      0.06

(i)  Compute the 99% 1-day Value-at-Risk (VaR) of the portfolio in value terms. Interpret your ndings.                                                                                    [4]

(ii)  [Type] It is well known that VaR is not a coherent risk measure because it

does not always satisfy the sub- additivity property. Explain in your own words using no more than one A4 page what is meant by the sub- additivity property, and why it is important in risk management.                                                 [4]

(b)  Consider ABC Insurance company that wishes to measure its risk exposure. Let Lt

denote independent and identically distributed annual aggregate losses that follow the exponential distribution with mean 10, i.e.  E(Lt ) = 10.  Note that the losses are positive.

Compute the Expected Shortfall (ES0.99 ) which is defined as follows:

 

ES0.99  = E [Lt |Lt  ≥ VaR0.99]

[6]

 

Question 5 [18 marks]

(a) Let Pt  denote the daily stock price for XYZ company at time t, which follows a

log-normal distribution.  A historical record of the prices over the past 8 days is available, leading to the following independent and identically distributed daily log- returns, Yt  = ln(Pt ) − ln(Pt1 ):

Y2  = −2.83,Y3  = −2.13,Y4  = 1.17,Y5  = −2.61,Y6  = −1.08,Y7  = 3.90,Y8  = 0.11

(i)  [Type] Describe in your own words using no more than one A4 page how Extreme Value Theory (EVT) can be used to estimate Value-at-Risk (VaR) using historical data on log-returns Yt .                                                           [5]

(ii)  [Type] Using no more than one A4 page, discuss advantages and drawbacks

of using EVT in this context.                                                                          [6]

(b)  [Type] Now consider EFG company whose daily log-returns Xt  can be described

as follows:

Xt  = µt + ut

where

P

µt  = c +X δiXti

i=1

ut  = σt"t           " F"      E("t ) = 0, Var("t ) = 1   8t

F"  is unknown.

S                             O                                               Q

σt(2)  = ! +X sut(2)s +X γout(2)oI[uto<0] +X βσt(2)q

s=1                        o=1                                          q=1

where P = 6, S = 3, O = 2, and Q = 3.  Suppose a historical record over the past 500 days is available. Describe in your own words using no more than one A4 page how EVT can be used to estimate 99% 1-day VaR for log-returns Xt . [7]

 

Question 6 [30 marks]

An insurance company is interested in estimating the frequency of claims in order to calculate an appropriate insurance premium. Let Xt  be a discrete random variable which represents the number of claims received in year t . The company considers a Poisson model to estimate the frequency of claims, i.e. X Poisson(λ).

After analysing the company’s insurance portfolio, the following historical record was obtained which shows the number of claims received over the last 10 years:

Year, t

1

2

3

4

5

6

7

8

9

10

Number of claims

1

3

1

0

1

1

2

2

3

2

The company knows that there has been a structural change in the frequency of claims at some unknown point in time ⌧ . This implies that the only change point has occurred at ⌧  so that observations X+1 , . . . ,X10   still come from a Poisson distribution but with a di↵erent parameter λ, i.e.:

Xt   

(a) Using a Gamma(1, 1) conjugate prior, derive the marginal likelihood of the

observations up to and including the change point p(x1 , . . . ,x|⌧ ).        [1]

(b) Similarly, derive the marginal likelihood of the observations to the right of the

change point p(x+1 , . . . ,x10 |⌧ ).                                       [1]

(c) Now the company has obtained further information regarding the possible location of the change point ⌧ . For instance, it is known that the change point can only occur at odd-numbered time points, i.e. ⌧  = k, where k  2 Ao    = {1, 3, 5, 7, 9}.  Furthermore, company’s prior beliefs have resulted in the following prior for ⌧ : p(⌧   = 1) = p(⌧   = 3) = p(⌧   = 5) =  and p(⌧  = 7) = p(⌧  = 9) =  . Considering that only one change point has oc- curred and using a conjugate Gamma(1, 1) prior for λ 1  and λ2 , compute pos- terior probabilities for ⌧  at all possible time points, i.e. p(⌧  = k|x1 , . . . ,x10 ), 8k 2 Ao . You may use results from (a) and (b).                       [18]

(d) Using posterior distribution p(⌧ = k|x1 , . . . ,x10 ) from (c), derive the posterior predictive distribution p(|x1 , . . . ,x10 ) based on available data.           [7]

(e) Assume that no further change points will take place in the future. Based on the posterior predictive distribution derived in (d), compute the probability of no claims being made in a single year in the future.                     [3]