Exam Feedback November 2018 Part 2 Exam Feedback

[firsova]

Clustering identifies outliers that do not have strong connections with the rest of the data. You can't predict PD with clustering only.

For example , you can use clustering to divide the applicants into 10 classes. But without any PD information about some applicants in each class. You will not know the PD of each class.

Well that depend on the details of the question which I do not remember. I might be wrong but I was thinking that the general idea is that regression is applied when you have a view about the characteristics of the data and you want to predict the dependant variable.
However the questions was about the data with unknown characteristics which makes clustering more reliable option to apply (in my humble opinion).

 

[highwayone]

Well that depend on the details of the question which I do not remember. I might be wrong but I was thinking that the general idea is that regression is applied when you have a view about the characteristics of the data and you want to predict the dependant variable.
However the questions was about the data with unknown characteristics which makes clustering more reliable option to apply (in my humble opinion).

This is how I interpret the question: A bank can't figure out PD only base on the application information. There should already have a model base on their existing customers to predict PD. And the application information is complex so CART may be the model the bank already have.

My interpretation may be wrong. But I want to emphasize that a bank can't figure out PD only base on the application information using clustering.

 

[shailthakker]

Was the binomial tree question with semi annual compounding?

 

[shailthakker]

As far as i remember the var came around 7 million and i marked C) C) Reject the VaR model...

yup me too, backtesting needs to be done 99%

 

[voichi]

Hello, from my part, I managed to finish but with no time left for checks.
Quant questions did not seem very difficult, many questions with hazard rates based on cds spreads, cumulative/ anual PDs, ES two or 3 questions,
incremental VaR, ho lee,
bond discounted value with 2 steps semianual comp, cva where anual pd had to be calculated,
raroc maximum interest to be paid on deposits to remain under target raroc,
nsfr - measures to take to meet nsfr limit if additional loans X are granted, var 99% model backtesting 95% ci to decide if model is good or bad (I got good model),
Liq exogenus adjusted var with variable spread,
information ration using IC and breath; alpha due to selection;
Qualitative questions quite difficult, heavy text, I remember 2 questions regarding unsmoothing procedure for iliquid assests, 2 questions related to usd vs ViX/ CIP deviation, biases affecting iliquid assets, fama french regression function with positive beta for hml and negative beta for smb (select hf type and impact of positive/neg beta could not decide on right answer after eliminating 2), quantile plots, volatility smile,
risk anomaly
contemporanus Vs lagged relationship between beta/ volatilities and returns, which var is highest betwen principal/ duration / cf mapping,
basel III var backtesting green vs yellow area,
mapping risk factors - FRA 6x12 mapped to 6 and 12 month 0 cupon bonds?,
ifrs9 Vs fasb,
risk culture lctm/ aig /sec ( i've chosen the one on lctm even though none seemed right to me),
validation (I guess correct answer was related to performing qualitative validation regarding model utilisation/ implementation in the sistem),
hedge fund risk sharing asimmetry,
repo vs rev-repo and investing/ selling bonds trading special vs general,
portfolio construction I ve also selected linear programming but maybe screening was correct answer. Don't know the answer for the one related to method to be used to estimate default on large sample many variables... i've chosen regression, other options were clustering, pca, and someth I ve never heard of

Am I the only one who had suggested 'pool insurance' for securitization CE? Waterfall and OC will be rendered useless - even for senior tranches - if loss defaults and LGD is expected to spike tremendously...

I answered pool insurance with same logic as yoy. It was mentioned also someting regarding increase in defauls and correllations I think.

 

[runal]

Looks like most people have found the exam quite challenging and can't really say for sure if they'll pass.
What do you think would be a cutoff score to pass ?
Inputs from previous takers would be much appreciated

 

[voichi]

As far as i remember the var came around 7 million and i marked C) C) Reject the VaR model...

You had to transform the VaR from 95% to 99% ci and 10 days, so it was higher than 9 million. This was my result.

 

[voichi]

Clustering identifies outliers that do not have strong connections with the rest of the data. You can't predict PD with clustering only.

For example , you can use clustering to divide the applicants into 10 classes. But without any PD information about some applicants in each class. You will not know the PD of each class.

I remember also the question asking for model for PD prediction using application data, big sample, many variables. I answered regression too. I don't remember the question referring to machine learning techniques.

 

[voichi]

I also opted for 'pool insurance', given that the question was clearly suggesting that defaults are going to increase and LGD will bevery very high. Also the question clearly asked for the "most" effective.. I believe pool insurance is the safest?

Same here. Given high LGD I thought OC wouldn't be effective, also the waterfall would not be very effective in case of high defaults and high correlations. I remember I specifically checked if smth was specified regarding internal vs external enhancements and there was no such specification as internal one.

 

[[email protected]]

You had to transform the VaR from 95% to 99% ci and 10 days, so it was higher than 9 million. This was my result.

I also converted 95% 1 day to 99% 10 day var and the qstn asked for Market Risk Charge (not VaR) if I remember correctly....

 

[matt6558]

Hello, from my part, I managed to finish but with no time left for checks.
Quant questions did not seem very difficult, many questions with hazard rates based on cds spreads, cumulative/ anual PDs, ES two or 3 questions,
incremental VaR, ho lee,
bond discounted value with 2 steps semianual comp, cva where anual pd had to be calculated,
raroc maximum interest to be paid on deposits to remain under target raroc,
nsfr - measures to take to meet nsfr limit if additional loans X are granted, var 99% model backtesting 95% ci to decide if model is good or bad (I got good model),
Liq exogenus adjusted var with variable spread,
information ration using IC and breath; alpha due to selection;
Qualitative questions quite difficult, heavy text, I remember 2 questions regarding unsmoothing procedure for iliquid assests, 2 questions related to usd vs ViX/ CIP deviation, biases affecting iliquid assets, fama french regression function with positive beta for hml and negative beta for smb (select hf type and impact of positive/neg beta could not decide on right answer after eliminating 2), quantile plots, volatility smile,
risk anomaly
contemporanus Vs lagged relationship between beta/ volatilities and returns, which var is highest betwen principal/ duration / cf mapping,
basel III var backtesting green vs yellow area,
mapping risk factors - FRA 6x12 mapped to 6 and 12 month 0 cupon bonds?,
ifrs9 Vs fasb,
risk culture lctm/ aig /sec ( i've chosen the one on lctm even though none seemed right to me),
validation (I guess correct answer was related to performing qualitative validation regarding model utilisation/ implementation in the sistem),
hedge fund risk sharing asimmetry,
repo vs rev-repo and investing/ selling bonds trading special vs general,
portfolio construction I ve also selected linear programming but maybe screening was correct answer. Don't know the answer for the one related to method to be used to estimate default on large sample many variables... i've chosen regression, other options were clustering, pca, and someth I ve never heard of

I answered pool insurance with same logic as yoy. It was mentioned also someting regarding increase in defauls and correllations I think.

wow good memory!!_ one quick question, I vaguely remember the "repo vs rev-repo and investing/ selling bonds trading special vs general" question, what was the question about? thank you!

 

[voichi]

I also converted 95% 1 day to 99% 10 day var and the qstn asked for Market Risk Charge (not VaR) if I remember correctly....

Indeed, MRC which is based on the 99% ci 10 days normal + st var, considering also max rule and multiplier. Maybe I was wrong on my calculations but I remember first geting somthing like 7 mil then adjusted to 99% ci 10 days and got higher than 9 mil.

 

[[email protected]]

Indeed, MRC which is based on the 99% ci 10 days normal + st var, considering also max rule and multiplier. Maybe I was wrong on my calculations but I remember first geting somthing like 7 mil then adjusted to 99% ci 10 days and got higher than 9 mil.

yes i also marked option B greater than 9 million.... option C was to reject at 95% confidence and option D was to accept at 99% confidence again if i remember it correctly...

 

[voichi]

wow good memory!!_ one quick question, I vaguely remember the "repo vs rev-repo and investing/ selling bonds trading special vs general" question, what was the question about? thank you!

I think it was about having some excess liquidity to place on short term and question was wether to enter a repo or rev-repo and whether to invest in bonds that trade special or ones at general rates.

 

[jkowalonek]

I think it was about having some excess liquidity to place on short term and question was wether to enter a repo or rev-repo and whether to invest in bonds that trade special or ones at general rates.[/QUOTE
I remember it was Commercial paper, and the answer was something about if you'd get a higher rate as General Collateral or Special.... I think...

 

[voichi]

Something like that, I don't remember exactly. General rate is higher than special rate thus better for the lender of cash and more expensive for the repo party who is borrowing money and placing the collateral.

 

[mamoo]

Ah...ok I see. I am not sure why I chose thinner tails then, beause I saw that the negative returns in the tails are much lower, the same as for the up tail. Then I might have been veery unattentive. thanks

i chose thin tails on both ends too.

 

[fxlprasetyo]

portfolio construction I ve also selected linear programming but maybe screening was correct answer.

The answer wouldn't be linear programming since, the option is "Linear Programming where transaction cost can't be incorporated explicitly", which is not true, unfortunately, which is my answer as well. Many answers required a very detail knowledge of the topic, as this was a tricky exam indeed.

 

[FreemanMI]

Does someone remember the result for the regression hedge...? Was it 380 mln or 105? As DV01 was half of the bond he wanted to hedge, i chose 380 mln

 

[jkowalonek]

Does someone remember the result for the regression hedge...? Was it 380 mln or 105? As DV01 was half of the bond he wanted to hedge, i chose 380 mln

I'm not sure what the answer was but I just did DV01 times the position / dv01 of the hedge... but that seemed to easy, but my answer was there.