Data

Question 1

What is a data dictionary and what is its purpose

Answer 1

1) A map of data assets where data is specified including the required metadata
2) Valuable to keep track of data, and the effort of making one is minimized if data editing is organized from the beginning

Question 2

Data element

Answer 2

Also called data field. An aspect of an individual or object that can take on varying values among individuals. Every piece of info in a database is a measurement of a data element.

Question 3

Heteroscedasticity

Answer 3

Describes data that does not have constant variance. (The variability of a variable is not consistent across the values of another variable, like time.) In a residual plot, this looks like a fan or cone shape

Question 4

Metadata

Answer 4

Descriptions of the fields in the database and their permissible values, as well as how they are created and limitations on their use, if known.

Question 5

Redundancy

Answer 5

A technique for obtaining high quality data. Ask for the same or similar information at least twice to reduce risk of errors and inaccuracy.
Ex: ask for email twice to make sure it is typed correctly
Ex: ask for age and date of birth

Question 6

When to implement redundancy

Answer 6

Only when a virtually error free result is required. Otherwise, the cost might outweigh the benefit

Question 7

Skewness. Positive vs negative

Answer 7

Describes a distribution’s departure from symmetry. Negative skew means the left-hand tail is longer. Positive skew means the right-hand tail is longer.

Question 8

Why is skewness important?

Answer 8

• Variance tends to understate the likelihood of loss if the distribution is skewed but assumed to be symmetric.
• We must consider skew to avoid taking higher-than-anticipated levels of risk and rejecting projects with understated likelihood of profits.

Question 9

Stationary

Answer 9

If a distribution is stationary, it means the parameters are stable over time

Question 10

TVaR

Answer 10

Tail value at risk. The expected loss given that the loss falls in the worst (1-alpha) part of the distribution

Question 11

Pros and cons of TVaR

Answer 11

• Pro: coherent risk measure
• Pro: describes the full tail of the distribution
• Con: difficult to calculate

Question 12

VaR

Answer 12

• Value at Risk
• The maximum loss that could occur with a specified probability over a given time horizon for a given distribution

Question 13

Portfolio VaR

Answer 13

The VaR of the entire portfolio

Question 14

Individual VaR

Answer 14

The VaR of one asset in the portfolio in isolation

Question 15

Diversified VaR

Answer 15

The portfolio VaR, taking into account diversification benefits

Question 16

Undiversified VaR

Answer 16

The sum of the individual VaRs in the portfolio when there is no short position and all correlations are unity

Question 17

Marginal VaR

Answer 17

The VaR that would be added for a unit increase in the investment in a particular asset

Question 18

Incremental VaR

Answer 18

The VaR that would be added to the portfolio VaR if the given investment adjustments were made to the portfolio

Question 19

Component VaR

Answer 19

A partition of the portfolio VaR that indicates how much the portfolio VaR would change (approximately) if the given asset was deleted from the portfolio

Question 20

Purpose of VaR

Answer 20

1. Identify the component (asset or BU or risk) that contributes most to the total risk
2. Pick the best hedges
3. Rank trades
4. Select the asset/project/BU that provides the best risk-return tradeoff

Question 21

How to calculate VaR

Answer 21

1. Empirical: the worst (1-alpha)% of results in the sample data
2. Parametric: assume that the data follows a statistical distribution and use that distribution to calculate VaR
3. Stochastic: apply the empirical method to thousands of simulations

Question 22

Pros and cons of VaR

Answer 22

• Pro: easy to understand
• Con: not a coherent risk measure
• Con: doesn’t describe the tail of the distribution

Question 23

Data quality

Answer 23

Refers to data’s “fitness for use.” The ability to fulfill the requirements of intended usage of data in a specific situation

Question 24

Why is data quality important?

Answer 24

High data quality can be a competitive advantage. Poor data quality can:
1. reduce customer satisfaction
2. reduce employee satisfaction (causing high turnover)
3. breed organizational mistrust
4. make it difficult or impossible to accurately determine the financial position of the business
5. make it difficult or impossible to calculate premium income and reserve required
6. waste time and resources investigating and fixing data issues

Question 25

Properties of high quality data

Answer 25

1. Relevance 2. Accuracy 3. Timeliness 4. Accessibility and clarity of results 5. Comparability 6. Coherence 7. Completeness

Question 26

How to obtain high quality data

Answer 26

1. Prevention: keep bad data out (most) 2. Detection: look for bad data already entered 3. Repair: let the bad data find you and fix things (least)

Question 27

How to prevent bad data from entering the dataset

Answer 27

- Edit data before it enters the database to prevent issues instead of fix them - Encourage staff to improve the data management process - Improve the data collection instrument - Improve the data collection method - Build in Redundancy

Question 28

How to look for bad data already in the dataset

Answer 28

- Deterministic tests - Probabilistic tests to detect outliers - Exploratory data analysis - Frequency counts - Two-way tabulations - Record Linkage

Question 29

Deterministic tests

Answer 29

- Range Test - If-Then Test - Ratio Control Test - Zero Control Test - Internal Consistency Test

Question 30

How to maintain high quality data

Answer 30

Data quality should: 1) be a regular item on the Management Board’s agenda 2) receive ongoing priority attention within the organization 3) be a structural component of operational management 4) be applied to the reporting and operational processes

Question 31

Pros and cons of standard deviation as a risk measure

Answer 31

Pro: easy to understand Con: not a coherent risk measure Con: doesn't describe the tail of the distribution well Con: considers upside and downside risk equal Con: underestimates risk if the underlying distribution is leptokurtic (thicker tails than normal dist)

Question 32

Correlation

Answer 32

The relationship between variables (like risks)

Question 33

Types of correlation measures

Answer 33

1) Pearson's Rho 2) Spearman's Rho 3) Kendall's Tau 4) Tail Correlation

Question 34

Pearson's Rho

Answer 34

The linear correlation coefficient. A value of 0 may not mean 0 correlation, only 0 linear correlation.

Question 35

Spearman's Rho

Answer 35

- Equal to Pearson's Rho in the uniform distribution. - Only the order of observations matters. Spearman's Rho is independent of the statistical distribution.

Question 36

Kendall's Tau

Answer 36

Calculated by comparing pairs of data points. If X and Y both in/decrease from one data point to another, the observations are concordant. If not, they are discordant.

Question 37

Tail correlation

Answer 37

Correlation between 2 variables may not be constant for all values. We can measure the correlation of the tails.

Question 38

How can you choose which correlation measure to use?

Answer 38

- If the variables are numeric and you care about the linear relationship specifically, choose Pearson's. - If any variables are ordinal, choose Spearman's or Kendall's. - If you care about the degree of deviation between data points, choose Spearman's. If not, Kendall's.

Question 39

Kurtosis

Answer 39

- Describes the degree of flatness of a distribution - An indication of the likelihood of extreme observations relative to those that would be expected with the normal distribution

Question 40

Types of kurtosis

Answer 40

- Mesokurtic distribution: normal distribution with kurtosis of 3 - Platykurtic distribution: thinner tails than the normal distribution, so kurtosis less than 3 - Leptokurtic distribution: thicker tails than the normal distribution, so kurtosis greater than 3

Question 41

Why is leptokurtosis important to look out for when quantifying risk?

Answer 41

It means the distribution has thicker tails than the normal distribution. If it is present and not properly accounted for, then the probability of extreme events will be underestimated.

Question 42

Spread

Answer 42

Often referred to as the difference between the returns on 2 assets

Question 43

Comonotonic

Answer 43

- 2 risks are comonotonic if one can be expressed as an increasing deterministic function of the other - If X lies at its q-quantile, Y will also be at its q-quantile - More than 2 risks can be comontonic

Question 44

Countermonotonic

Answer 44

- 2 risks are countermonotonic if one can be expressed as a strictly decreasing function of the other - If X lies at its q-quantile, Y will be at its (1-q)-quantile - More than 2 risks cannot be countermonotonic

Question 45

Credibility

Answer 45

For full credibility, there needs to be a sufficient number of observations. Technically, true full credibility can never be achieved. But we say full credibility exists at a given confidence level for a given distance from the expected value. There are 3 types of credibility: classic, Buhlmann, and Bayesian.

Question 46

Expected Shortfall

Answer 46

Sometimes considered the same as TVaR. While TVaR is the average value in the tail, Expected shortfall is the probability of loss multiplied by the expected loss given that a loss has occurred.

Question 47

Pros and cons of expected shortfall

Answer 47

It has many of the same benefits as TVaR, but it has little intuitive meaning.

Question 48

How to calculate expected shortfall

Answer 48

1. Empirical: sum the losses in the tail and divide by the total number of observations 2. Parametric: (1-alpha)xTVaR 3. Stochastic: apply the empirical approach to the output from the stochastic model

Question 49

Probability of Ruin

Answer 49

It is the probability that a given extreme loss will occur. The reciprocal of VaR because VaR is the max loss for a given level of confidence.

Question 50

What's wrong with probability of ruin?

Answer 50

It generally has the same limitations as VaR, and the assessment of loss if it occurs is not usually a priority. We use VaR and TVaR to determine the loss for which capital is required.

Question 51

Seasonality

Answer 51

Values are generally higher than an underlying trend at some points in the period and below it in others. To handle this, we can use an ARIMA model because seasonality is essentially an autoregressive process.

Question 52

Types of dependency

Answer 52

1) Immediate: direct immediate causal relationship 2) Time-lagged: delayed causal relationship 3) Feedback: variables interact with each other over time 4) Phase-shift: one variable affects another only after a change has reached a threshold

Question 53

Why is dependency important

Answer 53

Experience has shown that in many situations, dependencies in stressed situations are different than they are under normal situations. Ex: Low interest rates, high credit risk, and widespread panic among investors (a financial crisis) can happen in many countries at the same time and quickly deplete a firm's capital

Question 54

Ways to quantify dependency

Answer 54

1) Correlation matrix approach 2) Copula approach 3) Structured scenario approach 4) Multivariate distribution approach

Question 55

Correlation matrix approach to dependency

Answer 55

Assume linear correlation. Use a correlation matrix or a collection of matrices. Different matrices can be used for different percentiles to reflect higher correlation in tail events.