CCP Credit Scoring & Probability of Default 2 — Questions and Answers
Question 1: Which statistical technique is most commonly used to build logistic regression-based credit scorecards?
- Ordinary least squares regression
- Binary logistic regression (Correct answer)
- Principal component analysis
- K-means clustering
Correct answer: Binary logistic regression
Binary logistic regression models the probability of a binary outcome (default/non-default), making it the standard method for scorecard development.
Question 2: A credit model's Gini coefficient is 0.72. How does this compare to a random model?
- It performs 72% better than perfect prediction
- It is 72% worse than a random model
- It significantly outperforms a random model, which has a Gini of 0 (Correct answer)
- It equals a perfect model since Gini cannot exceed 0.72
Correct answer: It significantly outperforms a random model, which has a Gini of 0
A random model has a Gini of 0, so a Gini of 0.72 indicates strong discriminatory power well above random chance.
Question 3: What does the term 'scorecard points-to-double-the-odds' (PDO) represent?
- The score at which a borrower's income doubles
- The number of score points needed to double the odds of being a good credit risk (Correct answer)
- The probability that a doubling of score increases default risk
- The cut-off score for approving twice the normal volume
Correct answer: The number of score points needed to double the odds of being a good credit risk
PDO defines the score increment required to double the odds ratio of good-to-bad, calibrating the scorecard's scaling.
Question 4: Which variable transformation is typically applied to predictors before building a logistic scorecard to measure predictive power?
- Z-score normalization
- Weight of Evidence (WoE) transformation (Correct answer)
- Box-Cox transformation
- Min-max scaling
Correct answer: Weight of Evidence (WoE) transformation
WoE transformation converts categorical or binned continuous variables into log-odds format, enabling straightforward use in logistic regression scorecards.
Question 5: A lender observes that its scorecard's KS statistic has dropped from 42 to 28 over two years. What does this most likely indicate?
- Improved model discrimination due to better data quality
- Model drift — the scorecard no longer discriminates defaults as well (Correct answer)
- The population of applicants doubled in size
- The approval rate has increased by 14 percentage points
Correct answer: Model drift — the scorecard no longer discriminates defaults as well
A declining KS statistic signals model drift, meaning the scorecard has lost predictive power, likely due to population or economic shifts.
Question 6: In the context of credit scoring, what is 'reject inference'?
- Removing rejected variables from a model
- Estimating the performance of declined applicants to correct sample bias (Correct answer)
- Inferring a borrower's intent to repay from rejection letters
- Adjusting scores downward for borrowers who previously rejected credit
Correct answer: Estimating the performance of declined applicants to correct sample bias
Reject inference techniques attempt to estimate what the outcome would have been for declined applicants, correcting the sample selection bias in scorecard development.
Question 7: Which of the following best describes 'through-the-cycle' (TTC) probability of default?
- PD measured only at the peak of the economic cycle
- PD averaged across a full economic cycle to smooth out cyclical fluctuations (Correct answer)
- PD that increases monotonically through a borrower's credit lifecycle
- PD estimated only for revolving credit facilities
Correct answer: PD averaged across a full economic cycle to smooth out cyclical fluctuations
TTC PD averages default rates across an entire economic cycle, providing a more stable estimate used in Basel capital calculations.
CCP Credit Scoring & Probability of Default 2
Which statistical technique is most commonly used to build logistic regression-based credit scorecards?