Assuming time invariance and the Markov property, it is easy to calculate the transition matrix for any time period as P^n, where P is the given transition matrix for one period and n the number oftime periods that we need to compute the new transition matrix for.
However, when the new time period is less than the time period the matrix is available for, the only way to deriving a transition matrix for a partial period is to numerically calculate a matrix M such that M^n = P.
Therefore Choice 'b' is the correct answer. Taking cube roots of a matrix is not a possible operation, dividing by 3 gives a matrix meaningless in this context, and P x P x P will give us the transition matrix for 3 years, not 1/3rd of a year.

Recovery rates vary a great deal from year to year, and are difficult to predict. Therefore statement III is true.
Similarly, any attempt to predict these is hamstrung by a high standard error, which can be as high as the historical mean itself. The error does not cancel itself out due to the effect of the business cycle making the error directionally biased. Thus statement IV is false.
Statement II is true as these are all factors that make forecasting recovery rates for any credit risk model rather difficult. Statement I is false because recovery rates are difficult to predict and assumptions are not easy to make.

An 'extreme value' will be a value that will lie in the tails. We need to determine the distribution that will have the least weight in the tails so that the probability of exceeding this tail value is minimum across the given choices.
The t-distribution, a distribution with kurtosis > 3 and a normal mixture distribution are all distributions with tails fatter than that for a normal distribution. A normal distribution will have the 'thinnest' tails among the choices and therefore the lowest probability of exceeding a given tail event value.
A note about the t-distribution: Leptokurtic distributions (those that have kurtosis>3, ie kurtosis greater than that for a normal distribution) generally appear to have higher peaks on their PDF graphs. The t-distribution is flatter, and actually appears lower than a normal distribution, which may make one think that it has a lower kurtosis and therefore should have thinner tails than a normal distribution. But that is not so, and the "visual" inspection test fails for inferring the kurtosis from just looking a the shape of the distribution. The kurtosis of a t-distribution is given by the formula {3 + 6/(d - 4)}, where d is the degrees of freedom and d > 4. Therefore the kurtosis of a t-distribution is always greater than 3 as "6/(d-4)" will always be a positive number being added to 3. Therefore there is no conflict between a t-distribution having fatter tails than a normal distribution as it has a higher kurtosis, even though it appears 'lower' on a graph when superimposed with a normal distribution.

The credit migration approach to assessing portfolio credit risk involves obtaining a distribution of future portfolio values from the ratings migration matrix. First, the frequencies in the matrix are used as probabilities, and expected future values of the securities belonging to each rating category are calculated.
These are then discounted to the present using the discount rate appropriate to the 'future' rating category. This gives us a forward distribution of the value of each security in the portfolio. These are then combined using the default correlations between the issuers. The default correlation between the issuers is often proxied using asset returns, and recognizing that default occurs when asset values fall below a certain threshold. A distribution for the future value of the portfolio is generated using simulation, and from this distribution the Credit VaR can be calculated.
Thus, we need the migration matrix, the risk horizon from which the present values need to be calculated, and the forward yield curve or the discount curve for each rating category for the risk horizon. Thus, Choice 'd' is the correct answer.

For operational risk modeling, both frequency and severity distributions need to be modeled. Modeling severity involves finding an analytical distribution, such as log-normal or other that approximates the distribution best represented by known data - whether from the internal loss database, the external loss database or scenario data. A 'risk functional' is a measure of the deviation of the model distribution from the risk's actual severity distribution. It assigns a penalty value for the deviation, using a statistical measure, such as the KS distance (Kolmogorov-Smirnov distance).
The problem of finding the right distribution then becomes the problem of optimizing the risk functional. For example, if F is the model distribution, and G is the actual, or empirical severity distribution, and we are using the KS test, then the Risk Functional R is defined as follows:
A diagram of a function Description automatically generated with medium confidence

Note that supx stands for 'supremum', which is a more technical way of saying 'maximum'. In other words, we are calculating the maximum absolute KS distance between the two distributions. (Note that the KS distance is the max of the distance between identical percentiles of the two distributions using the CDFs of the two.) Once the risk functional is identified, we can minimize it to determine the best fitting distribution for severity.