Risk and Artificial Intelligence v1.0
Question 1
A financial planning and analysis manager is analyzing transaction-level operational loss data across multiple business units in a corporation. These data contain:
Irregularly shaped clusters of loss events driven by different operational processes.
Several outliers resulting from extreme loss events.
Which clustering technique is most appropriate for this analysis?
- A. Principal component analysis clustering
- B. Hierarchical clustering
- C. Density-based clustering (DBSCAN)
- D. K-means clustering
Answer : C
Question 2
A financial firm is using a reinforcement learning trading model. The team determines that the reward value should be updated at the end of each scenario (episode). They test two methods: the first computes the reward using simple summation, and the second computes the reward using summation of discounted values.
What learning method(s) are used in this case?
- A. Both methods are Monte Carlo.
- B. The first method is Temporal Difference, while the second one is Monte Carlo.
- C. The first method is Monte Carlo, while the second one is Temporal Difference.
- D. Both methods are Temporal Difference.
Answer : A
Question 3
An analyst is considering using ensemble methods to build a model for predicting corporate defaults.
Which of the following is a benefit of using ensemble methods?
- A. Using ensemble methods avoids underfitting.
- B. Using ensemble methods reduces individual model errors.
- C. Using ensemble methods enhances the interpretation of results.
- D. Using ensemble methods lowers computational costs.
Answer : B
Question 4
An analyst is training a model for classifying consumer loans into two categories: no-default (0) and default (1). The training data contains a mix of defaulted loans (1% of total loans) and non-defaulted loans (99% of total loans). The analyst determines that this is an unbalanced data set and is looking for solutions to address this issue.
Which approach would be appropriate to handle this situation?
- A. Randomly sample data to create a smaller data set for training purposes.
- B. Use a higher penalty for non-defaulted loans in the cost function used for fitting the model.
- C. Train the model with available data, since that reflects the actual default experience.
- D. Oversample the defaulted loans to increase its proportion in the training data.
Answer : D
Question 5
A team builds a predictive model that performs extremely well on the training dataset but shows noticeably worse accuracy on the validation dataset. The team decides to apply ridge regression (L2) regularization before retraining. After doing so, they observe that the validation accuracy improves.
Based on this scenario, which aspect of L2 regularization is likely playing a role in improving the model's performance?
- A. L2 regularization increases model flexibility by allowing larger coefficients.
- B. L2 regularization reduces model variance by penalizing large coefficients.
- C. L2 regularization changes irrelevant features’ coefficients to zero.
- D. L2 regularization eliminates multicollinearity by removing correlated features.
Answer : B
Question 6
A validation expert is examining whether a model is overfitting or underfitting the data used to train it.
Which of the following statements is most accurate regarding overfitting or underfitting?
- A. An underfitted model exhibits low bias and high variance in predictions.
- B. Overfitting occurs when a model is trained with too many features.
- C. Failure to include non-linear or interaction terms can lead to overfitting.
- D. Overfitting is less of an issue with machine learning models than with econometric models.
Answer : B
Question 7
A financial analyst is comparing different machine learning techniques.
Which of the following best illustrates a scenario where reinforcement learning would be more appropriate than other machine learning methods?
- A. When the system needs to infer logical rules from a set of axioms and apply them to deduce conclusions.
- B. When the system needs to learn an optimal sequence of actions in an environment.
- C. When the system needs to classify images into predefined categories based on a labeled dataset.
- D. When the system needs to identify clusters in data without any prior labeling or feedback.
Answer : B
Question 8
If the analyst believes that the coefficients on too many variables are reduced to 0, how could the parameters be changed to address this problem?
-
A.
- B. Increase the value of λ.
-
C.
- D. Decrease the value of λ.
Answer : D
Question 9
An analyst is using the agglomerative hierarchical clustering method to cluster data points A through F using single linkage. The current state of the distance matrix, after two steps in the clustering process, is shown below:
What would the next step in the clustering process be?
- A. Combine E and BC
- B. Combine AD and BC
- C. Combine E and F
- D. Combine B and C
Answer : C
Question 10
A data analyst implemented a deep learning technique to predict loan default by consumers using historical data of various consumer characteristics, and macro-economic variables.
Based on this information, what type of machine learning method is the analyst likely using?
- A. Regression random forest.
- B. Reinforcement learning.
- C. Principal components analysis.
- D. Neural network.
Answer : D
Question 11
An analyst finds that a decision tree model tends to perform worse than other machine learning methods. He decides to test if the random forests technique can improve the results.
Which of the following statements is correct regarding the random forest technique?
- A. The random forest technique allows for features with very strong predictive power to be ignored.
- B. The random forest technique is a special type of decision tree and is easily interpretable.
- C. The random forest technique is a boosting technique that combines the forecasts from many decision trees.
- D. The random forest technique can be used if the target is categorical but not continuous.
Answer : A
Question 12
A private wealth relationship manager wants to better understand clients’ views on some recently offered products. The manager sends out an anonymous questionnaire to all clients and the results include both labeled and unlabeled data.
In deciding between using self-training and co-training to analyze the data, which of the following statements about these two techniques is correct?
- A. Compared with self-training, co-training reduces the risk of overfitting.
- B. Both self-training and co-training require a minimum of 2 features.
- C. While self-training generally requires a clustering assumption, co-training does not.
- D. Self-training is a transductive method, but co-training is an inductive method.
Answer : A
Question 13
An analyst is selecting the activation function for a neural networks model.
Which of the following correctly represents a ReLU activation function?
-
A.
-
B.
-
C.
-
D.
Answer : D
Question 14
An analyst is considering the use of logistic regression for predicting consumer satisfaction level (Below expectations, Meets expectations, Exceeds expectations). Two quantitative variables reflecting behaviors (average rating, and discount applied) are available for use as independent variables.
Which of the following statements is correct regarding the use of logistic regression in this context?
- A. Logistic regression should be considered since there are no more than two independent variables.
- B. Logistic regression should be considered since this is a classification problem which provides a categorical outcome.
- C. Logistic regression should not be considered since both independent variables are continuous variables.
- D. Logistic regression should not be considered since satisfaction level is categorized into more than two groups.
Answer : B
Question 15
A data scientist at a university is developing a model to predict how students will perform in various courses based on their performance in previous courses. The faculty has asked that a high degree of explainability be part of any model developed for this purpose.
Which of the following statements correctly describes a technique that can be used to improve explainability?
- A. LIME (locally interpretable model-agnostic explanations) approximates complex models with simpler models locally.
- B. Principal Component Analysis reduces the dimensionality of the data by transforming the data into a set of orthogonal components.
- C. Shapley values can be used to scale the data used in the model to ensure the features are all contributing to the model’s prediction.
- D. LUCID (Locating Unfairness through Canonical Inverse Design) visualizes the activation patterns of different neural network layers.
Answer : A