NEWEST Predictive Analytics (PA) Exam |
Society of Actuaries (SOA) | ULTIMATE
EXAM WITH CORRECT ANSWERS AND
RATIONALES FOR CERTIFICATION
SUCCESS
1. In preparing data for a predictive model, a variable
with 30% missing values is best handled by:
A) Deleting all rows with missing values
B) Imputing the mean
C) Creating a separate missing indicator and
imputing a placeholder
D) Using a model that handles missingness natively
(e.g., tree-based)
E) Both C and D are valid approaches
Correct answer: E
Rationale: For high missingness, indicator +
imputation or native handling are appropriate.
2. A dataset has 10,000 rows and 500 columns. The
target variable is binary with 5% event rate. Which
resampling method is most appropriate for training?
,A) Random under-sampling of the majority class
B) Random over-sampling of the minority class
C) SMOTE (Synthetic Minority Over-sampling
Technique)
D) No resampling needed
E) Both A and C may be considered
Correct answer: E
Rationale: Imbalanced classification benefits from
resampling; SMOTE and under-sampling are
common.
3. In a decision tree, the cost complexity parameter
(α) in pruning controls:
A) The maximum depth of the tree
B) The minimum number of samples in a leaf
C) The trade-off between tree size and fit
D) The splitting criterion
E) The number of features considered at each split
Correct answer: C
Rationale: α penalizes tree size; larger α gives
smaller trees.
,4. A confusion matrix shows 100 true positives, 50
false negatives, 10 false positives, and 200 true
negatives. The precision is:
A) 100/150 = 0.667
B) 100/110 = 0.909
C) 100/300 = 0.333
D) 200/210 = 0.952
E) 100/200 = 0.5
Correct answer: B
Rationale: Precision = TP / (TP + FP) =
100/(100+10)=0.909.
5. The primary difference between bagging and
boosting is that:
A) Bagging builds models sequentially; boosting
builds in parallel
B) Boosting builds models sequentially, each
focusing on previous errors
C) Bagging uses all features; boosting uses random
subsets
D) Bagging reduces bias; boosting reduces variance
E) There is no difference
Correct answer: B
, Rationale: Boosting is sequential, weighting
misclassified observations.
6. Which of the following is NOT an appropriate
evaluation metric for a regression model predicting
insurance claim severity?
A) Mean absolute error (MAE)
B) Root mean squared error (RMSE)
C) Area under the ROC curve (AUC)
D) Mean absolute percentage error (MAPE)
E) R-squared
Correct answer: C
Rationale: AUC is for classification, not regression.
7. The "curse of dimensionality" refers to:
A) The difficulty of interpreting high-dimensional
models
B) The exponential increase in data needed to
maintain statistical significance as dimensions
increase
C) The increased computational time for high-
dimensional models
D) The tendency for overfitting in high dimensions
Society of Actuaries (SOA) | ULTIMATE
EXAM WITH CORRECT ANSWERS AND
RATIONALES FOR CERTIFICATION
SUCCESS
1. In preparing data for a predictive model, a variable
with 30% missing values is best handled by:
A) Deleting all rows with missing values
B) Imputing the mean
C) Creating a separate missing indicator and
imputing a placeholder
D) Using a model that handles missingness natively
(e.g., tree-based)
E) Both C and D are valid approaches
Correct answer: E
Rationale: For high missingness, indicator +
imputation or native handling are appropriate.
2. A dataset has 10,000 rows and 500 columns. The
target variable is binary with 5% event rate. Which
resampling method is most appropriate for training?
,A) Random under-sampling of the majority class
B) Random over-sampling of the minority class
C) SMOTE (Synthetic Minority Over-sampling
Technique)
D) No resampling needed
E) Both A and C may be considered
Correct answer: E
Rationale: Imbalanced classification benefits from
resampling; SMOTE and under-sampling are
common.
3. In a decision tree, the cost complexity parameter
(α) in pruning controls:
A) The maximum depth of the tree
B) The minimum number of samples in a leaf
C) The trade-off between tree size and fit
D) The splitting criterion
E) The number of features considered at each split
Correct answer: C
Rationale: α penalizes tree size; larger α gives
smaller trees.
,4. A confusion matrix shows 100 true positives, 50
false negatives, 10 false positives, and 200 true
negatives. The precision is:
A) 100/150 = 0.667
B) 100/110 = 0.909
C) 100/300 = 0.333
D) 200/210 = 0.952
E) 100/200 = 0.5
Correct answer: B
Rationale: Precision = TP / (TP + FP) =
100/(100+10)=0.909.
5. The primary difference between bagging and
boosting is that:
A) Bagging builds models sequentially; boosting
builds in parallel
B) Boosting builds models sequentially, each
focusing on previous errors
C) Bagging uses all features; boosting uses random
subsets
D) Bagging reduces bias; boosting reduces variance
E) There is no difference
Correct answer: B
, Rationale: Boosting is sequential, weighting
misclassified observations.
6. Which of the following is NOT an appropriate
evaluation metric for a regression model predicting
insurance claim severity?
A) Mean absolute error (MAE)
B) Root mean squared error (RMSE)
C) Area under the ROC curve (AUC)
D) Mean absolute percentage error (MAPE)
E) R-squared
Correct answer: C
Rationale: AUC is for classification, not regression.
7. The "curse of dimensionality" refers to:
A) The difficulty of interpreting high-dimensional
models
B) The exponential increase in data needed to
maintain statistical significance as dimensions
increase
C) The increased computational time for high-
dimensional models
D) The tendency for overfitting in high dimensions
