Page 1 of 110
SOA PA-PREDICTIVE ANALYTICS EXAM |
2026/2027 EDITION | VERIFIED QUESTIONS
WITH DETAILED ANSWERS AND RATIONALES |
BRAND NEW!
1. A health insurance company wants to build a model to predict
which policyholders will develop chronic conditions within the
next two years. The data science team has access to claims history,
pharmacy records, and demographic information. Before
beginning model development, the actuarial team should first:
A) Fit a generalized linear model with all available variables
B) Clearly define the business problem, identify the target variable,
and assess data suitability
C) Apply principal component analysis to reduce dimensionality
D) Build a random forest to identify the most important predictors
Expert Explanation: Problem definition is the critical first step in any
predictive analytics project. Before selecting techniques or building
models, the team must translate the vague business question into a
precise analytical problem, identify the outcome variable of interest,
and assess whether available data supports the intended analysis.
Jumping directly to modeling without this foundation risks misalignment
between analytical output and business needs.
,Page 2 of 110
2. An actuary is evaluating whether a predictive model should be
deployed for underwriting decisions. The model has strong
discriminatory power (high AUC) but relies on variables that may
serve as proxies for protected characteristics. Under the SOA Code
of Professional Conduct, the actuary's primary obligation is to:
A) Deploy the model since predictive accuracy is the paramount
consideration
B) Ensure the model complies with applicable law and does not
produce unfairly discriminatory outcomes
C) Remove the model and report the findings to the state insurance
department
D) Deploy the model but disclose the potential bias in the actuarial
communication
Expert Explanation: Precept 1 of the Code requires actuaries to act
honestly, with integrity, and in a manner that fulfills the profession's
responsibility to the public. ASOP No. 12 (Risk Classification) requires
compliance with applicable law and consideration of whether risk
characteristics may result in unfair discrimination. Strong predictive
performance does not justify deployment if the model violates legal or
ethical standards.
3. A dataset contains the variable "claim_amount" with values
ranging from $0 to $2,500,000, with a right-skewed distribution.
The actuary wants to use this variable as a target in a regression
model. Which transformation would be most appropriate to
address the skewness?
,Page 3 of 110
A) Standardization (z-score transformation)
B) Logarithmic transformation
C) Principal component analysis
D) One-hot encoding
Expert Explanation: For right-skewed continuous variables, a logarithmic
transformation compresses the upper tail and can make the distribution
more symmetric, which is often beneficial for regression modeling.
Standardization changes scale but not shape. PCA is a dimensionality
reduction technique, not a variable transformation for distributional
improvement. One-hot encoding is for categorical variables.
4. A predictive model for insurance fraud detection is being
developed. The dataset contains 500 fraud cases out of 250,000
total claims. The data scientist trains a logistic regression model
and achieves 99.7% accuracy. Why should this result be
interpreted with caution?
A) Logistic regression is not appropriate for fraud detection
B) Accuracy is misleading for imbalanced datasets because predicting
the majority class alone achieves similar performance
C) The model must have overfit because accuracy exceeds 95%
D) Fraud detection models require random forests, not logistic
regression
Expert Explanation: With only 0.2% fraud cases, a naive model that
predicts "no fraud" for every claim would achieve 99.8% accuracy. For
imbalanced classification problems, metrics such as precision, recall, F1-
score, or AUC-ROC provide more meaningful assessments of model
performance than overall accuracy.
, Page 4 of 110
5. An actuary is using K-means clustering to segment a portfolio of
life insurance policyholders based on age, income, and risk score.
Which statement about this technique is correct?
A) K-means requires the number of clusters to be determined by the
algorithm automatically
B) K-means partitions observations into K clusters by minimizing
within-cluster variance
C) K-means can only be applied to categorical variables
D) K-means produces overlapping clusters to capture uncertainty in
segment membership
Expert Explanation: K-means is an unsupervised learning algorithm that
partitions n observations into K clusters, where each observation
belongs to the cluster with the nearest mean (centroid). The algorithm
minimizes the sum of squared distances within clusters. The number of
clusters K must be specified by the analyst, often using techniques such
as the elbow method or silhouette analysis.
6. A general insurance company wants to model claim frequency for
auto insurance policies. The target variable is the number of
claims per policy per year, which takes values 0, 1, 2, ... The
appropriate GLM distribution family for this target is:
A) Gaussian (Normal)
B) Binomial
C) Poisson
D) Gamma
SOA PA-PREDICTIVE ANALYTICS EXAM |
2026/2027 EDITION | VERIFIED QUESTIONS
WITH DETAILED ANSWERS AND RATIONALES |
BRAND NEW!
1. A health insurance company wants to build a model to predict
which policyholders will develop chronic conditions within the
next two years. The data science team has access to claims history,
pharmacy records, and demographic information. Before
beginning model development, the actuarial team should first:
A) Fit a generalized linear model with all available variables
B) Clearly define the business problem, identify the target variable,
and assess data suitability
C) Apply principal component analysis to reduce dimensionality
D) Build a random forest to identify the most important predictors
Expert Explanation: Problem definition is the critical first step in any
predictive analytics project. Before selecting techniques or building
models, the team must translate the vague business question into a
precise analytical problem, identify the outcome variable of interest,
and assess whether available data supports the intended analysis.
Jumping directly to modeling without this foundation risks misalignment
between analytical output and business needs.
,Page 2 of 110
2. An actuary is evaluating whether a predictive model should be
deployed for underwriting decisions. The model has strong
discriminatory power (high AUC) but relies on variables that may
serve as proxies for protected characteristics. Under the SOA Code
of Professional Conduct, the actuary's primary obligation is to:
A) Deploy the model since predictive accuracy is the paramount
consideration
B) Ensure the model complies with applicable law and does not
produce unfairly discriminatory outcomes
C) Remove the model and report the findings to the state insurance
department
D) Deploy the model but disclose the potential bias in the actuarial
communication
Expert Explanation: Precept 1 of the Code requires actuaries to act
honestly, with integrity, and in a manner that fulfills the profession's
responsibility to the public. ASOP No. 12 (Risk Classification) requires
compliance with applicable law and consideration of whether risk
characteristics may result in unfair discrimination. Strong predictive
performance does not justify deployment if the model violates legal or
ethical standards.
3. A dataset contains the variable "claim_amount" with values
ranging from $0 to $2,500,000, with a right-skewed distribution.
The actuary wants to use this variable as a target in a regression
model. Which transformation would be most appropriate to
address the skewness?
,Page 3 of 110
A) Standardization (z-score transformation)
B) Logarithmic transformation
C) Principal component analysis
D) One-hot encoding
Expert Explanation: For right-skewed continuous variables, a logarithmic
transformation compresses the upper tail and can make the distribution
more symmetric, which is often beneficial for regression modeling.
Standardization changes scale but not shape. PCA is a dimensionality
reduction technique, not a variable transformation for distributional
improvement. One-hot encoding is for categorical variables.
4. A predictive model for insurance fraud detection is being
developed. The dataset contains 500 fraud cases out of 250,000
total claims. The data scientist trains a logistic regression model
and achieves 99.7% accuracy. Why should this result be
interpreted with caution?
A) Logistic regression is not appropriate for fraud detection
B) Accuracy is misleading for imbalanced datasets because predicting
the majority class alone achieves similar performance
C) The model must have overfit because accuracy exceeds 95%
D) Fraud detection models require random forests, not logistic
regression
Expert Explanation: With only 0.2% fraud cases, a naive model that
predicts "no fraud" for every claim would achieve 99.8% accuracy. For
imbalanced classification problems, metrics such as precision, recall, F1-
score, or AUC-ROC provide more meaningful assessments of model
performance than overall accuracy.
, Page 4 of 110
5. An actuary is using K-means clustering to segment a portfolio of
life insurance policyholders based on age, income, and risk score.
Which statement about this technique is correct?
A) K-means requires the number of clusters to be determined by the
algorithm automatically
B) K-means partitions observations into K clusters by minimizing
within-cluster variance
C) K-means can only be applied to categorical variables
D) K-means produces overlapping clusters to capture uncertainty in
segment membership
Expert Explanation: K-means is an unsupervised learning algorithm that
partitions n observations into K clusters, where each observation
belongs to the cluster with the nearest mean (centroid). The algorithm
minimizes the sum of squared distances within clusters. The number of
clusters K must be specified by the analyst, often using techniques such
as the elbow method or silhouette analysis.
6. A general insurance company wants to model claim frequency for
auto insurance policies. The target variable is the number of
claims per policy per year, which takes values 0, 1, 2, ... The
appropriate GLM distribution family for this target is:
A) Gaussian (Normal)
B) Binomial
C) Poisson
D) Gamma
