100+ Free PCPA Practice Questions

Pass your CAS Property & Casualty Predictive Analytics Exam exam on the first try — instant access, no signup required.

✓ No registration✓ No credit card✓ No hidden fees✓ Start practicing immediately

100+ Questions

100% Free

1 / 100

Question 1

Score: 0/0

An auto insurance dataset has 12% of `driver_age` values missing, and the missingness is unrelated to age, claim amount, or any other variable. Which missing-data mechanism best describes this?

MCAR — Missing Completely At Random

MAR — Missing At Random

MNAR — Missing Not At Random

Selection bias

to track

Same family resources

More CAS Casualty Actuarial Society Exams Prep

Continue through related practice pages, study guides, comparisons, and articles from the same exam family.

Practice Pages

CAS Exam 5 Basic Ratemaking and Estimating Claim Liabilities Practice100 questions CAS Exam 6 Regulation and Financial Reporting (US) Practice100 questions CAS Exam 7 Estimation of Policy Liabilities, Insurance Co Valuation, ERM Practice100 questions CAS Exam 8 Advanced Ratemaking Practice100 questions CAS Exam 9 Financial Risk and Rate of Return Practice100 questions CAS Exam MAS-I Modern Actuarial Statistics I Practice100 questions

2026 Statistics

Key Facts: PCPA Exam

Jan 1, 2026

ACAS Requirement Effective

CAS announcement

2 parts

Exam + Project

CAS PCPA pathway

~3 hrs

Exam Length

CAS exam page

25%

GLM (Heaviest Topic)

CAS PCPA outline

~$700-900

Exam Fee (each part)

CAS fee schedule

Exam first

Order of Sittings

CAS PCPA Project announcement

PCPA is the multiple-choice exam half of the new CAS predictive-analytics requirement that became effective January 1, 2026 (originally targeted November 1, 2025) for ACAS. The full PCPA pathway is two pieces — this multiple-choice exam plus the separate take-home PCPA Project — and the Exam must be passed before registering for the Project. Coverage emphasizes GLMs for P&C (Tweedie, Poisson with log-exposure offset, Gamma severity), tree-based models for pricing (Random Forest and XGBoost), model validation (lift charts, double lift, Gini, ROC/AUC), variable selection (LASSO, stepwise, PCA, embeddings), business communication (PDPs, SHAP, waterfalls), and ethics including the NAIC AI Bulletin, Colorado SB 21-169, NY DFS Circular 7, and ASOPs 12, 23, 41, and 56.

Sample PCPA Practice Questions

Try these sample questions to test your PCPA exam readiness. Each question includes a detailed explanation. Start the interactive quiz above for the full 100+ question experience with AI tutoring.

1An auto insurance dataset has 12% of `driver_age` values missing, and the missingness is unrelated to age, claim amount, or any other variable. Which missing-data mechanism best describes this?

A.MCAR — Missing Completely At Random

B.MAR — Missing At Random

C.MNAR — Missing Not At Random

D.Selection bias

Explanation: MCAR means the probability of missingness is independent of both observed and unobserved values. Because age is missing for reasons unrelated to any variable, simple deletion or mean imputation will not introduce bias.

2A workers' compensation dataset shows that older policyholders are systematically less likely to report `prior_claims_count` (it is missing more for ages 60+). Conditional on observed age, the missingness does not depend on the unobserved claim count itself. This pattern is best described as:

A.MCAR

B.MAR

C.MNAR

D.Survivorship bias

Explanation: MAR (Missing At Random) means the probability of missingness depends on observed variables (age) but not on the unobserved value itself. Multiple imputation (MICE) or regression-based imputation conditional on age will give unbiased estimates under MAR.

3A homeowners dataset has loss amounts that are heavily right-skewed with a long upper tail. Which transformation is most commonly applied before fitting a linear model on `loss_amount`?

A.Square transformation

B.Logarithmic transformation

C.Standardization (z-score)

D.Min-max scaling to [0,1]

Explanation: Log transformations compress right-skewed positive distributions and stabilize variance. For loss severity it is standard practice; alternatives include Box-Cox (positive only) or Yeo-Johnson (handles zeros and negatives).

4Which transformation is the appropriate generalization of Box-Cox when the predictor or response can take zero or negative values?

A.Log transformation with offset of 1

B.Yeo-Johnson transformation

C.Square-root transformation

D.Inverse (reciprocal) transformation

Explanation: Yeo-Johnson extends Box-Cox to handle zero and negative values by piecewise definition. Box-Cox is restricted to strictly positive inputs, so Yeo-Johnson is preferred for raw fields that may contain zeros (e.g., prior claims).

5An actuary uses the IQR method to flag outliers in `policy_premium`. The first and third quartiles are $400 and $1,200, respectively. The conventional outlier fences are at:

A.$0 and $1,600

B.$200 and $1,400

C.$-800 and $2,400

D.$-1,600 and $3,200

Explanation: IQR = Q3 − Q1 = 800. Tukey fences are Q1 − 1.5·IQR and Q3 + 1.5·IQR, giving 400 − 1200 = −800 and 1200 + 1200 = 2400. Values outside that range are flagged for review.

6Which method detects outliers by isolating observations using random splits on randomly selected features, with shorter average path lengths indicating anomalies?

A.Mahalanobis distance

B.Local Outlier Factor (LOF)

C.Isolation Forest

D.DBSCAN

Explanation: Isolation Forest builds many random binary trees; outliers are isolated faster (shorter paths) than typical points. It scales well to high dimensions and is widely used for outlier screening in P&C predictive modeling pipelines.

7Capping the top 1% and bottom 1% of a continuous predictor at the 99th and 1st percentile values is called:

A.Winsorization

B.Standardization

C.Trimming

D.Bootstrapping

Explanation: Winsorization replaces extreme values with the nearest preserved percentile rather than removing the records. Trimming, in contrast, deletes the extremes entirely, which reduces the sample size.

8For an auto liability frequency model, the most appropriate exposure measure is generally:

A.Number of policies

B.Earned car-years

C.Total written premium

D.Number of drivers in household

Explanation: Earned car-years matches the units of loss exposure for auto coverage and accounts for partial policy terms. Premium and policy counts confound rate level and tenure with true exposure.

9Which exposure base is conventional for workers' compensation predictive modeling?

A.Number of employees

B.Insured payroll (per $100)

C.Hours worked

D.Number of locations

Explanation: Workers' compensation uses payroll as the standard exposure base, typically expressed per $100 of payroll. It correlates strongly with both frequency and severity of injury claims and is the regulator-required rating basis.

10A product manager asks you to impute missing `vehicle_age` values using the mean of all observed ages. Which is the most important drawback to flag?

A.It increases the variance of `vehicle_age`

B.It artificially shrinks the variance and weakens correlations with other variables

C.It changes the mean of `vehicle_age`

D.It removes records with missing data

Explanation: Mean imputation underestimates variance and attenuates relationships with other variables, which can lead to biased downstream coefficients. Multiple imputation, KNN, or model-based imputation generally preserve covariance structure better.

About the PCPA Exam

The CAS Property & Casualty Predictive Analytics (PCPA) Exam is the multiple-choice component of the new ACAS predictive-analytics requirement, testing whether candidates can clean P&C data, fit GLM and tree-based models, validate them, communicate results, and apply current ethics and bias standards.

Assessment

Multiple-choice exam (varies). The PCPA pathway has two components — the multiple-choice PCPA Exam and the separate take-home PCPA Project; candidates must pass the Exam before registering for the Project.

Time Limit

Computer-based, ~3 hours

Passing Score

Scaled cut score

Exam Fee

~$700-900 (Exam) + $700-900 (Project — separate) (Casualty Actuarial Society (CAS))

PCPA Exam Content Outline

15%

Data Preparation & Exploratory Data Analysis

Diagnose missing-data mechanisms (MCAR, MAR, MNAR), apply imputation (mean/median, KNN, MICE), detect outliers (IQR, z-score, Isolation Forest), apply transformations (log, Box-Cox, Yeo-Johnson), winsorize, and choose appropriate exposure measures (car-years, payroll, sales).

25%

GLM Modeling for P&C

Build and interpret GLMs for P&C: Tweedie pure premium (1<p<2), Poisson with log(exposure) offset for frequency, Gamma with log link for severity. Use frequency × severity decomposition, canonical and practical link functions, deviance, AIC/BIC, and overdispersion remedies.

15%

Tree-Based Models for P&C Pricing

Apply Random Forest (mtry, ntree, OOB error, Gini and permutation importance) and XGBoost (eta, max_depth, gamma, lambda, alpha, subsample, colsample, early stopping) — including monotonic constraints for regulatory defensibility.

15%

Model Validation

Use train/validation/test splits, k-fold and stratified cross-validation, lift charts, double lift charts, Gini coefficient (twice the area between Lorenz and 45° line), ROC/AUC, calibration, and residual diagnostics (Pearson, deviance, normal Q-Q).

10%

Variable Selection & Interactions

Apply forward/backward/stepwise selection, LASSO (L1), Ridge (L2), Elastic Net, principal components, embeddings for high-cardinality categoricals, and meaningful predictor interactions (e.g., age × territory in auto, payroll × class in WC).

10%

Communication of Results

Translate model output for non-technical audiences with executive summaries, waterfall plots, partial dependence plots (PDP), individual conditional expectation (ICE) plots, SHAP values, and ASOP 41-compliant model documentation.

10%

Ethics & Bias in P&C Predictive Models

Apply the NAIC Model Bulletin on AI (Dec 2023), Colorado SB 21-169, NY DFS Circular 7 (2024), proxy-discrimination and disparate-impact analysis, and ASOPs 12 (Risk Classification), 23 (Data Quality), 41 (Communications), 56 (Modeling), plus the CAS Code of Professional Conduct.

How to Pass the PCPA Exam

What You Need to Know

Passing score: Scaled cut score
Assessment: Multiple-choice exam (varies). The PCPA pathway has two components — the multiple-choice PCPA Exam and the separate take-home PCPA Project; candidates must pass the Exam before registering for the Project.
Time limit: Computer-based, ~3 hours
Exam fee: ~$700-900 (Exam) + $700-900 (Project — separate)

Keys to Passing

Complete 500+ practice questions
Score 80%+ consistently before scheduling
Focus on highest-weighted sections
Use our AI tutor for tough concepts

PCPA Study Tips from Top Performers

1Master the GLM block first — it is the largest weighted topic. Be fast with Tweedie (1<p<2) for pure premium, Poisson with log(exposure) offset for frequency, and Gamma with log link for severity, and know how to interpret coefficients on the multiplicative scale.

2Drill validation: lift charts, double lift, Gini (twice area between Lorenz and 45° line), ROC/AUC, calibration, and residual diagnostics. Know what each diagnostic catches and which problems it CANNOT detect (e.g., Gini measures ranking, not calibration).

3Memorize the XGBoost hyperparameters and what they do — eta (learning rate), max_depth, gamma (min split loss), lambda (L2), alpha (L1), subsample, colsample_bytree, early_stopping_rounds — and how they trade bias against variance.

4Learn the regulatory landscape cold: NAIC AI Bulletin (Dec 2023), Colorado SB 21-169, NY DFS Circular 7 (2024), proxy discrimination, disparate impact, and ASOPs 12, 23, 41, 56 plus the CAS Code of Professional Conduct.

5Practice translating model output for non-technical audiences using PDPs, ICE plots, SHAP values, waterfall plots, and ASOP 41-style executive summaries. Communication is small in weight but easy to lose points on if framed poorly.

Frequently Asked Questions

What is the difference between the PCPA Exam and the PCPA Project?

PCPA is a two-part requirement. The PCPA Exam is a multiple-choice computer-based test (~3 hours) covering data prep, GLMs, tree-based models, validation, variable selection, communication, and ethics. The PCPA Project is a separate take-home capstone modeling assignment with a defined submission window. Candidates must pass the PCPA Exam BEFORE they can register for the PCPA Project. Both pieces are required for the new ACAS predictive-analytics requirement effective January 1, 2026 (originally targeted November 1, 2025).

When did the PCPA become a required ACAS exam?

The PCPA requirement was originally targeted for November 1, 2025 and went into effect January 1, 2026. Candidates pursuing ACAS designation under the new pathway need both the PCPA Exam and the PCPA Project, in addition to the other CAS exam requirements.

How much does the PCPA cost?

Each component (Exam and Project) is approximately $700-900, paid separately. The iCAS prep course is optional. Always confirm current fees on the CAS exam-registration page because pricing is updated periodically.

Which topics carry the most weight on the PCPA Exam?

GLM modeling for P&C is the largest section at roughly 25%, reflecting how central GLMs (Tweedie, Poisson frequency with log-exposure offset, Gamma severity) remain to ratemaking. Data preparation/EDA, tree-based models, and model validation each carry around 15%. Variable selection, communication, and ethics each carry roughly 10%.

Do I need to know the NAIC AI Bulletin and Colorado SB 21-169?

Yes. The Ethics & Bias section explicitly tests current P&C predictive-modeling regulation, including the NAIC Model Bulletin on Use of AI by Insurers (December 2023), Colorado SB 21-169 (algorithmic discrimination testing), NY DFS Circular Letter No. 7 of 2024, proxy discrimination and disparate-impact analysis, and the relevant ASOPs (12, 23, 41, 56).

Is MAS-II a prerequisite for the PCPA Exam?

MAS-II is not a formal prerequisite, but it is strongly recommended because PCPA assumes the underlying GLM, regression, and statistical-learning material covered there. Candidates who skip MAS-II usually need substantial extra preparation on link functions, deviance, model selection, and tree-based methods.