UUS Practice Tests⌕ Search
Actuarial Examinations · PCPA

CAS Property and Casualty Predictive Analytics Examination Practice Test

Predictive analytics in property and casualty insurance uses statistical learning to estimate claim frequency, severity, and loss cost from policy and claims data. The practice material samples model validation, exposure-aware comparisons, sampling adjustments, and monitoring for shifts in inputs or outcomes. These skills help actuaries assess whether predictions remain useful across policyholders, agencies, territories, and future periods.

Exam overview

Exam providerCasualty Actuarial Society
Exam codePCPA
CategoryActuarial Examinations

About the CAS Property and Casualty Predictive Analytics Examination Exam

Applied work includes choosing validation schemes that reflect how predictions will be used: time-ordered holdouts for future underwriting periods, grouped splits for agencies or other clusters, and cross-validation for assessing performance across folds. Analysts also evaluate whether a model represents variation in claim counts, permits nonlinear predictor effects, and avoids leakage from information unavailable at the scoring date. Portfolio comparisons may require accounting for policy limits or adjusting a disproportionate sample to represent the target business mix. For claims with unresolved costs, estimates must recognize uncertainty in ultimate severity rather than treating current paid amounts as final. Monitoring input distributions can flag production changes before outcomes mature; once claims develop, observed-versus-predicted results help assess calibration and ongoing performance.

Sample Practice Questions

Review a selection from this exam's published US Practice Tests practice bank.

Question 1

A portfolio contains policies with different limits. An analyst wants to compare loss amounts on a basis that accounts for the maximum amount the insurer could pay under each policy. Which quantity is most directly relevant?

  1. ALoss divided by the policy limit, interpreted with care when limits differ
  2. BLoss multiplied by the policy limit
  3. CThe policy limit minus the loss for every policy
  4. DThe portfolio-wide average loss, assigned equally to every policy
Show answerCorrect answer: A

Scaling loss by the applicable limit expresses the outcome relative to the policy's coverage capacity. This can aid comparisons across limits, though the interpretation depends on the claim and policy structure.

Question 2

A model produces a distribution of predicted claim counts for a group of policies. The observed total count is close to the predicted mean, but the observed variation across comparable groups is consistently larger than the model implies. Which feature of the data may be inadequately represented?

  1. AOverdispersion relative to the assumed count distribution
  2. BThe ordering of policy identification numbers
  3. CThe number of decimal places used to display predictions
  4. DThe distinction between a predictor and a response
Show answerCorrect answer: A

Overdispersion occurs when count outcomes vary more than the fitted distribution assumes. A model can match average counts while understating variability, which may affect uncertainty estimates and prediction intervals. The analyst should examine the distributional assumptions and consider an approach that accommodates the observed dispersion.

Question 3

A model is trained using historical claim records, but the mix of business written in production changes substantially. Which monitoring comparison most directly detects a change in the distribution of model inputs, even before outcome data are available?

  1. ACompare current predictor distributions with their development-period distributions
  2. BCompare the model's training loss with its training loss from the same run
  3. CRecalculate the development-period coefficient estimates without new data
  4. DCompare claim outcomes only among policies that were never scored
Show answerCorrect answer: A

Input-distribution monitoring, often called population stability or feature-drift monitoring, checks whether the characteristics of scored business have shifted from the development data. It can flag changes before outcomes mature, though it does not by itself prove that predictive performance has deteriorated.

Question 4

An analyst estimates a model using a sample drawn disproportionately from high-premium policies, although the target portfolio has a different mix. If the sampling design is known, what is a reasonable way to make aggregate estimates reflect the target portfolio?

  1. AWeight observations to represent their relative frequency in the target portfolio
  2. BGive every sampled observation a weight proportional to its observed loss
  3. CDiscard all high-premium observations before fitting
  4. DUse the sample's unweighted average as the target estimate
Show answerCorrect answer: A

Appropriate sampling or post-stratification weights can adjust the contribution of observations so that represented groups align more closely with the target population. The weights should reflect the sampling and target-population design, not observed outcomes.

Question 5

When using k-fold cross-validation, what is the basic procedure?

  1. APartition the data into k parts, fit on k−1 parts, evaluate on the remaining part, and rotate the held-out part
  2. BFit k models on the full dataset and report the best training score
  3. CHold out the same single observation in every fold
  4. DUse each fold only to select predictors and never to evaluate predictions
Show answerCorrect answer: A

In k-fold cross-validation, each fold serves once as the validation portion while the other folds are used for fitting. The resulting out-of-fold performance estimates are then summarized, subject to suitable splitting for the data structure.

Question 6

A team wants to estimate how a policyholder characteristic relates to loss cost while allowing the relationship to bend rather than remain a straight line. Which approach directly provides this flexibility?

  1. AAdd a nonlinear transformation or spline representation of the characteristic
  2. BReplace the response with a constant for every policy
  3. CRemove the intercept from every model
  4. DUse a random train-test split as the only model change
Show answerCorrect answer: A

Transformations such as polynomial terms or splines allow the fitted relationship to vary nonlinearly with a predictor. The added flexibility should be evaluated for stability and predictive performance rather than assumed to be beneficial.

Question 7

A claims model is trained using a predictor that is populated only after a claim has been settled. The model is intended to score policies at the time of sale. What is the main concern?

  1. AThe predictor has too few possible values
  2. BThe predictor would not be available at the intended prediction time and may cause leakage
  3. CThe predictor necessarily has a linear relationship with the outcome
  4. DThe predictor makes the response variable continuous
Show answerCorrect answer: B

A predictor must be available when the model is used. Information generated after settlement can reveal future outcomes or their consequences, creating target leakage and misleading apparent performance. The modeling data should be restricted to information that would genuinely be available at policy sale.

How to use this practice test

The free practice session draws up to 20 random questions from the published bank of 50 and displays results after completion. Use the questions to check how you distinguish validation designs, recognize data leakage and distribution shifts, and select comparisons suited to insurance exposures. Review missed answers by identifying the modeling assumption or data limitation involved.

Preparation tips

  • For each validation scenario, identify the deployment target and choose a split that prevents information from the same future period or agency leaking across training and evaluation.
  • Practice distinguishing input-distribution monitoring from outcome-based performance checks; the former can be performed before mature claims are available.
  • When comparing loss amounts, note policy limits and exposure differences, and determine whether a sample adjustment is needed to represent the target portfolio.
  • Check claim-count assumptions against observed dispersion, and consider whether a flexible predictor effect is appropriate when a straight-line relationship is inadequate.
  • For open claims, separate current payments from estimated ultimate costs and account for uncertainty in unresolved severity.
Official sourceOfficial exam page ↗Official exam guide ↗