The CAS made the PCPA requirement unconditional on January 1, 2026: every candidate pursuing the ACAS designation who had not already received it must now pass a two-hour computer-based exam and complete a separate graded take-home project before the credential is awarded. Candidates who earned their ACAS before that date are grandfathered for FCAS purposes, which means the practical weight of the requirement has landed on the largest active entering cohort the CAS has seen in the post-pandemic hiring surge. The first full class of mandatory PCPA candidates is now moving through the exam cycle, with project-window results from the January-through-April sitting period only beginning to surface.

Reviewing the PCPA sample exam questions and content outline alongside the predictive modeling workflows deployed in actual carrier pricing projects reveals something worth naming directly: the GLM diagnostics coverage on the CBT is noticeably more rigorous than what most candidates encounter in their first two years of rotation, and the take-home project expects a level of variable selection narrative that resembles a consulting deliverable more than a traditional exam answer. The candidates who pass on the first attempt are not necessarily the ones who know more statistics. They are the ones who understood what two distinct assessment instruments are actually measuring.

Two Components, One Sequenced Requirement

The PCPA requirement has two parts, and the sequencing is strict: a candidate must pass the computer-based exam before registering for the project. There is no path to the project without an exam pass.

The exam is a two-hour, 40-question computer-based test administered on demand year-round at Pearson VUE test centers. The question format is more varied than the standard actuarial CBT: alongside conventional multiple-choice items, the exam includes multiple-selection questions where more than one answer may be correct, matching questions, and fill-in-the-blank items. That format demands a kind of precision that candidates accustomed to eliminating wrong answers may find more difficult than a comparable multiple-choice exam. The registration fee is $300 per attempt, with up to three attempts permitted within any rolling 12-month window and a mandatory two-week waiting period between attempts. Preliminary results are available at the testing center immediately; final confirmed results appear in the CAS candidate portal approximately 15 days later.

The project runs on four quarterly windows per year. The June 2026 project window opened June 15, with registration having closed June 8. Each window gives candidates access to a dataset and a business problem. They then have approximately two weeks to build a predictive model and submit three deliverables: a technical report capped at 1,000 words, working model code in R, Python, or SAS, and up to five supporting tables or exhibits. The 1,000-word ceiling is enforced absolutely. Submissions that exceed it fail regardless of modeling quality. Project results take six to eight weeks after the submission deadline, which means that a candidate sitting the June window will not receive a result until August at the earliest. The project fee is $700 per attempt, bringing the combined single-attempt cost for both components to $1,000.

Generative AI tools are permitted as a consultation resource during the project, but the CAS requires that all submitted work be the candidate's own original output. Using an AI system to generate model code or draft the technical report wholesale would violate that requirement; using it to check syntax or talk through a distributional assumption is permitted.

What the CBT Covers in Practice

The exam content organizes into three areas: Dealing with Data, Model Diagnostics and Selection, and Model Interpretation and Presentation. The progression matters: the first area is foundation, the second is where most of the difficulty concentrates, and the third is where the exam connects modeling technique to regulatory and professional context.

The data section covers preparation work that candidates who have rotated through a pricing unit will recognize: outlier identification, missing data treatment, variable encoding strategies for categorical features, log transformation for right-skewed continuous variables, and exposure-adjusted rate construction. Candidates with MAS-I preparation have the statistical foundation for this material, even if the P&C application context is not covered directly in the MAS sequence. This section rewards practical familiarity with how insurance datasets are structured before a model sees them.

The model section is where the PCPA diverges sharply from prior exam content. GLM coverage goes substantially deeper than MAS-I. The exam tests Tweedie distributions for pure premium modeling, Poisson regression with log-exposure offset for frequency, and Gamma regression for severity, with explicit attention to link function selection and the diagnostic tools that reveal whether the chosen distributional assumption fits the data. A candidate who understands how to calculate a Pearson residual but cannot interpret a deviance residual plot, identify overdispersion from a quasi-Poisson model comparison, or explain why a log link is appropriate for severity while an identity link would not be is going to encounter real difficulty in the GLM diagnostics questions. These are the mechanics that most early-career pricing rotations touch only at the surface, where the candidate feeds data into an existing model framework rather than selecting and validating the framework from scratch.

Tree-based models appear with more structure than MAS-II provides. The exam covers Random Forest and XGBoost as distinct model families, testing not just what they are but how to tune them, how the bias-variance tradeoff operates at different depth settings, and how variable importance metrics differ between bagging ensembles and boosted ensembles. Variable selection methods include LASSO, with questions about its shrinkage-toward-zero behavior and how it differs from ridge regression in the P&C context; stepwise selection procedures and their documented overfitting tendencies; principal component analysis; and embedding-based approaches for high-cardinality categorical variables such as territory or vehicle class. A candidate who has used only stepwise GLM selection in a pricing rotation is likely to find the variable selection section the densest part of the exam.

The ethics and professional standards section is specific, not generic. The exam tests knowledge of ASOPs 12, 23, 41, and 56; the NAIC Model Bulletin on the Use of Algorithms and Predictive Models; Colorado SB 21-169; and New York DFS Circular Letter 7. ASOP 12 governs risk classification; ASOP 23 covers data quality standards for actuarial work; ASOP 41 addresses actuarial communications; ASOP 56 sets out model governance and documentation obligations. The Colorado and New York regulatory instruments address algorithmic bias testing, disparate impact analysis, and disclosure requirements for insurance AI and predictive model systems in rate and underwriting contexts. Exam questions in this section are situational: they describe a specific fact pattern and ask what a defined standard requires, not what general ethics principles suggest. A candidate who has not read these documents will struggle.

The Take-Home Project: Three Deliverables, One Defense

The project is graded by human examiners against a rubric, and the three-deliverable structure reveals what that rubric rewards. Code, exhibits, and the technical report each carry independent weight, and deficiencies in one cannot be offset by strength in another.

The code submission is evaluated on reproducibility and internal consistency, not just on whether it runs. Examiners check whether data cleaning steps are scripted rather than performed interactively before the session begins; whether model selection decisions are encoded in the script rather than made ad hoc and then described as if they were planned; and whether the modeling choices are consistent throughout the pipeline. A candidate who applies LASSO in the variable selection step but then describes the final model using stepwise inclusion logic in the report has introduced a logical gap that graders will identify, because the code and the report are read against each other.

The 1,000-Word Ceiling

The project technical report has a hard maximum of 1,000 words. Submissions that exceed this limit are automatically failed, regardless of modeling quality or insight. This is not a soft guideline; it is an enforcement mechanism. A report structured as a chronological narration of the modeling workflow will exhaust the budget describing data preparation and have no room left for the model selection rationale, which is the content graders most want to read.

The five-exhibit limit forces candidates to allocate deliberately. A lift chart comparing the selected model against a baseline, a variable importance plot, and two or three diagnostic exhibits is a more defensible allocation than five univariate predictor distributions. What examiners want to see is the evidence that supports the model selection decision, not a complete data description. Candidates who fill the exhibit budget with input-data visualizations are signaling that they did not think carefully about what a model reviewer needs to evaluate the submitted work.

The technical report is the most consequential deliverable for most candidates and the one most likely to fail on structure rather than substance. A report that narrates the modeling sequence chronologically will use most of the 1,000 words describing what the candidate did before reaching the question examiners are most interested in: why this model family over the alternatives, why this variable set given what LASSO and domain knowledge revealed, and what the validation diagnostics show about the model's fit to the business problem. The report is better structured as a defense of a decision than as a methodology journal entry.

In register, the project technical report resembles a consulting deliverable more than a traditional actuarial exam answer. A candidate who has written actuarial memoranda justifying rating plan modifications, or who has prepared model validation documentation for a state rate filing, has a relevant frame of reference. A candidate whose primary writing experience comes from MAS-II written-answer questions is likely to over-narrate the methodology and under-justify the choices. The word limit is the mechanism that forces the distinction: a MAS-II written answer rewards completeness; the PCPA technical report rewards concision in the service of a defensible position.

What MAS-I and MAS-II Carry Into PCPA

MAS-I provides the statistical foundation for a meaningful portion of the PCPA exam. The GLM framing, regression diagnostics, and model fit reasoning from MAS-I transfer directly, even if the specific P&C-applied distributions (Tweedie, Poisson with log-exposure offset) are not MAS-I material. The probability model intuition from MAS-I supports the distributional selection questions in the PCPA content. Candidates who took MAS-I seriously and understood the material rather than treating it as a pass-rate optimization problem have a real advantage in the GLM section of the PCPA.

MAS-II's machine learning curriculum provides partial preparation for the tree-based model content. The bias-variance decomposition, cross-validation methodology, and ensemble model framing from MAS-II map directly to PCPA exam questions. What MAS-II does not fully prepare candidates for is application-layer specificity: PCPA frames questions in a P&C insurance context, connecting model diagnostics to filing and regulatory implications, and the ethics section's link to ASOPs and state regulatory instruments has no direct MAS-II predecessor.

The CAS recommends completing MAS-I, MAS-II, and Exam 5 before attempting PCPA, and the sequencing has substantive rationale beyond scheduling convenience. Exam 5's ratemaking curriculum introduces the actuarial reasoning that connects model output to rate indications, which is the business context the PCPA project embeds its dataset in. A candidate who has not yet worked through classification ratemaking and individual risk rating will find the project's framing of the business problem less intuitive. The model they build will be technically coherent; the selection narrative in the report may not correctly identify what the model output is actually measuring in actuarial terms.

Where PCPA Stops

The PCPA closes the GLM and tree-based model gap between academic preparation and practitioner expectation. It does not close the full distance between the ACAS credential requirements and what advanced pricing or reserving roles now expect from candidates in the first 18 months of employment.

Real-time model scoring infrastructure, the deployment pipelines that move a trained GLM or gradient-boosted model into a production rating engine, is outside the PCPA scope. Passing PCPA demonstrates the ability to build and validate a model in a two-week project setting. It does not demonstrate the ability to version a model, monitor it for drift against incoming production data, or embed it in a filing submission package structured for state regulatory review under the NAIC AI evaluation framework. Those capabilities are what employers mean when they list model deployment or model risk management as separate requirements alongside the PCPA on a job posting.

Neural networks appear in the PCPA scope at the level of definitional awareness. Large language model integration, agentic AI systems in underwriting or claims workflows, and the vendor model evaluation protocols now appearing in carrier AI governance frameworks are beyond the PCPA scope entirely. A candidate who passes PCPA has not demonstrated the ability to evaluate a third-party vendor's embedded ML model for a claims triage tool, assess the documentation requirements under the NAIC's Exhibit C framework, or apply the governance standards that 12 states are now operationalizing through the AI evaluation pilot. Those expectations arrive post-credentialing, in the role itself.

The CAS AI competency framework and the Fast Track CE program for practicing actuaries address some of this gap, but they are continuing education obligations, not PCPA content. The PCPA is correctly positioned as a foundation credential for applied modeling; the post-ACAS development curve for model governance, deployment, and regulatory compliance is steeper than the exam syllabus suggests.

How Employers Are Reading PCPA Completion

The PCPA requirement has changed how some P&C employers structure early-career actuarial candidate evaluation. Entry-level and actuarial analyst postings at mid-market and large carriers began listing PCPA exam completion as preferred in mid-2025, when the January 2026 mandatory date was confirmed; since the requirement took effect, the frequency of that signal has increased. In some cases, both the exam and the project are listed as separate preferred qualifications, distinguishing candidates who have demonstrated CBT-level knowledge from those who have completed the full applied modeling cycle.

The signal function is clear to employers who run pricing rotations with small teams. A candidate who has passed the PCPA exam has demonstrated comfort with GLM diagnostics, tree-based model selection, and the regulatory ethics framework for predictive modeling in a standardized, externally verified setting. That translates into a shorter onboarding period for the model-building components of a pricing project, even if the candidate still lacks production deployment experience. A candidate who has also completed the project has demonstrated that the modeling knowledge extends to independent application under a deadline, with a written defense graded against professional standards.

Employers in reserving roles are beginning to apply a similar filter, particularly for positions that involve building or reviewing IBNR models with statistical components. The PCPA content overlaps meaningfully with the validation methodology relevant to stochastic reserving, and the ethics coverage of ASOP 23 (data quality) and ASOP 56 (model governance) is directly applicable to reserve model documentation.

First-Cohort Signals and Preparation Resources

The CAS has not yet published pass rate data for either the PCPA exam or the project from the January 2026 mandatory cohort. The project grading timeline means that results from the first full quarter of mandatory candidates, those who sat the exam in January and February 2026 and attempted the April project window, produced results only in June 2026. A full statistical picture of first-cohort outcomes will not be available until late 2026 at the earliest.

Early signals from candidates and preparation resource providers point to two recurring difficulty areas. The first is the GLM diagnostics section of the CBT, specifically the questions that require interpreting diagnostic output rather than reproducing a calculation. Candidates who memorized Pearson residual formulas for MAS-I but did not develop intuition about what a systematic pattern in deviance residuals indicates about distributional misspecification are encountering those questions as genuinely unfamiliar. The second is the project technical report word limit, which catches candidates who approach the report as a full methodology documentation rather than a model defense. The word ceiling does not penalize technical knowledge; it penalizes candidates who do not know what question the report is supposed to answer.

The CAS provides a free preparatory course, "Building a GLM from Start to Finish," currently available in R with Python and SAS versions in development. The course covers the hands-on model construction workflow and is a useful foundation for project preparation, though it focuses more on execution than on the written defense framing that the technical report requires. Commercial preparation providers including The Infinite Actuary offer structured PCPA preparation courses that cover both CBT content and project report strategy; candidates who have found commercial resources effective for MAS-I and MAS-II preparation are reporting similar utility for the PCPA exam component.

One alternative path exists for candidates with prior CAS Institute credit. Candidates who passed the CAS Institute's CSPA Course 3 (Predictive Modeling: Methods and Techniques) and its associated case study project may petition for a full PCPA waiver covering both components. The waiver applies specifically to CSPA Course 3 completion obtained directly; credit transferred from another exam waiver mechanism does not qualify. Candidates who obtained CSPA Course 3 credit through a waiver rather than by sitting the course should confirm their waiver eligibility before relying on it in their ACAS credential timeline.

The Credential Signal the PCPA Creates

The PCPA requirement is the most structurally significant change to the ACAS pathway in years. It inserts, for the first time, a standardized external assessment of hands-on modeling capability into a credential that previously tested only the statistical theory underlying those methods. The decision to make the project a separate, human-graded deliverable rather than a simulation-based module reflects the CAS's judgment that the ability to select a model family, defend the selection in writing, and produce reproducible code under a deadline cannot be adequately measured by a multiple-choice instrument alone.

For candidates, the practical consequence is that exam prep timelines need to account for the project's quarterly availability and the six-to-eight-week grading window. A candidate who passes the CBT in October faces a project window in December with results not available until February. Fitting both components into a credential timeline requires planning the PCPA sequence as a two-stage process a semester ahead, not treating the project as an on-demand extension of the exam. The candidates who miss their target ACAS date because of PCPA sequencing are, in the early cohort data, almost uniformly those who did not model the grading timeline when they sat the CBT.

For employers, the PCPA changes what the ACAS in progress credential communicates. A candidate's position in the PCPA sequence, exam-passed but project-pending versus both components complete, now carries information about practical modeling readiness that prior credential timelines did not surface. How employers weight that distinction in hiring decisions is becoming visible in job posting language, and the direction of travel is toward treating both components as meaningful separately rather than treating PCPA completion as a single binary credential event.

Actuaries who incorporate the PCPA's specific content rigor, particularly the ASOPs 23 and 56 documentation standards and the state regulatory instruments for model bias testing, into their ongoing practice rather than treating them as exam content to be remembered and then set aside will find those frameworks re-appearing in every meaningful model development project they encounter after credentialing. The exam tests them because the job requires them. That is the coherence the PCPA requirement was designed to enforce.

CAS Exams Actuarial Exams Exam Study

Further Reading

Sources