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 ACAS before that date are grandfathered for FCAS purposes, so the requirement has landed on the largest active entering cohort of the post-pandemic hiring surge. The first aggregate figures are now out: the Spring 2026 sitting passed 67.5% of 126 candidates.

Key Takeaways

  • The Spring 2026 CBT sitting passed 67.5% of 126 candidates, charted on the PCPA pass-rate page. Project-component results from the first mandatory cohort remain unpublished.
  • The two components are strictly sequenced: no candidate can register for the project without an exam pass, and the combined single-attempt cost is $1,000.
  • The technical report has a hard 1,000-word ceiling enforced absolutely. Submissions over it fail regardless of modeling quality, alongside a five-exhibit cap.
  • GLM coverage runs well past MAS-I, testing Tweedie, Poisson with log-exposure offset, and Gamma with explicit link function and diagnostic interpretation.
  • Project grading takes six to eight weeks after a quarterly submission deadline, which makes the credential timeline a two-stage plan rather than an on-demand extension of the exam.

The Two Components and Their Mechanics

The sequencing is strict and the cost structure differs sharply between the halves.

The exam is a two-hour, 40-question computer-based test administered on demand year-round at Pearson VUE centers. The format is more varied than a standard actuarial CBT: multiple-selection items where more than one answer is correct, matching questions, and fill-in-the-blank alongside conventional multiple choice. That mix removes the elimination strategies candidates carry over from other sittings. Registration is $300 per attempt, with up to three attempts in any rolling 12-month window and a two-week wait between them. Preliminary results appear at the center immediately; confirmed results reach the candidate portal roughly 15 days later.

The project runs on four quarterly windows. Each opens a dataset and a business problem, and gives candidates about two weeks to produce three deliverables: a technical report capped at 1,000 words, working code in R, Python, or SAS, and up to five supporting tables or exhibits. The fee is $700 per attempt. Results follow six to eight weeks after the deadline, so a June window returns a result in August at the earliest.

Generative AI is permitted as a consultation resource. Using it to check syntax or talk through a distributional assumption is allowed; using it to generate model code or draft the report would breach the requirement that submitted work be the candidate's own.

What the Assessment Measures That MAS-I and MAS-II Do Not

The exam content splits into data handling, model diagnostics and selection, and interpretation and presentation, and the difficulty concentrates in the middle.

MAS-I supplies the GLM framing, regression diagnostics, and fit reasoning, but the PCPA applies them at P&C-specific depth: Tweedie distributions for pure premium, Poisson regression with a log-exposure offset for frequency, Gamma for severity, with link function selection tested directly. A candidate who can compute a Pearson residual but cannot read a deviance residual plot, identify overdispersion from a quasi-Poisson comparison, or say why a log link suits severity where an identity link would not will find this section unfamiliar.

Tree-based content goes past MAS-II in structure rather than concept. Random Forest and XGBoost are treated as distinct families, with questions on tuning, how bias and variance move with depth, and how variable importance differs between bagging and boosting. Variable selection covers LASSO and its shrinkage behavior against ridge, stepwise procedures and their overfitting record, principal component analysis, and embedding approaches for high-cardinality categoricals such as territory or vehicle class.

The ethics section is situational rather than general. It tests 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, by describing a fact pattern and asking what a named standard requires. Exam 5's ratemaking material is the recommended companion for a different reason: it supplies the business context the project's dataset is framed in.

The 1,000-word ceiling is the grading mechanism, not an administrative limit. A report that narrates the workflow chronologically spends its budget on data preparation and arrives at model selection rationale with nothing left, which is the content graders are reading for. The five-exhibit cap works the same way: a lift chart against a baseline, a variable importance plot, and two or three diagnostics defend a selection, where five univariate predictor distributions describe an input. Code and report are read against each other, so applying LASSO in the script and describing stepwise inclusion in the report is a gap examiners will find.

What the Credential Does Not Reach

The PCPA closes the gap between statistical theory and applied model building. It does not reach what carriers now expect within the first 18 months in the role.

Model scoring infrastructure sits outside the scope entirely. Passing demonstrates that a candidate can build and validate a model in a two-week project setting. It does not demonstrate version control on a model, drift monitoring against incoming production data, or assembly of a filing package structured for state regulatory review. Those are the capabilities behind the separate model deployment and model risk management lines that appear alongside PCPA on job postings.

Neural networks appear at definitional level. Large language model integration, agentic systems in underwriting and claims, the NAIC Exhibit C documentation framework, and the vendor model evaluation protocols that 12 states are operationalizing through the AI evaluation pilot are not tested. A candidate holding the credential has not shown they can assess a third-party claims triage model. The CAS AI competency framework and Fast Track CE address part of that, but as continuing education rather than PCPA content.

The scheduling constraint is the one that reaches candidates first. Quarterly windows plus a six-to-eight-week grading cycle means a CBT pass in October meets a December window and a February result. That is two quarters between passing an exam and holding the credential, on a requirement that also carries a $1,000 single-attempt cost and a two-week wait between exam attempts. The sequencing has to be planned around a target ACAS date rather than picked up when the exam clears.

P&C Insurance CAS Exams Actuarial Exams Exam Study

Further Reading

Sources