CAS Exam MAS-II is a 4-hour exam and the fourth preliminary on the CAS path. You take it after Exams P and FM (which the SOA and CAS give jointly) and MAS-I, and before the written-answer exams that start with Exam 5. MAS-I covers probability models, statistics and GLMs, and MAS-II picks up from there, so between them they cover most of the statistics used in P&C actuarial work. Statistical learning is 40-50% of MAS-II, so that's the section to plan your time around. The other three are credibility (15-25%), linear mixed models (10-20%) and time series (15-25%).
There are 628 original practice questions on FreeFellow across all four sections, along with 21 written lessons with charts and worked examples, a formula sheet, all 12 CAS sample questions, and 126 questions from three past CAS exams (Fall 2018, Spring 2019 and Fall 2019), each with its source noted. You can try the samples without an account, and a free one opens the full MAS-II bank.
What MAS-II covers
MAS-II is the last exam on the CAS path without written answers. Along with standard multiple choice, the computer-based exam can include multiple-selection, point-and-click, fill-in and matching items. The CAS doesn't publish the question count, but its practice exams run 45 questions. The 2026 content outline weights the four sections like this:
| Section | Weight | Topics |
|---|---|---|
| A. Introduction to credibility | 15-25% | Classical limited-fluctuation credibility, Buhlmann, Buhlmann-Straub, Bayesian credibility with conjugate priors, credibility-weighted estimates for frequency, severity and aggregate loss |
| B. Linear mixed models | 10-20% | LMM assumptions, fixed vs random effects, REML vs ML estimation, BLUP, intraclass correlation (ICC), hierarchical and nested grouping, residual and Q-Q diagnostics, identifying variance components |
| C. Statistical learning | 40-50% | KNN, decision trees and pruning, tree ensembles (random forest, boosting), PCA, K-means and hierarchical clustering, reading neural network output, predictive-accuracy measures (lift, Gini, AUROC), comparing models with a double lift chart |
| D. Time series with constant variance | 15-25% | AR, MA and ARIMA models, ACF and PACF diagnostics, stationarity, deterministic vs stochastic trends, seasonality through regression and seasonal differencing, building forecasts, reading prediction intervals |
Credibility comes first in the outline and introduces the Bayesian thinking that shows up across the exam. Don't skim it because the weight looks small. It's 15-25% of the exam and unlike anything on MAS-I, and Bayesian credibility (conjugate priors, posterior derivations) is the most-missed topic in candidates' post-mortems. People who skim it lose more points there than in the much bigger statistical learning section.
Linear mixed models carry the least weight but they're dense, and you'll spend more time than you expect learning to read lmerMod-style output. Time series is mostly recognizing the model framework and building forecasts.
MAS-II vs Exam SRM
If you're still choosing between the SOA and CAS, SOA Exam SRM is the closest SOA exam on statistical learning, but the coverage is different:
| MAS-II (CAS) | SRM (SOA) | |
|---|---|---|
| Credibility | 15-25% | Not covered |
| Linear mixed models | 10-20% | Not covered |
| KNN, trees, PCA, clustering | Part of statistical learning (40-50%) | Decision trees 20-25%, PCA and clustering 10-15%, KNN inside linear models |
| Time series | 15-25% | 10-15% |
| Neural networks | Reading output | Not covered |
| Focus | P&C insurance | Any line of business |
If you've passed SRM, you have a head start on KNN, PCA, clustering and time series, but credibility and linear mixed models are new. Coming the other way, SRM goes deeper on tree ensembles than MAS-II does, so that part needs fresh study.
Study hours and a week-by-week plan
There's no official hours figure from the CAS, but candidates and prep providers generally put it at 150 to 200 hours if you have a statistics or applied math degree with regression and machine learning coursework, 250 to 300 if you have solid math and some stats (engineering, econ, an actuarial science undergrad), and 300 to 400 if you've had little stats or machine learning, which covers most career changers and people coming from finance. Most people take 16 weeks at 15 to 25 hours a week. A lot of the statistical learning questions ask you to read output (PCA scree plots, LMM diagnostic plots, ROC and lift curves, double lift charts), and that gets better with regular practice, so studying most days works better than weekend cramming. People who pass on the first attempt usually keep studying 5 days a week through the last 4 weeks.
I'd split the 16 weeks like this:
- Weeks 1 to 3: credibility. Lessons plus 25 to 35 questions a day on classical, Buhlmann, Buhlmann-Straub and Bayesian credibility. Bayesian is the heaviest part, so leave extra time for the conjugate-prior tables.
- Weeks 4 to 5: linear mixed models. Fixed vs random effects, ICC, REML and BLUP shrinkage. Practice reading lmerMod output until picking out the variance components is automatic.
- Weeks 6 to 12: statistical learning. KNN and trees (weeks 6 and 7), ensembles (8), PCA and clustering (9 and 10), neural networks (11), then predictive-accuracy measures including the double lift chart (12). Lift and Gini calculations come up the most.
- Weeks 13 to 14: time series. The ARIMA framework, identifying models from the ACF and PACF, stationarity, seasonality and forecasting. It's computational, and practice recognizing AR(p) and MA(q) models from the diagnostic plots pays off.
- Weeks 15 to 16: mixed practice and past papers. Take both FreeFellow practice exams under realistic conditions, then work the 126 past-paper questions and aim for 70% or better before you sit.
Most of this is free. The 628 questions use the CAS answer-range format on calculation items ("Less than X / At least X, but less than Y / ... / At least W"), the same shape as the real exam, and each has a step-by-step solution and notes on the answer choices. The 21 lessons cover every learning outcome with credibility-factor curves, KNN decision boundaries, decision tree diagrams, PCA scree plots, clustering plots, ROC and lift curves, double lift charts and ACF and PACF correlograms, and you can read them on the web or save them offline. The MAS-II formula sheet has the credibility formulas, ICC, tree summary statistics, PCA loadings, the K-means update step, AUROC and Gini, and ARIMA forecast equations, and it downloads as a print-ready PDF.
The 12 sample questions are reproduced word for word with attribution, including the Systolic Blood Pressure case study two of them use. The 126 past-paper questions are also verbatim, with worked solutions written for FreeFellow, and they include the Spring 2019 case-study supplement (Systolic Blood Pressure again, which Fall 2018 also refers to). The Fall 2019 Warranty Payments case study isn't publicly available, so the two Fall 2019 questions that use it come with a placeholder explaining that. Spring 2018 isn't included because the CAS hasn't released that paper.
The practice exams are part of Fellow, along with spaced-repetition flashcards, topic analytics and a study plan, for $39 a month or $79 a quarter per track. The $199 annual plan is Fellow Plus, which also includes Common traps. Prices are in USD, and each subscription covers one credential track. Check pricing for your track and country.
If you haven't taken MAS-I yet, read the MAS-I study guide first. If you have, the free MAS-II practice questions are the place to start. SOA vs CAS helps if you're still choosing a path, and the best free actuarial prep resources covers the other actuarial exams.
FreeFellow is not affiliated with the Casualty Actuarial Society. CAS does not endorse, promote, review, or warrant the accuracy or quality of FreeFellow products or services.