Every released Exam PA sitting from October 2023 on is free on FreeFellow as a graded walkthrough: October 2023, April 2024, October 2024, April 2025, October 2025 and April 2026. There are also 500 free concept-check questions and 22 free lessons. I'm Jeffrey Ting, an FSA and CFA® charterholder, and I built FreeFellow so good exam prep could be free. Further down, I work one bank question and comment on one released task.

What Exam PA is, and why practice looks different

Exam PA (Predictive Analytics) isn't multiple choice. It's a 3.5-hour open-response exam: you're handed a business data problem and type written responses to specific tasks in Microsoft Word, with Microsoft Excel also available, and you're graded on the quality and correctness of what you write. It's an Associate-level (ASA) exam, and most candidates take it after SRM.

Since the April 2023 administration, R and RStudio aren't available during the exam. All the relevant R code and output are in the exam materials, and your job is to read and interpret that output, not write or run code under time pressure. The skills that matter are reading model output, reasoning about it, and writing clear, complete answers to what each task asks.

So PA prep can't look like prep for the multiple-choice exams. There's no giant answer bank that mirrors it. What transfers is writing responses to real released tasks and then grading yourself strictly against the rubric.

The SOA's schedule lists the next sittings as April 13-16, 2027 and October 12-15, 2027, both computer-based. On November 1, 2026, its study page still linked the October 2026 syllabus, so look for an April 2027 version before you start. The study guide covers the syllabus and a timeline.

What's free for PA

Samples are open without signup, and the full free content needs a free account, with no credit card. The full original question bank is free with an account, subject to usage limits. Worked solutions, written lessons, mixed practice, and your readiness score stay free.

Each released administration is reproduced verbatim from the SOA's published Project Statements and Model Solutions. For every task you write your own answer first, then compare it with the model solution, the graders' comments and a part-by-part rubric. You'll find them at the PA walkthroughs. The SOA's Past Exams and Solutions page also lists April 2023 and older sittings, which FreeFellow doesn't carry, and October 2026 wasn't posted as of November 1, 2026.

Because each walkthrough comes from a real released sitting, it maps to the exam almost one to one. The Project Statement is the business problem candidates faced, the tasks are the same tasks, the model solution is the answer the SOA published, and the graders' comments tell you what they valued and penalized. I've split each sitting into its tasks so you can write one at a time, and turned the model solution and comments into a rubric you can score against. The rubric is FreeFellow's, built on the SOA's printed point values, and it isn't SOA grading. FreeFellow's own notes sit in a separate block labeled FreeFellow commentary.

The 500 concept-check questions are short items across the syllabus for drilling the underlying ideas quickly. They're separate from the written exam; use them to confirm you understand a concept before you try to write about it under exam conditions.

The 22 lessons, plus a formula and technique sheet and a glossary, give you the conceptual backbone: what each method does, when to use it, how to read its output and how to talk about it precisely. It's all under the free Exam PA practice page.

How to practice the writing

This is the part candidates skip, and it's what decides your score. Reading a model solution and nodding along won't get you there, so write your own answer cold first.

Here's the loop I recommend:

  1. Open a task from a released sitting. Read the Project Statement context and the specific ask.
  2. Type your full written response, the way you would in Word on exam day, without peeking at the solution. If the task points to R output, work from that output as you would in the real exam.
  3. Only then reveal the model solution, the graders' comments and the rubric.
  4. Score yourself part by part. Where you lost marks, you usually knew the method but didn't say what the rubric wanted, stated a result without the required justification, or skipped a step the graders explicitly rewarded.
  5. Rewrite the weak parts until your answer would earn the points.

For grading, free users get a copy-to-AI prompt for each task, so you can paste your answer and the rubric into your own ChatGPT or Claude and get a structured critique. A free account also gets three AI-graded attempts for the exam. Fellow includes five AI-graded attempts a day inside the platform, against the same rubric. Fellow Plus removes Fellow's 5-attempt daily allowance on exams with AI grading. Grading rate and usage limits still apply. Either way, the standard comes from the published model solution and graders' comments, not from my opinion. The AI scores each part against the rubric, criterion by criterion. It can't tell you how an SOA grader would mark your answer.

Across all the released sittings you'll start to recognize the recurring task types: describe a variable and its issues, justify a modeling choice, interpret a coefficient or a tree split, compare models, recommend and defend. The exam reuses these shapes even as the data changes.

A realistic plan is to get through the concept checks and a few lessons first so the methods are solid, then spend most of your time writing task responses from the released sittings, oldest to newest. Save one full recent sitting to do timed near the end, so you get a feel for the 3.5-hour pacing.

The April 2026 exam's instructions say tasks are graded on how you reasoned, as far as it shows in what you wrote, plus your conclusions and your presentation, and the syllabus gives communication some weight in most tasks. Unless a subtask says otherwise you're writing for the graders, so technical language is fine. Answers go in Word, because graders won't look at an uploaded Excel file.

The question below is from FreeFellow's own concept-check bank. The real exam has no multiple choice, so after the answer I say what I'd want in a written response. That part's my opinion and doesn't reproduce SOA grading.

A predictor the engineers can't explain

A predictive model for equipment breakdown at Northfork Utilities identifies "time since last inspection" as its single strongest predictor. The reliability engineers on the review panel find this surprising and cannot explain a physical mechanism. Which next step is the MOST appropriate response before relying on the model?

  • A. Investigate with the engineers how and when inspections are scheduled, to determine whether the variable encodes a hidden operational process
  • B. Accept the result without discussion, because the data-driven ranking outranks expert opinion
  • C. Remove the predictor, since a variable engineers cannot explain must be spurious
  • D. Increase the training set size until the predictor's importance ranking stabilizes
  • E. Rescale the predictor so its importance is comparable to the other variables

The answer is A. If troubled equipment gets inspected sooner, or a repair resets the inspection date, the variable records how the utility operates as much as how equipment wears out. C is the tempting choice, but engineers being surprised doesn't make a predictor spurious.

A weak written answer says to consult the engineers because domain knowledge matters. A strong one names a specific way the variable could mislead, says what it would check (the scheduling rules, and whether that date is known when the model scores), and says what each result means for keeping the variable. If the score will drive inspections, it also says that acting on the score changes the variable.

The SOA's task: April 2026 Exam PA, Task 5

This is a 5-point task, and the SOA posts the exam and model solution on its past exams page. In my words, you get a fitted linear model and a residual plot for a target that's a percentage. The work is a critique of the distribution and link, a better choice with reasons, then what happens to the meaning of the coefficients. Try it before you read my notes.

My commentary on that task (not from the SOA)

None of this is SOA wording or SOA grading. I'd start with the target. A percentage has a floor and a ceiling, and that fact drives the first two parts. I wouldn't want an answer that lists the Gaussian assumptions from memory, then recommends a family bounded only from below, which deals with negative predictions and still allows values over 100. I'd want each complaint tied to something you can see in the plot, a family and link that respect both limits, and a line on how the target gets rescaled. On the last part, write a sentence for each model, since the task asks for a comparison. Then read the graders' comments in the SOA's solution for what candidates did.

FreeFellow is independent of the Society of Actuaries. The materials I reproduce are the published Project Statements and Model Solutions, and I'd encourage you to read the source documents directly as well.