Exam PA is a 3.5-hour open-response exam. You get a business data problem with a set of tasks, and you type your answers in Microsoft Word, with Excel also available. There's no multiple choice. It's an Associate-level (ASA) exam, and most candidates take it after Statistics for Risk Modeling (SRM).

Since the April 2023 sitting, R and RStudio aren't available in the exam. You don't write or run code. The R code and output you need are in the exam materials, and your job is to read them, interpret them and explain what they mean. So what you're practicing is reading output correctly and writing a clear, justified answer around it.

The exam tests three things at once: whether you can follow the predictive modeling workflow end to end, whether you interpret model output and statistical results correctly, and whether you can explain your reasoning to a business stakeholder who doesn't know the math. The third is where a lot of otherwise strong candidates lose points.

What the syllabus covers

The content follows the stages of a predictive analytics project. Read the SOA's current syllabus for the exact scope of your sitting, but at a high level it covers:

  • Problem framing and data: the business problem, the target variable, what the stakeholder needs, and data quality issues like missing values, outliers and the limits of the data
  • Exploratory analysis and feature generation: distributions and relationships, transforming variables, combining or binning levels, and a reason for every choice
  • Supervised models: generalized linear models (GLMs) are central, including link functions, distributions and coefficient interpretation, and decision trees, random forests and boosted trees also appear
  • Unsupervised methods: clustering such as k-means, and dimension reduction such as principal components analysis, as tools for exploration and feature creation
  • Model evaluation and selection: train/test splits, cross-validation, performance metrics, the bias-variance tradeoff, and why one model suits the business goal better
  • Communication: findings and recommendations written for a non-technical reader

I wouldn't obsess over exact weights. PA is built around one integrated case rather than separate topic buckets, and a given sitting can lean on GLMs or on trees depending on the project statement, so treat the whole workflow as fair game.

Hours and a timeline

I'd plan on roughly 100 to 150 hours over 10 to 14 weeks. Where you land depends on how comfortable you are with the SRM material and how fast you write under time pressure.

  1. Weeks 1 to 3: rebuild the statistics. Review GLMs, trees, regularization, resampling and unsupervised methods. This goes quickly if SRM is fresh; if it's been a while, slow down here.
  2. Weeks 4 to 7: learn to read output. Work through R code and output until you can look at a model summary, a confusion matrix or a variable importance plot and say right away what it means for the business problem.
  3. Weeks 8 to 12: full written practice. Do complete past sittings under timed conditions, write real Word responses, then grade yourself against the model solution and rubric.
  4. Final week: tighten how you structure justifications and recommendations, and fix the recurring point leaks your self-grading turned up.

How to practice, and where points go

PA has no multiple-choice bank to grind. What you're training is writing a complete, well-justified answer to an open-ended task and then judging it against the rubric.

That's how I built FreeFellow's PA prep, which isn't connected to the SOA. It centers on free past-exam walkthroughs: for every released sitting from October 2023 on, FreeFellow reproduces the exam from the SOA-published Project Statements and Model Solutions. You write your own answer to each task, then compare it with the model solution, the graders' comments and a point-by-point rubric. There are also 22 free lessons and 500 free concept-check questions for the underlying ideas, all at /free/exam-pa/. AI grading of your written answers is part of Fellow, at five a day, and Fellow Plus removes that daily allowance, though grading rate and usage limits still apply. A free account gets three graded attempts to try it.

My routine:

  • Write your full response before you open the model solution. If you peek early, you train recognition instead of production.
  • Grade against the rubric line by line, and only give yourself credit for what a grader could point to, not for what you meant.
  • Keep a list of the points you miss. The patterns show up fast: not tying a choice back to the business goal, stating a result without interpreting it, skipping the limitation.
  • Practice explaining output someone else produced, since you won't run code.
  • Time yourself. Writing is slower than you think, and you need to know how much you can produce in 3.5 hours.

The points candidates lose most often come from answering the math instead of the business problem. Tasks are framed around a stakeholder's need, and a correct calculation that never connects to that need leaves points on the table. Every transformation, dropped variable and model choice needs a stated reason, because an unjustified choice is marked incomplete even when it's reasonable.

Saying a coefficient is significant isn't enough; explain its direction, its practical meaning and why it matters here. Rubrics often give points for saying what the data can't tell you or where bias could come in. Long, disorganized answers that never reach a clear recommendation cost you, so practice a short summary a non-technical reader could act on. And don't build your prep around live coding, because the exam gives you the output.