What Exam PA actually tests

Exam PA (Predictive Analytics) is an Associate-level (ASA) exam that most candidates take after Statistics for Risk Modeling (SRM). I want to be clear about the format up front, because it surprises people: PA is not a multiple-choice exam. It is a 3.5-hour open-response exam. You are handed a business data problem with a set of specified tasks, and you type your written responses in Microsoft Word. Microsoft Excel is also available to you.

Here is the part that changed how I prepare. As of the April 2023 administration, R and RStudio are not available in the exam. You do not write or run code during the sitting. Instead, all relevant R code and output are provided in the exam materials, and your job is to read, interpret, and explain them. That shifts the skill from coding under pressure to reading output correctly and writing a clear, justified narrative around it.

So the exam is really testing three things at once. First, can you follow the predictive modeling workflow end to end. Second, can you interpret model output and statistical results correctly. Third, can you communicate your reasoning to a business stakeholder who does not know the math. That third piece is where a lot of otherwise strong candidates lose points.

Blueprint and topic weights

The SOA publishes the syllabus and the topics it draws from, and I recommend reading their current version for the exact scope for your sitting. At a high level, the content maps to the stages of a predictive analytics project:

  • Problem framing and data. Understanding the business problem, identifying the target variable, and reasoning about what a stakeholder needs. This includes data quality, missing values, outliers, and the limitations of the data.
  • Exploratory data analysis and feature generation. Examining distributions and relationships, transforming variables, combining or binning levels, and justifying every choice you make.
  • Supervised models. Generalized linear models (GLMs) are central, including link functions, distributions, and interpretation of coefficients. Tree-based methods such as decision trees, random forests, and boosted trees also appear.
  • Unsupervised methods. Clustering (for example k-means) and dimension reduction such as principal components analysis show up as tools for exploration and feature creation.
  • Model evaluation and selection. Train/test splits, cross-validation, performance metrics, and the bias-variance tradeoff, plus explaining why one model is preferable for the stated business goal.
  • Communication. Summarizing findings and recommendations for a non-technical audience.

I would not obsess over exact percentage weights, because the SOA structures PA around an integrated case rather than isolated topic buckets. Any single sitting can lean heavier on GLMs or on trees depending on the project statement. Treat the whole workflow as fair game.

Study hours and a realistic timeline

My planning number is roughly 100 to 150 hours across about 10 to 14 weeks. Where you land in that range depends on how comfortable you are with the SRM material and how fast you write under time pressure.

A timeline I have found workable:

  1. Weeks 1 to 3: rebuild the statistics. Review GLMs, trees, regularization, resampling, and unsupervised methods. If SRM is fresh, this goes quickly. If it has been a while, slow down here.
  2. Weeks 4 to 7: learn to read the output. Work through provided R code and output until you can look at a model summary, a confusion matrix, or a variable importance plot and immediately say what it means and why it matters for the business problem.
  3. Weeks 8 to 12: full written practice. Do complete past sittings under timed conditions, writing real Word responses, then grade yourself against the model solution and rubric.
  4. Final week: refine your writing. Tighten how you structure justifications and recommendations. Fix recurring point leaks you have identified in your self-grading.

A practice strategy built around writing, not grinding

Because PA has no traditional multiple-choice bank, grinding questions is the wrong mental model. The thing you are training is the ability to write a complete, well-justified answer to an open-ended task, then judge it honestly.

That is exactly how I structured FreeFellow's PA prep. It centers on free AI-graded past-exam walkthroughs. For every released sitting from October 2023 on, we reproduce the exam from the SOA-published Project Statements and Model Solutions. You write your own answer to each task, then compare it against the model solution, the graders' comments, and a point-by-point rubric. Alongside that there are 22 free lessons and 500 free concept-check questions to lock in the underlying ideas. You can start at /free/exam-pa/.

To be clear, FreeFellow is independent and is not connected to the SOA. The value of the self-grading loop is that it forces you to see the gap between what you wrote and what earns points.

My concrete routine:

  • Write first, look later. Draft 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. Do not give yourself credit for what you meant. Give credit only for what a grader could point to.
  • Log your leaks. Keep a running list of the points you miss. Patterns emerge fast: forgetting to tie a choice back to the business goal, stating a result without interpreting it, or skipping the limitation.
  • Practice the interpretation of provided output. Since you will not run code, get fluent at reading someone else's output and explaining it.
  • Time yourself. Writing is slower than you think. Learn how much you can produce in 3.5 hours.

Common mistakes I see

Answering the math, not the business problem. Tasks are framed around a stakeholder's need. If you compute something correctly but never connect it to that need, you leave points on the table.

Not justifying choices. Every transformation, every variable dropped, every model selected needs a stated reason. Unjustified choices are treated as incomplete even when they are reasonable.

Vague interpretation. Saying a coefficient is significant is not enough. Explain the direction, the practical meaning, and why it matters here.

Ignoring data limitations. Rubrics frequently award points for acknowledging what the data cannot tell you or where bias could enter.

Weak communication. Long, disorganized answers that never surface a clear recommendation cost you. Practice writing a crisp summary a non-technical reader could act on.

Assuming you will code in the exam. You will not. Do not build your prep around live coding when the exam gives you the output to interpret.

Frequently Asked Questions

What is the pass rate for Exam PA?

The SOA publishes pass rates by sitting, and Exam PA has historically landed in a range around the low-to-mid 50 percent area, though it moves from one administration to the next. I would not anchor on a single figure. Check the pass rate the SOA posts for the specific sitting you are targeting, and study to the rubric rather than to a number.

How many hours should I study for Exam PA?

I budget roughly 100 to 150 hours over about 10 to 14 weeks. Because the exam is written response, a large share of that time should go to actually writing answers and comparing them to model solutions, not just reading. If your writing is slow or your statistics feel rusty from SRM, add hours on the higher end.

What topics does Exam PA cover?

The exam covers the full predictive modeling workflow: understanding the business problem, exploring and cleaning data, feature generation, and building and evaluating models such as generalized linear models (GLMs), decision trees, random forests, and unsupervised methods like clustering and principal components. It also tests communication of results to a non-technical audience.

What is the format of Exam PA?

It is a 3.5-hour open-response exam. You receive a business data problem and a set of specified tasks, and you type your answers in Microsoft Word. Microsoft Excel is available. As of April 2023, R and RStudio are not in the exam, so all R code and output you need are provided for you to read and interpret.

Should I take SRM before Exam PA?

Yes, most candidates take Statistics for Risk Modeling (SRM) first. SRM builds the statistical learning foundation (GLMs, trees, regularization, model selection) that PA assumes you already understand. PA then tests whether you can apply that foundation to a real business problem and write it up clearly.