Earlier this year I passed the Society of Actuaries’ Statistics for Risk Modeling (SRM) exam — my first SOA exam, taken while working full time as an actuarial associate. While preparing, I found plenty of resource lists and forum threads, but what I really wanted was a simple explanation of what the exam actually tests and how the topics fit together. So that’s what this post is.
This is the first of three posts about my SRM experience. Here I’ll cover the exam itself. Part 2 is about study materials, and Part 3 is the study approach I’d recommend.
What does “SRM” stand for?
SRM is Statistics for Risk Modeling, one of the exams on the SOA’s pathway to the Associate (ASA) designation. It introduces the statistical-learning tools actuaries use to understand data, build models, make predictions, and judge whether a model is any good. It’s also the foundation for the Predictive Analytics (PA) exam, which goes deeper into the same territory.
In plain language, SRM is about questions like these:
- What type of model fits this problem?
- How do I tell whether a model fits the data well?
- When is a simple, interpretable model better than a flexible one?
- How do I validate a model and avoid overfitting?
- What does this output actually mean for the business?
Exam format at a glance
According to the September 2026 SOA syllabus:
| Item | Current format |
|---|---|
| Length | 3 hours 30 minutes |
| Questions | 35 multiple-choice, five options each |
| Delivery | Computer-based testing (CBT) |
| R programming | No coding, but R output may appear for interpretation |
| Assumed knowledge | Calculus, Exam P probability, VEE Mathematical Statistics |
| Leads to | SRM is a prerequisite for Exam PA |
A few unscored pilot questions may be mixed in, and you won’t know which ones they are — so treat every question as real. Unanswered questions count as wrong; the syllabus itself tells you to answer everything.
One nice feature of the CBT format: an unofficial pass/fail result is emailed within about an hour of your appointment. (Procedures change, so confirm the details for your own sitting.)
What topics are tested?
Five areas, with weights from the September 2026 syllabus:
| Topic | Syllabus weight |
|---|---|
| Basics of statistical learning | 5–10% |
| Linear models | 40–50% |
| Time series models | 10–15% |
| Decision trees | 20–25% |
| Unsupervised learning | 10–15% |
Basics of statistical learning is the smallest section on paper, but it supplies the vocabulary for everything else — supervised vs. unsupervised, regression vs. classification, bias–variance trade-off, training and test sets, cross-validation. If overfitting or cross-validation feel fuzzy, every later topic gets harder than it needs to be.
Linear models are the centre of gravity: ordinary least squares, GLMs, link functions, diagnostics, transformations, interactions, hypothesis tests, AIC and BIC, prediction intervals, regularization, and k-nearest neighbours. At 40–50% of the exam, this area deserves the largest share of your study time — and it demands more than memorized formulas. You have to interpret coefficients, diagnostics, and model-selection output.
Time series covers random walks, stationarity, autocorrelation, exponential smoothing, autoregressive models, and ARCH models, plus computing and interpreting forecasts and their intervals.
Decision trees covers how classification and regression trees are grown and pruned, why pruning controls overfitting, and the ensemble methods — bagging, boosting, random forests. The exam likes asking how these compare with linear models: trees capture nonlinearity and interactions naturally; linear models are easier to interpret.
Unsupervised learning means principal component analysis and clustering — how principal components are built and read, how k-means and hierarchical clustering work, and how you choose the number of clusters.
Is SRM a calculation exam or a conceptual exam?
Both, but more conceptual than you might expect. Some questions are calculations, but many test whether you can distinguish similar methods, check assumptions, read output, or pick the right model for a scenario. The syllabus keeps using verbs like compare, describe, select, and interpret — that’s a strong hint about how to study.
Knowing a formula is useful. Knowing when it applies, what each input means, and how the result changes a modeling decision is what actually gets tested.
Do I need to know R?
You don’t write any R code. But the syllabus says questions may present R output for interpretation, so you should be comfortable reading a model summary and picking out coefficients, standard errors, test statistics, p-values, residual information, AIC, and predictions. The goal isn’t to become an R programmer — it’s to translate output into a correct conclusion.
Who is ready to take SRM?
The SOA assumes calculus, probability at the level of Exam P, and mathematical statistics at the VEE level. In practice, you’re in a reasonable position to start if you’re comfortable with random variables and common distributions, expectation and variance, estimation and hypothesis testing, confidence intervals, and basic regression. If those feel rusty, a short review first will save you time later.
My impression after passing
Coming from a data-science background — I work with Python and R and had already used random forests, PCA, and clustering in real projects — a lot of the terminology felt like home. What surprised me was how little that familiarity was worth on its own. The exam doesn’t ask whether you’ve used a random forest; it asks precisely how bagging differs from a random forest, or which assumption a residual plot is violating. Before studying I expected a maths-heavy exam. By exam day I’d understood it’s really a “which model, and why” exam — and the interpretation questions are where preparation pays off.
Final thoughts
SRM introduces a broad toolkit, and the volume of terminology can feel intimidating at first. It gets manageable once you see the connections: model fit, prediction, validation, interpretation, and controlling overfitting run through almost every topic. Start with the current syllabus, read the learning outcomes carefully, and give linear models the time their weight demands.
Next in the series: The study materials question — what’s actually worth using for SRM
Sources: SOA SRM study page · September 2026 syllabus · SOA exam-day information. Exam structures and syllabi change — always check the syllabus for your own sitting.