Passing SRM takes more than remembering formulas. You need to know how the models differ, read statistical output, spot violated assumptions, and make sensible modeling choices under time pressure — 35 questions in three and a half hours.
This is the final post in my SRM series (Part 1: what the exam is · Part 2: materials). Having passed, this is the strategy I’d hand to a friend starting tomorrow. It’s not the only way to pass and it’s certainly not a guarantee — I prepared around a full-time actuarial job, and plenty of my plan bent to reality along the way. But the principles held up.
1. Begin with the syllabus, not the study manual
Download the syllabus for your sitting and turn each learning outcome into a checklist item. A lesson is done only when you can perform the verb in the outcome: define, calculate, compare, select, describe, interpret.
That distinction matters more than it sounds. Reading an explanation of cross-validation is not the same as being able to compare k-fold with leave-one-out. Memorizing the definition of AIC is not the same as using AIC to pick between two models. The exam tests the verb, not the reading.
2. Allocate time according to the exam weights
Linear models are 40–50% of the exam; decision trees another 20–25%. Two-thirds of the marks live in those two areas. A reasonable starting allocation:
| Topic | Share of study time |
|---|---|
| Basics of statistical learning | 10% |
| Linear models | 40% |
| Time series | 15% |
| Decision trees | 20% |
| Unsupervised learning | 15% |
Adjust toward your weak spots as practice results come in — but don’t let a favourite low-weight topic eat time that belongs to linear models. (Mine was clustering, which I already knew from work and kept “reviewing” because it felt good. It was worth at most five questions.)
3. Learn every topic in three passes
For each syllabus area:
Pass A — understand the idea. What problem does the method solve? What does it assume? Why would someone choose it? Explain it out loud without notes.
Pass B — do the calculation. Practise the mechanical steps the syllabus expects, and know which direction the answer should move when an input changes.
Pass C — interpret the result. Read the tables, plots, and R output. Translate them into a statistical conclusion, then into plain business language. This is the pass most people skip, and it’s where memorized knowledge becomes exam-ready knowledge.
4. Study comparisons, not isolated definitions
SRM is full of methods designed to be confused with each other. Build short comparison tables for the classic pairs: supervised vs. unsupervised, training vs. test error, lasso vs. ridge, confidence vs. prediction intervals, bagging vs. random forests vs. boosting, k-means vs. hierarchical clustering. For each, record the purpose, key assumptions, tuning parameters, and the effect on bias and variance.
These sheets earn their keep twice: building them forces the distinctions into your head, and rereading them in the final week replaces hours of chapter review.
5. Give linear models special treatment
The biggest topic deserves to be broken into smaller units: OLS assumptions and coefficient interpretation; residuals and diagnostic plots; transformations and interactions; hypothesis tests and AIC/BIC model selection; confidence and prediction intervals; GLMs, distributions, and link functions; regularization and KNN.
Don’t move on just because the formulas look familiar. The real checks are: can you recognize a violated assumption from a plot, interpret an interaction correctly, choose an appropriate link function, and explain what a modeling choice does to the results?
6. Practice reading R output
No coding is required, but R output shows up. Get comfortable identifying coefficient estimates and their signs, standard errors and p-values, reference categories for categorical predictors, residual and fit information, AIC, and fitted values with their intervals.
When reviewing worked examples, cover the explanation and write your own conclusion first. Passive reading builds recognition; the exam requires retrieval.
7. Keep an error log
This was the single highest-value habit of my preparation. An error log turns practice questions into a personalized syllabus. For every meaningful miss, record:
| Field | Example |
|---|---|
| Topic | Decision trees |
| Error type | Concept / recall / interpretation / calculation / timing |
| What I did | Confused bagging with random feature selection |
| Correct idea | Random forests use a random subset of predictors at each split |
| Prevention | Added the distinction to my ensembles comparison sheet |
Review the log weekly. If the same error shows up twice, go back to the lesson and drill a small targeted set before returning to mixed practice. The turning point in my own preparation was exactly this: I noticed my practice scores had plateaued, realized I was redoing questions without diagnosing them, and switched from “more questions” to “understand every miss.” Scores moved within two weeks.
8. Move from topic practice to mixed practice
Topic-based questions quietly tell you which method to use before you’ve read the problem. Mixed sets remove that clue — and the real exam is one giant mixed set. A sensible progression: untimed questions after each lesson → timed topic sets → mixed quizzes over completed topics → full 35-question sets → timed simulations under exam conditions.
Review each practice exam at least as seriously as you sat it. For every question you weren’t sure about — including lucky guesses — explain why the right answer is right and why the others are wrong.
A flexible eight-week schedule
A template, not a prescription. More time: spread it out and add review days. Less time: compress the first learning pass, but never cut mixed practice or mocks.
| Week | Main focus | Practice target |
|---|---|---|
| 1 | Statistical-learning basics; start linear models | Short untimed topic sets |
| 2 | OLS, interpretation, diagnostics | Topic quizzes + error log |
| 3 | Model selection, transformations, intervals | Timed linear-model sets |
| 4 | GLMs, regularization, KNN | First mixed quiz |
| 5 | Time series | Topic sets + cumulative review |
| 6 | Decision trees and ensembles | Mixed quiz; revisit linear models |
| 7 | PCA and clustering | First 35-question timed sets |
| 8 | Full review | Simulations, error-log review, light recap |
How to know when you’re ready
Don’t hang your confidence on one practice score — look for a pattern. You’re close when you can finish mixed sets at the required pace, explain why the wrong options are wrong, interpret unfamiliar output without panicking, score consistently rather than swinging between extremes, and end a simulation with time left to revisit flagged questions.
Exam-day time management
Three and a half hours over 35 questions averages six minutes each — but no question deserves exactly six minutes. On the first pass, bank the ones you’re sure of, make a provisional choice and flag anything that needs more work, and never let one ugly calculation swallow the time for three conceptual questions. Leave time to revisit flags and confirm every question has an answer — unanswered questions score as wrong, so answer everything. And don’t switch an answer just because one letter has appeared four times in a row; the syllabus explicitly says the answer key owes you no pattern.
Before the appointment, check the current SOA exam-day information for ID and calculator rules.
Mistakes to avoid
The ones I either made or watched others make: spending weeks producing beautiful notes instead of doing questions; memorizing formulas without ever interpreting a result; ignoring lucky guesses (a right answer you can’t explain is a gap wearing a disguise); taking mock after mock without diagnosing them; giving every topic equal time in defiance of the weights; and buying more resources instead of finishing the ones you have.
Final thoughts
You don’t need to feel perfect on every topic to sit SRM. You need a reliable process: follow the current syllabus, weight your effort where the marks are, practise interpretation as hard as calculation, review mistakes deliberately, and simulate the exam before the real thing. Consistency beats marathon sessions — especially if, like me, you’re studying around a job.
Good luck. Next stop for me: Exam FM in August.
Sources: SOA SRM study page · September 2026 syllabus · SOA exam-day information. One candidate’s approach, not a guarantee — verify all details for your own sitting.