200+ Free SOA SRM Practice Questions

Pass your SOA Exam SRM Statistics for Risk Modeling exam on the first try — instant access, no signup required.

✓ No registration✓ No credit card✓ No hidden fees✓ Start practicing immediately

200+ Questions

100% Free

1 / 200

Question 1

Score: 0/0

An actuary uses prior claims, driver age, territory, and vehicle type to predict next year's claim count for each policyholder. What type of learning problem is this?

Supervised regression

Supervised classification

Unsupervised clustering

Dimensionality reduction

to track

2026 Statistics

Key Facts: SOA SRM Exam

Multiple-Choice Questions

SOA syllabus

3.5 hrs

Time Limit

SOA syllabus

Grade 6

Passing Standard

SOA grading scale

$357

2026 Exam Fee

SOA fee page

40-50%

Linear Models Weight

SOA syllabus

Prereq for PA

Pathway Role

SOA syllabus

SOA Exam SRM is a 35-question, 3.5-hour preliminary exam. The heaviest weight is Linear Models (40-50%), followed by Decision Trees (20-25%), with Time Series and Unsupervised Learning each at 10-15% and Basics of Statistical Learning at 5-10%. SOA uses scaled grading, and a 6 or higher is a passing result.

Sample SOA SRM Practice Questions

Try these sample questions to test your SOA SRM exam readiness. Each question includes a detailed explanation. Start the interactive quiz above for the full 200+ question experience with AI tutoring.

1An actuary uses prior claims, driver age, territory, and vehicle type to predict next year's claim count for each policyholder. What type of learning problem is this?

A.Supervised regression

B.Supervised classification

C.Unsupervised clustering

D.Dimensionality reduction

Explanation: The model uses historical observations with a known response variable, so it is supervised learning. Because the target is a numerical count rather than a class label, the task is regression. Clustering and dimensionality reduction are unsupervised methods because they do not rely on a labeled response.

2A company segments policyholders into groups with similar purchasing behavior but does not supply any target variable to the algorithm. Which description is most accurate?

A.Binary classification

B.Unsupervised learning

C.Supervised learning

D.Regression modeling

Explanation: No response variable is provided, so the algorithm is learning structure directly from the predictors. That is the defining feature of unsupervised learning. Regression and classification both require labeled outcomes.

3An insurer wants to predict whether a policyholder will lapse within 12 months. Which modeling label best fits the response variable?

A.Clustering

B.Principal component analysis

C.Classification

D.Regression

Explanation: Lapse status is a categorical outcome, typically coded as yes or no. Predicting a category is a classification problem. Regression is reserved for quantitative outcomes, while clustering and PCA are unsupervised techniques.

4Which measure is usually more appropriate for evaluating prediction accuracy when the response variable is continuous?

A.Misclassification rate

B.Gini impurity

C.Silhouette width

D.Test-set mean squared error

Explanation: For a continuous response, squared-error measures directly quantify how far predictions are from actual values. Misclassification rate is for categorical outcomes, Gini impurity is a tree-splitting criterion, and silhouette width is used in clustering. The test set is preferred because it reflects out-of-sample accuracy.

5A model has extremely low training error but performs poorly on new data. Which statement best explains the problem?

A.The model likely has high variance and is overfitting

B.The model likely has high bias and is underfitting

C.The model is necessarily unbiased

D.The model must be unsupervised

Explanation: Very strong in-sample performance paired with weak out-of-sample performance is the classic sign of overfitting. Overfit models tend to adapt too closely to noise in the training sample, which produces high variance. High bias would usually show up as poor performance even on the training data.

6Which statement best describes the bias-variance tradeoff as model flexibility increases?

A.Neither bias nor variance changes systematically

B.Bias tends to decrease while variance tends to increase

C.Both bias and variance tend to decrease

D.Bias tends to increase while variance tends to decrease

Explanation: More flexible methods can capture complicated patterns, which usually reduces bias. The cost is that predictions become more sensitive to the particular training sample, so variance tends to rise. Model selection tries to balance those two forces to minimize test error.

7Why is a single training-set error usually not a reliable estimate of future prediction accuracy?

A.It only applies to unsupervised methods

B.It ignores the response variable entirely

C.It is optimistically biased because the model was fit on that same data

D.It always exceeds test error

Explanation: Training error is computed on the observations used to estimate the model, so it typically looks too favorable. That makes it an overly optimistic estimate of generalization performance. A held-out test set or cross-validation is needed for a better out-of-sample assessment.

8A data scientist wants to estimate out-of-sample error using all observations efficiently because the data set is small. Which resampling approach is most consistent with that goal?

A.Ignoring validation entirely

B.Using training error only

C.Creating a permanently tiny test set of one observation

D.k-fold cross-validation

Explanation: k-fold cross-validation repeatedly trains on most of the data and validates on a held-out fold, so each observation contributes to both training and validation. That generally uses limited data more efficiently than a single split. Relying only on training error would understate prediction error.

9What is the main difference between leave-one-out cross-validation and 10-fold cross-validation?

A.LOOCV uses one observation per validation set, while 10-fold uses larger validation blocks

B.LOOCV is unsupervised, while 10-fold is supervised

C.LOOCV estimates training error, while 10-fold estimates test error

D.10-fold can only be used for classification problems

Explanation: In LOOCV, the model is refit n times, each time leaving out just one observation. In 10-fold cross-validation, the data are partitioned into ten groups and one full group is held out at a time. Both methods target out-of-sample performance, but LOOCV is usually more computationally intensive.

10Which statement about a training/test split is correct?

A.The test set should contain only the response variable

B.The test set should be reserved for final performance evaluation after model tuning

C.The test set should be used repeatedly to choose hyperparameters

D.The test set replaces the need for training data

Explanation: If the test set is used during model tuning, it stops being an independent measure of future performance. A proper workflow uses training data for fitting, often cross-validation for tuning, and the test set only for the final unbiased check. That separation helps prevent data leakage.

About the SOA SRM Exam

Exam SRM tests the core statistical learning methods actuaries use before moving into predictive analytics. The 2026 syllabus emphasizes linear models most heavily, then decision trees, time series, and unsupervised learning, with a smaller foundational section on model assessment and the bias-variance tradeoff.

Questions

35 scored questions

Time Limit

3 hours 30 minutes

Passing Score

Grade 6 or higher (scaled)

Exam Fee

$357 (Society of Actuaries (SOA))

SOA SRM Exam Content Outline

5-10%

Basics of Statistical Learning

Supervised vs. unsupervised learning, model accuracy, bias-variance tradeoff, and cross-validation methods.

40-50%

Linear Models

OLS and GLMs, exponential-family links, diagnostics, model selection, intervals, regularization, and K-nearest neighbors.

10-15%

Time Series Models

Stationarity, autocorrelation, random walks, exponential smoothing, autoregressive models, and ARCH/GARCH concepts.

20-25%

Decision Trees

Tree construction, pruning, regression and classification trees, bagging, boosting, random forests, and tree-vs-linear-model tradeoffs.

10-15%

Unsupervised Learning Techniques

Principal components, loading interpretation, explained variance, k-means clustering, hierarchical clustering, and cluster-count selection.

How to Pass the SOA SRM Exam

What You Need to Know

Passing score: Grade 6 or higher (scaled)
Exam length: 35 questions
Time limit: 3 hours 30 minutes
Exam fee: $357

Keys to Passing

Complete 500+ practice questions
Score 80%+ consistently before scheduling
Focus on highest-weighted sections
Use our AI tutor for tough concepts

SOA SRM Study Tips from Top Performers

1Treat linear models as the anchor domain: GLMs, diagnostics, AIC/BIC, and coefficient interpretation should be automatic.

2Practice translating between model scale and business scale, especially with log links, logits, and interaction terms.

3Use timed sets to build comfort with short calculations such as AR(1) forecasts, PCA variance explained, and tree-leaf predictions.

4When reviewing misses, classify them as assumption, interpretation, or model-selection mistakes so the pattern is visible.

5Finish with mixed sets that force you to switch quickly between linear models, trees, time series, and clustering.

Frequently Asked Questions

How many questions are on the SOA SRM exam?

The January 2026 SOA SRM syllabus states that Exam SRM is a 35-question multiple-choice computer-based exam. SOA also notes that a small number of pilot questions may appear and are not scored. For pacing, candidates should still plan for a 35-question, 3.5-hour sitting.

What score do you need to pass SOA Exam SRM?

SOA reports preliminary exam grades on a 0-10 scale, and a grade of 6 or higher is a pass. The exact effective pass mark varies by administration rather than being a single permanent percentage. That means you should think in terms of scaled passing performance, not a fixed published cutoff.

Which topic matters most on SRM?

Linear Models is by far the largest domain at 40-50% of the exam, so it should anchor your study plan. Decision Trees is the next biggest block at 20-25%. Time Series and Unsupervised Learning are each 10-15%, while Basics of Statistical Learning is 5-10%.

What changed for SRM in 2026?

The official SOA 2026 fee page lists the standard Exam SRM fee at $357, and SOA announced a temporary refund/transfer policy expansion through December 31, 2026 that includes SRM under specified conditions. The latest exam-specific update on the SRM page itself remains the July 11, 2025 sample-question update rather than a new 2026 syllabus restructuring.

How should I study for SOA SRM?

Start with linear models because they drive almost half the exam, then add decision trees and time-series forecasting, and finish with PCA/clustering review. Timed mixed-topic practice is important because SRM rewards interpretation and model-selection judgment as much as memorized definitions. Use detailed review on every miss so you can connect diagnostics, assumptions, and business context.