PracticeBlogFlashcardsEspañol
All Practice Exams

100+ Free ATPA Practice Questions

Pass your SOA Exam ATPA - Advanced Topics in Predictive Analytics exam on the first try — instant access, no signup required.

✓ No registration✓ No credit card✓ No hidden fees✓ Start practicing immediately
100+ Questions
100% Free
1 / 100
Question 1
Score: 0/0

In a Generalized Linear Model, what is the role of the link function?

A
B
C
D
to track
2026 Statistics

Key Facts: ATPA Exam

~$1,800

Typical Fee

SOA Exam ATPA page

~6 weeks

Take-Home Window

SOA ATPA structure

8 modules

e-Learning Components

SOA ATPA syllabus

Pass/Fail

Grading

SOA

PA + SRM

Required Prior Exams

SOA ASA pathway

100-200 hrs

Typical Prep Time

Candidate self-reports

ATPA is the post-PA capstone in the SOA's predictive analytics track. Candidates work through eight self-paced e-Learning modules and complete an open-book take-home assessment over roughly six weeks. The fee is approximately $1,800. Topics include a GLM refresher (link functions, exponential family, deviance), decision trees and random forests, gradient boosting with XGBoost-style hyperparameters, neural network basics, K-means and hierarchical clustering, principal components analysis, communication and reporting of predictive models, and ethics, bias, and model risk management aligned with ASOPs 23, 41, and 56 and the NAIC AI Bulletin (December 2023). Prerequisites include VEE Math Stats, Exam P, Exam SRM, and Exam PA.

Sample ATPA Practice Questions

Try these sample questions to test your ATPA exam readiness. Each question includes a detailed explanation. Start the interactive quiz above for the full 100+ question experience with AI tutoring.

1In a Generalized Linear Model, what is the role of the link function?
A.It rescales the response variable to have unit variance.
B.It connects the linear predictor to the expected value of the response.
C.It transforms predictors so they become orthogonal to one another.
D.It selects the loss function used during gradient descent.
Explanation: The link function g satisfies g(E[Y|X]) = X*beta, mapping the mean of the response to the linear predictor. This lets a GLM model bounded or non-Gaussian responses (counts, binary, positive) while keeping the right-hand side linear in the parameters.
2Which link function is the canonical link for a Bernoulli response in a GLM?
A.Identity link
B.Log link
C.Logit link
D.Inverse link
Explanation: For the Bernoulli (binomial with n=1) distribution, the canonical link derived from the exponential-family form is the logit link, log(p/(1-p)). The probit link is also used for binary data but is not the canonical link.
3An actuary fits a Poisson GLM for claim frequency with a log link and includes log(exposure) as an offset. What does the offset accomplish?
A.It forces the coefficient on exposure to equal one so that fitted counts scale linearly with exposure.
B.It regularizes the intercept toward zero.
C.It changes the response from frequency to severity.
D.It removes exposure from the model entirely.
Explanation: An offset adds a known term to the linear predictor with an implicit coefficient of 1. With a log link and log(exposure) offset, expected counts scale proportionally with exposure, so the remaining coefficients model the rate per unit of exposure.
4Which family of distributions does the GLM exponential family NOT include in its standard form?
A.Normal
B.Gamma
C.Cauchy
D.Poisson
Explanation: Normal, Gamma, Poisson, binomial, and inverse Gaussian all belong to the exponential dispersion family used by GLMs. The Cauchy distribution does not have a moment-generating function and is not a member of this family.
5Two nested GLMs are compared. Model A has 8 parameters and deviance 320; Model B nests A and has 11 parameters and deviance 305. Using a likelihood-ratio approach with chi-square, what is the test statistic?
A.3 with 11 degrees of freedom
B.15 with 3 degrees of freedom
C.15 with 8 degrees of freedom
D.305 with 11 degrees of freedom
Explanation: The likelihood-ratio statistic for nested GLMs is the difference in deviances, 320 - 305 = 15, with degrees of freedom equal to the difference in parameters, 11 - 8 = 3. Compare 15 to a chi-square with 3 df to decide whether the extra parameters are justified.
6Which model selection criterion penalizes the log-likelihood by 2k, where k is the number of parameters?
A.BIC
B.AIC
C.Adjusted R-squared
D.Mallows Cp
Explanation: AIC = -2 log L + 2k. BIC instead uses k * log(n) as its penalty, which grows with sample size and tends to favor smaller models than AIC.
7An actuary models pure premium with a Tweedie GLM (1 < p < 2) and a log link. Why is Tweedie attractive for this task?
A.It has a closed-form solution that is faster than OLS.
B.It models a positive mass at zero combined with a continuous positive component.
C.It guarantees that residuals are normally distributed.
D.It is the only family that supports a log link.
Explanation: Tweedie distributions with 1 < p < 2 are compound Poisson-Gamma, naturally giving a point mass at zero (no claim) plus a continuous positive distribution for claim amounts. That matches the structure of pure premium per exposure unit.
8In a binary classification GLM, an actuary observes that fitted probabilities cluster heavily near 0 and 1, the Hessian is nearly singular, and one coefficient grows without bound during fitting. What is the most likely diagnosis?
A.Heteroskedasticity
B.Quasi-complete or complete separation
C.Underdispersion
D.Multicollinearity among numeric predictors
Explanation: When a predictor (or combination) perfectly or nearly perfectly separates the two classes, the maximum likelihood estimate diverges and the standard errors blow up. Common fixes include penalized logistic regression (ridge), Firth's correction, or merging sparse categories.
9An actuary fits a Gamma GLM with a log link to claim severity. How should a coefficient of 0.18 on a binary predictor be interpreted?
A.Expected severity is 0.18 higher when the predictor equals 1.
B.Expected severity is multiplied by exp(0.18), about 1.20, when the predictor equals 1.
C.The probability of a claim is 18% higher when the predictor equals 1.
D.The variance of severity is reduced by 18% when the predictor equals 1.
Explanation: With a log link, exp(beta) gives the multiplicative effect on the expected response. A coefficient of 0.18 multiplies expected severity by exp(0.18) approximately 1.197, a roughly 20% lift.
10In a Poisson GLM the dispersion statistic (Pearson chi-square divided by residual degrees of freedom) is 3.2. What does this suggest?
A.The model fits very well; no further action is needed.
B.Overdispersion; standard errors are likely understated and a quasi-Poisson or negative binomial may be preferred.
C.Underdispersion; the Poisson assumption is too flexible.
D.Perfect separation in the response.
Explanation: A Poisson model assumes mean equals variance, so a dispersion statistic well above 1 indicates overdispersion. Quasi-Poisson scales standard errors; negative binomial introduces an extra variance parameter and is often a better generative match.

About the ATPA Exam

SOA Exam ATPA Advanced Topics in Predictive Analytics is a required ASA-pathway component delivered as a series of e-Learning modules plus a multi-week take-home assessment. It builds on Exam SRM and Exam PA, deepening coverage of GLMs, tree-based models, gradient boosting, neural networks, unsupervised learning, model communication, and ethical model governance.

Assessment

Take-home assessment + 8 e-Learning modules

Time Limit

~6 weeks (open-book)

Passing Score

Pass/Fail (graded by SOA)

Exam Fee

~$1,800 (Society of Actuaries (SOA))

ATPA Exam Content Outline

15%

GLM Refresher

Link functions (logit, probit, log, identity), the exponential family and canonical parameter, deviance, AIC and BIC for model comparison, offsets, and coefficient interpretation.

15%

Decision Trees and Random Forests

CART splitting criteria (Gini, entropy, MSE), cost-complexity pruning, the bias-variance tradeoff, bagging, mtry, out-of-bag (OOB) error, and feature importance interpretation.

15%

Gradient Boosting and Ensemble Methods

XGBoost hyperparameters (learning_rate, max_depth, n_estimators, subsample, colsample_bytree, gamma, lambda, alpha), early stopping, stochastic gradient boosting, stacking, and ensemble selection.

10%

Neural Networks and Deep Learning Basics

Forward and backpropagation, activation functions (ReLU, sigmoid, tanh, softmax), Adam and SGD optimizers, dropout and weight decay, vanishing gradients, and feature scaling.

10%

Cluster Analysis

K-means (within-cluster sum of squares, elbow method, silhouette), hierarchical clustering (single, complete, average, Ward linkage), dendrogram cutting, and density-based methods such as DBSCAN.

10%

Principal Components Analysis (PCA)

Eigenvalues and variance explained, scree plots, loadings interpretation, varimax rotation, scaling requirements, SVD computation, and using PCA components in downstream supervised models.

15%

Communication and Reporting Predictive Models

Executive summaries, partial dependence and SHAP plots, ROC, calibration, lift and gains charts, confusion matrices, F1, and precision/recall framing for non-technical and technical audiences.

10%

Ethics, Bias, and Model Risk Management

ASOPs 23, 41, and 56; the SOA Code of Professional Conduct; fairness metrics (demographic parity, equal opportunity, equalized odds, four-fifths rule); model governance, drift monitoring, and the NAIC AI Bulletin (December 2023).

How to Pass the ATPA Exam

What You Need to Know

  • Passing score: Pass/Fail (graded by SOA)
  • Assessment: Take-home assessment + 8 e-Learning modules
  • Time limit: ~6 weeks (open-book)
  • Exam fee: ~$1,800

Keys to Passing

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

ATPA Study Tips from Top Performers

1Refresh your Exam PA notes on GLMs, link functions, deviance, and CART before tackling the ATPA modules - the assessment assumes that PA-level fluency.
2Build at least one end-to-end XGBoost workflow on a tabular dataset, tuning learning_rate, max_depth, subsample, colsample_bytree, and using early stopping.
3Practice writing one-page executive summaries that lead with a recommendation, follow with key drivers, then validation, then limitations - that is the ATPA-graded structure.
4Memorize the scope of ASOPs 23 (Data Quality), 41 (Communications), and 56 (Modeling), plus the December 2023 NAIC AI Bulletin's expectations for AI governance.
5Use cross-validation that respects time order on any time-series example, and pre-commit to fairness metrics before reviewing group-level results to avoid p-hacking the audit.

Frequently Asked Questions

Is SOA Exam ATPA open-book?

Yes. ATPA is administered as a take-home assessment that runs roughly six weeks and is open-book in the sense that candidates may consult permitted reference materials. The work submitted, however, must be the candidate's own and must comply with SOA's exam policies and the Code of Professional Conduct. Misuse, including unauthorized collaboration, can lead to disqualification and disciplinary action.

What does ATPA cost and how is it administered?

ATPA is delivered through the SOA's e-Learning platform (eight modules) and a take-home assessment, with a typical fee of about $1,800. Pricing and exact module count can change between sittings, so always confirm current cost and structure on the official SOA Exam ATPA page before registering.

What are the prerequisites for SOA Exam ATPA?

ATPA assumes the foundation built by VEE Mathematical Statistics, Exam P (Probability), Exam SRM (Statistics for Risk Modeling), and Exam PA (Predictive Analytics). The syllabus expects working knowledge of GLMs and tree-based models from PA, so most candidates take ATPA after passing PA.

How does ATPA differ from Exam PA?

Exam PA is a proctored Prometric written-answer exam focused on the standard predictive analytics workflow. ATPA is an e-Learning plus take-home assessment that goes deeper on advanced topics including gradient boosting (XGBoost-style hyperparameters), neural networks, advanced unsupervised learning, communication of complex models, and ethics, bias, and model risk management.

How is ATPA graded?

ATPA results are reported as Pass/Fail by the SOA. Graders evaluate the take-home submission against the published learning objectives, including modeling rigor, validation, communication clarity, and adherence to professional standards. Detailed numeric grading is not published the same way exam scores are reported for traditional Prometric sittings.

How long does it take to prepare for ATPA?

Most candidates spend roughly 100 to 200 study hours after Exam PA, in addition to the time required to complete the eight e-Learning modules and the assessment itself. Those new to gradient boosting, neural networks, or formal ethics topics (ASOPs 23, 41, 56; NAIC AI Bulletin) typically need the higher end of that range.