Associate of the Society of Actuaries (ASA/FSA) Exam PA — Predictive Analytics Syllabus

Exam PA — Predictive Analytics syllabus — full chapter and topic list

Expand any chapter to see its topics and sub-topics. This is the whole examinable outline for Exam PA — Predictive Analytics in Associate of the Society of Actuaries (ASA/FSA), not a summary of it.

1. 1

   Problem Definition and Data Preparation

   3 topics

   • Framing the Business Problem
     • Stakeholder objectives and constraints
     • Defining the target variable
   • Data Quality and Cleaning
     • Handling missing values
     • Outlier detection and treatment
     • Data dictionary and validity checks
   • Exploratory Data Analysis
     • Univariate and bivariate exploration
     • Visualization for insight
2. 2

   Feature Engineering and Variable Transformation

   3 topics

   • Handling Categorical Variables
     • Factor levels and combining categories
     • Dummy/one-hot encoding
   • Transforming Numeric Variables
     • Log and power transformations
     • Binning and standardization
   • Dimensionality Reduction in Practice
     • PCA for feature creation
     • Interaction and derived features
3. 3

   Model Construction and Selection

   3 topics

   • Building GLMs
     • Distribution and link selection
     • Interaction terms and offsets
   • Tree-Based Models in Practice
     • Single trees vs. random forests
     • Hyperparameter tuning
   • Model Validation
     • Train/test splits and cross-validation
     • Performance metrics by problem type
4. 4

   Interpretation and Communication

   3 topics

   • Interpreting Model Output
     • Coefficients and relativities
     • Variable importance
   • Communicating Results to Stakeholders
     • Executive summaries
     • Limitations and assumptions
   • Ethical and Professional Considerations
     • Bias, fairness, and proxy discrimination
     • Data privacy and governance

Exam PA — Predictive Analytics flashcards for Associate of the Society of Actuaries (ASA/FSA)

23 of 66 cards from the Exam PA — Predictive Analytics deck — real questions with worked answers.

1. In the SOA Exam PA framing process, what is the difference between a business problem and a predictive modeling problem?

   A business problem is the stakeholder's real-world objective (e.g., reduce customer churn). A predictive modeling problem is the technical translation of that objective into a specific target variable, predictors, and model type (e.g., predict the probability a policyholder lapses within 12 months). Framing converts the former into the latter.

2. What are the key components that must be specified when framing a predictive analytics problem?

   The business objective, the target (response) variable, candidate predictors, the unit of observation (granularity), the success metric, available data sources, constraints (time, regulatory, interpretability), and how the model output will be used and by whom.

3. Distinguish a supervised from an unsupervised learning problem in framing.

   Supervised learning has a labeled target variable to predict (regression for numeric, classification for categorical). Unsupervised learning has no target variable and instead finds structure in the data (e.g., clustering, principal components / dimensionality reduction).

4. What is the difference between a prediction-focused and an interpretation-focused modeling goal, and how does it affect model choice?

   A prediction-focused goal prioritizes accuracy on new data and may favor complex 'black-box' models (e.g., boosted trees). An interpretation-focused goal prioritizes understanding driver effects and favors transparent models (e.g., GLMs). The goal drives the accuracy-vs-interpretability trade-off.

5. Define a 'stakeholder' in the context of an actuarial predictive analytics project.

   Any party with an interest in or affected by the model and its results, such as management, underwriters, regulators, customers, and IT. Identifying stakeholders early shapes the success metrics, interpretability needs, and ethical considerations.

6. What is the difference between structured and unstructured data?

   Structured data is organized into a defined tabular format with rows (records) and columns (fields), e.g., a database table. Unstructured data has no predefined schema, e.g., free text, images, or audio, and typically requires extra processing to use in a model.

7. List common data quality issues that must be addressed before modeling.

   Missing values, outliers and erroneous entries, duplicate records, inconsistent formatting or units, mislabeled categories, structural errors (typos), and data leakage. Each can bias estimates or invalidate the model.

8. What is data leakage and why is it dangerous?

   Data leakage occurs when information that would not be available at prediction time (e.g., the target itself or a post-outcome variable) is included among the predictors. It produces unrealistically good validation performance and a model that fails in production.

9. Compare the three common missing-data mechanisms: MCAR, MAR, and MNAR.

   MCAR (Missing Completely At Random): missingness is independent of all data. MAR (Missing At Random): missingness depends only on observed variables. MNAR (Missing Not At Random): missingness depends on the unobserved value itself. MNAR is the hardest to handle and can bias results even after imputation.

10. What are common strategies for handling missing values?

    Deletion (listwise/complete-case or dropping a sparse variable), imputation (mean/median/mode, regression, or k-NN), creating a 'missing' indicator flag, or using models that handle missingness natively (e.g., some tree implementations). Choice depends on the missingness mechanism and amount.

11. How can outliers be detected during data cleaning?

    Visually via boxplots and histograms, statistically via z-scores ($|z| > 3$) or the IQR rule (points beyond $Q_1 - 1.5\,\mathrm{IQR}$ or $Q_3 + 1.5\,\mathrm{IQR}$), and through domain-knowledge range checks.

12. State the IQR rule for identifying outliers.

    With $\mathrm{IQR} = Q_3 - Q_1$, a value is flagged as an outlier if it lies below $Q_1 - 1.5 \times \mathrm{IQR}$ or above $Q_3 + 1.5 \times \mathrm{IQR}$. Using a multiplier of $3$ instead identifies 'extreme' outliers.

13. What is the goal of Exploratory Data Analysis (EDA)?

    To understand the data before modeling: examine distributions, detect data quality issues, identify relationships between predictors and the target, spot collinearity, and generate hypotheses that inform feature engineering and model selection.

14. Distinguish univariate, bivariate, and multivariate EDA.

    Univariate analysis examines one variable at a time (distribution, central tendency, spread). Bivariate analysis examines the relationship between two variables (e.g., predictor vs target). Multivariate analysis examines three or more variables jointly, including interactions and collinearity.

15. Which graph types are appropriate for each variable-type combination in bivariate EDA?

    Numeric vs numeric: scatterplot. Numeric vs categorical: boxplots or grouped histograms. Categorical vs categorical: contingency table, stacked/grouped bar chart. Target vs predictor plots guide feature transformation.

16. Define skewness and explain how right-skew affects modeling of a numeric variable.

    Skewness measures asymmetry of a distribution. A right- (positive-) skewed variable has a long upper tail (mean > median), common for monetary amounts. It can disproportionately influence linear models, so a log transform is often applied to reduce skew.

17. What is the difference between correlation and causation in EDA?

    Correlation measures the strength of a (linear) association between two variables; causation means one variable directly produces a change in the other. EDA reveals correlations, but causation requires controlled reasoning or experiments. Models built on correlation can mislead if interpreted causally.

18. Why must categorical variables be encoded before use in a GLM?

    GLMs require numeric inputs, so non-numeric (nominal) categories must be converted to numbers. Encoding (typically dummy/one-hot) lets each level contribute its own coefficient without implying a false numeric ordering among levels.

19. Explain one-hot (dummy) encoding and the dummy-variable trap.

    One-hot encoding creates a binary indicator column for each level of a categorical variable. For a variable with $k$ levels, only $k-1$ dummies are used (one level is the baseline/reference); including all $k$ plus an intercept causes perfect collinearity — the 'dummy-variable trap'.

20. What is the difference between a nominal and an ordinal categorical variable?

    Nominal variables have unordered categories (e.g., region, color). Ordinal variables have a meaningful order (e.g., low/medium/high). Ordinal variables can be encoded with integers preserving order; nominal variables should not be integer-coded as that imposes a false ordering.

21. What is the purpose of combining (collapsing) factor levels, and when is it appropriate?

    Combining sparse or similar levels reduces the number of parameters, improves stability, and avoids overfitting on rare categories. It is appropriate when levels have few observations or statistically/practically similar target behavior, often guided by EDA and domain knowledge.

22. What is target (mean) encoding and what is its main risk?

    Target encoding replaces each category level with a summary statistic of the target for that level (e.g., its mean response). Its main risk is target leakage and overfitting, especially for rare levels; it should be computed using only training data, often with smoothing or cross-validation.

23. Why are numeric variables transformed, and name common transformations.

    Transformations reduce skew, stabilize variance, linearize relationships, and reduce the influence of outliers. Common ones: logarithm, square root, Box-Cox, standardization (z-score), min-max scaling, binning/discretization, and polynomial terms.

See more Exam PA — Predictive Analytics flashcards →

Planning Exam PA — Predictive Analytics for Associate of the Society of Actuaries (ASA/FSA)

Exam PA — Predictive Analytics (Associate of the Society of Actuaries (ASA/FSA)) FAQ