SOA Exam PA (Predictive Analytics) Glossary
33 essential terms and definitions for SOA Exam PA (Predictive Analytics). Each definition is written for exam preparation, covering the concepts as they are tested on the 2026 syllabus.
B
- Bivariate analysis
- Bivariate analysis examines the relationship between two variables, typically a predictor against the target, using scatter plots, side-by-side box plots, or grouped means to judge whether and how the predictor should enter the model.
- Box plot
- A box plot displays a distribution's median, interquartile range, and outliers. Side-by-side box plots of the target across levels of a factor are one of the most common Exam PA exhibits, and tasks ask what the plot implies for including that factor in the model.
- Binning
- Binning converts a continuous variable into a categorical one by grouping its range into intervals. It captures nonlinear effects in a GLM at the cost of losing within-bin information and adding parameters, a tradeoff Exam PA tasks frequently ask candidates to argue.
- Bias-variance tradeoff
- The bias-variance tradeoff says simpler models miss real patterns (bias) while more flexible models chase noise (variance); total prediction error is minimized somewhere in between. It is the organizing argument behind pruning, regularization, and validation-based model choice.
C
- Confusion matrix
- A confusion matrix cross-tabulates predicted versus actual classes at a chosen cutoff, yielding accuracy, sensitivity, and specificity. Tasks often ask which error type matters more for the business problem and how the cutoff should move in response.
- Cross-validation
- Cross-validation estimates out-of-sample performance by repeatedly training on part of the data and scoring on the held-out remainder, most commonly in k folds. It underlies the cp table for trees and the penalty choice in regularized GLMs.
D
- Data dictionary
- The data dictionary is the exam-provided listing of every variable in the dataset with its type, range, and meaning. Tasks expect candidates to reference it when flagging data-quality concerns, proposing transformations, or questioning whether a variable is appropriate to use.
- Dummy (indicator) variable
- A dummy variable encodes one level of a categorical factor as 1 and everything else as 0. A factor with k levels enters a GLM as k minus 1 dummies against a baseline level, and coefficient interpretation is always relative to that baseline.
- Deviance
- Deviance measures how far a fitted GLM is from a saturated model that fits the data perfectly; lower is better. Comparing deviance across nested models, or penalized versions of it like AIC, supports the model-selection arguments Exam PA tasks ask for.
- Decision tree
- A decision tree predicts by recursively splitting the data on predictor values into increasingly homogeneous nodes. Trees are easy to read and present to stakeholders but unstable and prone to overfitting, the tradeoff Exam PA tasks most often probe.
E
- Elbow plot
- An elbow plot graphs within-cluster variation against the number of clusters; the bend where additional clusters stop yielding much improvement suggests a reasonable k. Interpreting one of these exhibits, including its ambiguity, has appeared on released Exam PA sittings.
F
- Feature engineering
- Feature engineering creates or reshapes predictors to make patterns learnable: combining sparse factor levels, binning a continuous variable, creating interaction terms, or deriving ratios. Tasks grade the reasoning for each change, not only the change itself.
G
- Generalized linear model (GLM)
- A GLM relates a function of the target's mean to a linear combination of predictors, with the response following an exponential-family distribution such as Poisson, gamma, or binomial. GLMs are the largest topic on Exam PA, and tasks stress distribution choice, link choice, and coefficient interpretation.
- Gradient boosting machine (GBM)
- A GBM builds trees sequentially, with each new tree fit to the errors of the ensemble so far, controlled by a learning rate, tree depth, and the number of trees. It is often the most accurate tree-based method on Exam PA but the most sensitive to tuning and overfitting.
H
- Hierarchical clustering
- Hierarchical clustering builds a nested tree of clusters by successively merging the closest groups under a chosen linkage, displayed as a dendrogram. Cutting the dendrogram at a height selects the number of clusters after the fact, unlike k-means.
K
- K-means clustering
- K-means partitions observations into k clusters by iteratively assigning each point to the nearest centroid and recomputing centroids, minimizing total within-cluster variation. The number of clusters is chosen by the analyst, often with an elbow plot.
L
- Log transformation
- A log transformation replaces a positive right-skewed variable with its natural logarithm, compressing the long right tail so relationships become closer to linear and variance more stable. Exam tasks expect a written justification, not just the transformation itself.
- Link function
- The link function connects the mean of the response to the linear predictor in a GLM. The log link makes effects multiplicative (natural for claim frequency and severity); the logit link maps probabilities for binary targets. Justifying the link in writing is routinely worth points.
O
- Outlier
- An outlier is an observation far from the bulk of the data. Exam responses are expected to weigh explanations (data error, legitimate extreme value), the effect on the model being fit, and a defensible treatment: keep, cap, transform, or remove with justification.
- Offset
- An offset is a predictor entered with its coefficient fixed at 1, most often log exposure in a log-link count model, so the model predicts a rate per unit of exposure rather than a raw count.
P
- Proxy variable
- A proxy variable is a predictor that stands in for a characteristic the modeler cannot or should not use directly, such as a rating variable that correlates with a protected class. Identifying proxy concerns and their regulatory or ethical implications is a recurring problem-definition task.
- Principal component analysis (PCA)
- PCA re-expresses correlated numeric variables as uncorrelated principal components ordered by the variance they explain, enabling dimension reduction. Exam tasks focus on interpreting loadings, choosing how many components to retain, and explaining the cost in interpretability.
- PCA loading
- A loading is the weight a principal component places on an original variable. Reading a loadings table to name what a component represents (for example, an overall size effect versus a contrast between variable groups) is a standard Exam PA skill.
- Pruning and the complexity parameter
- Pruning cuts back a grown tree to prevent overfitting. In R's rpart output, the complexity parameter (cp) table shows cross-validated error by tree size, and candidates are expected to pick and defend a cp value from it.
R
- Regression vs classification
- Regression problems predict a numeric target (claim cost, length of stay); classification problems predict a categorical target (lapse or persist, fraud or legitimate). Framing the business problem as one or the other is a scored step on Exam PA because it constrains the model family, the evaluation metric, and how results are communicated.
- Regularization
- Regularization shrinks GLM coefficients toward zero by penalizing their size, trading a little bias for less variance. LASSO (an L1 penalty) can set coefficients exactly to zero and so performs variable selection; ridge (L2) shrinks without eliminating. The penalty weight is chosen by cross-validation.
- ROC curve and AUC
- The ROC curve plots sensitivity against one minus specificity across all cutoffs, and the area under it (AUC) summarizes a classifier's ability to rank positives above negatives. An AUC of 0.5 is no better than chance; 1.0 is perfect ranking.
- Random forest
- A random forest averages many trees grown on bootstrap samples, with each split choosing from a random subset of predictors so the trees decorrelate. It usually outperforms a single tree at the cost of interpretability, leaving variable importance as the main lens into the model.
S
- Skewness
- Skewness is asymmetry in a distribution. Insurance amounts are typically right-skewed (a long tail of large values), which motivates log transformations, gamma or Tweedie response distributions in GLMs, and caution with means as summaries.
T
- Target variable
- The target variable (or response) is the quantity a predictive model estimates, chosen during problem definition so that predicting it actually answers the business question. Its type (count, amount, binary outcome) drives the choice between regression and classification and the candidate model families.
- Train/test split
- A train/test split holds out part of the data so the model is judged on observations it never saw. On Exam PA the split is given in the R output; candidates must resist evaluating models on training performance and say why that would mislead.
U
- Univariate analysis
- Univariate analysis examines one variable at a time, using summary statistics, histograms, and bar charts to assess distribution shape, central tendency, spread, missingness, and outliers before any modeling begins.
V
- Variable importance
- Variable importance scores rank predictors by their contribution to a tree ensemble's fit, restoring a measure of interpretability to random forests and GBMs. Exam tasks ask candidates to read an importance plot and reconcile it with business intuition.