SOA PAPredictive Analytics
Syllabus learning objectives
Paraphrased from the SOA syllabus so the study-plan builder and practice sets track the topics you will actually be examined on. Weights are the official topic ranges.
Problem definition & data exploration
- Translate a business problem into a well-defined predictive modeling question with clear target and success metrics.
- Assess data quality, identify missing values, outliers and data leakage risks before modeling.
- Perform exploratory data analysis, including univariate and bivariate summaries and visualizations, to inform feature choices.
- Engineer and transform predictor variables (binning, interactions, encoding of categorical variables) appropriately for the chosen model.
Model building: GLMs, trees, regularization
- Specify and fit a GLM with an appropriate distribution and link function for the target variable.
- Build and tune decision-tree-based models (single trees, random forests, gradient boosting) with cross-validation.
- Apply regularization techniques and variable-selection procedures to control overfitting.
- Compare candidate models using appropriate metrics and select a final model justified by business and statistical criteria.
Validation & business communication
- Validate a fitted model on hold-out or cross-validated data and assess stability of results.
- Interpret and communicate model results (coefficients, relativities, variable importance) to a non-technical business audience.
- Recommend business actions supported by the model output, articulating limitations, assumptions and risks.
- Document the modeling process (data, decisions, diagnostics) in a clear written report.
Lecture videos for this exam
Open the full video library →Webinar: Machine Learning in Reserving on 15 July 2025
International Actuarial Association · Machine Learning in Actuarial Work
Machine Learning to Predict Underwriting Decisions for Life and Health Insurance – ICA2023
Actuaries Institute · Machine Learning in Actuarial Work
6.0001 Introduction to Computer Science and Programming in Python. Fall 2016
MIT OpenCourseWare · Programming (Python) · playlist
Neural networks
3Blue1Brown · Neural Networks & Deep Learning · playlist
MIT 6.S191: Introduction to Deep Learning
Alexander Amini · Neural Networks & Deep Learning · playlist
Overview
PA is a project-style exam: you receive a business problem, a dataset description, and R output, and write a report with model choices and recommendations. The grading rewards judgment and communication as much as technique.
- Duration
- 3.5 hours (plus 15-minute read)
- Questions
- One case study, multi-part written response
- Style
- Computer-based, R output provided
- Passing
- Scaled 6 of 10
Syllabus map
Key formulas
Key ratios: RMSE, MAE, AUC, deviance; Gini / lift: sort by prediction, cumulative capture of actuals.
Link/distribution choices: gamma-log for positive skewed severities; Poisson-log with exposure offset for counts; binomial-logit for binary; Tweedie for pure premium.
Study strategy
Practice the SOA's released PA projects under timed conditions, writing the report exactly as graders expect (executive summary first).
Learn to critique a data field: leakage, granularity, missingness, target encoding.
Develop a repeatable structure: problem → data → method → validation → recommendation → limitations.
Common traps
Recommending a model without explaining it in business terms.
Ignoring variables that leak the target (e.g., claim payment when predicting claim occurrence).
Comparing models on training rather than test performance.