Predictive Modeling FAQ
Clear answers for curious, rigorous readers — the questions this site keeps circling in research, Exam PA prep, and real consulting work.
1. What is predictive modeling, in plain language?
It is the craft of using historical data to estimate something useful about the future or the unseen — a probability of claim, an expected loss, a price, a readmission risk — while being explicit about uncertainty and the decisions the estimate will support.
2. How is this different from “fortune telling”?
Fortune telling sells certainty. Predictive modeling sells calibrated uncertainty: loss functions, holdout tests, residual diagnostics, governance, and the humility to say when the data cannot answer. Actuaries win trust by being wrong in measured, documented ways — not by never being wrong.
3. GLM, GBM, neural net — how do I choose?
GLMs shine when interpretability, regulatory narrative, and sparse signal matter. Tree ensembles / GBMs capture interactions and nonlinearities with less feature engineering. Neural nets earn their keep on unstructured inputs (text, images) or huge tabular regimes. Start with the decision and the audience, not the trendiest acronym.
4. What should I check before I trust a lift chart?
Leakage, temporal ordering, segment stability, calibration (not only discrimination), and whether the business action implied by the ranking is actually feasible. A beautiful ROC curve that cannot change a workflow is decoration.
5. How do prediction intervals differ from confidence intervals?
Confidence intervals speak about parameters (or sometimes mean response). Prediction intervals speak about a new observation. Quantile regression and conformal methods are practical tools when you need “where will the next point fall?” rather than “where is the mean?” — a theme in the SOA research write-up on this site.
6. What is data leakage, and why does it haunt Exam PA stories?
Leakage is accidental information from the future (or from the label itself) sneaking into features. It makes offline metrics glow and production models stumble. Time-based splits, careful feature dictionaries, and a paranoid friend in code review are the cure.
7. How do class imbalance and missingness change the story?
Rare events need metrics beyond accuracy (PR curves, expected cost, proper scoring rules). Missingness needs a model of why it is missing — mean imputation that ignores mechanism can invent bias. Weighting, resampling, and explicit missing indicators each have a place; none is universal.
8. Where should I practice hands-on?
Try the interactive Predictive Modeling App, the SOA research narrative, GLM case studies under actuarial exams, and the PA-R Study Manual for structured Exam PA practice.
9. How does this connect to quantum and language pages?
Quantum literacy sharpens how you talk about uncertainty and representation. Local language models (Dolphin) speed exploration — but they do not replace validation. Predictive judgment stays with you.
10. Still stuck?
Browse the Open Questions Hub, skim About, or reach out via Contact. The best modeling questions are usually decision questions in disguise.