Predictive Models
Notes and demos for classical predictive models — GLM / ML siblings of the PA-R study track, modeling app, and SOA research.
A predictive model is a decision aid with an error distribution. Report both, or you are selling point estimates dressed as certainty.
What’s here
Modeling app
Cost-complexity pruning and the reproducible tree workflow live on the modeling-app page. The repository is the code.
GLM to GBM arc
Start with interpretable baselines; graduate to trees and boosting when the residual story demands it — not before.
Uncertainty
Pair averages with intervals. The SOA research page shows quantile GBM in anger; this folder is the quieter practice ground.
Study loop
- Define the decision. Pricing, underwriting triage, reserving support — the metric follows the job.
- Baseline ruthlessly. A well-specified GLM beats a mis-tuned black box in production more often than conference talks admit.
- Validate by slice. Overall AUC can hide sparse-segment failure. Check tails and thin cells.
- Close the loop. Read the FAQ, open the modeling app, then the SOA notebook — theory → practice → uncertainty.
Start with the modeling app, then the SOA research page for intervals.