2026 Virtual Predictive Analytics Bootcamp

Session 1 — Foundations & Elementary GLMs

Tuesday, September 8, 2026 · 12:00–2:30 PM ET · Instructor lead: Dani

Framing

The opening session gets everyone to a shared statistical baseline (univariate stats, OLS, bootstrap) and then moves immediately into the GLM machinery actuaries use every day. Survey feedback flagged the original opening as fast for less-experienced attendees, so the most foundational material moves into pre-work and the live session stays focused on the conceptual leap from OLS to GLM.

Learning objectives

• Frame predictive modeling against causal inference and articulate when each is appropriate.
• Use bootstrap resampling to quantify uncertainty in a risk measure.
• Move from OLS to GLMs and choose between Gamma and Poisson likelihoods for typical actuarial targets (severity vs. frequency).
• Fit, diagnose, and interpret an elementary GLM in R and/or Python.

Live session (150 min)

• Welcome, course logistics, AI-tooling ground rules (10 min)
• Concept block: prediction vs. causation; the role of uncertainty (20 min)
• Concept block: OLS → GLMs (link, family, deviance) — worked example (35 min)
• Break (10 min)
• Hands-on case study: Auto collision severity (Gamma) and claim frequency (Poisson) — students drive in breakouts, instructors circulate (60 min)
• Debrief, common pitfalls, AI-assisted code review demo, Q&A (15 min)

Asynchronous modules (~5 hours total)

• Pre-work video series — Statistics refresher: univariate, multivariate, correlation, OLS (≈60 min, segmented)
• Pre-work video — Bootstrapping for actuaries, with notebook walkthrough (≈30 min)
• Pre-work reading — Short note: prediction vs. causal inference for actuaries
• Pre-work quiz — 10-question self-check covering OLS interpretation and bootstrap mechanics
• Post-session tutorial (Markdown lab) — Self-paced GLM lab: deJong-Heller diabetes data; extension exercise on commercial liability (≈90 min hands-on)
• Post-session code walkthrough video — Line-by-line walkthrough of the GLM tutorial, R and Python tracks (≈45 min)
• Optional deep-dive video — Diagnostics: deviance residuals, Pearson residuals, leverage plots (≈30 min)

AI integration

• Demo: using an AI coding assistant (Claude / Copilot / ChatGPT) to generate exploratory data analysis scaffolding from a dataset description, then critiquing the output.
• Prompt library: short, vetted prompts students can re-use for EDA, link-function selection, and residual diagnostics.
• Discussion: where AI helps vs. where it confidently mis-specifies the model — examples of plausible-but-wrong GLM family choices.

Session 2 — Modern Regression: Advanced GLMs, Regularization & GAMs

Tuesday, September 22, 2026 · 12:00–2:30 PM ET · Instructor lead: Peng

Framing

Session 2 covers the modern regression toolkit an actuary uses to build a defensible model: advanced GLM machinery (Tweedie, frequency-severity, AIC), the bias-variance/regularization lens (Ridge, LASSO, cross-validation), and non-linear extensions via splines and Generalized Additive Models (GAMs). The live session frames these as one continuous workflow — fit, evaluate out-of-sample, regularize, relax linearity — and uses async modules to give each technique its full technical treatment.

Learning objectives

• Specify and fit a Tweedie / compound Poisson-Gamma model for pure-premium prediction.
• Distinguish in-sample fit from out-of-sample generalization; use AIC and cross-validation appropriately.
• Apply Ridge and LASSO penalties and read the regularization path.
• Build polynomial, spline, and GAM models and explain when each is preferable to a linear GLM.
• Fit a penalized GLM and a GAM end-to-end and interpret the selected effects.

Live session (150 min)

• Recap + AI-prompt sharing from Session 1 (10 min)
• Frequency-severity and Tweedie modeling — worked example on French MTPL data (25 min)
• Bias-variance, optimism, cross-validation, and regularization intuition (25 min)
• Relaxing linearity: polynomials → splines → GAMs (25 min)
• Break (10 min)
• Hands-on case study: build a pure-premium model — start with a regularized GLM, then add a GAM smoother — in breakouts (45 min)
• Debrief; AI-assisted model comparison demo; Q&A (10 min)

Asynchronous modules (~5 hours total)

• Pre-work video — Optimism, AIC, and why training error lies (≈25 min)
• Pre-work video — Ridge and LASSO visualized: the regularization path (≈25 min)
• Pre-work video — Splines and GAMs visualized (≈30 min)
• Pre-work reading — Selected pages from ISLR (CV, Ridge, LASSO); short note on GAMs in actuarial pricing
• Pre-work quiz — Self-check on bias-variance, CV, and effective degrees of freedom
• Post-session tutorial (Markdown lab) — Penalized GLM lab: French MTPL revisited, Allstate loss data extension (≈90 min)
• Post-session tutorial (Markdown lab) — GAM lab: mgcv (R) and PyGAM (Python), pricing application (≈60 min)
• Post-session code walkthrough video — End-to-end walkthrough of both labs (≈45 min)

AI integration

• Workflow demo: use an AI assistant to draft a cross-validation harness, then verify it against a reference implementation.
• Exercise: ask an AI to suggest features, interactions, and smooth terms; evaluate which suggestions actually improve out-of-sample deviance.
• Caveat module: regularization paths, leakage, and how AI suggestions can quietly cause data leakage in CV — common failure modes.

Session 3 — Machine Learning: From Algorithmic Learners to AI

Tuesday, October 6, 2026 · 12:00–2:30 PM ET · Instructor lead: Peng / Dani

Framing

Session 3 tells the story of machine learning as a single arc — from Breiman's algorithmic culture (k-NN, SVMs, decision trees) to the ensembles that dominated tabular ML in the 2010s (random forests, gradient boosting), and finally to the representation-learning paradigm that defines modern AI. The aim is to give actuaries a coherent mental map of where the field has gone rather than a deep dive into any one method; trees and boosting are treated as a brief but important stop on that arc, not the centerpiece. The session sets up Session 4, where the AI half of the title is fully unpacked.

Learning objectives

• Articulate Breiman's two cultures and the shift from model-first to algorithm-first thinking.
• Sketch the intuition behind k-NN, SVMs, decision trees, and boosting; identify when each remains a sensible default.
• Explain what changes when learners stop relying on hand-crafted features and start learning representations from data.
• Interpret MAPE, MSE, confusion matrices, ROC curves, AUC, and calibration plots.
• Evaluate the same actuarial task as both a regression and a classification problem and choose appropriately.

Live session (150 min)

• Recap + AI-prompt sharing from Session 2 (10 min)
• Breiman's two cultures and what algorithmic learning buys you that regression doesn't (20 min)
• Algorithmic learners: k-NN, SVM, decision trees — intuition and trade-offs (25 min)
• Ensembles in one breath: bagging, random forests, gradient boosting — the practitioner workhorses (15 min)
• The bridge to AI: from hand-crafted features to learned representations (15 min)
• Break (10 min)
• Hands-on case study: compare a GLM (Session 2), an SVM, and a gradient-boosted model on a frequency task; discuss what each "sees" in the data (45 min)
• Classification metrics & calibration debrief; AI-assisted model comparison demo; Q&A (10 min)

Asynchronous modules (~6 hours total)

• Pre-work reading — Breiman, “Statistical Modeling: The Two Cultures” (excerpt + summary)
• Pre-work reading — Short piece on the conceptual arc from algorithmic learners to representation learning
• Pre-work video — k-NN, SVMs, and decision trees — an actuary's quick tour (≈30 min)
• Pre-work video — Bagging, random forests, and boosting in 25 minutes
• Pre-work video — Classification metrics deep dive: ROC, AUC, calibration (≈30 min)
• Pre-work quiz — Self-check on algorithmic learners, ensembling, and confusion-matrix arithmetic
• Post-session tutorial (Markdown lab) — Compare SVM, random forest, and XGBoost against a baseline GLM on a claim-frequency task; small grid search (≈75 min)
• Post-session tutorial (Markdown lab) — Frequency modeling as a classification task — model card and stakeholder summary (≈60 min)
• Post-session code walkthrough video — End-to-end walkthrough of both labs in R and Python (≈60 min)

AI integration

• Demo: have an AI assistant generate three competing models (logistic GLM, SVM, gradient-boosted) and a single benchmarking harness; students audit and run the comparison.
• Exercise: ask an AI to map the algorithmic-learners-to-AI arc onto a real actuarial workflow at the student's company; critique the answer.
• Discussion: how LLMs summarize ML results, and where their language is misleading (e.g., conflating precision, accuracy, and calibration).

Session 4 — Neural Networks in Actuarial Science

Tuesday, October 13, 2026 · 12:00–2:30 PM ET · Instructor lead: Dani

Framing

Session 4 is dedicated to neural networks in actuarial science: not just "what is a neural net", but where they earn their place in a working actuary's toolkit. We move quickly from perceptron → MLP → deep learning, then spend the bulk of the live session on Combined Actuarial Neural Networks (CANNs) and concrete deployments in pricing, reserving, and mortality. The session also gives a brief, practitioner-level sketch of transformers / LLMs and where each is (and is not) relevant to actuarial work.

Learning objectives

• Build a small feed-forward neural network and reason about depth, width, activation, and regularization choices.
• Explain the architecture and motivation of Combined Actuarial Neural Networks (CANNs) and identify when they outperform a pure GLM or pure deep model.
• Survey concrete neural-network applications in pricing, reserving, and mortality forecasting.
• Sketch the architectures behind LLMs and discuss where they realistically fit in actuarial workflows (text, claims notes, regulatory summarization) vs. where they don't.
• Choose between tree-based and neural approaches based on data shape, interpretability needs, and operational constraints.

Live session (150 min)

• Recap + AI-prompt sharing from Session 3 (10 min)
• Perceptron → MLP → deep learning, framed as flexible regression (25 min)
• Combined Actuarial Neural Networks (CANNs): structure, training, interpretation (25 min)
• Applications in actuarial science: case studies in pricing, reserving, and mortality (25 min)
• Break (10 min)
• Hands-on case study: build a CANN for pure-premium prediction; compare lift and calibration against the GLM and the XGBoost from prior sessions (45 min)
• LLMs & transformers — a brief practitioner sketch and where they fit; Q&A (10 min)

Asynchronous modules (~6 hours total)

• Pre-work video — Neural networks from scratch (visual, minimal calculus) (≈35 min)
• Pre-work video — From neural networks to CANNs: the actuarial twist (≈30 min)
• Pre-work video — Transformers and LLMs in 25 minutes — what an actuary needs to know
• Pre-work reading — Wuthrich & Merz: CANN paper (assigned excerpt); one applied paper on deep reserving or deep mortality
• Pre-work quiz — Self-check on activations, backprop intuition, and CANN structure
• Post-session tutorial (Markdown lab) — CANN pricing lab in Keras and PyTorch (≈90 min)
• Post-session tutorial (Markdown lab) — Deep reserving exercise on a simulated triangle dataset (≈60 min)
• Post-session code walkthrough video — Notebook walkthrough for each lab, with discussion of training pitfalls (≈60 min)
• Optional deep-dive — Recorded conversation: "When would I actually use a neural net at work?" — practitioner panel (≈30 min)

AI integration

• Demo: ask an AI assistant to translate an XGBoost model into a PyTorch neural net of similar capacity; discuss what's preserved and what's lost.
• Exercise: use an AI to generate SHAP-style explanations of the trained network and a critique of those explanations.
• Discussion: when an LLM-generated training loop is fine, and when it silently introduces a bug (label leakage, wrong loss, missing eval set, overfitting via early stopping mis-configuration).

Session 5 — Responsible AI: Applications, Limits & Outlook

Tuesday, October 20, 2026 · 12:00–2:30 PM ET · Instructor lead: Dani / Peng

Framing

The closing session covers responsible-use questions: now that we can build these models, when should we, and what does deployment actually look like? It combines the practical/regulatory layer (black-box concerns, explainable AI, fairness) with a closing perspective on AI's limits, common "snake-oil" patterns, and the deployment realities of putting ML into a regulated insurance environment. This is the responsible-use and outlook content that survey respondents most clearly wanted more of.

Learning objectives

• Apply at least two explainability techniques (SHAP, PDPs, surrogate models) to a fitted ML model.
• Define and compute common fairness metrics; articulate trade-offs against accuracy.
• Identify common pitfalls and "AI snake oil" patterns in industry pitches and vendor tooling.
• Sketch a responsible deployment plan: monitoring, governance, human-in-the-loop, AI tool policy.

Live session (150 min)

• Black-box challenge in regulated insurance contexts; current regulatory landscape (20 min)
• Explainable AI: SHAP, partial dependence, surrogate models — hands-on demo (25 min)
• Fairness criteria, why they conflict, and the accuracy-fairness frontier (25 min)
• Break (10 min)
• Hands-on case study: auto data with protected characteristics — trade off fairness and accuracy in breakouts (45 min)
• Limitations & AI snake oil; deployment, monitoring, and AI tool policy in actuarial teams (20 min)
• Wrap-up, next steps, alumni community (5 min)

Asynchronous modules (~5 hours total)

• Pre-work reading — Algorithmic bias playbook excerpt; recent insurance-regulation overview (CO, NY, NAIC)
• Pre-work reading — Selected piece on AI snake oil and where ML fails
• Pre-work video — Explainable AI: SHAP, PDPs, surrogate models (≈30 min)
• Pre-work video — Fairness criteria: where they come from and why they conflict (≈30 min)
• Pre-work quiz — Self-check on fairness criteria and their incompatibility results
• Post-session tutorial (Markdown lab) — Fairness lab: extended COMPAS / auto-pricing exercise with multiple fairness criteria (≈90 min)
• Post-session tutorial (Markdown lab) — Model-card and governance template — build for the model trained in Session 4 (≈45 min)
• Capstone reflection — Two-page memo: how the participant would integrate ML and AI tooling into their team's modeling workflow, with safeguards (≈90 min)
• Optional deep-dive — Recorded discussion: regulator perspectives on ML in insurance (≈30 min)

AI integration

• Demo: AI-assisted model-card generation for the trained ML model — what to keep, what to discard.
• Discussion: using LLMs to draft regulatory disclosure language for ML pricing models, and the review steps that must remain human.
• Closing reflection: an honest map of where AI is genuinely useful in the actuarial workflow, and where it is a distraction or a liability.