AI 470 · Year 4 · Semester 2 · 3 credits · AI & Machine Learning

Reinforcement Learning for Dynamic Pricing, Hedging & Claims Triage

Master sequential decision-making under uncertainty, market frictions, and regulatory constraints: from dynamic pricing bandits to deep hedging and automated claims triage.

Full practice exam · 30 Q · 120 min

Season 1 · 8 episodes

The Policy Gradient

An ambitious actuarial data scientist must modernize a century-old insurer's legacy pricing, hedging, and claims engines using reinforcement learning before catastrophic churn and regulatory penalties trigger insolvency.

Protagonist · Maya Lin, FSA, Head of Algorithmic Actuarial Systems at Aethelgard Life & Casualty.

Setting · The trading desk, underwriting war room, and regulatory hearing chambers of Aethelgard Life & Casualty in Chicago.

Stakes · A multi-million dollar hedging deficit on legacy variable annuities, collapsing retention in personal auto from predatory competitor pricing, and imminent regulatory sanctions under ASOP 12 and VM-21.

Pick an episode

Episodes

Season finale: Maya deploys a provably safe, regulatory-shielded, off-policy validated reinforcement learning suite across renewal pricing, variable annuity hedging, and claims routing, satisfying the Insurance Commissioner and restoring statutory solvency.

From the lecturer: Welcome to AI 470. In this course, we bridge modern deep reinforcement learning with core actuarial science. You will move beyond static regression and single-period optimisations into dynamic, multi-period sequential decisions: pricing policy renewals under customer elasticity, hedging complex variable annuity guarantees with neural networks under market frictions, and routing claims between straight-through processing and specialized fraud investigation. Have a dedicated notebook for pencil-and-paper derivations of Bellman equations, policy gradients, and doubly robust estimators alongside a Python environment for training environments. This course feeds directly into advanced actuarial exams—including SOA QFI-QF for quantitative hedging under friction, SOA ILA-LPM/LAM for embedded guarantee risk management, and SOA ERM/CFE-SDM for dynamic capital allocation and enterprise decision governance. Approach every algorithm through a dual lens: algorithmic convergence in code and compliance with actuarial standards of practice (such as ASOP No. 12 on risk classification and Solvency II/ORSA capital constraints). Treat reward design not merely as an engineering parameter, but as the mathematical formalisation of solvency, profitability, and customer fairness.

Cold open

Eighteen million dollars in auto premium vanished overnight. Maya Lin stares at the quarterly lapse report while the Chief Actuary demands an answer: why did our static generalized linear model hike rates on our most profitable cohort just as digital competitors dropped theirs?

preparing lecture…

Transcript

Formulate multi-period insurance pricing, hedging, and claims triage problems as Markov Decision Processes (MDPs)
Derive the Bellman expectation and Bellman optimality equations for state-value and action-value functions
Prove the contraction mapping property of the Bellman operator and implement Value Iteration and Policy Iteration
Quantify long-term policyholder lifetime value and adverse selection dynamics under regulatory and capital constraints