AI 460 · Year 4 · Semester 1 · 3 credits · AI & Machine Learning

Deep Learning for Time Series, Mortality & Claims Forecasting

Bridging classical actuarial demographic and reserving baselines with modern neural sequence architectures and rigorous uncertainty calibration.

Full practice exam · 30 Q · 120 min

Season 1 · 8 episodes

The Reserve Frontier

When a catastrophic multi-billion-dollar longevity block and under-reserved commercial liability book threaten Aegis Reinsurance with insolvency, newly appointed quantitative lead Maya Lin must replace rigid actuarial heuristics with modern deep sequence models, rigorous backtesting, and distribution-free conformal guarantees before the state insurance commissioner shuts down the firm.

Protagonist · Maya Lin, FSA, Head of Quantitative Analytics at Aegis Global Reinsurance

Setting · Aegis Reinsurance Global Headquarters, Boston; high-stakes actuarial war rooms, regulatory examination chambers, and high-performance computing clusters.

Stakes · A looming four-hundred-million-dollar solvency shortfall under dynamic regulatory stress testing, risk of rating downgrade to junk, and potential regulatory takeover if mortality improvements and long-tail casualty reserves cannot be accurately modeled.

Pick an episode

Episodes

Season finale: Maya presents an ASOP 56-compliant, credibility-weighted deep ensemble to the state insurance commissioner and the board of directors, proving exact finite-sample reserve coverage and securing Aegis's solvency margin.

From the lecturer: Welcome to AI 460. In this advanced senior film series, we bridge classical actuarial demography and claims reserving with modern neural sequence architectures. Keep pencil, paper, and a scratch Python notebook open: you will hand-derive parameter estimation algorithms (singular value decomposition for Lee–Carter, maximum likelihood for Cairns–Blake–Dowd, attention projection matrices, and pinball loss gradients) before implementing them computationally on empirical mortality surfaces and claims triangles. This course directly fortifies your preparation for SOA Exam ALTAM (parametric mortality projections, longevity risk) and CAS Exam MAS-II (time series, state-space representations, statistical learning). Rather than treating deep learning as an opaque black box, our focus is rigorous calibration, uncertainty quantification, and defensive backtesting against standard deterministic and stochastic baselines like the Mack chain ladder. Maintain a dedicated model validation ledger throughout the semester to track empirical coverage rates, continuous ranked probability scores, and computational stability across rolling train-test splits.

Cold open

It is seven in the morning on Maya's first day as Head of Quantitative Analytics at Aegis Re. On her screen flashes an urgent internal audit notice: the firm's flagship annuity portfolio is bleeding twelve million dollars a quarter because actual retiree mortality is improving far faster than their static twenty-year-old actuarial tables predicted. The Chief Risk Officer gives her six weeks before the state insurance commissioner audits their capital adequacy. If she cannot build a dynamic projection model that captures generational shifts, Aegis faces a catastrophic capital surcharge.

preparing lecture…

Transcript

Formulate the classical Lee–Carter mortality model, state its identification constraints, and estimate its parameters using Singular Value Decomposition (SVD).
Model and project the time index using an ARIMA(0,1,0) random walk with drift to generate projected mortality surfaces.
Formulate the Cairns–Blake–Dowd (CBD) two-factor model for post-retirement mortality and contrast its logit structure with Lee–Carter.
Bridge classical bilinear factor models to low-rank neural matrix factorisation architectures.