A mortality basis built only by extrapolating the last twenty years of history is, in 2026, an assumption about the future that the future has stopped supporting. The shift in mortality and morbidity work is no longer cosmetic — it is structural, and the actuarial teams that have not yet noticed will see it first in their assumption committees.

By Meyr Kruger, FASSA, Actuary

Classical stochastic mortality models remain important because they are transparent and mathematically coherent. The frontier is broader: combining demographic trends, socio-economic factors, climate, pollution, pandemics, healthcare access, mental health, chronic disease, medical inflation and machine learning into a basis that does more than fit the past.

Why simple extrapolation is under pressure

Traditional models extrapolate historical mortality by age, sex, calendar year and cohort. Several forces have weakened that approach:

  • pandemics and post-pandemic health effects;
  • widening socio-economic mortality differences;
  • climate-related heat and pollution impacts;
  • changing healthcare utilisation;
  • obesity, diabetes and cardiovascular trends;
  • mental health and substance-use impacts;
  • medical technology and treatment advances;
  • migration and data-quality changes.

A modern basis should not simply “fit the past”. It should identify which drivers are stable, which are changing, and which require explicit scenario testing.

Multi-task learning, and what it actually gives you

De Mori and co-authors apply multi-task neural networks to mortality forecasting using data from seventeen countries. The interesting property of multi-task learning is that related populations share information while still learning distinct patterns. It is useful where a single population has limited data, or where common global trends interact with local features.

The actuarial challenge is interpretability. A mortality model used for annuities, life cover or capital modelling must be explainable enough for governance. Neural networks have a place, but only with diagnostics, back-testing, sensitivity analysis, scenario comparison and clear documentation of limitations. Where AI tools are used in any of this, we apply the controls on our How we use AI page.

Climate, air pollution and health

Climate-health modelling is now one of the most material emerging areas in life and health. Heat, cold, wildfire smoke, particulate matter, floods and vector changes affect mortality and morbidity in measurable ways. The SOA’s research examines this from a mortality angle and a companion morbidity angle, with the explicit caveat that traditional assessments may understate the future burden.

For South Africa, local evidence matters more here than almost anywhere. Climate impact depends on geography, infrastructure, housing, occupational exposure, disease burden and healthcare access. ASSA’s framework for South African life and health insurers is the right starting point because it is designed for the local environment rather than imported wholesale.

Morbidity is becoming operational

Health insurers and medical schemes have large volumes of claims data — diagnoses, procedures, admissions, pharmacy claims, chronic registrations, provider networks, authorisations, utilisation patterns. This enables more sophisticated morbidity modelling, and also makes the modelling harder to do honestly.

Morbidity is affected by disease incidence, treatment-seeking behaviour, provider practice, benefit design, coding changes, fraud, hospital capacity, medical technology and affordability. A spike in claims can reflect worse health, greater access, coding behaviour or changed benefit rules — they are not the same problem with the same solution. A decline in utilisation after a flood may reflect disrupted access rather than improved health. Morbidity models therefore need to combine actuarial analysis with operational and clinical context. Models that skip that step produce confident, false answers.

What modern analytics actually supports

A modern mortality and morbidity programme supports pricing and repricing of life, disability, critical illness and health products; annuity and longevity risk management; medical scheme contribution setting; chronic disease and wellness strategy; underwriting rule review; claims management and fraud detection; climate stress testing for life and health portfolios; reinsurance negotiation; experience investigations by geography, socio-economic segment and product design; and IFRS 17 assumption updates and sensitivity analysis.

The biggest value comes when experience investigations stop being once-a-year set pieces and become repeatable, automated and connected to assumption governance.

Governance for long-tail assumptions

Mortality and morbidity assumptions have long-term financial consequences. Good governance is the same governance any other material assumption deserves — clear data definitions and reconciliations; separation between observed experience and selected assumptions; documented expert judgement; sensitivity testing for key drivers; scenario testing for climate, pandemics and medical inflation; monitoring of actual-versus-expected; documented model limitations; and independent review for material assumptions. Where machine learning is involved, the framework extends naturally to feature selection, bias, explainability, drift and reproducibility.

The South African opportunity

South African insurers have unique modelling needs. Mortality and morbidity here are influenced by inequality, HIV and TB history, non-communicable disease, trauma, healthcare access, climate vulnerability and affordability. Imported assumptions are insufficient. Local data, local climate research and local actuarial judgement are not optional inputs to a credible basis — they are the basis. The opportunity is to build assumption frameworks that are transparent, evidence-based and responsive to emerging risks, with technology that can handle the data volume, version the assumptions and connect experience investigations directly into production models.

If you want help building an experience-investigation engine your assumption committee will trust, our Finance Modernisation practice does the data, model and governance work as a single engagement.

Sources