Few variables weigh as heavily on a pension system as longevity. Living longer than expected is excellent news for the individual and a financial challenge for the collective: every additional year of life expectancy makes a defined benefit plan's obligations more expensive. Modeling well how long pensioners will live, and how much uncertainty surrounds that projection, is one of the core tasks of the pension actuary, and that is precisely where artificial intelligence and machine learning are starting to add value.
The problem of modeling longevity
Traditionally, longevity risk is addressed with mortality tables and dynamic models that project how mortality changes over time. The Lee-Carter model and its extensions are classic references for capturing the trend of mortality improvement by age and calendar year.
These models work, but they have limits: they tend to rely on few variables, assume relatively rigid structures and may poorly capture complex interactions or regime changes. This is where modern techniques come in.
How machine learning helps
Machine learning offers tools that complement, rather than replace, the classical actuarial toolkit. Some applications:
- Better mortality projections. Methods such as boosted trees (gradient boosting), random forests or neural networks can capture nonlinear relationships and interactions among age, sex, period and cohort that parametric models simplify.
- Use of more information. When sufficient and appropriate data exist, these methods can incorporate additional covariates (for example socioeconomic level or region) to distinguish mortality patterns across subgroups.
- Tailored projections. Instead of applying a single average table to the whole population, it is possible to build projections better fitted to the characteristics of a plan's specific group, which reduces the error of using averages that represent no one.
- Pattern detection and validation. The same methods can be used to detect anomalies in the data, validate assumptions and stress mortality improvement scenarios.
The connection with projected unit credit and NIF D-3
None of this is an abstract exercise: it feeds directly into the valuation of obligations. Under the projected unit credit method, which is prescribed by NIF D-3 in Mexico and by IAS 19 internationally, each participant's future benefits are projected and discounted to present value, attributing a credit to each period of service.
The mortality assumption, and in particular its future improvement component, is a key input to that projection. A table that underestimates longevity artificially reduces the Defined Benefit Obligation; one that overestimates it inflates it. To the extent that machine learning makes it possible to select and support the demographic assumption more soundly, it contributes to a more faithful valuation of the DBO and of the service cost. AI does not change the method; it improves the quality of one of the method's inputs.
The limits: where prudence rules
That said, NIF D-3 and actuarial practice require assumptions that are a best estimate, supportable and mutually consistent. This places significant constraints on the use of sophisticated models:
- Data. Machine learning is demanding in both quantity and quality of data. Many pension plans in Mexico have small populations with little mortality experience of their own. With scarce data, a complex model overfits and produces unstable or misleading projections.
- Interpretability. An actuary must be able to explain and justify the assumptions before auditors, regulators and governing bodies. A model that works as a black box is hard to defend, even if it predicts well in sample. Transparency is not optional in a regulated environment.
- Regulatory prudence. Assumptions must be reasonable and unbiased, avoiding optimism that understates the liability. Any method, however advanced, is subordinate to this principle and to documented professional judgment.
- Stability over time. Valuations are compared year over year. A model whose results shift sharply with small changes in the data complicates the interpretation of remeasurements and creates noise in the financial statements.
A mature use: power with judgment
The balanced reading is that AI and machine learning are powerful allies for informing longevity assumptions, not for replacing actuarial judgment or the regulatory framework. Their natural place is in exploration, cross validation, pattern detection and improving projections when data are sufficient, always with an interpretable model and a final assumption the actuary can support.
Professional bodies are keeping pace with this evolution. The Society of Actuaries, the Casualty Actuarial Society and the Institute and Faculty of Actuaries publish research on mortality models, data science and their prudent application to longevity risk. It is worth following that literature to adopt new tools without losing the rigor that pension valuation has always demanded.