The Emperor's New Clothes, Part I

There is emerging hype about the application of artificial intelligence (AI) to mortality analysis, specifically the use of machine learning via neural networks.  In this note I provide a counter-example that illustrates why the human element is an absolutely indispensable part of actuarial work, and why I think it always will be.

In Richards and Macdonald (2024, Case Study A.6) we documented a real example of building a mortality model for annuitant mortality.  One of the data fields available was whether an annuity had an attaching spouse’s benefit, i.e. a contingent annuity that commences on death of the first life. Using the supplied data, statistical modelling suggested that the presence or absence of a spouse’s benefit was highly predictive of mortality.  The direction of the effect was plausible, with single-life annuitants exhibiting higher mortality.  However, the effect was similar in strength to an annuitant's sex, which caused me to wonder if there was data corruption.  An IT analyst familiar with the administration system assured me (twice) that there was no corruption.  If the effect were real, it would have been a major new risk factor for longevity that could easily be incorporated into annuity pricing.  I started drafting a note to the senior actuary.

However, if there is one thing I have learned about humans it is that they will say whatever is necessary to avoid additional work.  In this case, I had a theory as to what was causing the presence or absence of a spouse's benefit to be such a significant risk factor for annuitant mortality.  If the main annuitant died and the spouse was discovered to have pre-deceased them, the record of the contingent benefit might be removed during death processing. This would make the annuity look like a single-life annuity from outset, thus distorting the apparent mortality differential.  A third round of probing proved that my suspicions were correct, that the corruption existed...and that I had been lied to.  I binned my note to the senior actuary.

Where does this leave machine learning and neural networks?  Pretty much naked in the cold. When analysing mortality experience data, the actuary must be constantly aware of data quality (or the lack of it).  A statistical model provides a modest number of interpretable parameters, whereas a neural network has thousands of weight parameters in hidden layers.  The parameter estimates in a statistical model can be compared across risk factors to see if they violate sensible expectations, whereas the weights in a neural network have no such simple interpretation.  The estimates in a statistical model also come with standard errors to assess if a factor is meaningful or not.

The problem with using a neural network for actuarial experience analysis is that it creates a dangerous mental distance between the actuary and the data.  Data sets in actuarial work often have subtle quality issues, and uncritically feeding data into a neural network risks "garbage in, garbage out". Mortality analysis is an interactive, step-wise process best carried out by an experienced - and suspicious! - human being.