As the latest edition of the Longevity Bulletin, The machine learning issue, is published, Matthew Edwards, Senior Director at WTW and Editor of the Longevity Bulletin, explores the application of machine learning to mortality and associated risks
Actuarial literature, artificial intelligence, audio book, civil war, dark star, deterministic variable, ethical debt, free cover, guaranteed bonus, Microsoft Works, military intelligence, surrender value, virtual reality, working party: it’s a pleasant and amusing task to compile lists of contemporary oxymorons. Should we include ‘machine learning’ in such a list? While ‘machine’ and ‘learning’ do not seem necessarily contradictory, the phrase does exhibit a degree of dissonance.
Let’s take a look at this idea: when we apply a machine learning (ML) algorithm to a dataset, to what extent is the ‘machine’ really learning anything?
Many ML approaches involve some form of random forest, whereby the algorithm will find a factor that shows explanatory power in respect of the modelled quantity (for instance, age would be explanatory of mortality in almost any mortality data). That factor will then be split in two (for instance, above and below age 73), with a small degree of mortality prediction provided by that one split. The algorithm then starts again, seeking to explain another part of the residual mortality (ie the mortality effects over and beyond that explained by the previous split) with a further split. And so on…
This would be a laborious process if done manually, but the computer ‘machine’ is able to move through large datasets quickly, generally ending up with highly predictive models.
Quite apart from the philosophical or semantic question of whether the machine itself has actually learned anything, can the ‘machine’ under typical random forest approaches even learn that age is generally a continuous variable in mortality contexts and not a series of upwards zigzags? But going beyond that, to what extent can we, as modellers, learn anything? We may now have a far more predictive model than we would otherwise have had (assuming the model passes the appropriate tests against an independent holdout dataset), but how can we learn from that model?
This raises one of the central problems with machine learning: its considerable ‘black box’ opacity. We started with a list of oxymorons, and we could perhaps have added the term ‘free lunch’. Machine learning should come with something of a ‘no free lunch’ warning. The extra predictiveness tends to come at the expense of greatly reduced comprehensibility, as well as practical problems with implementation of the results (quite apart from the need for very large and rich datasets in the first place).
While there are ways to understand better the information implicit in ML models, for instance through ‘partial dependency plots’ which provide approximate views of the impact of any one factor in the model, interpretability remains a substantial challenge. This unfortunately increases the associated model risk, which is already a non-trivial problem with even apparently simple forms of modelling – especially if we understand model risk in its broadest possible sense to encompass aspects such as potential data bias at the input stage, or stakeholder misunderstanding and misapplication at the output stage.
These interesting characteristics of ML provide excellent opportunities for actuaries to enhance what data scientists can themselves offer. We can rely on the data scientists to find predictive models, but this leaves a crucial area where actuaries can operate. Actuaries can help in areas such as model risk, or communication of the modelling mechanisms, and associated strengths and weaknesses. Given the domain or context in question, we can also help with the choice of trade-offs: where do we want our model to lie on the transparency v predictiveness ‘trade-off axis’?
How can we ensure the model outputs are useable, if the ML has been done with the purpose of better assumptions to feed into a cash flow or equivalent model? How can we best tackle missing data (given that data imputation cannot really ‘create’ information), or allow the model to feed profitably off our a priori knowledge (such as the continuous nature of factors such as age)?
Machine learning offers great opportunities to extract predictive patterns from large datasets, whether in mortality, health or other domains, but that ‘extraction’ needs to be done with a clear purpose in mind, and awareness of the various associated dangers. Actuaries can learn from machine learning but, most of all, actuaries can compensate for the fact that the machines’ ‘learning’ is not really learning in the sense of understanding. It’s a fascinating area to work in!
Interested in finding out more? Don’t miss the webinar on Monday 11 September from 12.30–13.30. Register for webinar
You can read the latest edition of the Longevity Bulletin at: Our journals and research publications
