[manifold]

There's broadly two approaches actuaries use (I can't speak for this particular site which uses UN projections):

- One is to look at the average level of mortality improvements in the past at a population (or cohort) level and project those forwards. There are lots of models that do this, for example based on some combination of age, sex, year of birth and calendar year. You make the assumption that even though the specific advances in the past won't be repeated, future advances will follow a similar trend. The Lee-Carter model is an easy to understand example, although it's not particularly cutting edge.

- The other is to look at individual mortality factors (falling smoking levels cause less instances of cancer, etc) and project those forwards based on a mixture of historic advances and expert judgment. With these models you have to take into account that the people not dying of e.g. cancer are now at risk of dying from something else. These models are a lot harder to create because they rely a lot more on expertise. You also don't get information about the whole population, so you would end up using a combination of this with the first model.

Also, the site sounds like it is using period (i.e calendar year) life expectancy rather than cohort, so it might not be projecting ANY future improvements. I'd need to find their actual data source to be sure.