Black swan events happen, and he's correct that there's simply insufficient data to model for them (in his example by introducing an arbitrary constraint on years), which is why they remain black swans. Can also be tagged outliers.
Industry work has to move forward anyway, so you use the data you have and pull out meaningful situational information. Call it the non- black swan portion of the data, which typically represents well over 90%, and you can successfully build predictive models within this area. You also aren't constrained by significance. In my industry, if we were constrained to even .05 significance, no models would be built.
Then you loosely try to identify the conditions within which a black swan event is likelier to happen (albeit still unpredictable) and push off risk through other levers....i.e. reallocation of investments to safer instruments, deployment of additional safety officers, rein in loans within certain risk tiers, etc..
All this to say, situationally you can know which causes more deaths.