Here is a paper done by statisticians trying to do more sophisticated modeling to estimate true death rate. Cliff notes summary: what it did was build a transmission model that also adjusted for asymptomatic cases using the cruise ship data and then model the time to death which respectively account for the two key biases of the naive mortality estimate by adjusting for (1) undercounting mild cases and (2) undercounting mortalities of cases that will lead to mortalities but hadn’t yet. Factor (1) biases mortality estimate upward and is why people say it is not that bad but (2) actually biases downward and suggests it is worse than reported. Their analysis attempts to account for both with best data available to them at the time. It does not account for all factors but tries to adjust for these two key biases.