Average Centipawn Loss --> MEDIAN Centipawn Loss

Since the loser of a game usually has a huge loss near the end of the game, the "average" centipawn loss gets kind of meaningless.
I think a "median" centipawn loss, or even 90-percentile, or "average after removing a few total blunders" or whatever would be more useful.
Thoughts?

Median centipawn loss is a meaningless statistic. All it tells you is that the majority of your moves were okay, which is true for the vast majority of games. Since inaccuracies, mistakes, and blunders are determined by centipawn loss anyway, median tells you nothing that you can't learn from the number of inaccuracies, mistakes, and blunders.

Right, but it makes a difference if the median is 8 or 38.
My point is just that the *average* is not very useful.
Then what about this:
- Separate out the two worst blunders (top-2 centipawn loss)
- Indicate the average of the *rest* (discarding top-2)

E.g.: Losses: 6.50, 1.50, avg of rest: 0.20

@trebe said in #1:
> Since the loser of a game usually has a huge loss near the end of the game, the "average" centipawn loss gets kind of meaningless.
> I think a "median" centipawn loss, or even 90-percentile, or "average after removing a few total blunders" or whatever would be more useful.
> Thoughts?

Years ago I agreed so strongly with this that I prototyped it, ultimately discovering that the function I needed to determine the median from a collection wasn't available in the version of the library we were using at the time.