A bug

Ok then i leave my 2 cents

The only way to fix the horzain effect is to increase nodes which lichess does not want to do (it is what caused those spikes in the graphs)

The only way to fix accuracy would be to change the formula not to be depended on acpl on
Which i suggest you join the discussion for that on the discord.

Hey @TastyKing

I found an example from an elite game that's relevant to the topic of horizon effect :)

The brilliancy where Duda beat Giri last week makes the computer analysis go wrong:



The brilliant 28.Rg7 gets marked as a blunder when it's in fact forced mate in 11.
The reason is that when calculating at lower depths Stockfish thinks it goes from like +12 to +4 (sees the sacrifice but not the mate).

With the rest of the game, Duda still got more accuracy and less ACPL, but I thought that single move was funny and also illustrative of the horizon effect.

--

game imported here: lichess.org/bF9jOwSt#54

@pepellou said in #32:
> Hey @TastyKing
>
> I found an example from an elite game that's relevant to the topic of horizon effect :)
>
> The brilliancy where Duda beat Giri last week makes the computer analysis go wrong:
>
>
>
> The brilliant 28.Rg7 gets marked as a blunder when it's in fact forced mate in 11.
> The reason is that when calculating at lower depths Stockfish thinks it goes from like +12 to +4 (sees the sacrifice but not the mate).
>
> With the rest of the game, Duda still got more accuracy and less ACPL, but I thought that single move was funny and also illustrative of the horizon effect.
>
> --
>
> game imported here:

Cool, I feel like lichess should fix this horizon effect to prevent incorrect analyzation.

@TastyKing said in #33:
> I feel like lichess should fix this horizon effect

Can't be really fixed for every scenario. Imagine it was a mate in 20, or 40 (there are positions like that), ...

A balance needs to be picked on how deep you want to go and still offer a computer analysis in a reasonable time since the user requests it.

You can definitely go deeper and reduce these rare cases to appearing even more rarely, but that's more cost in resources used and more time the user will have to wait for the analysis.

I think the lesson here is we should bear in mind that the computer analysis is preliminary and we can still go analyze each position deeper with the engine.