Your network blocks the Lichess assets!

lichess.org
Donate

King's Gambit vs. Queen's Gambit: A Statistical Breakdown by Elo

Hi! Nice article. What are the stats for the QGA you mentioned but didn ́t include? Thanks!

Hi! Nice article. What are the stats for the QGA you mentioned but didn ́t include? Thanks!

maybe 2500+ use the K's Gambit against lower rated players to, and save the sounder stuff for more closely rated players. I would like to see a graph for what you said about the queen's gambit accepted.

maybe 2500+ use the K's Gambit against lower rated players to, and save the sounder stuff for more closely rated players. I would like to see a graph for what you said about the queen's gambit accepted.

Very interesting graphs ! Thank for the share.

One stat that can be interesting to analyse also is the time spent by opponent the respond. Something that could feed the "surprise " graph. But I don't know if does data are available ?

Very interesting graphs ! Thank for the share. One stat that can be interesting to analyse also is the time spent by opponent the respond. Something that could feed the "surprise " graph. But I don't know if does data are available ?

as someone who only learmed ruy lopez as main poening and is now perfecting it i hate going against kings gambit. i simply have no clue what to do. all i now is do not eat the pawn at the start. i really gotta look into it more

as someone who only learmed ruy lopez as main poening and is now perfecting it i hate going against kings gambit. i simply have no clue what to do. all i now is do not eat the pawn at the start. i really gotta look into it more

This is incredible analysis... can we do this for EVERY opening, and compare them side-by-side?

This is incredible analysis... can we do this for EVERY opening, and compare them side-by-side?

@LKama said in #8:

Hello! Thank you for the high-value feedback input!

  1. On the "Expected Elo Gain" metric:

There is a difference in the calculation depending on the mode:

  • In the hunt and line modes, the goal is to find high-impact moves. The formula incorporates reachability to discount the value of lines that are hard to get on the board.
  • In the plot mode, the goal is to show the raw performance of a line if you get it. The formula is simpler: (Win % - Loss %) * Elo_Factor * 100, which gives the expected Elo change after 100 games.

I will clarify this in the README.

This metric is not a predictor that you will gain that much Elo. It simply measures the performance of the average player at that level with that opening. To achieve those results, you'd need to match their proficiency, and as you gain Elo, you'd face stronger opponents, changing the calculation.

  1. On the Opening Explorer Data Bias:

The blog post you linked is excellent.
Unfortunately, since my tool uses the same Lichess API as the website's explorer, it inherits the same potential biases.

The core problem, as the author points out, is that stronger players might self-select into certain openings, skewing the stats for a given Elo bucket.

The author's proposed fix is interesting, but it seems to introduce a different kind of bias (and I can’t implement it with the opening API alone, the code would need to download the entire list of games). I've considered trying to "correct" the data using popularity stats, but that feels like a very risky way of manipulating data where I might have to hand-tune correction factors.

For now, I'm moving forward with the current method because it is at least transparent and consistent with the source data on Lichess. It provides a broad picture for comparing openings, even with this known caveat. I have to say that I was surprised by how much the win rates can change depending on how the games are grouped.

I wish the opening API had an option to filter games where opponents were more than, say, 50 rating points from one another.
@LKama said in #8:

  1. On the "Expected Elo Gain" metric:

There is a difference in the calculation depending on the mode:

  • In the hunt and line modes, the goal is to find high-impact moves. The formula incorporates reachability to discount the value of lines that are hard to get on the board.
  • In the plot mode, the goal is to show the raw performance of a line if you get it. The formula is simpler: (Win % - Loss %) * Elo_Factor * 100, which gives the expected Elo change after 100 games.

I will clarify this in the README.

This metric is not a predictor that you will gain that much Elo. It simply measures the performance of the average player at that level with that opening. To achieve those results, you'd need to match their proficiency, and as you gain Elo, you'd face stronger opponents, changing the calculation.

  1. On the Opening Explorer Data Bias:

The blog post you linked is excellent.
Unfortunately, since my tool uses the same Lichess API as the website's explorer, it inherits the same potential biases.

The core problem, as the author points out, is that stronger players might self-select into certain openings, skewing the stats for a given Elo bucket.

The author's proposed fix is interesting, but it seems to introduce a different kind of bias (and I can’t implement it with the opening API alone, the code would need to download the entire list of games). I've considered trying to "correct" the data using popularity stats, but that feels like a very risky way of manipulating data where I might have to hand-tune correction factors.

For now, I'm moving forward with the current method because it is at least transparent and consistent with the source data on Lichess. It provides a broad picture for comparing openings, even with this known caveat. I have to say that I was surprised by how much the win rates can change depending on how the games are grouped.

I wish the opening API had an option to filter games where opponents were more than, say, 50 rating points from one another.
4
I have only found it possible to compensate for the relative ratings of the White and Black players using databases available elsewhere (say chesstempo/chessbase etc. databases for games of different levels that gives the average ELO of the players of a colour and the ELO performance). Additionally the other obvious issue with the way the data is presented is that you need to correct for the average K factor at the different ratings. At 1500 there is a disproportional number of players with a high K factor because they have very few games played. The odd success of the Kings gambit at this rating is likely an artifact of this. Great work, and very promising steps towards a revolutionary tool.

@LKama said in #8: > Hello! Thank you for the high-value feedback input! > > 1. On the "Expected Elo Gain" metric: > > There is a difference in the calculation depending on the mode: > - In the hunt and line modes, the goal is to find high-impact moves. The formula incorporates reachability to discount the value of lines that are hard to get on the board. > - In the plot mode, the goal is to show the raw performance of a line if you get it. The formula is simpler: (Win % - Loss %) * Elo_Factor * 100, which gives the expected Elo change after 100 games. > > I will clarify this in the README. > > This metric is not a predictor that you will gain that much Elo. It simply measures the performance of the average player at that level with that opening. To achieve those results, you'd need to match their proficiency, and as you gain Elo, you'd face stronger opponents, changing the calculation. > > 2. On the Opening Explorer Data Bias: > > The blog post you linked is excellent. > Unfortunately, since my tool uses the same Lichess API as the website's explorer, it inherits the same potential biases. > > The core problem, as the author points out, is that stronger players might self-select into certain openings, skewing the stats for a given Elo bucket. > > The author's proposed fix is interesting, but it seems to introduce a different kind of bias (and I can’t implement it with the opening API alone, the code would need to download the entire list of games). I've considered trying to "correct" the data using popularity stats, but that feels like a very risky way of manipulating data where I might have to hand-tune correction factors. > > For now, I'm moving forward with the current method because it is at least transparent and consistent with the source data on Lichess. It provides a broad picture for comparing openings, even with this known caveat. I have to say that I was surprised by how much the win rates can change depending on how the games are grouped. > > I wish the opening API had an option to filter games where opponents were more than, say, 50 rating points from one another. @LKama said in #8: > 1. On the "Expected Elo Gain" metric: > > There is a difference in the calculation depending on the mode: > - In the hunt and line modes, the goal is to find high-impact moves. The formula incorporates reachability to discount the value of lines that are hard to get on the board. > - In the plot mode, the goal is to show the raw performance of a line if you get it. The formula is simpler: (Win % - Loss %) * Elo_Factor * 100, which gives the expected Elo change after 100 games. > > I will clarify this in the README. > > This metric is not a predictor that you will gain that much Elo. It simply measures the performance of the average player at that level with that opening. To achieve those results, you'd need to match their proficiency, and as you gain Elo, you'd face stronger opponents, changing the calculation. > > 2. On the Opening Explorer Data Bias: > > The blog post you linked is excellent. > Unfortunately, since my tool uses the same Lichess API as the website's explorer, it inherits the same potential biases. > > The core problem, as the author points out, is that stronger players might self-select into certain openings, skewing the stats for a given Elo bucket. > > The author's proposed fix is interesting, but it seems to introduce a different kind of bias (and I can’t implement it with the opening API alone, the code would need to download the entire list of games). I've considered trying to "correct" the data using popularity stats, but that feels like a very risky way of manipulating data where I might have to hand-tune correction factors. > > For now, I'm moving forward with the current method because it is at least transparent and consistent with the source data on Lichess. It provides a broad picture for comparing openings, even with this known caveat. I have to say that I was surprised by how much the win rates can change depending on how the games are grouped. > > I wish the opening API had an option to filter games where opponents were more than, say, 50 rating points from one another. > 4 I have only found it possible to compensate for the relative ratings of the White and Black players using databases available elsewhere (say chesstempo/chessbase etc. databases for games of different levels that gives the average ELO of the players of a colour and the ELO performance). Additionally the other obvious issue with the way the data is presented is that you need to correct for the average K factor at the different ratings. At 1500 there is a disproportional number of players with a high K factor because they have very few games played. The odd success of the Kings gambit at this rating is likely an artifact of this. Great work, and very promising steps towards a revolutionary tool.

Did not think of the fact that a fraction of 1500 players have a very uncertain ranking (initial rank with high k factor). Very good input thanks.

Thanks for the questions and the requests. I'm currently very busy at work but I'll come back with new analysis soon (I have a simple idea that I need to test). Best!

Did not think of the fact that a fraction of 1500 players have a very uncertain ranking (initial rank with high k factor). Very good input thanks. Thanks for the questions and the requests. I'm currently very busy at work but I'll come back with new analysis soon (I have a simple idea that I need to test). Best!

"Are the charts easy to understand?" * Yes, but a bit small. CTRL+ is necessary to read them.
"What could be improved?" * Openings do not matter much for rapid. Maybe process classical games from the master's database?
"Would you use a website with these kinds of interactive stats?" * No. Why? To tell me what to play in rapid in my rating bracket? Openings should be the same in all time controls and against all opponents and geared against the strongest opponents.
"What features would you like?" * None I can think of.
"What openings you'd want to see compared?" * Maybe 1 b3 vs. 1 g3
"Are the opening logos a good or bad idea?" * Just clutter, not useful or informative.

"Are the charts easy to understand?" * Yes, but a bit small. CTRL+ is necessary to read them. "What could be improved?" * Openings do not matter much for rapid. Maybe process classical games from the master's database? "Would you use a website with these kinds of interactive stats?" * No. Why? To tell me what to play in rapid in my rating bracket? Openings should be the same in all time controls and against all opponents and geared against the strongest opponents. "What features would you like?" * None I can think of. "What openings you'd want to see compared?" * Maybe 1 b3 vs. 1 g3 "Are the opening logos a good or bad idea?" * Just clutter, not useful or informative.

@LKama said in #17:

Did not think of the fact that a fraction of 1500 players have a very uncertain ranking (initial rank with high k factor). Very good input thanks.

Thanks for the questions and the requests. I'm currently very busy at work but I'll come back with new analysis soon (I have a simple idea that I need to test). Best!
Ah great, very good work. I'm not sure how easy to implement the average rating calc for white and black for a position. If you can do that then the rest is fairly easy. Your automated predicted outcome calculation coupled with forced nature of the play along with how unusual ones proposed line is essentially an automated version of how many players and authors put together repertories. Really excellent ground breaking work! Well done!

@LKama said in #17: > Did not think of the fact that a fraction of 1500 players have a very uncertain ranking (initial rank with high k factor). Very good input thanks. > > Thanks for the questions and the requests. I'm currently very busy at work but I'll come back with new analysis soon (I have a simple idea that I need to test). Best! Ah great, very good work. I'm not sure how easy to implement the average rating calc for white and black for a position. If you can do that then the rest is fairly easy. Your automated predicted outcome calculation coupled with forced nature of the play along with how unusual ones proposed line is essentially an automated version of how many players and authors put together repertories. Really excellent ground breaking work! Well done!