lichess.org
Donate

Predicting the Winner of Tournaments

Prediction is about the future.
Here is my prediction for the upcoming Candidates' Tournament:

Nakamura 8.5
Caruana 8
Praggnanandhaa, Giri, Wei: 7
Sindarov 6.5
Esipenko, Bluebaum 6

Prediction is about the future. Here is my prediction for the upcoming Candidates' Tournament: Nakamura 8.5 Caruana 8 Praggnanandhaa, Giri, Wei: 7 Sindarov 6.5 Esipenko, Bluebaum 6

Between the previous post and this post... there isn't much mention of "which players play which openings and how well" and "what are the odds of player X getting white against player Y". If there were gambling there would be incentives to generate accurate predictions... sigh.

Between the previous post and this post... there isn't much mention of "which players play which openings and how well" and "what are the odds of player X getting white against player Y". If there were gambling there would be incentives to generate accurate predictions... sigh.

@Toadofsky said ^

Between the previous post and this post... there isn't much mention of "which players play which openings and how well" and "what are the odds of player X getting white against player Y". If there were gambling there would be incentives to generate accurate predictions... sigh.

I just focused on the results of games and tournaments in the last two posts, so I didn't pay any attention to the openings. I might do that in the future, but my big problem with openings is always that the categorisation is very difficult. For example, saying that the Spanish is one big opening seems like it's way too general, but at the same time, there are so many variations that one player won't have many classical games in most lines, which makes a statistical comparison difficult.

Regarding the pairings, I only focused on round robin tournaments for now, where the pairings are known beforehand. For a Swiss tournament, one would have to generate the pairings after each simulated round, which I'll probably do in the future.

@Toadofsky said [^](/forum/redirect/post/jeEPGJOk) > Between the previous post and this post... there isn't much mention of "which players play which openings and how well" and "what are the odds of player X getting white against player Y". If there were gambling there would be incentives to generate accurate predictions... sigh. I just focused on the results of games and tournaments in the last two posts, so I didn't pay any attention to the openings. I might do that in the future, but my big problem with openings is always that the categorisation is very difficult. For example, saying that the Spanish is one big opening seems like it's way too general, but at the same time, there are so many variations that one player won't have many classical games in most lines, which makes a statistical comparison difficult. Regarding the pairings, I only focused on round robin tournaments for now, where the pairings are known beforehand. For a Swiss tournament, one would have to generate the pairings after each simulated round, which I'll probably do in the future.

@jk_182

Are the players probabilities of winning simply ranked in order of their Elos?

@jk_182 Are the players probabilities of winning simply ranked in order of their Elos?

Fair enough; the categorization work is crudely performed by ECO codes, however there are many practical challenges (including that players play different openings in different circumstances and against different opponents). There are many interesting research questions here, but unless people actually start gambling on this sort of thing the research won't be started anytime soon.

Fair enough; the categorization work is crudely performed by ECO codes, however there are many practical challenges (including that players play different openings in different circumstances and against different opponents). There are many interesting research questions here, but unless people actually start gambling on this sort of thing the research won't be started anytime soon.

It seems to me these types of exercises are a fool's errand. Using a computer to run monte-carlos when a back of the envelope calculation is enough to reveal the obvious: in a field of players all of whom are of nearly equal strength, the 1st order probability is that each player has 1/N chance of being the winner. So all the 2nd and 3rd order effects (slight differences in ELO, current form, historical records between players etc) can be thought as (1/N + epsilon), where clearly epsilon << 1/N.

Put another way, if each player has 16% chance of winning, does it really help to create a list saying "this player actually has 18%, this one 15%, this one 14%...) ?

Perhaps a better question is: using this model (before tournament begins) - and comparing to eventual winner... how many tournaments need to be studied before concluding that this is worth pursuing? :-)

It seems to me these types of exercises are a fool's errand. Using a computer to run monte-carlos when a back of the envelope calculation is enough to reveal the obvious: in a field of players all of whom are of nearly equal strength, the 1st order probability is that each player has 1/N chance of being the winner. So all the 2nd and 3rd order effects (slight differences in ELO, current form, historical records between players etc) can be thought as (1/N + epsilon), where clearly epsilon << 1/N. Put another way, if each player has 16% chance of winning, does it really help to create a list saying "this player actually has 18%, this one 15%, this one 14%...) ? Perhaps a better question is: using this model (before tournament begins) - and comparing to eventual winner... how many tournaments need to be studied before concluding that this is worth pursuing? :-)

@RuyLopez1000 said ^

@jk_182

Are the players probabilities of winning simply ranked in order of their Elos?

They are ranked according to their ratings, as the game prediction is based on the ratings. The non-obvious part is how an Elo advantage actually translates into the chances to win the tournament

@RuyLopez1000 said [^](/forum/redirect/post/46SCbOKf) > @jk_182 > > Are the players probabilities of winning simply ranked in order of their Elos? They are ranked according to their ratings, as the game prediction is based on the ratings. The non-obvious part is how an Elo advantage actually translates into the chances to win the tournament

@PerfectPatzer said ^

It seems to me these types of exercises are a fool's errand. Using a computer to run monte-carlos when a back of the envelope calculation is enough to reveal the obvious: in a field of players all of whom are of nearly equal strength, the 1st order probability is that each player has 1/N chance of being the winner. So all the 2nd and 3rd order effects (slight differences in ELO, current form, historical records between players etc) can be thought as (1/N + epsilon), where clearly epsilon << 1/N.

Put another way, if each player has 16% chance of winning, does it really help to create a list saying "this player actually has 18%, this one 15%, this one 14%...) ?

Perhaps a better question is: using this model (before tournament begins) - and comparing to eventual winner... how many tournaments need to be studied before concluding that this is worth pursuing? :-)

I don't think that your 1/N + epsilon assumption is correct for a typical chess tournament, since the skill differences between players are just too big. The clearest example would be Carlsen, who won around 50% of his classical tournaments while he was still playing actively, and most of these tournaments had 10 or more players in them.

I tested it on a few tournaments, where the prediction was correct 20%, which was roughly in line with numbers given by my model, but the sample size was small

@PerfectPatzer said [^](/forum/redirect/post/yD4nRvUh) > It seems to me these types of exercises are a fool's errand. Using a computer to run monte-carlos when a back of the envelope calculation is enough to reveal the obvious: in a field of players all of whom are of nearly equal strength, the 1st order probability is that each player has 1/N chance of being the winner. So all the 2nd and 3rd order effects (slight differences in ELO, current form, historical records between players etc) can be thought as (1/N + epsilon), where clearly epsilon << 1/N. > > Put another way, if each player has 16% chance of winning, does it really help to create a list saying "this player actually has 18%, this one 15%, this one 14%...) ? > > Perhaps a better question is: using this model (before tournament begins) - and comparing to eventual winner... how many tournaments need to be studied before concluding that this is worth pursuing? :-) I don't think that your 1/N + epsilon assumption is correct for a typical chess tournament, since the skill differences between players are just too big. The clearest example would be Carlsen, who won around 50% of his classical tournaments while he was still playing actively, and most of these tournaments had 10 or more players in them. I tested it on a few tournaments, where the prediction was correct 20%, which was roughly in line with numbers given by my model, but the sample size was small

@PerfectPatzer said ^

Perhaps a better question is: using this model (before tournament begins) - and comparing to eventual winner... how many tournaments need to be studied before concluding that this is worth pursuing? :-)

As a player, I'd appreciate a crystal ball which accurately predicts what my opponents will play to maximize their winning chances. Every player would appreciate the same.

@PerfectPatzer said [^](/forum/redirect/post/yD4nRvUh) > Perhaps a better question is: using this model (before tournament begins) - and comparing to eventual winner... how many tournaments need to be studied before concluding that this is worth pursuing? :-) As a player, I'd appreciate a crystal ball which accurately predicts what my opponents will play to maximize their winning chances. Every player would appreciate the same.