Revised European go ratings

[gennan]

Here you can compare the different functions: http://fooplot.com

Red is the EGD winrate prediction, black is my fit to observed winrates, blue is Ales' fit to observed winrates.

[Pio2001]

I have linked to this discussion in the french forum http://go-on.forumactif.com/t3078-echel ... e-revision

Sylvain answered that the european ratings are linked to the national ratings in the following way (my translation) :

- Jean is 12 kyu in both the european and french lists.
- Jean registers in a tournament "A" at 12 kyu and wins 5 games out of 5.
- Results of the tournament A are taken into account in the french list. With the special adjustment iteration used in the french system, Jean is now 9 kyu in the french list.
- The results are also taken into account in the european list. The european rating of Jean is now 11 kyu.
- Jean registers as 9 kyu in the tournament B, and wins 2 games out of 5.
- Results of tournament B are taken into account in the french list. Jean looses some points, but remains 9 kyu.
- Results of tournament B are taken into account in the european list. His registration rank being 2 levels above his best european rank, his european level is resetted to 9 kyu before the calcuation of the variation of the tournament B (reset policy). He looses some points, but remains 9 kyu in the european list.

What Sylvain is saying is that in this example, it is the special rules of the french list that have an effect on the european list.
Here, the reset policy is not the ability of Jean to choose his registration rank, it is the automatic adjustments of the french ratings that are copied into the european list.

So the question is to what extend does the european list owe to its own rating calculation, and to what extend is it a mere compilation of the various national lists.

[gennan]

I have linked to this discussion in the french forum http://go-on.forumactif.com/t3078-echel ... e-revision

Sylvain answered that the european ratings are linked to the national ratings in the following way (my translation) :

- Jean is 12 kyu in both the european and french lists.
- Jean registers in a tournament "A" at 12 kyu and wins 5 games out of 5.
- Results of the tournament A are taken into account in the french list. With the special adjustment iteration used in the french system, Jean is now 9 kyu in the french list.
- The results are also taken into account in the european list. The european rating of Jean is now 11 kyu.
- Jean registers as 9 kyu in the tournament B, and wins 2 games out of 5.
- Results of tournament B are taken into account in the french list. Jean looses some points, but remains 9 kyu.
- Results of tournament B are taken into account in the european list. His registration rank being 2 levels above his best european rank, his european level is resetted to 9 kyu before the calcuation of the variation of the tournament B (reset policy). He looses some points, but remains 9 kyu in the european list.

What Sylvain is saying is that in this example, it is the special rules of the french list that have an effect on the european list.
Here, the reset policy is not the ability of Jean to choose his registration rank, it is the automatic adjustments of the french ratings that are copied into the european list.

So the question is to what extend does the european list owe to its own rating calculation, and to what extend is it a mere compilation of the various national lists.

As long as the declared rank does not trigger a reset, the European list (and the revised list) ows only to its own rating calculations.
When the declared ranks triggers a reset, the European list (and the revised list) ows to that declared rank. I don't think the system needs to know the reason for the promotion. In some countries it's a self-promotion and in other countries it's a promotion by the National Go Association.

In your example, Jean is 12k in tournament A and he is 9k in tournament B (in all the systems). The only difference between the French rating list and the European rating list is the Jean's rank between tournament A and B. That doesn't mean much for the rating system, but if it's important for Jean's National Go Association, they could send an email to the EGD manager to adjust Jean's rank to 9k before tournament B (I assume that would be the normal procedure).

If this happens all the time and if it's really important to adjust ranks between tournaments, the National Go Association could ask the EGD manager to make a special web page where the National Go Association can manually reset their players ranks without the EGD manager's help.

[gennan]

I have linked to this discussion in the french forum http://go-on.forumactif.com/t3078-echel ... e-revision

I just read the discussion (using Google Translate). I thank you for calling attention to my investigation and explaining my findings so clearly!

[gennan]

What Sylvain is saying is that in this example, it is the special rules of the french list that have an effect on the european list.
Here, the reset policy is not the ability of Jean to choose his registration rank, it is the automatic adjustments of the french ratings that are copied into the european list.

So the question is to what extend does the european list owe to its own rating calculation, and to what extend is it a mere compilation of the various national lists.

From your description of the French system it seems that one of its purposes is to be a platform where the FFG officially publishes player promotions.

I don't think that's a purpose of the European list. It just follows players' declared ranks and their tournament games results.

Also, I don't think various national lists exist (I know only of France and Belgium). As far as I know, most countries just use the European list.

[gennan]

I have linked to this discussion in the french forum http://go-on.forumactif.com/t3078-echel ... e-revision

I have one correction to your explanation on that forum:

You state that I found that two times 3 stones handicap equals 6 stones handicap, but that is not really what I found. I found that two times 3 stones handicap equals two times 250 rating points difference = 500 points, which corresponds to 5.5 stones handicap (6 stones handicap with komi for white), not 6 stones handicap.

The EGD and the revised system use this formula: Rating difference = (handicap - 0.5) * 100, because theoretically, the first handicap stone is only half a move advantage: it's the same as an even game without komi. My finding is that this formula is consistent with the observed statistics of handicap games (although I cannot say this with high accuracy / confidence, because the data is a bit too sparse for high handicaps).

[Pio2001]

Also, I don't think various national lists exist (I know only of France and Belgium). As far as I know, most countries just use the European list.

Ah, that's good.
My concern was that we can't tell a "declared rank" from a "national rating overwriting a European rating".
If all resets would have come from national calculations (and not from players declaration), and if all countries suddenly stop using their rating system and replace it with yours, then all resets would have disappeared from the system.

In France, it has not been allowed to register in a tournament at a different rank than the one published by the French federation (if it exists) since 2010.
It is possible to ask for a reevaluation if the rank of a player becomes too different from his/her real strength, but this is a heavy process : the reevaluation can only by asked by the administrator of a club, and must be supported by sgf files proving the real level of the player.
In 2017, a total of 10 reevaluations have been made. There were 8 in 2016, and 22 in 2015. All other "resets" are just the French system calculating a different rating than the European one.

From your description of the French system it seems that one of its purposes is to be a platform where the FFG officially publishes player promotions.

The French rating system was created in 1981 and have been the official rating list used in France since then. It has undergone several transformations. The use of an ELO-Style core calculation, borrowed from the EGF, came in 2010.

I have been in charge of the pairing and the results of local tournaments for two years. I discovered the existence of the European rating list in my pairing software, And I learned that someone was calculating my rating in the European list just now, reading this thread.
Since I never communicate our tournament results to anyone else than the French federation, I was assuming that someone was just copying the French ratings into the European list !

[gennan]

Since I never communicate our tournament results to anyone else than the French federation, I was assuming that someone was just copying the French ratings into the European list !

I assume that the FFG communicates your tournament results (and all other French tournament results) to the EGD.

[WinDigo]

Also, I don't think various national lists exist (I know only of France and Belgium). As far as I know, most countries just use the European list.

Ukraine also has its own ranking system. As far as I know, there are three major differences from EGF ratings:

1. Dropping for more than 100 points is allowed.
2. After-tournament ratings are calculated sequentially, i.e. ratings "change" during the tournament. However, this doesn't affect seedings.
3. There is a term "abnormal tournament result". If player performed in tournament far better than his(her) rating, then he(she) gets a new one (the result of calculation) and the whole tournament is recalculated. This protects other players from losing too many points. Also, the "abnormality threshold" is different for every initial rating.

[Uberdude]

Also, I don't think various national lists exist (I know only of France and Belgium). As far as I know, most countries just use the European list.

In Britain we use the European rating system, but we used to apply a linear transform to convert an EGF rating to a British dan/kyu rank. We would look at the average rating of a player declared as a 1 dan in Europe (it was a bit less than the theoretical 2100) and also the average rating gap between declared ranks (not sure exactly over what range, it was a bit less than the theoretical 100) and then apply extrapolate these, e.g. 1 dan is 2070 and 95 points per rank so 3 dan is 2070 + 2*95 = 2260 rather than 2300. However, this adjustment was abandoned earlier this year.

[gennan]

Also, I don't think various national lists exist (I know only of France and Belgium). As far as I know, most countries just use the European list.

In Britain we use the European rating system, but we used to apply a linear transform to convert an EGF rating to a British dan/kyu rank. We would look at the average rating of a player declared as a 1 dan in Europe (it was a bit less than the theoretical 2100) and also the average rating gap between declared ranks (not sure exactly over what range, it was a bit less than the theoretical 100) and then apply extrapolate these, e.g. 1 dan is 2070 and 95 points per rank so 3 dan is 2070 + 2*95 = 2260 rather than 2300. However, this adjustment was abandoned earlier this year.

Interesting. So the BGA detected this issue: http://goratings.eu/Histograms/Histogra ... country=UK and used a transformation to get something like http://goratings.eu/Histograms/Histogra ... country=UK. It's a nice workaround, but I think it's preferrable to improve the source system instead, so that workarounds like this are not needed.

BTW, I think the required 95 points per ranks is a sign of the contraction effect caused by the unrealisitic winrate predictions of the EGD. I think that without rank resets, handicap games and new player ranks countering this contraction, over time the EGD ranks spacing would tend to shrink to about 50 points per rank (except for high dans). See http://goratings.eu/Probabilities which determines the spacing between ranks from even game winrates.

I used the revised rank spacing in the scaling of the vertical axis of my player rating histories, for example http://goratings.eu/Home/History?PIN=10449472.

[gennan]

Also, I don't think various national lists exist (I know only of France and Belgium). As far as I know, most countries just use the European list.

Ukraine also has its own ranking system. As far as I know, there are three major differences from EGF ratings:

1. Dropping for more than 100 points is allowed.
2. After-tournament ratings are calculated sequentially, i.e. ratings "change" during the tournament. However, this doesn't affect seedings.
3. There is a term "abnormal tournament result". If player performed in tournament far better than his(her) rating, then he(she) gets a new one (the result of calculation) and the whole tournament is recalculated. This protects other players from losing too many points. Also, the "abnormality threshold" is different for every initial rating.

Thanks for your info!

1. I can understand that the 100 point drop limit helps the EGD against deflation, because in combination with the large K factor that the EGD uses for lower ranks, it basically inflates lower ranks (in a somewhat convoluted way). The revised system also allows dropping more than 100 points, but because of the smaller K factor for lower ranks in the revised system, it hardly matters (even lower ranks don't drop 100 points that often in the revised system). But it just seemed like a very arbitrary rule to me, so I left it out of the revised system, like in the Ukrainian system.

2. Currently, the revised system also updates the participant rating sequentially for each game in a tournament. But it only does that while processing the games for each participant separately. For their opponents, it uses the pre-tournament rating. But perhaps I'll change this procedure a bit (see point 3).

3. That sounds very similar to the French system: iterating to fix anomalies. This is not how standard Elo systems work. It introduces an element of systems like Glicko or Rémi Coulom's WRH system. It's interesting and I have no problem with deviating more from Elo, but it does feel a bit arbitrary to me to distinguish between normal and abnormal results. I feel that it should be a smooth transition and not some threshold that triggers a different algorithm for a few participants.
Perhaps this can be achieved by updating each participant's K factor from the likelyhood of their results when processing a tournament. Also, a resulting increase from the normal K factor could be used to reduce their opponents rating changes in a post processing step. This would be similar to the rating reset mechanism currently used by the revised system.

[Pio2001]

Hi Gennan,
In the french federation, we are wondering if out players are ranked according to their level difference in handicap stones.

Would your dataset allow to draw the corresponding graph from french games played since 2012 (the french rating system was changed in 2010) ?

[gennan]

Hi Gennan,
In the french federation, we are wondering if out players are ranked according to their level difference in handicap stones.

Would your dataset allow to draw the corresponding graph from french games played since 2012 (the french rating system was changed in 2010) ?

It's possible to extract statistics for that, but the question would have to be specified more precisely. There is also the question of eliminating undesired biases.

For example, if one compares results from French players to the European average around the year 2014, the French players from 6k to 4d are underrated by about 50 points (half a stone, comparing http://goratings.eu/Histograms/Histogra ... country=FR with http://goratings.eu/Histograms/Histogra ... 4&country=).
I think this kind of bias will skew the results so much that your question cannot be meaningfully answered. My guess is that the only clear conclusion will be that on average, French players can give other European players of the same rank a handicap of no komi.

But perhaps you don't want to compare French players to the rest of Europe?

It's possible to extract statistics only for handicap games between French players. Preferably higher handicap games to detect long range trends. But the EGD does not contain all that many higher handicap games. If one narrows the data down to handicap games between French players after 2010, I think there won't be enough data points to get reliable statistics. Even when I use all handicap games from the EGD, the statistics for higher handicaps are quite noisy (for example http://goratings.eu/Probabilities/P_ObservedRevised?h=5).

But I don't think the EGD ratings should even be used to determine ranks. Ranks are in principle determined by handicap games and I guess that most handicap games are informal games that never enter the EGD. The EGD collects mostly even game results from tournaments and then it guestimates ranks from even game statistics with a formula that doesn't even match with the EGD statistics.

[gennan]

I think a rating system is useful to track tournament results of stronger players, but the mapping between rating and rank should be more fluid. A rating system can compute ratings for players that play in tournaments regularly, but it should just register the associated declared ranks. By continually updating the correlations between ratings and declared ranks, it could generate mappings that change over time (and the mapping could even be different for different countries).