Error fixed in Go4Go database.

Are you able to write the data on your home page in a suitable format that I can fetch from an URL. I want to create a page on Go4Go containing the top ranked players. I will of course give credit and links back to your site.

Remi: The lists look VERY dubious to me, apart from maybe the top one or two. For example, the list for 2000-01-01 has Hyeon Mi-chin in 8th place. That beggars belief, and on checking I see that GoGoD has just 4 games for her in 1999 and she scored just 1-3 against rather undistinguished opposition. Given that she'd only been a pro for 4 years and took that long to get from 1-dan to 2-dan, previous results are hardly likely to have bumped her up. That seems confirmed by known results from GoGoD prior to 1999 - she scored 2-8.

That suggests to me something fishy about the maths, but even if the maths is correct (e.g. all players above her had even worse results - hardly seems possible, though), surely a minimum number of games should be required. I believe 30 is the usual criterion?

What have I missed?

The females you thought were male are indeed male in the GoGoD onomasticon. You will need to be aware also of a couple of cases where there exist both male and female pros with the same westernised name. Tang Yi is the only one I think of at the moment.

Thank you John for your feedback.

In the go4go database, she has a 6-1 result from 1997 to 2001, that's why her 2000 rating is good.

I agree that 7 games can't produce a very accurate result. I will try to rank players based on the lower bound of the confidence interval instead. That should prevent a lucky player from entering the top.

The current rating list is available to anybody in json format at that url:
http://www.goratings.org/ratings.json
Anybody should feel free to use this data (and cite goratings.org as the source).

macelee: I sent an email to you with more problems that need to be fixed.

Hi Remi,

Looks like there's a linking issue on the main page (http://www.goratings.org/). If I click on the Ke Jie (#1) for example, it takes me to Li Zhangyuan (http://www.goratings.org/players/1154.html).

The history page works fine.

_________________
Go Books

Thanks for your report. Please refresh the index page to clear your cache. I changed the player ids to match those of go4go.net.

Awesome, thanks!

_________________
Go Books

Rémi, can you please reformat the output a bit. It is not a valid JSON at the moment. Try the validator at jsonlint.com

It should be OK now.

Remi: The 2000 rating list is headed 2000-01-01. So not only should it not reflect the results for 2001, surely it shouldn't include the results for 2000 either. On that basis, go4go has just two games, for 1997, scoring 1-1. It's well known I'm not a numbers guy, so what's the trap I've fallen into?

Another male-female pair to watch for is Yi Chi-hyeon.

Rémi is the expert, but the way I understand it is that in WHR, for the year 2000, all games from all years effect her rating. It's just that those from 2000 affect it most strongly. Games from 1 year before and 1 year after still have an effect but less. And games from 2 years ago too but even less.

I took John's point to be, rather, that for a rating list labeled as "January 1, 2000", games after that date should not be included in the calculation.

_________________
be immersed

That position is not necessarily logical. The fact that we can't use games that have not yet been played in order to estimate someone's strength as of today, is quite different from claiming that games played today should not be used in estimating someone's strength as of yesterday. It seems just as reasonable to include games played one month (or year) after a particular date as it is to use games played one month (or year) before that date if they are available.

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Kind of fuzzy either way, in my opinion. On one hand, you might argue that games from the following year provide additional data for estimating the player's strength around that time.

But this assumes that the player's games from 2001 are good data points to estimate someone's strength in 2000. Maybe it's true if they haven't improved (or gotten worse) much in that year.

But if their strength is changing over the course of the year, it's a bad assumption.

_________________
be immersed

But equally bad whether you look forward or back. So basically we are SOL.

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Well, I thought that the point of this historical ratings was to have a range. So when I read a category labeled as '2000-01-01', I interpreted this to be a snapshot of the rank at that point in time (i.e. 2000-01-01 was the end of the range).

I thought that was why the '2015-01-01' category has a different set of results than the one on the main page - even though it's still 2015, I thought that the listing from '2015-01-01' was a snapshot ending at that point in time.

If that interpretation is correct, I don't think it's good to include data from after the snapshot.

_________________
be immersed

This, and other points made by Kirby, reflect what I think is a valid point of view. And, by the merest fluke, there was some support from that on the very day (but afterwards) that I posted my doubts.

I discovered (to my horror) that a very famous (or, better, notorious) game by Inoue Genan Inseki against Honinbo Jowa was not in the GoGoD database. It is now! In the course of transcribing it, my eye naturally wandered to the surrounding commentary. The significance of this game was that Jowa had included it, along with some others from the same short period, in a now famous book. Genan was incensed because Jowa had not included any games of his (Genan's) from the period just after, when he had made a significant improvement. He believed therefore that Jowa was trying to belittle him.

Now if we follow the grading-list maths argument, and we could view only the Elo-type number and not the actual games, yes, it sounds plausible to say that Genan's poor grading in 1822 based on that year alone could be massaged by using his games in 1823 and 1824 and so better reflect his true strength in 1822. That is, Genan was worrying about nothing but the merest blip.

But Genan himself would say hooey to that, and if you look at the actual games you'd be inclined to agree with him. The famous game in question is famous/notorious because it shows Genan (then pretty strong at 5-dan) getting into all sorts of bad-shape tangles under 6-dan Jowa's relentlessly forensic investigation of his weaknesses. But most of all, it features an eye-popping example of a "White to live and die" situation - White can live but in the process wipes out a huge territory of his own, so loses the game. Even amateurs, reading Jowa's book, could not fail to be swayed by such an egregious example. Furthermore, Genan himself claimed he did make a big improvement after that game (and who are we to gainsay a man who reached Meijin status not too long after.) To repeat, it would apparently have been quite wrong, in Genan's view, to massage his 1822 rating with 1824 games.

In the first case I was talking about, Hyeon Mi-chin's, even on the figures from go4go alone, I posit that her case might be similar, simply because in the relevant period she apparently did not improve much, i.e. took a rather long time to go from 1-dan to 2-dan. As it happens, and as I have already mentioned, the extra games for that period that GoGoD has, also suggest her rating should be downgraded. On top of all of that, it is well known that players do go through several slumps in a long career. Why try to hide them?

At any rate, it seems to me there are two scenarios: (1) a player's rating can be massaged by using later results, presumably in the hope of ironing out irregularities, or smoothing the curve, and (2) a snapshot based on games up to a specific time but not beyond can be presented.

I'm in mirkest glen on this. I've read books like Freakonomics and I'm aware of statistical traps it is possible to fall into unthinkingly. But my intuition (that word again!) still tells me that for historical assessments at least, the raw snapshot is more reliable (even though it can produce glitches such as red-eye) than cropping and photoshopping the image.

The WHR rating algorithm has retroactive effect: the rating of one day depends on games of the past and games of the future. That is what makes it more accurate than incremental rating algorithms.

It's just curve smoothing. If a player has a good performance in year n-1, a bad year n, and a good year n+1, the curve will dip in year n, and rise in year n+1 either way, just more extremely in the case excluding future information.

I understand the intuitiveness of excluding future data, and the desirability of lowering variance by including it.

I don't necessarily buy the argument that just because there are more published losses of an ancient player in year n, and more published wins in year n+1, that the player magically improved significantly. First, old game records are incomplete, and there could be a bias in those that survived. Second, performance is naturally variable, and just because a player's performance dips in year n doesn't mean they got worse at go. Nor does an improvement in year n+1 mean they got better. Of course, we are all, even pros, subject to a recency bias.

Overall, I like the variability reduction, and appreciate Remi's WHR approach.

_________________
- Brady
Want to see videos of low-dan mistakes and what to learn from them? Brady's Blunders

This seems to me to be a variant of the precision versus accuracy debate. I'd have thought precision was more relevant in historical ratings (as here). In other words, we want to precisely know how good someone's peak was, not with the peak worn down by performances when a player was past his prime.