dudeski_robinson
3 months ago
#5
Right now, what I do is (a) calculate the FIDE-Lichess gap for each player, (b) calculate the standard deviation of that gap, and (c) drop all observations where the FIDE-Lichess gap is over 2 standard deviations away from the sample median.
Does not your blitz rating of 1577 (69 points lower than your Fide prediction, when the median gives a blitz rating as 78 points higher for a difference of 147 points). Does this not fall outside 2 standard deviations away from the median? The formulas prediction is a Fide rating of 1646.
What was the point of calculating the median ratings if not to use them in developing the formula?
@mdinnerspace To answer your post #320.
If we calculate Median(Blitz - Fide) for different subsets, we get the following results:
Full sample: 78
Low ratings (800-1778): 1
Medium ratings (1779-2050): 71
High ratings (2051---): 128.5
As you can see, the median gap between Fide and Blitz ratings changes for different skill levels. Lichess overestimates the skills of better players to a greater extent.
The point of calculating the median was as a "first-cut" descriptive statistic. That is standard practice; that result was just meant as an overview, and was not designed as a predictive exercise. That's what the formula was for.
EDIT: I chose the low/medium/high cut-points based on quantiles of the sample of players with reported Fide ratings, such that there are about the same number of players in each group.
>As you can see, the median gap between Fide and Blitz ratings changes for different skill levels. Lichess overestimates the skills of better players to a greater extent.
Lichess makes estimations of players skills? Are you speaking of a players rating? Suggesting the higher rated a player is, the more the rating is inflated above his Fide rating/ or what?
@mdinnerspace The data say that for higher rated players, Lichess ratings are worse approximations of their self-reported Fide ratings. There seems to be more "ratings inflation" at the higher end of the skill spectrum.
That's what the different medians show. That's what the raw data plots show. That's what the formula takes into account (thanks to the 187 constant that you dislike).
How do you interpret this 187 constant being added to both lower rated players (where it's value is 20% of the total and higher rated players where it is only 10% of the total?) how can the same number be added to both? Should it not be written as a fraction? (a variable written as a fraction of the blitz and classical combined.)
Thinking about your explanation I see where adding 187 to lower rated players awards them more points in the resulting prediction and less points(percentage wise) for the higher rated players. This is consistent with the premise that the higher rated players ratings are inflated to a larger degree than the lower rated players. This is an "assumption" imo that can not be justified by relying on voluntary information in profiles for developing a reliable conversion formula.
To the issue of inflation. Much can be said the same for lower rated players, more of which have never played OTB. They are Time specialists. Know only blitz and faster time controls. They are very good at making fast decisions without too many serious mistakes. They progress and achieve ratings above 1800 in Blitz. Put them in a 90 minute game and they will lose rather handily, probably in less than 10 minutes to an established 1400 OTB rated player. It is these blitz specialists that have an inflated online rating above their Fide rating as much, if not more so than much higher rated players who posses both skill sets.
Your argument about blitz skills and ratings inflation sounds quite reasonable to me, actually. I buy it.
But that's precisely what the formula takes into account!
In any case, I'm done explaining the math and the role of the constant to you. Every single person who came to this thread and engaged with the problem has sided against you. Like they say: "When you think everyone around is an X, maybe you're the X".
I added a paragraph above about the 187 constant. You choose not to give an explanation about how it was derived, but I think I figured it out.
You know quite well that there have been many posts that disagree with your formula as being a useful prediction. I am far from the only one. A minority for sure, but the X comment seems beneath you.
That's right.
So the only valid criticism that remains is this: The data are self-reports, which could be unreliable.
If you re-read post #1, I was very clear in acknowledging this particular limitation of this dataset.
You've written dozens of posts claiming that there were (non-existent) mathematical problems with the formula. Those claims are simply wrong.
Now, you fall back on a critique that I had transparently acknowledged in the original post. Basically, all of your interventions could have been replaced by "Thanks for the effort, but I don't trust the self-reports, and I won't be using the formula myself." Seems like this would have saved us all a lot of grief.
Grief?
There is a mathamatical problem with the formula. I stand by this claim. It does not make predictions that are consistent. Your claim that it is inconsequential that the predictions for lower and higher rated players need not be consistent because certain factors exist; I do not agree with.
Sorry if you experienced any grief over this controversial topic (that any formula can make such a prediction). It has never been my intention. Again an apology. It is an interesting theoretical debate. Imo, a better approach and improved data may have served better. Your expertise in the field is far greater than mine. However, the possible pitfalls I observe remain valid.
I think you have been more than patient/tolerant of my criticisms. You've done a lot of well intentioned work, much of it enlightening.