oh ok:-)
the problem is that on other websites(e.g schacharena.de)
i have:1431 classicla and 1680 blitz???
the problem is that on other websites(e.g schacharena.de)
i have:1431 classicla and 1680 blitz???
The "formula" is proven to be faulty. Results given are not mathematically consistent from nor uniform from the low end to high end spectrum.
The OP could simply have said, with his research, players stated FIDE OTB in profiles averages to be 78 points lower than their blitz rating and 169 points lower than their classical rating here at Lichess.
Of course he based this on a "median", averaging the ratings, which as in your case and the majority of other users do not fall into.
Myself, I do not play blitz. My skill at the fast time controls sucks. If I played a few games, with the low rating acheived, the "formula" would miss my OTB rating by 400+ points. Most players fall out of the "median" in one way or another, making the "formula" useless, besides the fact the equation is only accurate within 10 points for ratings of 2100-2200.
The OP could simply have said, with his research, players stated FIDE OTB in profiles averages to be 78 points lower than their blitz rating and 169 points lower than their classical rating here at Lichess.
Of course he based this on a "median", averaging the ratings, which as in your case and the majority of other users do not fall into.
Myself, I do not play blitz. My skill at the fast time controls sucks. If I played a few games, with the low rating acheived, the "formula" would miss my OTB rating by 400+ points. Most players fall out of the "median" in one way or another, making the "formula" useless, besides the fact the equation is only accurate within 10 points for ratings of 2100-2200.
I'm done posting here. It is my opinion the OP did a disservice for new players for will entering OTB tournaments for the 1st time.
My suggestion is not to put any stock in internet blitz ratings. The only comparison that may hold a slight validity is if someone has played a 100's of correspondence games over a few months time.
My suggestion is not to put any stock in internet blitz ratings. The only comparison that may hold a slight validity is if someone has played a 100's of correspondence games over a few months time.
@mdinnerspace Thank goodness!
(Also, correspondence games would probably be even worse predictors of OTB strength, especially under the official ICCF rules where engines are allowed.)
(Also, correspondence games would probably be even worse predictors of OTB strength, especially under the official ICCF rules where engines are allowed.)
Thanks @dudeski_robinson for this reg analysis! I agree with you that outcomes are "sensitive to the strategy used to eliminate outliers and fake FIDE reports". And also agree with "Treat this as a good informed guess". Many users commenting here tend to ignore that the formula comes with an error in it.
What's the cleaning you came up with? Also, could the dataset you were working off be made available?
Cheers,
Marco
What's the cleaning you came up with? Also, could the dataset you were working off be made available?
Cheers,
Marco
Question: If someone properly evaluates the data e.g. eliminates outliers and uses only players with a certain amount of games and so on. Then he concludes:
"* A typical (median) user's FIDE rating tends to be 78 points lower than her Lichess Blitz rating
* A typical (median) user's FIDE rating tends to be 169 points lower than her Lichess Classical rating"
Why is the estimation not something like
Elo = (Lichess Blitz - 78 + Lichess Classical -169) / 2
or converted into
Elo = ((Lichess Blitz + Lichess Classical ) / 2 ) - 123.5
"* A typical (median) user's FIDE rating tends to be 78 points lower than her Lichess Blitz rating
* A typical (median) user's FIDE rating tends to be 169 points lower than her Lichess Classical rating"
Why is the estimation not something like
Elo = (Lichess Blitz - 78 + Lichess Classical -169) / 2
or converted into
Elo = ((Lichess Blitz + Lichess Classical ) / 2 ) - 123.5
@marcolom Thanks. I'm planning a new post on a different ratings-related topic, and will try to include a link to the full dataset.
@Sarg0n
Your formula:
FIDE = -123.5 + .5 Classical + .5 Blitz
My formula:
FIDE = 187 + .38 Classical + .38 Blitz
Since blitz ratings are more strongly correlated with FIDE ratings, the regression model gives the blitz rating a higher "weight" (coefficient). In your formula, you implicitly give classical and blitz equal weights.
Also, the regression model uses multiple imputation to handle cases where players play only one kind of game (i.e., classical OR blitz).
In the end, it simply comes down to an empirical fact: Your formula does not minimize in-sample (squared) prediction errors. In particular, it makes a lot of mistakes for lower-rated players.
Remember the cranky serial poster who just left this thread? He proposed a very similar approach to yours, not realizing that his predictions could be re-stated using exactly the same formula as mine (but with different weights). Unsurprisingly, the two formulas gave different predictions (especially at low ratings levels). The point is that his prediction formula (and yours) are inferior, because they make bad in-sample predictions. This is the whole goal of using least squares regression: minimizing prediction error.
@Sarg0n
Your formula:
FIDE = -123.5 + .5 Classical + .5 Blitz
My formula:
FIDE = 187 + .38 Classical + .38 Blitz
Since blitz ratings are more strongly correlated with FIDE ratings, the regression model gives the blitz rating a higher "weight" (coefficient). In your formula, you implicitly give classical and blitz equal weights.
Also, the regression model uses multiple imputation to handle cases where players play only one kind of game (i.e., classical OR blitz).
In the end, it simply comes down to an empirical fact: Your formula does not minimize in-sample (squared) prediction errors. In particular, it makes a lot of mistakes for lower-rated players.
Remember the cranky serial poster who just left this thread? He proposed a very similar approach to yours, not realizing that his predictions could be re-stated using exactly the same formula as mine (but with different weights). Unsurprisingly, the two formulas gave different predictions (especially at low ratings levels). The point is that his prediction formula (and yours) are inferior, because they make bad in-sample predictions. This is the whole goal of using least squares regression: minimizing prediction error.
Ok. First I was sure that you used min-error-fitting then I wasn‘t. I was somewhat bamboozled. So overall yours works better, no doubt.
Of course then everything is SQR(scientific^2) :D
Of course then everything is SQR(scientific^2) :D
@dudeski_robinson
Copied from your 1st post:
FIDE Rating = 187 + Lichess Classical Rating X 0.38 + Lichess Blitz Rating X 0.48
Now you're changing it to Blitz as .38 ???
Which is it?
How do you figure a blitz rating is more representative rating than a classical rating? This makes no sense.
You never did explain what hat you pulled this "187" from.
Next time try doing a "proof" of your formula by plugging in the numbers (as I did). Based on your claim of a players blitz rating is 78 points lower and classical is 169 points lower, this is easy to do. You will find the ONLY rating at which your formula is consistent is for approx. 2080 FIDE. As the rating becomes lower, your formula progressively predits a higher number, to 180+ points higher as FIDE 1180+ when the equation should equal FIDE 1000.
Copied from your 1st post:
FIDE Rating = 187 + Lichess Classical Rating X 0.38 + Lichess Blitz Rating X 0.48
Now you're changing it to Blitz as .38 ???
Which is it?
How do you figure a blitz rating is more representative rating than a classical rating? This makes no sense.
You never did explain what hat you pulled this "187" from.
Next time try doing a "proof" of your formula by plugging in the numbers (as I did). Based on your claim of a players blitz rating is 78 points lower and classical is 169 points lower, this is easy to do. You will find the ONLY rating at which your formula is consistent is for approx. 2080 FIDE. As the rating becomes lower, your formula progressively predits a higher number, to 180+ points higher as FIDE 1180+ when the equation should equal FIDE 1000.
@mdinnerspace
One last time before I give up.
In post #62, you proposed two formulas:
A: FIDE = -78 + 1 * Blitz + 0 * Classical
B: FIDE = -169 + 0 * Blitz + 1 * Classical
In post #76, @Sarg0n proposed a third formula:
C: FIDE = -123.5 + 0.5 x Blitz + 0.5 * Classical
In post #1, I proposed a fourth formula:
D: FIDE = 187 + 0.48 * Blitz + 0.38 * Classical
Using any of these formulas, we can calculate the squared prediction error like this:
Predicted FIDE = 187 + 0.48 * Blitz + 0.38 * Classical
Squared error = (Actual FIDE - Predicted FIDE)^2
The first observation (user profile) in my dataset looks like this:
* Username: 2Ap
* FIDE rating: 2078
* Blitz rating: 2217
* Classical rating: 2239
For this specific user, the squared prediction error with each formula is:
A : (2078 - (-78 + 1 * 2217 + 0 * 2239))^2 = 3721
B: (2078 - (-169 + 0 * 2217 + 1 * 2239))^2 = 64
C: (2078 - (-123.5 + 0.5 * 2217 + 0.5 * 2239))^2 = 702.25
D: (2078 - (187 + 0.38 * 2217 + 0.48 * 2239))^2 = 685.39
In that example, it looks like your formula B is the most accurate (followed by mine), since it produces the smallest squared prediction error. So the question becomes: Is your formula B most accurate overall, or was the case of player 2Ap a fluke?
To answer this, we calculate the squared prediction error for each of the 2807 players in the sample. Then, we take the sum of those squared errors to see which formula is more precise OVERALL.
The sum of squared errors per formula are as follows:
A: 400100865
B: 405032490
C: 387459623
D: 366554861
Out of the four proposed formulas, your two formulas are the worst at predicting FIDE ratings. The best one is the one I proposed in the original post. In fact, it is mathematically impossible to find different weights that produce smaller total squared errors than mine. That’s a feature of the least squares regression estimator that I used.
And to answer your question directly, I didn’t pull the 187 number out of nowhere. 187 is the estimated intercept of the linear regression model. This was explicit in the original post, and the intercept is a well-known concept which I teach to all my first-year intro to stats students.
Your calculations aren't a "proof" of anything, except your lack of understanding.
One last time before I give up.
In post #62, you proposed two formulas:
A: FIDE = -78 + 1 * Blitz + 0 * Classical
B: FIDE = -169 + 0 * Blitz + 1 * Classical
In post #76, @Sarg0n proposed a third formula:
C: FIDE = -123.5 + 0.5 x Blitz + 0.5 * Classical
In post #1, I proposed a fourth formula:
D: FIDE = 187 + 0.48 * Blitz + 0.38 * Classical
Using any of these formulas, we can calculate the squared prediction error like this:
Predicted FIDE = 187 + 0.48 * Blitz + 0.38 * Classical
Squared error = (Actual FIDE - Predicted FIDE)^2
The first observation (user profile) in my dataset looks like this:
* Username: 2Ap
* FIDE rating: 2078
* Blitz rating: 2217
* Classical rating: 2239
For this specific user, the squared prediction error with each formula is:
A : (2078 - (-78 + 1 * 2217 + 0 * 2239))^2 = 3721
B: (2078 - (-169 + 0 * 2217 + 1 * 2239))^2 = 64
C: (2078 - (-123.5 + 0.5 * 2217 + 0.5 * 2239))^2 = 702.25
D: (2078 - (187 + 0.38 * 2217 + 0.48 * 2239))^2 = 685.39
In that example, it looks like your formula B is the most accurate (followed by mine), since it produces the smallest squared prediction error. So the question becomes: Is your formula B most accurate overall, or was the case of player 2Ap a fluke?
To answer this, we calculate the squared prediction error for each of the 2807 players in the sample. Then, we take the sum of those squared errors to see which formula is more precise OVERALL.
The sum of squared errors per formula are as follows:
A: 400100865
B: 405032490
C: 387459623
D: 366554861
Out of the four proposed formulas, your two formulas are the worst at predicting FIDE ratings. The best one is the one I proposed in the original post. In fact, it is mathematically impossible to find different weights that produce smaller total squared errors than mine. That’s a feature of the least squares regression estimator that I used.
And to answer your question directly, I didn’t pull the 187 number out of nowhere. 187 is the estimated intercept of the linear regression model. This was explicit in the original post, and the intercept is a well-known concept which I teach to all my first-year intro to stats students.
Your calculations aren't a "proof" of anything, except your lack of understanding.
This topic has been archived and can no longer be replied to.