the world of the ELO rating system Mariia Koroliuk Nicholas R Moloney
the world of the ELO rating system
Mariia Koroliuk
Nicholas
R Moloney
ELO rating in the real life
screenshot from the movie “The Social Network”
A -rating Ra B -rating Rb
- expected score (for A)
S - gained score (for A)a
- new rating for A
vs.
…and in the chess
experience
K = 40 (30 before july 2014,25…)
30 games 18th birthday
2300
K = 20 K = 10
2400
K factor (FIDE official since July 2014)
(whatever happens first)
issues
• Game activity versus protecting one's rating
• Selective pairing
• Ratings inflation and deflation. Example, Around 1979 there was only one active player (Anatoly Karpov) with rating >2700, September 2012 - guess? 44
DataData comes from https://ratings.fide.com that
publish the records every 3 months since 2001 and every months since September 2012.
https://ratings.fide.com
Distribution of different ratings
Data
Typical evolution over time
time (1=3 months period since Jan 2001)
ELO
ratin
g
selected players , that plays actively and around 30 years old at t=0
AR(1)?
Linear around (0,1/2) with m=0.0014
theoretical prediction: coefficient: 1-mlk=0.78 for L=10 (games) K=15 (average parameters)
We proved, that the score evolution for stable active player is an AR(1) process both empirically and theoretically with parameter alpha of 1-mKL (m- scope of the curve, L -average amount of games per time interval) and variance that depends on how much player vary in good days and bad days.
Estimation Results 1 -variance
stronger players plays more stable
Estimation Results 2 -alpha
predicted level
only regular players born from 1970 to 1975
?
WHO IS THIS GUY?
Palac,
Mladen
time
ELO
ratin
g
played a lot!
around 50 games
Recalculate alpha:
coefficient: 1-mlk=0.4 for L=50 (games) K=10
diffe
renc
e fro
m th
e m
ax
learn
ing p
eriod sta
ble period (AR(1))
decline
Learning periodWe also proved, that if person enters the list with some fixed stable level starting from smaller level learning much evolve in exponential speed.
However, predicted speed of learning is always higher that empirical speed.
The explanation is that We are improving while learning.
extreme values
Is there some possible maximum?
What are the chances for a really-really strong player (more
that 3000) being born in next year?
Generalized Pareto Model: really good fit- helps to explore events, that never
happened
data stops here,
but the model could be
continued
What is the chances
the event of person
with 3000 ELO will appear?
p=4e-10
exampleconditional probability on p>2600
1-(1-p)^n=0.1 —> nlog(1-p)=log(0.9)
—> n=2E8
now there is p=7e4 players in FiDE list.
on whole earth,
there is 7E9 people (which means
that 1 out of 100000
plays chess professionally)
so there need to be 1E10 professional players -> 1E12 people
SADLY, there was only 1E11 people ever on the earth
NOt to add on a sad point, as was said:
“A theory has only the alternative of being
right or wrong. A model has a third
possibility: it may be right, but irrelevant.”
and not every prediction is a true one.
THank you and time for questions