Pepe.Chess's blog

By Pepe.Chess, history, 7 years ago, In English

Hello

After like more than one year of the new rating change formula , I observed some things that should be mentioned. I don't know if somebody else think of them as not fair.

When you are a div2 participant and you manage to get like a top 3 rank. Your rating is gonna go far like 2050-2100 (and I've encountered people with like 2150). The opposite goes the same , when you are like red or yellow and you finish a contest miserably your rating will change by (-200 , -300 even more). My rating was dropped by like 250 the last round :v . But on the other side there were contests that I got my rating increased because of the opposite effect (and still not finding it fair).

Like as I remember from the past (div1 : 1700) when you are a div2 contestant even if you manage to get 1st rank , your rating won't exceed 1750. the same applies to div1 when you are like the worst contestant in a round even if you are very good , your rating won't drop that much (it even underneath the div1 limit by a little gap).

I actually find this formula kind of unfair. If you are a coder who is expected to be ranked like (150-200) in a round. and you manage to be the 1st. It doesn't mean that you will be in the top 50 every single round. A lot of times luck does its job. Like I believe if tourist enters a combined round and submits a solution then his laptop broke and he skips the round he will go down to yellow :D. I saw a lot of people dropped very hard because they failed in a combined round and were gone to the bottom of the ranklist (that happened to me twice) , and once I jumped from the middle of Candidate to International (and I find this unfair).

The whole point is , if you are a div2 contestant and you get like 2100 after a perfect round , (there are people that I believe suffering in div1 to get that 2100) so you should compete against them to get it , not against div2 participants. The same applies for all colors (even if you're yellow and you did a perfect round like top 5 , It's not fair to go like 2500 in the same way) , Ok you should be rated as a grandmaster , but to go further you should proof that you are a grandmaster in few consecutive rounds. The same applies for dropping down.

I thought of sharing this here in CF , hope there are some people who agrees.

  • Vote: I like it
  • +215
  • Vote: I do not like it

»
7 years ago, # |
  Vote: I like it +10 Vote: I do not like it

Agreed bro

I also have same feeling.I was on the positive side of that effect(still I dont feel I deserve so much of it). My case two months back was go back to Div 2 get and increase and get some life in div 1 for two more contests.

If its possible and they feel necessary they can choose some eps(<1) and multiply to all changes in ratings and then change(though just a personal opinion).

»
7 years ago, # |
  Vote: I like it +89 Vote: I do not like it

I completely agree with you.

In this formula, mi is the geometric mean of the seed (computed from old rating) and the actual place, so roughly speaking, a half of mi comes from the old rating and the other half comes from the performance of a new contest. Then, new rating is the average of R (computed from mi) and the old rating, so again very roughly speaking, 1 / 4 of the rating comes from the performance in the most recent contest, and 3 / 4 comes from older contests.

This ratio doesn't depend on the number of contests you've participated. However, in my opinion 1 / 4 looks too large in case you've participated in hundreds of contests. For example you continue getting the performance of 2000 for hundreds of contests and get the rating of 2000. Then you get the performance of 3000 (1000) for three matches. This formula thinks that you improved (declined) a lot and gives you the rating of 2578 (1422), but I think it's more natural to assume that you just get lucky (unlucky) in some recent contests.

In TopCoder, this ratio is 0.6 for newcomers and converges to 0.18 as you participate in more competitions. In addition to that TC has the rating cap and avoids too much change. In AtCoder, this ratio is 1 for newcomers (but subtract a constant to compensate that) and converges to 0.1.