Lots of highly active bullet + blitz players in the 2200-2700 range have noticed sudden 100-200 point spikes in rating since around August.
Many of us have discussed it and concluded that (a) rating increases far exceed any potential skill increase, this is particularly observable in 6+ month previously inactive players who've experienced a several hundred point jump when they resumed play (2) more points seem to be coming into the system (3) a gradual increase still seems to be occurring rather than stabilizing (at least, not yet).
Almost every player in this range is showing the same phenomena.
From the analysis I've done:
(1) Ratings equivalent to percentiles (e.g. rating that = 99th percentile) have been steadily increasing over the past 6 months.
(2) The degree of this is extremely unlikely to be explained by an increase in lower rated players coming into the pool
(3) It is probably not statistically explained by the change to the minimum rating floor.
(4) Recently, we have also observed several times in matches of 20-50 games between players of near equivalent ratings (with relatively low rds close to 50) having a near equal score will result in a net increase in points. Both players will gain rating after the match. E.g. started at 2600 each, played 50 bullet games with an equal score, then both end up at approx 2606. In the github issue @IsaVulpes posted below, he references several examples of this.
We suspect that changes to rd decay earlier in the year unbalanced the glicko2 system and may have caused points to be generated out of thin air, or at least a trend towards a much higher stabilization point.
Perhaps someone with a background in statistics could comment on how these changes might affect the system?
I don't have the programming expertise to do it personally, but think it would be a good idea to do deeper data analysis to assess the degree of these changes and if it will stabilize. Then, maybe look at reverting any tweaks to the system.
Otherwise everyone may end up 3000+ rated. Which honestly does sound like an attractive prospect, but there's also the inevitable heartbreak when you realize all your friends are now 3000 rated too.
github.com/ornicar/lila/issues/5673 Here is a Github Issue with some examples of matches just generating rating points
E: Okay, this has been confirmed by thibault, and is apparently intended. See his comment in the Issue
One has to look at the whole population. Is it shifting to higher values? Or is it more wide-spread? Or multiple modes?
I would be careful with the term "inflation" and "hyper" seems exaggerated. Or do you mean "hyperrating inflation"? :D
@Sarg0n Thibault literally confirmed that inflation is happening (and also stated that it is there on purpose), as I had written in my post already.
Not sure what you are now talking about in regards to "look at whole population" and "be careful with the term inflation"..
Yeah, but does the whole population (=distribution) shift to higher values? Or just the tip gets higher? So everyone is affected or just the top seed? Only above d50? You? Me?
I heard that the bottom was lowered recently. So not everyone will have an inflated value. Right?
People often speak of inflation when only a certain part is inflated but not the whole distribution.
PS: and „hyper“ remains grossly exaggerated
I think it is certain that anyone playing in the 2100-2400 range as been inflated. I was always bating around 2100-2300 [closed account]. This year it is rare fro me to drop below 2300. Yet, if anything, I suspect my chess has been deteriorating.
@Sarg0n Thibault clearly explained the way the gradual adjustment algorithm works and clarified that it is almost at the p50 = 1500 stabilization point. There is no need to be pedantic.
The term "Hyperinflation" is (1) a cautionary nod towards historical lessons we should learn but that are often forgotten and (2) it is a perfectly appropriate quantitative descriptor.
In terms of (1) Any curious student of history will remember incidents stretching back from coinage devaluation in the roman crisis of the third century, to recent examples often taught in schools such as hyperinflation in the weimar republic 1920-1923. Hyperinflation occurs when there is a continuing increase in the amount of printed money in circulation that is not supported by growth in the output of goods and services, leading to a shift in their respective prices. The feedback between the the increase in the money supply and the demand-pull inflation may tempt a government to print more money. Since consumers expect continued inflation, they will buy more now to avoid paying a higher price later. The interaction between these factors leads to a vicious perpetuating cycle that is not easily halted.
Hyperinflation in a rating pool is where ratings are devalued with respect to skill. Since players participating in the glicko2 system will tend to credit their inflationary rating gains to skill, they are more likely to hold onto them and peak sit (in a similar way to consumers buying more goods/services). You can see this trend represented in an increase in spikes at xx00 ratings in the pool distribution graphs over the past 6 months.
Since the glicko2 system is a zero sum game where one player peak sitting will deprive the total pool of rating points, this leads to a situation where more algorithmically induced elo points may need to be introduced to the system to counteract this effect and shift the percentiles, leading to further rating devaluation. However, since the stabilization criteria of p=50 has nearly been reached, psychologically reactive effects to inflationary trends such as a pronounced increase in peak sitting will also diminish and stabilize to their prior norms.
In terms of (2) hyperflation is usually defined in economics as an inflation rate exceeding 50% per month or up to 5-10% in any given day. If we model the potential for inflationary increaeses in rating on a per player basis, we can see that it may reach or exceed this threshold. In @IsaVulpes thorough analysis linked above, in a specific example he demonstrated a net rating gain of 0.25 points added to the system per game, or 0.125 points per player per game. Let us assume that this algorithmic augmentation remains relatively constant with rating increases. Let us also assume that both player A and player B start at 2600, score equally and play hyperbullet (for more validity since this estimate was derived from the bullet pool). 2600 + 5% = 2730. 130/0.125 points per game = 1040 games necessary to reach 2730. The toal amont of hyperbullet games possible in 24 hours will be approximately 1 minute per game = 60 x 24 games = 1440. We can see that the 1040 games required is well within the threshold of significance to meet the criteria of "hyperinflation". It even gives both players time to grab some toast, watch a few TV shows and still inflate their ratings by 5%.
Please note that in this example, Player A and Player B play each other exclusively. If we introduce a 2rd, 3rd, 4th, etc player to the theoretical pool, the per player inflationary effects of the same total amount of games are diminished in proportion to the amount of players. Since rating (and monetary) systems tend to incentives agents to exploit the systems rules for maximum self benefit, we should define our criteria to protect against rating hyperinflation in the potentially worst case scenarios. This was the reason for my choice of word.
However, since all this was an intentional adjustment that is about to stabilize, we can all sleep easy.
@SoWeakAtThis I can also confirm that my ratings are up approximately 150 points despite no increase in skill and (likely) lifestyle induced slow, yet steady, brain deterioration.
Warm regards, Burrower 🙏
For me, I saw a notable uptick a few weeks ago. Idk if this is part of that or just a skill increase however.
I am +200 than my usual rating in Bullet and Blitz. After the rating fix, will the ratings go down or remain where they are now?
@Burrower No one is disputing that the rating pool on lichess has gone through several stages of inflation and deflation over the years. I however, also take issue with the term "hyperinflation" and copypasting / typing out some wikipediaesque definition of inflation is not a reason to use the prefix hyper. You yourself seem to know quite well that hyperinflation is typically defined as an increase of over 50% per month , we had an increase of something like 3% per month.
The rapid / classical modes have also been subject to inflation and have peaked arround 2016 with 2200 not being in the top 1%.