Project ELO (Dataset released)

[davesaz]

You have me twice, once capitalized and once without.

What is the starting point for each quarter? Does it roll over from quarter to quarter or start fresh? (rolling over is best, meaning 2017 Q1 starting elo is the ending elo from 2016 Q4)
Are opponent elo's updated per game or use the starting elo for the quarter? (per game is best, if the record of games is in date order)
Does the town / mafia / average calculation all use the average as the starting point and for the opponent, or use the opponent's town when mafia and mafia when town? (either way works as long as it is consistent)
Also how do you handle 3p and multiball? (this is a tough one. elo started with chess which doesn't have 3p)

The glicko numbers are universally high from the spot check sampling I did, which indicates the elo numbers are not converging like they should. That could reflect the randomness inherent in role draws.
It would be very helpful to have #games played, opponent elo as each alignment columns, to help verify the output is correct.

If sufficient data were available, a more telling set of statistics would be lynched as town/scum, NK'd as town/scum, survived as town/scum. It would be challenging to determine a meaningful formula...

[Realeo]

1. Noted. The earliest I can respond is end May
2. Q1 is January - March. Q2 is March - June. Judged from start time.
3. Roll over
4. The nature of Glicko algorithm is that it is calculated per quartre
5. Individual elo vs average opponent elo
6. Yes, we handle multiball
7. The indication of converging can be seen from the

+

8. #games played may not help as
<a> there is going to be not covered games (ie. Open games)
<b> It is better for you to directly check the data.txt instead your suggestion for more specific fixing.

[Realeo]

7. The indication of converging can be seen from the +

Which is why I silently want to kill LicketyQuickety

top 39.3 % of players

for creating such a misleading information. You can't tell Ircher is at top 39.3% when he has a huge elo deviation.

[LicketyQuickety]

7. The indication of converging can be seen from the +

Which is why I silently want to kill LicketyQuickety

top 39.3 % of players

for creating such a misleading information. You can't tell Ircher is at top 39.3% when he has a huge elo deviation.

I might have been a math guy once, but those days are long past.

[Ircher]

While is my deviation so high? And top 39%? That's way higher than expected.

[mhsmith0]

While is my deviation so high? And top 39%? That's way higher than expected.

1) Small sample size most likely (IIRC you have like one scum game in newbies and zero in mini normals?)
2) If you have a winning record you'll probably be slightly above avg. I don't recall if scum results are over-weighted or not tho.

[mastina]

For the record, I'm also listed twice, as Mastin2 and mastin2.

[Ircher]

I think we could include non-bastard mini themes, perhaps at a lower k-factor.

A lot of mini themes are decently balanced imo and checked by at least 1 reviewer.

[LicketyQuickety]

I think we could include non-bastard mini themes, perhaps at a lower k-factor.

A lot of mini themes are decently balanced imo and checked by at least 1 reviewer.

No, mini-theme's are the worst balanced games on this site that I know of.

[mhsmith0]

I think we could include non-bastard mini themes, perhaps at a lower k-factor.

A lot of mini themes are decently balanced imo and checked by at least 1 reviewer.

And probably a lot of them aren't

Ultimately it's a question of variance vs bias. Extending the sample to much noisier data introduces the bias inherent in the fact that some people will have drawn formats more friendly to their alignment while others will have gotten hosed by the same. You can justify it either way I guess; I was concerned enough by the issue that I didn't even want to deal with the format variance of mini normals, looking only at newbie games, but others can reasonably be much more permissive on the issue.

[mhsmith0]

I think we could include non-bastard mini themes, perhaps at a lower k-factor.

A lot of mini themes are decently balanced imo and checked by at least 1 reviewer.

No, mini-theme's are the worst balanced games on this site that I know of.

Probably large themes are worse balanced, but then again they're probably just generally hard to balance.

[Realeo]

My only praying is somebody created a clean data for Open Game. It was a disaster.

[Dunnstral]

And probably a lot of them aren't

[LicketyQuickety]

Realeo, when is Q1 going to be complete?

[Realeo]

When I no longer have a life.

[LicketyQuickety]

When I no longer have a life.

Fair answer I suppose. LOL