Play
Multi-Player
Coins
Community
Settings
Help
Community   Maps   Forum   Mail   Ladders   Clans   Recent Games
Sign In | Sign Up
<< Back to Ladder Forum   

Posts 1 - 30 of 38   1  2  Next >>   
Rio 2016 Stallympics: 5/31/2016 20:44:15


kynte
Level 51
Report
Hey!

Just recently ran a script on currently ranked 1v1 ladder teams to get some data about stalling. I took every currently ranked team, and from their non-expired games that lasted longer than 2 turns and ended in elimination or surrender (as opposed to a boot) I compared how many turns their wins last relative to their losses, and how much time their wins take compared to their losses. Keep in mind that your win is someone else's loss (and vice versa).

Basically, I was looking at discrepancies between how long your wins take relative to your losses. Finally, I ranked all these teams based on how much time their losses took longer than their wins, expressed as a percentage of their average game speed. I went with this rather than absolute time to reflect the pace players play at- if you're taking 3 days longer to finish your losses than your wins, that's much worse if you finish your games in 1 day vs. 20 days.

Anyhow, here's the data presented as a makeshift table with the headings in row 1: https://gist.github.com/anonymous/9688db10e17e50cb82c388691b23e0f4

It should be accurate as of about 30 minutes ago.

Here's what each of the headings means:

RANK: A player/team's rank on this list
PLAYER NAME: Player name (ugly encoding, I know)
LADDER ID: Ladder team ID (you can append this to "https://www.warlight.net/LadderTeam?LadderTeamID=" to generate their team URL)
CLAN NAME: The name of their clan (again, ugly encoding)
L-WPCT: The difference between time taken on average to lose (LOSS FINISH TIME) and the time taken on average to win (WIN FINISH TIME), expressed as a percent of the time taken to finish a game on average (AVG FINISH TIME).
RNK: A team's current rank on the 1v1 ladder
RTNG: Team's current rating on the 1v1 ladder
CT: Number of games I looked at
W: Number of wins
L: Number of losses
L-W FINISH TIME: The difference between time taken on average to lose (LOSS FINISH TIME) and the time taken on average to win (WIN FINISH TIME)
WIN FINISH TIME: Average time from turn 1 to the final turn of all game in which the listed team won
LOSS FINISH TIME: Average time from turn 1 to the final turn of all game in which the listed team lost
AVG FINISH TIME: Average time from turn 1 to the final turn of all games by the listed team
L-WTRN: The difference between the turns taken on average to lose (L-TRN) and the turns taken on average to win (W-TRN)
W-TRN: The number of turns taken on average to win
L-TRN: The number of turns taken on average to lose
AVGTURN: The number of turns taken on average to complete a game

Edited 5/31/2016 20:55:26
Rio 2016 Stallympics: 5/31/2016 21:07:12


kynte
Level 51
Report
Looking at the data:

First off, it's a bit hilarious that alhazi ranks at the very top with 488.67%. That means that, in the time difference between an average alhazi loss and an average alhazi win, alhazi has enough time to start and finish nearly 5 average alhazi 1v1 ladder games.

An average alhazi win takes about 2 and a half days. An average alhazi loss? Nearly 26 days. Pretty big difference. His losses also take 8 turns longer than his wins.

However, there's some serious caveats. For starters, we only had a sample size of 5 losses and 45 wins- 5 losses that could've easily occurred simultaneously while alhazi took 2 10-day vacations. My "finish" times were actually the time taken from the pick round to the final turn, so I didn't factor in any time taken to surrender, and my start times were actually the turn 1 times- because I couldn't find where to get distribution turn times from the query game API and had some issues scraping actual end times from the ladder (that I've since fixed). I also ended up intentionally leaving off the pick times because you have no idea who's winning/losing during the pick stage and time-skew based on your opponents (taking longer when your opponent's got a decent rating) would likely be most exaggerated during the pick stage.

Moreover, looking at alhazi again, you'll notice that he's got more than 5 non-expired losses on the 1v1 ladder- but a lot of them weren't included because they ended in alhazi getting booted. Why? I wanted to keep weird edge cases from throwing off the data (e.g., wins by boot or losses by boot that occurred when a player was on vacation IRL).

Most importantly, this shouldn't be taken as conclusive evidence of stalling (instead, it's hopefully a starting point). You don't have full control over the start-end time of a game- your opponent's part of the equation, too, and they might just be playing more slowly when they're winning against you to avoid screwing things up. If there's a significant skill difference between the players you lost against and the players you won against, and you tend to play more slowly against higher rated players, that could explain why you take about twice as long to finish the game when you're losing. And of course, we're looking at games in hindsight- it might not have been evident to you that you were losing, and you just pressed on.

Let's take Buns as an example:
RANK                              PLAYER NAME (LADDER ID;                                      CLAN NAME):   L-WPCT RNK RTNG  CT   W   L          L-W FINISH TIME          WIN FINISH TIME         LOSS FINISH TIME          AVG FINISH TIME L-WTRN  W-TRN  L-TRN AVGTRN

 29.                                  Buns157 (ID: 10115; clan:                        Icelandic Turtles):   45.89%   1 2291  38  32   6   1 day, 23:29:02.500000         3 days, 23:59:11  5 days, 23:28:13.500000   4 days, 7:29:01.921052   1.95   9.88  11.83  10.18


He finishes wins in about 4 days and losses in about 6, leaving a difference of about 2 days or about half a game- and a turn difference of about 2 more turns taken per loss. But we're looking at a sample size of just 6 losses here, and those could've been slower because of Buns' opponents realizing they were in a winning position against the ladder #1 and playing more slowly to not screw things up. Plus people could just be surrendering to Buns faster than they should- your opponents have more control over when your wins end than you do, after all.


And last but not least, a note about percentages: That 45.89% doesn't meant that Buns finishes 46% of a new game during a loss. Keep in mind that games are concurrent and that turns/time to finish varies. If you've got 5 games going on at once, that extra time could mean you've finished 4 wins with the extra time while waiting on the loss, made some progress in 4 other ongoing games, or something in between.

Negative percentages mean your wins tend to take longer than you're stalling. This doesn't mean you're "reverse-stalling"- just that your average opponent takes longer to lose than you do. So, in a way, you stall less than the people you play against.

Edited 5/31/2016 21:55:53
Rio 2016 Stallympics: 5/31/2016 21:45:32


Жұқтыру
Level 55
Report
What's stopping you from taking a bigger sample size?
Rio 2016 Stallympics: 5/31/2016 21:52:48


Min34 
Level 58
Report
Moreover, looking at alhazi again, you'll notice that he's got more than 5 non-expired losses on the 1v1 ladder- but a lot of them weren't included because they ended in alhazi getting booted. Why? I wanted to keep weird edge cases from throwing off the data (e.g., wins by boot or losses by boot that occurred when a player was on vacation IRL)

In the case of Alhazi you can probably include those games as well. We all know he is a massive staller, if he got booted it was because he didn`t want to play that game anymore, not because he wasn`t able to take his turn in time.
Rio 2016 Stallympics: 5/31/2016 21:53:33


kynte
Level 51
Report
Good question.

Mainly that I've sampled every possible game that meets my criteria. If I wanted a bigger sample, I'd have to include one or more of the following:

- expired games
- games that ended before Turn 2
- games that ended on a boot

I didn't include expired games because their impact isn't felt on current ladder ratings (stalling's all about the effect, right? we can't really quantify or police intent here) and because they're too old to accurately represent a player's behavior/habits.

Games that ended before Turn 2 were excluded as they'd randomly skew the times for either losses/wins and because, before Turn 2, you generally do not have sufficient information to determine who's winning/losing. This was done to mainly weed out first/second-turn surrenders after looking at the top players' ladder histories.

Games that ended in a boot were excluded for a similar reason- they're too random to reflect a player's tendency (you could get booted while winning or while losing) and would just skew the data.

If I wanted to increase the sample size for players like alhazi, Buns, and that one person from Vitrix with only 4 games analyzed, I'd likely just include expired games that ended before a certain date (so setting a new expiration threshold), change the turn cutoff to Turn 1 (since there's some cases where you have the intel to know you're very likely to lose), and perhaps include games that ended in boot after a certain # of turns and when certain thresholds were met (e.g., booted player had 20% less income and 20% fewer armies than the winning player)- although that would still be an issue as you don't control when you get booted, and this is much easier to analyze when players have control over how long they take to lose (through the surrender button).


Beyond that, sample size isn't an issue for most players. I just included all players that had at least 1 win and 1 loss because we could just ignore the weird cases if we wanted to.


In the case of Alhazi you can probably include those games as well. We all know he is a massive staller, if he got booted it was because he didn`t want to play that game anymore, not because he wasn`t able to take his turn in time


True. I didn't notice alhazi's boot history until after I'd written the script and it'd produced the data. That said, it'd be hard to programmatically catch these cases with enough precision for me to be comfortable, given current data (what thresholds do I set?).

Edited 5/31/2016 21:54:44
Rio 2016 Stallympics: 5/31/2016 22:13:24


TBest 
Level 59
Report
Yaaas, I am a top ten staller!

Or, we might need some smart ecuations stuff to find real stalling. Hm. Ther is so many factors to consider. Maybe manully looking at games will reamin the best way to catch stallers /:
Rio 2016 Stallympics: 5/31/2016 22:23:04


Zephyrum 
Level 60
Report
@Knyte, a negative % means you take your turns faster when you are losing?

292.                                 Zephyrum (ID: 12790; clan:        Coalition Of Role Playing Friends):  -43.90% 281 1343  31  14  17 
Rio 2016 Stallympics: 5/31/2016 22:29:07


zażółć gęślą jaźń
Level 55
Report
It surely does. I have negative % as well and I often feel my opponents are taking too long in lost games...

Great stuff like always. Alhazi's percentage is hilarious :D
Rio 2016 Stallympics: 5/31/2016 22:40:39


kynte
Level 51
Report
@Zephyrum: Remember that your opponent controls how long your wins take more than you do. It likely means you're less of a staller than your average opponent.

@TBest: This was just a preliminary look. I didn't want to use some contrived thresholds to determine when people were "losing" vs. "winning" or "even" and instead just went with what we already know- which games ended with wins, and which games ended with losses- and how long they ended up going. Lots of limitations here- it's looking at long-term behavior rather than single scenarios (so you can hide some stalling if your opponents stall too), the analysis is comparative (less "do you stall?" and more "how do you compare with others?"), and loosely defines stalling as taking longer to lose than you do to win (which catches stalling, but also potentially some other things- it judges your behavior by effect, not intent).

In your case, you take about 3 more turns to lose than to win, so perhaps you're just slower than average to surrender. Wanted to get input from people like you and Buns again so I could figure out which potential confounding variables to weed out/etc. in the next step.

Manually looking at games isn't an efficient way to catch stalling on the 1v1 ladder, anyhow. You can't analyze the whole ladder quickly, nor can you easily track time rather than # of turns (stalling takes place in longer turns at least as much as it does in more turns + time matters more on the ladder than # of turns).

That said, manually looking at your unexpired losses doesn't exactly invalidate the results.


Other ideas I've had:

- finding a threshold for which the experimental probability of a comeback is about 5%, seeing how long you continue after hitting that threshold

- getting your income/army ratio to your opponent at the end of a loss (so some of the conditions that it would take for you to surrender, ignoring map geography)

- a similar analysis as this one but adjusting for opponent rating in some way

Edited 5/31/2016 22:46:24
Rio 2016 Stallympics: 5/31/2016 22:53:10


Жұқтыру
Level 55
Report
I'm in the top 10 players that stall the least :)
Rio 2016 Stallympics: 5/31/2016 22:54:52

[wolf]japan77
Level 56
Report
I think the threshold analysis would probably be the best of those options. Income or army ratio may not work so well, as map position generally impacts whether or not I surrender at a given point(I have surrendered at times when I was ahead in income due to my opponents basically being in position to break all my bonuses within the next few turns.) Adjusting for opponent rating may not work well either, as top players may just have too few losses for you to get a validly representative sample.

However, the major issue in that case would becoming up with such a threshold.
Rio 2016 Stallympics: 5/31/2016 22:55:25

[wolf]japan77
Level 56
Report
@xbpty
It's all relative comparisons at this point.
Rio 2016 Stallympics: 5/31/2016 22:55:56


Жұқтыру
Level 55
Report
It's all relative comparisons at this point.


I'm still right, though (as far as data go).

Edited 5/31/2016 22:56:13
Rio 2016 Stallympics: 5/31/2016 23:01:05


kynte
Level 51
Report
Well, you stall the least relative to your opponents, based on this analysis. Opponent pool may have a significant impact here.

Another facet here is that you take 5 days to win and 2 days to lose, averaging about 13 turns for both wins and losses. So games where you win tend to have turns that take 150% longer on average, and that could be on your part or your opponents'.

Also it's not necessarily a positive thing to have a very negative rating- you could, for example, just be playing some games more quickly than you should be and losing as a result.

Edited 5/31/2016 23:03:42
Rio 2016 Stallympics: 5/31/2016 23:40:49


{Shredtail2}
Level 56
Report
[NVM misread shit]

Edited 5/31/2016 23:42:01
Rio 2016 Stallympics: 6/1/2016 00:12:06


TeamGuns 
Level 58
Report
Stalhazi should get a gold medal for that. Would be good if you had a second list with expired games so we could see the all-time winners.
Rio 2016 Stallympics: 6/1/2016 00:35:07


skull11244
Level 58
Report
idk i think Ellhoir should get a medal for being the least likely to care if he's winning or losing.
Rio 2016 Stallympics: 6/1/2016 01:09:16


Semicedevine
Level 59
Report
287. Semicedevine (ID: 8870; clan: The Lost Wolves): -43.42% 51 1864 100 60 40 -20:57:09.625000 2 days, 8:37:55.350000 1 day, 11:40:45.725000 2 days, 0:15:03.500000 -1.58 11.68 10.10 11.05

this looks so fake lool

i mean look at how im playing my ladder games it is a recipe for stalling lol
Rio 2016 Stallympics: 6/1/2016 01:25:02

[wolf]japan77
Level 56
Report
It may be that everyone else around you is stalling even more so. remember all rankings are relative atm.
Rio 2016 Stallympics: 6/1/2016 01:28:04


LeQuébécois_Benoit
Level 59
Report
I agree with TBest, show some examples where I stalled because I am in top 10 of your list. I play slower when I lose for sure because I have less enthousiasm about the games I am losing. If that makes me a staller so be it, but you'll see checking my ladder games that I don't wait until my opponent have the rest of the map to surrender...
Rio 2016 Stallympics: 6/1/2016 02:04:41


TeamGuns 
Level 58
Report
Maybe there could be a ranking that mixes ammount of turns taken when lose against the ammount of turns taken when you win, and cross those in some way with the ammount of time.

Bc I really don't think all of these are really stalling, I for once also take more time to play when I'm losing against when I'm winning, simply bc I want to think more. But I will never do that to gain any unfair advantage on the ladder and I will always surrender when all hope is lost.
Rio 2016 Stallympics: 6/1/2016 02:18:40


{Shredtail2}
Level 56
Report
I usually just play all my turns on games when I have less than a day left until boot. A bad habit, I know, but it has stuck to me since I had some crazy amounts of schoolwork that has kept me busy. I don't intentionally stall games to gain any sort of advantage, as I also surrender when I feel I lost.
Rio 2016 Stallympics: 6/1/2016 02:34:53


kynte
Level 51
Report
I play slower when I lose for sure because I have less enthousiasm about the games I am losing. If that makes me a staller so be it


Like I said in the early posts, this rating measures effect, not intent. If you play slower when you're losing, that falls within my definition of stalling as you draw out your losses and delay their impact on your rating.

, I for once also take more time to play when I'm losing against when I'm winning


That's true for probably everyone. However, keep in mind that your wins are your opponents' losses.

If you're playing slowly when you're losing more than your opponents tend to, you end up with a positive percentage. Simple as that.

Edited 6/1/2016 02:37:52
Rio 2016 Stallympics: 6/1/2016 02:50:11


Peixoto
Level 63
Report
Took a while until I found me O_o
Rio 2016 Stallympics: 6/1/2016 14:24:20


Beren • apex 
Level 62
Report
Ctrl+F, young grasshopper.
Rio 2016 Stallympics: 6/1/2016 20:31:00


TBest 
Level 59
Report
@Knyte,

Wanted to get input from people like you and Buns again so I could figure out which potential confounding variables to weed out/etc. in the next step.


See, I wanted to give this istead of a sarcastic response, but then I started looking into what variables that was.... and that is a lot. Practically speeking, the following factor is what I would try to look into.

1. Rating.
a) Look at the avg. opponent rating for won and lost games. If the difference is too big, the data can not be used to determine stalling. If rating is close (within 100 points?) then the data is more relevant. This method has several obvious disadvantages, but might be interesting either way. At least it would show a difference for my case :p This works better on players who have been active in the ladder for more then a expiration period (5 months, continuously)
b) Look at the average time it takes a X rating player playing Y rating player to achieve Z result, across the whole ladder. Then compere it to any individual player to see how long time they take to loose, compered to the norm. Don't think this would work to well, tbh.
c) Only consider games that is within X (150?) rating of a players own rating, then do the calculations you have already done for win/loss time.

Armies, income

Use the same rules as Seasonal ladder does to determine ties, and apply to the last turn in every game. Compere avg. win number, with avg. loss number. Then rank players, based on the difference.


Anyway, just some ideas. All the methods listed above have flaws but they might be interesting nevertheless. Mainely, I see the challenge being how you include all metrics in one ranking way. As you see the Rating methods only accounts for time taken, blantetly assuming ther is some sort of avg. game comparison that can be made. While Armies, Income in my suggestion don't account for time taken, only turns. What I fear is that one might quickely have a small sample size if combining both.
Rio 2016 Stallympics: 6/1/2016 20:34:49


Beren • apex 
Level 62
Report
Look at the avg. opponent rating for won and lost games. If the difference is too big, the data can not be used to determine stalling.


I disagree a little bit here. Wins against severely lower rated players and losses against much higher rated opponents could be discounted, but losses to low ranked players are the most likely games to be stalled.
Rio 2016 Stallympics: 6/1/2016 20:41:30


Fan the Apostle
Level 56
Report
I actively try not to stall.
Rio 2016 Stallympics: 6/1/2016 21:50:08


kynte
Level 51
Report
As far as rating goes, perhaps it'd be viable to find for each team a function that best approximates how much time they'd take for a given player given their rating and use that instead of loss - win. Alternatively, find the functions l(r), w(r), and a(r) that reflect loss, win, and average times (respectively) given an opponent's rating, and compare those (this is a place where it would get tricky unless you force the functions to be in the same family). This would likely also require me to trim out a bunch of teams that don't have decent (10w, 10l?) sample sizes as well as outliers.

Using time taken is imho much more valuable than turns because stalling also involves taking longer turns, and the # of turns taken has no direct impact on your rating- stalling is only effective if you make the game take more time, not more turns.

I ignored opponent rating in the first run-through and assumed that the ladder's doing a decent job as far as pairing and rating goes.

Edited 6/1/2016 21:51:39
Rio 2016 Stallympics: 6/2/2016 07:57:02

HotBeachBum
Level 62
Report
I love this data! Great job. I pride myself in being the highest rated player with the lowest stalling component, always playing every game at the same pace, and the current hottest player on the ladder over the past 3 months plus. Watch out, here I come....
Posts 1 - 30 of 38   1  2  Next >>   

Contact | About WarLight | Play Risk Online | Multiplayer Strategy Game | Challenge Friends, Win Money | Skill Game | Terms of Service