Coli Shuffle Winner Predictions

[lessigfan]

I put this in the coli shuffle thread, but I thought I'd put it here as well, so you guys can see it.

I snagged & cleaned up the data for all the shuffles to run a few regressions. I tried a few theories as to model specification, but eventually just did stepwise regression. If that doesn't make sense, let me put it this way: I had the statistics program figure out the best model.*

I was curious what made you more/less likely to win a shuffle - level, class, number of charms, PvP rating, HP, spirit, or the same attributes of your allies. For yourself, only PvP rating & class** was strongly predictive of how well you'd do in a shuffle, HP was marginally predictive. For your allies, again, only their PvP rating & class was relevant, and maybe HP. But not their level, number of charms, or spirit.

Anyhow, plugging in current team values into the fancy shmancy regression coefficients, I ginned up some predictions for this shuffle. Not the predictions I'd have made personally, but my opinions aren't relevant to this post.

The 3 teams predicted to do worst:

1. Valsu, Astrolabia, & Powerslave
   
2. Slickal, Gothmogged, & Sheira
   
3. Medefe, Slalderma, & Flame123

The 3 teams predicted to do best:

1. Tobenoid, Qqchampion, & Badmotivator
   
2. Godfat, Toben, & Giab
   
3. Mikka, Flavor, & Byakuren

If you're interested in the data or R code, let me know and I'll put it up on a share somewhere.***

*Note: Never use stepwise on real data. If you ever do stepwise on real data without **at least** cross validating, I will hunt you down. I also did logistic stepwise regression, since the dependent variable in my model was the win percentage sheira reports here, and it is bounded between 0 and 1, but the results - as far as significance & relative impact - were basically the same. I didn't bother looking for heteroskedasticity (which may be a problem), collinearity (near multicollinearity is certainly a problem), or any other gauss markov problems - so don't put a lot of stock in the validity of anything here.

** The class order wasn't strongly predictive, but here is the list from most likely to win a shuffle to least: Inquisitor, Warden, Summoner, Slayer, Assassin, Deathknight, Champion, Paladin, Warlord, Berserker, Pyromaniac, & Hierarch. I'm certain this list is affected by outliers that win repeatedly, like Chaofan or Toben. I'm doubly skeptical because Paladin is midway in the list, whereas some of those lower in the list are much stronger classes. Sheira & Sparky did a much more thorough analysis of PvP ratings v. class, and those results seemed more right to me.

*** I was also curious about whether charms, level, class, HP, or spirit are predictive of your PvP rating. More charms & HP were correlated with higher PvP ratings. Every 10 more charms was correlated with ~3 extra pvp points, and every 1 extra hp was correlated ~6 extra pvp points. Class was marginally relevant; level and spirit weren't important. The data is only of shufflers, so only about 400 observations, many are repeat individuals, and not much variety in levels, so I don't put much stock in this [hence it is a footnote of sorts]

Anyone else have a prediction?

[Howie Feltersnatch]

In your calculations, did you take into account total charms at all? Not that total charms are really that important other than high totals usually mean well developed characters and lots of epic charms.

[MeDeFe]

And with Flame123 being replaced by Soktebel, you may have to recalculate the three groups likely to do the worst. Our new teammate has a twice as high rating as our old one.

[Howie Feltersnatch]

Flame is just a kiai donk

[MeDeFe]

The guy is inactive and was replaced, is what I'm saying.

[Howie Feltersnatch]

I saw that, I just meant that he is kind of a statistical anomaly since his rating sucks, his hp is decent (i think) but with good teammates can be deadly just running as a kadonk (which is all he can do)

[lessigfan]

So I did this again, and if you want the excel sheet or R code, let me know and I'll send it over. The first question is how accurate was the last set of predictions?

Final Order v. Predicted Order v. Scoreboard % ordered from first to last

T03, T11, T07, T12, T08, T10, T06, T04, T09, T01, T02, T05 (too few fights to count)
T06, T10, T01, T11, T08, T12, T07, T09, T04, T02, T03, T05
T11, T03, T07, T12, T04, T10, T09, T08, T02, T06, T01, T05

Not very accurate. It got most teams +/- 2 spots at the bottom, but was bad about picking winners. The biggest mistake was predicting Shiera/Gothmogged/Slickal's team to take 2nd to last, when they took a commanding first. The prediction was probably no better than random chance. So the models predictive power? Bunk.

Since the model was so bad, lets do it again! I did the same thing as before (stepwise logit & GLM regressions), so I should get very similarly bad predictions this time. Again, I found only class & rating mattered. Here are the results:

Top 2 Teams:

• Howie/Tobenoid/Kamil
  
• Mikka/Zenetar/Drakon

Bottom 2 Teams:

• Medefe/Sukelywalker/Vestigate
  
• Valsu/Ruune/Lessigfan

I also thought of another way to think about the shuffle - instead of considering team attributes together, analyze only the past performance & attributes of each indivual alone, score them, then add the individual scores for each team. This led to some different results:

Top 2 Teams:

• Mikka/Zenetar/Drakon
  
• Godfat/Jeeves/Rebelscum

Bottom 2 Teams:

• Howie/Tobenoid/Kamil
  
• Medefe/Sukelywalker/Vestigate

It is a bit strange that the howie/tobenoid/kamil team shows up as first in the first regression and dead last in the individual regression. That alone should convince anyone that the results have no validity, if the Shiera/Gothmogged/Slickal outcome last time hadn't already!

So what is the take-away? Statistically, Mikka's team is a favorite and Medefe's team is an underdog. But last time the underdog won the whole thing, so who knows?

My guess? Whichever team can run null barbs strategy best

[MeDeFe]

And? How's it looking for January?

[lessigfan]

I'm still on vacation, so I probably won't be able to run the numbers until the shuffle is halfway done. Look for new predictions for February, assuming I don't get a job between now and then

[Howie Feltersnatch]

Have you tried looking here for jobs?

[TheJackal]

Have you tried looking here for jobs?

One wonders how Howie knew about that site... Bookmarked too I bet

[Sponsored content]