Talk Fantasy Football

sann wrote:Tier 1: Chaos Dwarf, Dwarf, Wood Elves, Skaven, Norse, Lizardmen, Orc, Undead, Amazon, Dark Elves

Tier 2: Chaos, Human, Khemri, Pact, Slann, High Elves, Nurgle, Necromantic, Pro Elves

Tier 3: Halflings, Gobbos, Vamps, Ogres, Underworld

most comon tiers i see in france/belgium/around is the following:

sann wrote:Tier 1: Chaos Dwarf, Dwarf, Wood Elves, Skaven, Norse, Lizardmen, Orc, Undead, Amazon, Dark Elves, Necromantic

Tier 2: Chaos, Human, Khemri, Pact, Slann, High Elves, Nurgle, Pro Elves, Underworld, Vampires

Tier 3: Halflings, Gobbos, Ogres

Sometimes I saw 4 tiers too.

Tier 1: Chaos Dwarf, Dwarf, Wood Elves, Skaven, Norse, Lizardmen, Orc, Undead, Amazon, Dark Elves, Necromantic

Tier 2: Chaos, Human, , Pact, High Elves, Pro Elves

Tier 3: Khemri, Nurgle, Underworld, Vampires, Slann

Tier 4: Halflings, Gobbos, Ogres, LIzardmen (all skinks), Underworld (only goblins)

Wulfyn used k-means analysis on NAF 2012/13 data to come up with the following tiers:

T1 (55.26) : Wood Elves, Undead, Dark Elves, Lizardmen
T2 (51.87) : Amazons, Elves, Norse, Dwarves, Chaos Dwarves, Necromantic, Skaven
T3 (48.93) : High Elves, Slann, Orc, Humans
T4 (45.54) : Khemri, Chaos Pact, Nurgle, Vampires, Underworld, Chaos
T5 (33.91) : Halflings, Goblins, Ogres

dode74 wrote:Wulfyn used k-means analysis on NAF 2012/13 data to come up with the following tiers:

T1 (55.26) : Wood Elves, Undead, Dark Elves, Lizardmen
T2 (51.87) : Amazons, Elves, Norse, Dwarves, Chaos Dwarves, Necromantic, Skaven
T3 (48.93) : High Elves, Slann, Orc, Humans
T4 (45.54) : Khemri, Chaos Pact, Nurgle, Vampires, Underworld, Chaos
T5 (33.91) : Halflings, Goblins, Ogres

Interesting!

dode74 wrote:Wulfyn used k-means analysis on NAF 2012/13 data to come up with the following tiers:

T1 (55.26) : Wood Elves, Undead, Dark Elves, Lizardmen
T2 (51.87) : Amazons, Elves, Norse, Dwarves, Chaos Dwarves, Necromantic, Skaven
T3 (48.93) : High Elves, Slann, Orc, Humans
T4 (45.54) : Khemri, Chaos Pact, Nurgle, Vampires, Underworld, Chaos
T5 (33.91) : Halflings, Goblins, Ogres

This one I liked!

dode74 wrote:Wulfyn used k-means analysis on NAF 2012/13 data to come up with the following tiers:

T1 (55.26) : Wood Elves, Undead, Dark Elves, Lizardmen
T2 (51.87) : Amazons, Elves, Norse, Dwarves, Chaos Dwarves, Necromantic, Skaven
T3 (48.93) : High Elves, Slann, Orc, Humans
T4 (45.54) : Khemri, Chaos Pact, Nurgle, Vampires, Underworld, Chaos
T5 (33.91) : Halflings, Goblins, Ogres

Which is the problem with lies, damn lies and statistics.

The numbers for Necromantic (and possibly Elves) are skewed by the fact that they are always the best team in tier 2 and so get bonus skills on tier 1 races. As a consequence, people usually only take them (or take them more often) at tournaments where they get a leg up.

The numbers for Orcs and Humans have always been considered skewed because they are "out of the box" teams and so are considered vanilla and more likely to be taken by rookies, nosediving the average. Orcs in particular are so, so, so clearly a better team than Elves or Necros in the hands of a good coach.

Dark Elves is another interesting one. IMO they are one of the best teams, and the numbers support this. However, this comes with two large caveats. One, they are an extremely difficult team to play well, but are a team with one of the highest ceilings for getting juice from your decision making. Even some of the best coaches in the world struggle with them at first.* Rookie coaches will often try them once and do horribly. As such, Dark Elves are most often used by good coaches and so I would guess their data contains the fewest games by rookies dragging the average down.

Two, they are only any good in tournaments with a medium to high number of front loaded skills. In tournaments with few skills or one skill per game, they are extremely average, probably worse than Humans. As such, few people take them in tournaments that don't suit them. Running a tournament with these tiers would disadvantage DEs and probably make them unplayable. For example, DEs vs Orcs is a really tough game; fair but if it favours anyone it is the Orcs. There is no way DEs should be playing that match up 2-3 skills down.


*Contrast this with say Undead. They are definitely in the top tier, but I would suggest their numbers are high both because they are good but also because they are forgiving. You can just beat people up with Mummies and get away with a lot of things, so even rookie coaches can win plenty of games with them.

dode74 wrote:Wulfyn used k-means analysis on NAF 2012/13 data to come up with the following tiers:

T1 (55.26) : Wood Elves, Undead, Dark Elves, Lizardmen
T2 (51.87) : Amazons, Elves, Norse, Dwarves, Chaos Dwarves, Necromantic, Skaven
T3 (48.93) : High Elves, Slann, Orc, Humans
T4 (45.54) : Khemri, Chaos Pact, Nurgle, Vampires, Underworld, Chaos
T5 (33.91) : Halflings, Goblins, Ogres

All for statistical analysis. Should be noted however that k means takes k (number of clusters) as an input. No evidence that 5 is the "best", just what was chosen in this case.

Which is the problem with lies, damn lies and statistics.

Which is a stupid quote often made by people who have no understanding of statistics
It's misuse of statistics which is the problem, not statistics itself. That's what the quote means: lies backed up by poor use of statistics are the worst sort. See homeopathy, for example.

The numbers for Necromantic (and possibly Elves) are skewed by the fact that they are always the best team in tier 2 and so get bonus skills on tier 1 races. As a consequence, people usually only take them (or take them more often) at tournaments where they get a leg up.

Do you have any actual evidence that the numbers are skewed, or is this supposition?

The numbers for Orcs and Humans have always been considered skewed because they are "out of the box" teams and so are considered vanilla and more likely to be taken by rookies, nosediving the average. Orcs in particular are so, so, so clearly a better team than Elves or Necros in the hands of a good coach.

Same question as above.

Dark Elves are most often used by good coaches and so I would guess their data contains the fewest games by rookies dragging the average down.

And again...

The reality is as the data shows. You can make all the suppositions you like for the reasons for the data being what it is, but it is what it is. How you use the data is another matter entirely, but making assumptions about the reasons and working on the basis of those assumptions will skew any further data more than simply working to the data will.

mubo - Wulfyn tried 4 and found 5 fit better. The rule of thumb for the number of clusters is root(n/2), which is root(12), so 3-4. With 5 you get all the centroids within the CI95 for their data. You could have more, but it seems to me that the aim would be to minimize the number of tiers where possible - otherwise you might as well have 24 tiers.

dode74 wrote: mubo - Wulfyn tried 4 and found 5 fit better. The rule of thumb for the number of clusters is root(n/2), which is root(12), so 3-4. With 5 you get all the centroids within the CI95 for their data. You could have more, but it seems to me that the aim would be to minimize the number of tiers where possible - otherwise you might as well have 24 tiers.

root(n/2) is not a rule of thumb here. Not sure where that comes from.

I may be mistaken, but I think Dan just did k means on a single dimension (win ratio). Meaning it's not even kmeans, it's just saying lets chose k values and minimize how far each is from nearest observation. I suspect 5 fits better than 4 for exactly the reason you hint at with k=24; As you increase the number of clusters the variance of each group will go down. I would wager that 6 has a lower SS than 5.

When you have a group of 5/6 observations it's not possible to get a good estimate of sd and pretty unlikely that any single value will fall outside 2 of these (95% CI, I assume normal dist here). Your CI example is absolutely not a good reason to believe k=5 over k=4. I don't want to debate the relative merits of 3 vs 4 vs 5. But I think to use this analysis as some kind of evidence that there are 5 tiers is pretty misleading in my opinion.

On a wider point, we all know that coach ability affects win %.

I would suggest that the onus is on the analyst to:
a) control for biases from uneven skill packages
b) control for biases regarding coach strength
c) control for biases arising from swiss (ie 'flings play more 'flings than other races do)
or
c) show there is no bias
rather than for others to find evidence for biases.

Hi Dode,

a) I started with what I thought was a lighthearted comment, hence the smiley. Chill. I know Wulfyn, he is in my real life BB league and I have chatted with him many times on the subject and greatly appreciate his input.
b) Yes these are my suppositions from understanding both the numbers and where the numbers come from. I.e. knowing what the hell I am talking about, not just seeing some numbers on the internet. It was an effort to help explain what might be going on rather than just lapping up the figures as previous comments seemed to.
c) What mubo said.

Good point re rules and regs, it was meant to be "common rules", or "example rules" but maybe that wasn't clear.

Well done Mike, less than a month in and you've created a spark of interest that will run and run...

please note: I'm going to Skew more data by playing Necro in my next tourney, thus reducing their win ratio

Oh and I don't mean in any way to disparage in any way what Wulfryn has done, I think it's a really important service to the community and should be applauded. I'm just pointing out the pitfalls of accepting an analysis blindly.

If we are to meaningfully look at WR in NAF games you should look at biases too. I think it's very possible that we have a race that is very weak in the hands of an inexperienced player, but very strong in the hands of a good one. Simple WR doesn't capture that at all (Joe's point).

Also, I would trust the WR ratio more if we just looked at R1. Swiss will have the effect of compressing your WRs, as woodies will tend to play more woodies, and if they do play T2/3 in a later round they are probably in the hands of a very good coach.

mubo - I agree coach ability affects win%. What we can't say is that there is any difference between which levels of coach skill play which races. We can't even measure coach skill since individuals' win ratios will be determined at least in part by the race they play. All we can reasonably do is take as large and "clean" a dataset as we can.
[quote][/I think it's very possible that we have a race that is very weak in the hands of an inexperienced player, but very strong in the hands of a good one. Simple WR doesn't capture that at all (Joe's point).[quote]Of course, but that will not have a large effect on the overall data unless a race in particular is played more by one "skill level" of coaches over another.

joemanji - yeah, I got you were lighthearted about it and thought my initial response was too - hence the smiley
That said, the "knowing what the hell I am talking about, not just seeing some numbers on the internet" comment comes across as somewhat unwarranted. I also know where the numbers come from and understand what they mean. It's based on individual match-level data.