Races and classes by streakability

[Berder]

Green dots are wins, red dots are losses. Races/classes are sorted by a custom scoring system, which is like winrate of those races/classes, but is adjusted for difficulty. For instance, a race that's often played with berserkers loses a little score compared to a race that has more games from the harder classes. The percentages in the margins are winrate.



Method:
I took the data on streaks from http://crawl.akrasiac.org/scoring/streaks.html. I looked at "attempted streak" games, which are either games in a streak or the streak breaker, excluding the first in streak. This chart looks at games that had the potential to be at least 4th in streak - so the streak breaker of a 3-streak or longer, or the 4th or later win in a 4-streak or longer.


Edit note 1: excluded unfair streak-breakers from lazy-life.ddo because someone (HLA says it wasn't him) deliberately ruined other people's streaks on that server
(here's the old one that included lazy-life.ddo)
(pre edit note 1 version: http://i.imgur.com/V5y9lMk.png )

Edit note 2: Re-ordered columns to account for some races being played easier classes than other races (and vice versa) (see debate towards end of page 1 of thread). Specifically,

First, I scored races. To do this, first I scored how well that race did with each class, compared to the average for that class with other races. A race's score for a class is the winrate of that race+class combo, minus the winrate of all games played with that class that weren't done with that race. Then, the race's final score is the average of its score for each class, weighted by the number of games it played in each class.

Then I did a similar procedure (with race/class swapped) to score classes.

(pre-edit note 2 version: http://i.imgur.com/tmU1U6u.png )

[XuaXua]

Thanked before someone disputes your numbers.

[tasonir]

While the visual format is nice, it's not very glanceable, I'd list races/classes in order like in the previous thread?

Also while I'm not disputing berserkers are strong you have to admit tedronai's berserker streak is definitely boosting that He even went out on an EE, so no Be death!

[Berder]

While the visual format is nice, it's not very glanceable, I'd list races/classes in order like in the previous thread?

They are in order - columns and rows are sorted by winrate.

[Sandman25]

IMHO this proves that streak statistics is meaningless in crawl, Fe is the best species, Hu is the worst and CK is better than Fi

[Berder]

Fixed display issue.

[elmdor]

This information is interesting and I like the visualization. But I wonder why you limited the data to streaks of 4 or more. There are about as many 3-streaks as 4,5,6... etc- streaks combined. And 2-streaks add up to more than all the rest of the others. What is your reason for not examining the rest of the of data?

[Berder]

IMHO this proves that streak statistics is meaningless in crawl, Fe is the best species, Hu is the worst and CK is better than Fi

I think Fe is legitimately in a top position. All a felid has to do is make it to xl3 and he has an extra life, and he has fast movement speed to help him do that. Once you have an extra life, if you're streaking and really don't want to die, you're probably safe.

Fi is artificially low because so many people broke their streaks with HuFi. If you excluded humans Fi would be a lot better. CK legitimately seems to do okay, though I do see more of the stronger races in the CK column.

Hu is bad largely because of HuFi.

Elmdor, I excluded shorter streaks because I wanted to get as close to optimal play as possible. Shorter streaks might be from weaker players. Here is the chart generated from all streaks, including 2- and 3-streaks:

[Rast]

All this proves is that if you're already a strong player, and you're in a streak you care a lot about, you'll be willing to play in an ultra-careful, ultra-tedious way, and then the extra speed and lives of Felid trump all other considerations.

Also, the way tournaments reward scoring encourages the very best players to play weak roles on strong races, and strong roles on weak races. IF players were trying to just win, without concern for repeating roles and races, the table would be 80% zerkers.

[Sandman25]

I have another explanation - steaks combine powerful species with weak background like MiCK or TrWr and vice versa. It does not make much sense to look at CK win rate without analyzing species used.
Probably a better approach would be to create a separate chart for every species/background but it is a lot of work and is not really helpful, we already know that TrMo/GhMo are the best Mo etc.

[Berder]

I have another explanation - steaks combine powerful species with weak background like MiCK or TrWr and vice versa. It does not make much sense to look at CK win rate without analyzing species used.
Probably a better approach would be to create a separate chart for every species/background but it is a lot of work and is not really helpful, we already know that TrMo/GhMo are the best Mo etc.

The way to compensate for this would be to alter the ordering of cols and rows so it's not solely based on winrate, but also considers the difficulty of the combos played. I could use a variant of Elo for this, or maybe some other system.

But that's why the chart has the dots. You don't have to go solely by the ordering of rows and cols; you can look and see, e.g. HuFi is weak but human mages aren't bad.

[elmdor]

Thanks for this new chart. I understand you're trying to get a sense of which races/backgrounds are better for streaking, assuming high player skill. But I'm not convinced that limiting it to streaks >= 4 really limits the sample to "optimal play."

Sure, weaker players are going to be less likely to streak than elite players. But low streak length isn't necessarily due to lower player skill. It might be because a certain player's #1 priority at that time wasn't simply maximizing their streak length. They might have been trying to run a specific build for tourney reasons, or trying to finish new races for greatplayer, or just play builds that seemed fun or challenging or different, rather than grind out a MiBe because it seemed like the best way to extend their streak...

(Also: I wonder if this many people picked HuFi because they mistakenly thought it would be a safe, "boring" streaking choice..?)

[daggaz]

You should convert the graphic display into a simple percentage. This has the advantage of being much easier to read, and while it can be interesting to see how popular a particular combo was, it is easier and makes more sense to normalize the winrate when comparing combo's. You could still show n, like MiBe 97% n=27. You could even compile the statistics for each streak length, and make a final weighted average (such that a 8 win streak is worth more than a 2 win streak).

But otherwise pretty cool.

[XuaXua]

Also you can make percentage text colored if you still feel you need to.

[Berder]

You should convert the graphic display into a simple percentage. This has the advantage of being much easier to read, and while it can be interesting to see how popular a particular combo was, it is easier and makes more sense to normalize the winrate when comparing combo's.

The problem is that it's much less clear. See http://i.imgur.com/tjCnnqG.png. It does make sense for the totals in the margins for race/class winrates to be percentages, but for the combo entries it's very hard to tell anything at a glance.

You could even compile the statistics for each streak length, and make a final weighted average (such that a 8 win streak is worth more than a 2 win streak).

I'm not sure what you mean.

[Dole]

viewtopic.php?f=17&t=14327

Most of those HuFi's shouldn't count.

[Berder]

https://crawl.develz.org/tavern/viewtopic.php?f=17&t=14327

Most of those HuFi's shouldn't count.

Thank you - I've now updated the original post to exclude streak breakers from the LLD server, unless the streak also had games on the LLD server. They were all HuFi except for one TrMo. As you can see, Human is still near the bottom, but Fighter leaped up to near the top.

[and into]

This is cool. And I'm not surprised by Fe's position at all.

[Sandman25]

This is cool. And I'm not surprised by Fe's position at all.

This is explained by Be IMHO, too many Fe were Be. If we remove Be, Fe will not be that great.

What about GhMo? Why is it only slightly above 50%? Even MuCK has higher win rate

[and into]

This is cool. And I'm not surprised by Fe's position at all.

This is explained by Be IMHO, too many Fe were Be. If we remove Be, Fe will not be that great.

What about GhMo? Why is it only slightly above 50%? Even MuCK has higher win rate

Fe is pretty solid across the board in comparison to other species. Going from the board with "small streakers," and discounting FeBe, you still have Felids getting greater than 75% (that involved a bit of counting tiny boxes on a small monitor so ymmv but Fe come out looking good either way).

As for GhMo and MuCK, I'd just say that obviously these statistics are not an unmediated expression of the strength of these species (so I'd recommend new players not draw hasty conclusions and go out there trying to streak GhWr and GhAE). The statistics show how "streakable" these characters proved to be for the players that chose to play them; thus, the stats say something about these species and backgrounds, but also say something about the level of skill of those who tend to choose those species, at the same time. (Even at high levels of play there are meaningful differentiations in skill level.)

[tasonir]

While Mo is a good start for Gh, imho Gh is one of the weaker races. Really bad int and dex, not much upside, basically a bad fighter (-1 to most apts, although they have +1 fighting) who can't really branch out into much else easily. At least if you take UC you get +1. Also they can't transmutate, which kills one of the perks of having taken UC.

I'd still say they're better than MuCK's of course, but they aren't good.

[Sandman25]

Well, GhAE is actually great but your main point stands - it tells more about players than about combos and I believe this explains Fe anomaly:
1) Fe is considered a challenging species so it is played in long streaks only and we all know that long streaks are done by really great players
2) only a few players love/play Fe and probably they are real masters with it

PS. We should probably change the queries to limit them with "player=elliptic"

[tasonir]

Don't exclude hyperelliptic, I hear he's just as good as elliptic is.

[Sandman25]

While Mo is a good start for Gh, imho Gh is one of the weaker races. Really bad int and dex, not much upside, basically a bad fighter (-1 to most apts, although they have +1 fighting) who can't really branch out into much else easily. At least if you take UC you get +1. Also they can't transmutate, which kills one of the perks of having taken UC.

I'd still say they're better than MuCK's of course, but they aren't good.

Is it really you tasonir? The species has +1 UC and you are not happy with it
Please take a look at GhMo cell (not Gh row), it has lower win rate than MuMo/DEMo/HEMo/OpMo/NaMo/DrMo etc.

[Sandman25]

Don't exclude hyperelliptic, I hear he's just as good as elliptic is.

What does "hyper" mean by the way?

Edit. Just found it in learndb, sounds fun.

[and into]

Well, GhAE is actually great but your main point stands - it tells more about players than about combos and I believe this explains Fe anomaly:
1) Fe is considered a challenging species so it is played in long streaks only and we all know that long streaks are done by really great players
2) only a few players love/play Fe and probably they are real masters with it

PS. We should probably change the queries to limit them with "player=elliptic"

I don't think GhAE is great for new players, that's what I was saying.

You raise a good point with item 2. Some backgrounds and species have a particular niche appeal, and we are dealing with small enough sample sizes that arbitrary choices of particular good and prolific players can definitely have an impact.

All the same, I'd stand by the assertion that Fe is a strong species for winning (not for fun IMHO), though arbitrary factors could be inflating its performance on the above boards.

[Sandman25]

I don't think GhAE is great for new players, that's what I was saying.

You raise a good point with item 2. Some backgrounds and species have a particular niche appeal, and we are dealing with small enough sample sizes that arbitrary choices of particular good and prolific players can definitely have an impact.

All the same, I'd stand by the assertion that Fe is a strong species for winning (not for fun IMHO), though arbitrary factors could be inflating its performance on the above boards.

My understanding is that that table was created from 4+ games streaks only so it is not really related to new players. I have a GhAE win but it was part of a 3 games streak so it didn't make it into the table unfortunately.
Item 2 was easy to write, I love Fe personally (it's my second most won species right after Mu) and indeed it is not that hard after getting some experience with it (initially I was dying a lot on D1-3 to find which monsters Fe should avoid, after that Fe becomes easy due to extra lives as you wrote).

[tasonir]

While Mo is a good start for Gh, imho Gh is one of the weaker races. Really bad int and dex, not much upside, basically a bad fighter (-1 to most apts, although they have +1 fighting) who can't really branch out into much else easily. At least if you take UC you get +1. Also they can't transmutate, which kills one of the perks of having taken UC.

I'd still say they're better than MuCK's of course, but they aren't good.

Is it really you tasonir? The species has +1 UC and you are not happy with it
Please take a look at GhMo cell (not Gh row), it has lower win rate than MuMo/DEMo/HEMo/OpMo/NaMo/DrMo etc.

It's really me, I know, I'm sorry I like the +1 UC, it's just not enough to make ghouls actually good. Ghouls are the 3rd/4th least won race, depending on if you include 2-3 game streaks. If you have to play a ghoul, you will probably think of monk as being the default background, because of claws and terrible int. So people fire up a lot of GhMo's compared to other backgrounds? There was only one GhEE, and there's 8 NaEE's. And ghouls are supposed to be good at earth magic (+1 apt). Nagas have a lot more choices, if you picked monk, it's probably because you wanted to play a monk, not because you picked ghoul and had to come up with some background they aren't horrible at.

I'm not entirely convinced of this argument myself, it seems a bit made up to fit the data, but it's about all I can think of. For the record I've only won one ghoul, I used necromancer as the background for regen/early pain attack, although I did use unarmed for my melee. Paired with Qaz for ranged damage/general power that doesn't require dex or int. Played that way, the character does eventually become rather strong, but early game before you can spam upheaval with decent invocations is rough. Later on haste, but I never really learned much high end necromancy. I'm probably not an expert on the race, although I'm not sure if we have any ghoul-focused players around here.

[Sandman25]

Well, as wrote in another thread I played 3 Ghs online (GhHu of Oka with bows+UC, GhWr of Ash with UC, GhAE of MAkleb with UC) and won all of them (15,5,15 runes). Also I won GhGl of Makhleb (staves) and GhIE of Dith (Ice + UC) offline, is it enough to be expert?
I don't see why low Gh's Dex matters, it usually relies on AC due to high Str, it is easy early game due to immunity to poison and high UC damage, it even has +10% HP like Mi or HO.
Yes, I remember how I created a thread asking how to play Gh when I was a new player because I was rotting too much but we are talking about players who streak 4+ games here (I don't even have such streaks on my Sandman25 account), they should not have problems with Gh. Seriously, I don't understand how it is possible to have 50% win rate with GhMo and 100% win rate with HaMo

[radinms]

Ghoul Monk of Chei is my favorite combo. Chei gives power to evade enemy's attacks to low dex ghoul.
And chei ghoul doesn't have a trouble to kill an enemy.
UC+1 Claw1 + Str/Dex+15 is very powerful, so there is no problem.Transmutation is not necessary because you have lich form!

recently my ghoul chei (15runes)
http://lazy-life.ddo.jp:8080/morgue/rad ... 044925.txt

[Magipi]

imho Gh is one of the weaker races.

This might be true, but imho Gh is still a lot better than Fe. So to interpret the data, some additional explanation is required.

[daggaz]

The problem is that it's much less clear. See http://i.imgur.com/tjCnnqG.png. It does make sense for the totals in the margins for race/class winrates to be percentages, but for the combo entries it's very hard to tell anything at a glance.

While it's less visually appealing, I still find it easier to read with the actual numbers. But then again, I can see the matrix.

[Bloax]

If you're already a very good player then streaking can be summed up as a competition in patience, much like going for highscores.
Except obviously streaking is about the patience for playing carefully instead of patience for splatting repeatedly with no end in sight.

So of course the race with high movement speed and multiple lives will be an excellent choice if you have infinite patience.
most of the time you don't, however

[tabstorm]

It's interesting that HO, Mf, Tr are so low and Ds, Og, Na are so high. Na is middling, I guess, but I would expect them to be near last because of their movement speed sometimes killing you. It could just be a sample size thing, though.

[Magipi]

Since this thread came out, I've been trying to figure out whether this method makes sense or not. Now I'm leaning towards "not".

Is there any such think as "streakability"? Is it really any different then winability (or how to call it)? I cannot imagine any reasoning for this. And even the particulars of the method are very shaky: taking the games of streaks and streak-enders, and then dividing the 2 numbers and calling it "winrate", double wow.

Of course, to point out that the methodology is strange is not enough to explain why the results are strange. And the result are really strange, I will even call them obviously false.

[Berder]

Is there any such think as "streakability"? Is it really any different then winability (or how to call it)? I cannot imagine any reasoning for this.

Streakability is the ability to win a combo, given that you're a strong player doing his best not to die no matter what.

And even the particulars of the method are very shaky: taking the games of streaks and streak-enders, and then dividing the 2 numbers and calling it "winrate", double wow.

You don't get it; it's simple.

First of all it is objectively a winrate. There's a body of games I have selected, and I'm seeing what fraction of them were won. So that's not at all mysterious.

Second, the reason I selected that particular body of games is that they are likely to represent players who are doing their best to win. If a game has the potential to extend a streak, it's likely that the player knows this and is trying his best to extend the streak with it, downplaying other considerations like score. The player is more likely to bother to use "streaking" tactics, and to play with more attention and less recklessness.

The games I selected are the ones that have the potential to extend a streak. Games in a streak other than the first do extend a streak, and streak-breakers could have extended a streak but failed to.

Of course, to point out that the methodology is strange is not enough to explain why the results are strange. And the result are really strange, I will even call them obviously false.

When the facts seem impossible, your preconceptions need adjusting.

Also, note that races and classes in my grid are often separated by only a few percentage points of winrate. If two races or classes are separated by only a 3% difference in winrate, there's not enough information to say which is better, just that they're about the same.

[Sandman25]

1) OpTm, MuCj - oops, MuCj is dead
2) OpTm, MuCj - oops, MuCj is dead again
3) OpTm, MuCj, FeBe - oops, FeBe is dead
4) OpTm - oops, OpTm is dead
5) OpTm, MuCj, FeBe, HaHu - oops, HaHu is dead
6) OpTm, MuCj, FeBe, HaHu, GhMo - oops, GhMo is dead

The chart will take into account only last 5 games and will show 0% win rate for GhMo (which is true in this example) and 100% wins for OpTm, MuCj, FeBe and HaHu (which is false).
Also it is known that when going for long streaks players start with hard combos and continue with easy combos which complicates it even more.

tl/dr It is not objectively win rate because it ignores many games.

[Berder]

1) OpTm, MuCj - oops, MuCj is dead
2) OpTm, MuCj - oops, MuCj is dead again
3) OpTm, MuCj, FeBe - oops, FeBe is dead
4) OpTm - oops, OpTm is dead
5) OpTm, MuCj, FeBe, HaHu - oops, HaHu is dead
6) OpTm, MuCj, FeBe, HaHu, GhMo - oops, GhMo is dead

The chart will take into account only last 5 games and will show 0% win rate for GhMo (which is true in this example) and 100% wins for OpTm, MuCj, FeBe and HaHu (which is false).

That isn't quite right. Like you said, GhMo was attempted once and lost, so it should have a 0% win rate, and it would. OpTm wouldn't be counted at all since it's the first in streak. FeBe and HaHu are streak breakers so they wouldn't have 100% win rate. (The streaks page includes all streaks by a player, not only their longest). You do have a point with MuCj, because it did not break a streak but did have a loss which wouldn't be counted. That would only be a real concern if certain races and classes were more likely to be specifically the 2nd game in the streak as opposed to 3rd 4th etc, but I doubt that is the case.

To correct for your point about MuCj I could exclude games that are 2nd in streak as well.

Also it is known that when going for long streaks players start with hard combos and continue with easy combos which complicates it even more.

It's not relevant where in the streak the game occurs (unless it's 2nd in streak, as above), only how often it was attempted in streak and how often it won. Also, myself, I'm already looking for a combo that I can confidently win, by my 2nd in streak.

[Sandman25]

Oh, I missed that you are looking at all streaks in updated version. Then it is not that bad indeed, not win rate of course, but probably close enough.

Personally I had a special strategy for streaks only when playing during last tournament. Otherwise I play what I wanted to play no matter if it is a streak or not.

[XuaXua]

Felids and Octopodes are the best streaking race since they don't wear any clothes.

[Berder]

Sandman, incorporating your point about MuCj, I excluded 2nd in streak as well. I also looked at only streaks >=4 for this one, because otherwise the large amount of berserkers breaking weak players' streaks would skew the results. This I think is the best chart so far.



edit: fixed bug

[tasonir]

Tm isn't last anymore! Tm isn't last anymore! Tm isn't last anymore!

To comment on some of magipi's concerns, this chart is simply winrate, pure and simple. What makes it unique is the way it addresses the two problems with winrate:

1) How to limit the data to games by players that are good at crawl
2) How to ensure that those players who are good at crawl are intentionally playing to win, not to speedrun, not playing casually/carelessly.

These are the main problems with win rate, and this attempts to solve them by using streaks. Seems to work fairly well with the updates, although having two or fives times as much data would be nice. Give it a few more years

[Sandman25]

I excluded 2nd in streak as well. I also looked at only streaks >=4 for this one

I don't understand how it is useful (my post about OpTm, MuCj etc. applies fully now, correct?)

Can you please remove Be? They add little to the table IMHO, a good player can win Be of anything easily so it has more to do with player skill and being unlucky with banishment than with power of species.

[Berder]

I don't understand how it is useful (my post about OpTm, MuCj etc. applies fully now, correct?)

Well, maybe it wasn't the point you intended to make, but my previous methods would indeed have failed to count the MuCj loss in your example (and only the MuCj), because one win followed by one loss wouldn't show up on the streaks page. The solution to not being able to count "2nd in streak losses" is not to count 2nd in streak wins, either.

Can you please remove Be? They add little to the table IMHO, a good player can win Be of anything easily so it has more to do with player skill and being unlucky with banishment than with power of species.

Heh, I think it's good that they are at the top, as it confirms expectations. It's a sanity check.

Here's the chart with the amendments from my previous post but including all streaks, not limiting to 4+. Actually don't use this one.
http://i.imgur.com/nJxiufq.png

[Sandman25]

Well, maybe it wasn't the point you intended to make, but my previous methods would indeed have failed to count the MuCj loss in your example (and only the MuCj), because one win followed by one loss wouldn't show up on the streaks page. The solution to not being able to count "2nd in streak losses" is not to count 2nd in streak wins, either.

I mean that if you analyze only streaks with length 4+, then you ignore all those FeBe and MuCj deaths in shorter streaks.

Heh, I think it's good that they are at the top, as it confirms expectations. It's a sanity check.

It's informational noise IMHO. Some species look much stronger than they are just because they have higher percentage of Be (just look at Fe, Te or Ko, for example).

[Berder]

It's informational noise IMHO. Some species look much stronger than they are just because they have higher percentage of Be (just look at Fe, Te or Ko, for example).

Well, if I simply removed Be, it wouldn't account for the fact that there are other classes that are easier than others, which also can skew the race results. (And ditto for races: some races are easier than others, which can skew the class results)

Instead, I've tried to compensate for it by assigning scores to races that take into account how easy or difficult the classes they played were (and vice versa). Let me explain the method I used to order the rows and columns of the following chart.

First, I scored races. To do this, first I scored how well that race did with each class, compared to the average for that class with other races. A race's score for a class is the winrate of that race+class combo, minus the winrate of all games played with that class that weren't done with that race. Then, the race's final score is the average of its score for each class, weighted by the number of games it played in each class.

Then I did a similar procedure (with race/class swapped) to score classes.

The rows and columns of the following chart are sorted by the scores above (not shown).

[tasonir]

edit: fixed bug

Was that bug "TM wasn't last anymore"?

Also, on the ghoul topics, I never responded: Yes sandman, 5 wins probably makes you a ghoul expert, or at the very least more of a ghoul expert than I am. And yes radinms, since one of the main limitations of ghouls is very low dex and int, picking chei basically removes those problems entirely, and lets you build a more much flexible character, with high dodging, better casting, etc. This is another reason I love Chei - all of the sudden everyone is a high dodging race with good spellpower My one ghoul was actually Qaz, so maybe I'll do a chei ghoul soon.

[Berder]

edit: fixed bug

Was that bug "TM wasn't last anymore"?

I was counting some games I shouldn't have, for column ordering purposes (the dots were still accurate).

[Rast]

I think your methodology has too much focus on games by the very best players.

I suggest looking only at games which would potentially be exactly the third in a streak. I think that's a good balance between games the player is definitely trying to streak vs games where the player is going full unfun FeSu mode to avoid death.

[celem]

I think your methodology has too much focus on games by the very best players..

About a page down from the top post he also posts the same graph for all streaks, 2+. I also felt it was a more interesting metric.
As usual another intriguing bit of number-crunching from Berder, though as you say, again initially skewed towards the upper-end of the playerbase.

I do get why you do this Berder, it provides a sample from optimal players and should therefore provide better insight into relative class-power. But the existence of tournaments, and the fact that some of you are masochists, with a fondness for playing obnoxious awkward combos because you can (cough greaterpode cough), leads these numbers to also be a bit skewed by the fact that the players you are looking at can make most things work. I would argue this promotes the impact of the RNG itself in what you end up looking at. When a streak breaks on D2 to adder poison, because none of the 4 pots you have found yet are curing after quaff-id...thats not really the combo or player's fault. Smart play, but not today. While generalised data sets might include massive amounts of beginner fail and so on, most of us understand this and will see it in the data and account for it ourselves, or you can mention it. The threads tend to be lively and well-debated, give us rawer data.

TLDR Good stuff, but continue to post both 'elite player' and more generalised data as you did here, it increases its validity.

The numbers are roughly where I expected them. I was interested to see no trys for DEWz and no wins for FoFi in the first table, seems people like DECj more though I always thought Wz had the best starter book. In the more general table suddenly FoFi is the most popular choice for bugs by some margin and DEWz is a strong choice for that race too.

This makes sense and validates my first point, those are my wins for those species, and DEWz is my highest scoring (only 15r) win, which appropriately enough began my first streak.