Hero Farm Dependency

Sorry for the deadspace there.  Went down for over a week with flu-like symptoms.  It’s also why my personal contacts have gone so long without responses, so if you messaged me in the past two weeks, rest assured I’m not ignoring you.

Anyway, here’s the chart for the day.


This is a recreation of a test I did several months ago with 6.74 data.  There are some adjustments I’m considering for the future, but for now it’s a straight replication because I wanted to see how it would behave across two different samples.

1. How the Test Works

The basic idea of the test is to try to determine which heroes see the biggest win rate increases through getting more farm.  To accomplish this I one-by-one take each hero’s games and sort them by the ending CS/min.  I create five groups representing 20% slices of the hero’s performances, and I find the win percentage of each slice.  I then express each slice by the difference between its win percentage and the hero’s base win percentage in the sample.  For example, Chen’s results for this test were

51.50% -21.42% -11.78% 3.56% 8.22% 21.65%

This means that his overall win rate was 51.5%, his bottom 20% win rate was 30.08% (51.5 – 21.42), and his top 20% win rate was 73.15% (51.5 + 21.65).

Every hero follows this kind of pattern where the better CS performances have a better win rate.  It’s very likely that the underlying causation works in both directions.  That is to say that having a higher CS can put you in a winning position just as much as being in a winning position makes it easier to get a higher CS.  However, the assumption I’m making is that the latter effect is much less hero dependent than the former, so that if we examine the strengths of the pattern on every hero the heroes with the highest discrepancies between the worst and best CS performances are likely the heroes who are most dependent upon farm in order to have a good game.  In other words, it’s a measurement of “carry-ness,” which should be considered quite distinct from semi-carries who tend to be more dependent upon exp advantages.

Why use CS/min instead of the easily available GPM?  My belief is that CS/min is a better pure measurement for gold dependency.  If you have a hero that regularly has games with low CS, their best GPM performances will be the ones where they got the most kill bounty.  But in these cases they would be winning because they terrorized their opponents into submission, and the high GPM was just an incidental residue from their early K/D/A prowess.  There’s also the consideration that the GPM of the winning team is inflated by the building gold they accumulate prior to the destruction of the ancient, and that the value of these building kills in GPM is very much dependent on the duration of the game.

It should be mentioned that CS/min isn’t perfect.  It could be that certain are better at sweeping up CS at the end of winning games which inflates their results.  That being said, two examples of heroes that might be well positioned to inflate their CS totals are Luna and Gyrocopter, but the two also end up on completely different ends of the CS dependency spectrum for traditional carries, so perhaps this effect isn’t as pervasive as some might fear.

Finally, I’ve decided in making this chart that I would use Very High data exclusively.  It’s my opinion that Very High play is the closest to the true potential of a given hero, and that the results from the lower bracket more often represent the idiosyncrasies of that level of play.  I did collect lower skill level data and will talk a bit about the interesting patterns.  I’ll also include links to the data at the bottom of the post.

2. Basic Analysis

The way I interpret the results, the top 40 spots are almost entirely what I would consider carries.  Carries are almost exclusively heroes with kits built around getting significant right-click scaling.  The only exception is Doombringer 19, whose continued presence in the top 20 admittedly confounds me.

At the very top end we have heroes like Anti-Mage at 1 and Faceless Void at 2 who are the traditional hard carries.  As we move down we move into heroes who are still carry threats but often ones with weaker scaling and a stronger early game.  Juggernaut at 23 is a classic example.  These heroes are still very much farm dependent, but they often have certain strong abilities that can turn a game even when their farm to that point has been less than impressive.  A team built around Juggernaut is certainly valid, but they will want to push for an earlier victory if up against an Anti-Mage team.

One interesting point is the presence of Huskar at 4.  I feel it’s a mistake to classify Huskar as a ganker.  This is something that works at lower levels of play where people position poorly and don’t understand the strength of his passive, but for Huskar to win consistently against higher caliber competition he really needs to farm in the early game to have the survivability to withstand enemy focus fire and the damage to be a threat even without his passive ramped up.

While I still classify the 30s as carries, the benefit from farm is much weaker here.  I feel that it’s often a mistake to build a 4 protect 1 team around these heroes unless you have a specific plan for early game dominance.  Chaos Knight at 37 is an example of a hero who could work as a 1 in a team built around fighting early and often, but for heroes like Gyrocopter at 35 and Naga Siren at 36, they often benefit heavily from being the secondary carry in a two carry lineup.

It should be noted that the top end is mostly agility heroes (hopefully I didn’t miscolor a hero, but I’m sure I will eventually).  In fact, there are only 6 agility heroes outside of the top 40: Razor(46), Bounty Hunter(49), Mirana(52), and surprisingly Venomancer(53) come in as being relatively farm dependent but not true carry threats.  Vengeful Spirit (71) and Nyx Assassin (90) are the only agility heroes I would classify as being not farm dependent at all.

By contrast, only two intelligence heroes qualify as outright carries: Storm Spirit at 24 and Outworld Devourer at 26.  Both heroes have fairly unique forms of right click scaling.  Storm Spirit can convert excess mana regen into Overload procs through short zips with his ult.  Outworld converts intelligence into damage directly through the mana scaling on his orb and through the damage multiplier on his ult.  Interestingly, neither qualifies as a hard carry, likely in a large part because of how much of their damage you can ignore through intelligent BKB use.  Also noteworthy is the absence of Silencer, who possesses an orb similar to Outworld’s.

40 through ~60 represents heroes that usually qualify as item dependent semi-carries, though the sorting is much more nebulous at this point.  It features a lot of popular intelligence semi-carries with strong right click potential such as Queen of Pain(42), Nature’s Prophet(44), Invoker(48) and Silencer(58).  It also includes some of the more item dependent int semi-carries like Necrolyte(45), Death Prophet(50), Leshrac(56), and Tinker(57).

One hero with interestingly low scores once again is Slardar at 59.  Valve lists him as a carry and I’m increasingly of the opinion that this is an egregious mislabeling.  He appears to only have a mild farm dependency, and really has more in common with Centaur Warrunner in role and Bounty Hunter via similar ults.  Slardar should be seen more as an initiator and damage amp than an actual carry.

There are many strong semicarries who fall into the bottom 40 or so spots, like Zeus at 72, Magnus at 81, and Nyx Assassin at 90.  One of the big things to take from this is to not think of yourself as a farm priority even if you happen to be solo mid.  The lower your hero is on this list, the more you need to prioritize ganking, and the less you should prioritize your own personal item build.  Certainly farm long enough to get the levels you need, make sure to build your essential items (Blink + Mana source on Magnus), and have a plan for making the most impact from the gold that you do acquire.  But still keep in mind that your ability to win the game hinges on your ability use and not your items.  If you’re playing Nyx Assassin and it’s 15 minutes in, consider buying a bunch of wards to take the load off of your supports.  Chances are, getting an earlier Mek on your Visage(64) will do more for your team’s chances of victory than upgrading your Dagon ASAP.

One last fact of note.  In every single skill bracket in both samples, the lowest farm dependency has always been Keeper of the Light.  I personally believe this to be driven by KotL being one of the easiest supports to wrack up a significant amount of CS using Illuminate.  What you should take from this is that it’s very likely that KotL can hinder his team by using Illuminate to take farm from higher priority teammates.

3.  The New Additions

8 new heroes have been added since my first test, and there were a couple e-mails wondering where they fell.  For the most part the results are unsurprising.

Medusa at 9 and Meepo at 11 definitely fall in the hard carry range.  Troll Warlord is a tad softer at 16, and can probably be considered comparable to post-6.74 Ursa who falls in at 18.

Slark comes in at 31.  I consider his farm profile pretty similar to Riki at 34.

Timbersaw at 41 puts him among the most gold dependent of the heroes with no significant right click scaling.  Unsurprising for a hero so dependent upon fixing his mana problems.

Centaur Warrunner comes in at 51.  This puts him in pretty similar company to other strength heroes like Spiritbreaker(47), Beastmaster(55), and Slardar(59).

Tusk at 63 I feel is most similar to Brewmaster one spot earlier at 62.  Other similar strength heroes are Pudge and Axe at 65 and 66 respectively.

Finally, we have Magnus at 81, which is surprisingly low.  My suspicion is that Magnus’ success hinges far more around proper ability use than it does his personal farm.  There’s probably a significant number of people who play him and can last hit well enough with an early Bottle and Shockwave but do not use Reverse Polarity and Impale well enough to convert that farm into victories.  Conversely, a good Magnus with just a Blink Dagger and some cheap mana is a huge initiation threat regardless of whether he gets any more CS in the game.

4. Changes from 6.74

The changes were actually relatively subdued, which I guess speaks well for the consistency of the system.  Ursa sees a small spike, which reflects perhaps a greater dependency on buying mobility now that Overpower has a significant cast time.  Alchemist actually sees one of the largest spikes.  Not terribly surprising given his significant improvements since 6.74.  Another example of recent buffs reflected in farm dependency is Clockwerk, though his improvements were much more modest than Alchemist due to Alchemist being a relatively hard carry.  Interestingly, the effect did not work in reverse for Morphling who maintained a similar farm profile despite the post-TI2 gutting.  Sven also saw a rather significant boost in ranking, moving up 27 spots.  Spiritbreaker saw one of the most significant declines in farm dependency, which actually drives home the point that being farm dependent is a double-edged sword.  Spiritbreakers recent patch changes made him less effective as a carry, but in exchange he received a significantly improved early game.  The net effect has been a huge boost to his public win rate because he’s much more capable now of winning games before farm even becomes a factor.

Overall though, it’s difficult to say much here with any degree of certainty.  The 6.74 sample is much smaller, so it’s also noisier.  I feel more confident in the fidelity of the 6.77b sample, but it’s not clearcut whether the differences between the two samples are due to performance shifts or just sample noise from the smaller sample.

5. Changes between skill levels

Some of the Normal results in particular are rather bizarre.  I don’t think they’re a valid representation of hero capabilities, but separately they are an interesting look into the behavior shifts that occur as you move through the brackets.

For starters, farm dependency is much weaker in Normal.  This shouldn’t be surprising because CS/min is significantly lower in the lowest bracket, but it’s worth noting.

On one end of the spectrum, we have Broodmother.  In Very High and High her farm dependency ranking is 27 and 39 respectively.  In Normal, it jumps to an astounding 91.  Despite this, her win rate does not actually fluctuate much between the brackets.   This exact pattern is repeated in the 6.74 sample, so it doesn’t appear to be a fluke.

Necrolyte is an equally interesting case, and one that’s much easier to explain.  Is Necrolyte an effective support?  The answer appears to depend on the skill level you’re playing at.  Necrolyte’s Very High farm dependency is ranked 45; in Normal, it is ranked 84.  This is despite the fact that Necrolyte’s overall win rate in Normal is the 2nd highest in the sample.  My belief is that Necrolyte has an extremely effective kit against lower level players.  On top of this, he’s a high priority target that needs to be focused down early.  In lower level games this doesn’t happen, so farming up tanky items like a Mek or a Shiva’s isn’t really all that necessary.  Stronger teams will put you through a much more intense survivability check in teamfights, and consequentially Necrolyte’s farm dependency shoots up in higher level games.  That being said, he’s still not a true carry, so when someone asks whether he’s a support or a carry, the answer is simply “no.”

Tinker follows a similar pattern to Necrolyte for what appears to be very different reasons.  My suspicion is that Tinker is only farm dependent if you know how to use Rearm effectively.  Lower level Tinker players simply do not, and as a result his Normal farm dependency is 97, second only to KotL in his eternal 98th spot.

Finally we have Enigma, who goes from 92 -> 54.  What happens here is, I think, somewhat similar to Tinker.  Enigma’s farm exists to enable Black Hole, and if you can’t hit your Black Holes, none of the farm in the world will do you any good.  At the same time, Enigma has one of the most superficially productive jungles, so Normal players are tempted to waste way too much time using Eidolons to farm instead of ganking or pushing.

Meanwhile, 3 interesting cases on the other end of the spectrum.

Bounty Hunter goes from 49 in Very High to 18 in Normal.  To some extent I feel that extremely high Normal farm dependency performances are to some extent a competency check.  Normal teams will not regularly shut down Bounty Hunter’s farm in lane, and the successful Bounty Hunter’s will be the ones capable of taking advantage of that, even if his actual carry scaling is relatively weak.

Templar Assassin is another striking example of this phenomena.  Her Very High farm dependency comes in at 30, but in High it shoots up to 13 and Normal to 7.  She doesn’t have especially high power growth, but she is definitely dependent upon momentum and nearly worthless if not given an early farm priority.  Her inflated farm dependency in High and Normal simply reflects an incredibly high failure rate in games where teams use her incorrectly.

This effect gets repeated in a lot of middling carries, but one particularly telling example is Bloodseeker.  Let’s first point out that Bloodseeker actually has a rather high farm dependency in Very High at 17.  He often gets compared (unfavorably) to Nightstalker(39), but I think that this is a mistake.  Despite their superficial similarities, Bloodseeker appears to be the much stronger carry of the two, which is both a blessing and a curse.

But while a ranking of 17 is relatively high, Bloodseeker’s rank in High and Normal is 3 and 2 respectively.  Again, what I feel is happening here is a competence check.  Bloodseeker’s early strength is his ability to heal himself off last hits.  Last-hitting is not an especially developed skill in these brackets, so being able to last hit at all becomes an important check to whether or not Bloodseeker will have a strong early game.  Interestingly, Necrolyte has a similar mechanic in Sadist, but his farm dependency skyrockets in the other direction.  I believe it’s safe to say that Bloodseeker is much more dependent on Blood Bath than Necrolyte is on Sadist.

6. Links

All the data for the 6.77b sample can be found here: https://docs.google.com/spreadsheet/ccc?key=0AoNi7mtSTYNzdHhYQW1rVVpNbFBUaFZFeTdzc2xnVWc#gid=0

It’s a bit outdated, but if you want to look at the 6.74 data you can go here: https://docs.google.com/spreadsheet/ccc?key=0AoNi7mtSTYNzdGJKelk5b1c5Sk54WEFtVU85YUZrYUE#gid=8


23 Responses to Hero Farm Dependency

  1. jimmydorry says:

    Nicely done. We should see a lot of data soon for you to play with.


    • phantasmal says:

      Possibly, but I don’t even have the hard drive space to hold the 75 gig version!

      I’m also a little concerned that the incompleteness of his skill data might be unavoidably not random. That’s not a knock against him, just an unfortunate side effect of the current design of the API.

      • Aardvarki says:

        At the very least, the Very High skill games are either all accounted for or when missing, missing in time-based chunks (due to API failures) and therefore more or less statistically random. The High skill games? Yeah, you got me on that one – the longest duration games are the ones I’m most likely to miss out on being able to label as High skill properly (I leave them null, along with the other 90+% of games I don’t retrieve skill data on). However, if you want to analyze only the top skill players (Very High), I’m very confident my script is working.

        I like seeing what you’ve been able to do with this! Keep up the good work!

      • phantasmal says:

        One thing I have to admit is that the more I do this the more I find High to be a less necessary bracket to look at. Occasionally there are interesting things there, but most of the time Very High vs not Very High tells most of the story. So perhaps it’s good enough, but finding the space for a dump that large is still a bit of a challenge!

      • jimmydorry says:

        A direct download is up btw if you find the space:

  2. brian says:

    Templar Assassin is another striking example of this phenomena. Her Very High farm dependency comes in at 30, but in High it shoots up to 13 and Very High to 7

    Did you mean Normal is at 7?

  3. xdv says:

    I’d say the KOTL’s farm doesn’t get as penalized as most heroes in losing games, as he can stand in the base and safely use illuminate to hold off enemy waves, which explains his relatively better CS in losing games and hence flatter slope. In fact turtling in the base with a KOTL might net him more CS than actively pushing enemy towers with his team.

  4. BishopsGuest says:

    Have you considered a general linear model rather than binning your data?

    A logistic regression might be a more informative analysis. Say Win ~ GPM + Length of game for each hero. I expect there is even enough data to do something like Win rate ~ Hero*(GPM +Game Length)

    • phantasmal says:

      It’s come up as a way to approach it, and it likely has potential, but setting it up is a bit beyond my comfort zone for now.

      Back in the 6.74 set, someone did approach me wanting to try something out with a straight CS/min regression. It was a bit of a struggle at the start, as the regression definitely didn’t behave like expected. My suspicion is that to get it to work you have to do some trickery to assume an exponential relationship (and this would probably also be true with GPM). He had started to introduce something along this line and the results were improving significantly, but he then disappeared before we could talk about it any further.

      The binning avoids this complication so it’s a nice quick and dirty test, but it definitely loses something in the simplification.

      • BishopsGuest says:

        I would be happy to take a look at it if you are willing to share your raw data. You should have my email from the comment.

        I am not surprised that the straight up regression did not work. This is a binary win/loss response and a straight regression is not the proper tool to use there. Since dota is a team game and we are just looking at individual heros there will be quite a bit of over dispersion in any model though.

        I am not really attached to GPM, just using it as a stand in variable. CS/min or just CS and min with an interaction term might be interesting as well. Working out total item value might be interesting too.

    • deviousalpha says:

      The only issue with using GPM is that it will scale with towers / kills / assists. If this is to be a pure look at farming then creep kills is the right stat to be looking atm.

      • BishopsGuest says:

        Yes, though it would be interesting to look at snowball potential as well. Rather than just farm dependance, item dependance.

  5. deviousalpha says:

    Doom does have right click scaling through eating an Alpha Wolf (which is the standard creep to choose). He gets the critical strike passive from the Alpha wolf 20% chance of 2x damage. And the packleader’s aura which is a flat 30% damage boost.

    He is also able to farm very quickly and get bonus gold through devour and midas if he goes for it. Explains his place as a carry and a top 20 or so.

    • concept says:

      There is that but there is also the fact that, in general, his ability to impact the game changes dramatically with farm. This includes not only right-click but also tankiness and other factors, such as prolonged item effect or usage. Doom is one those heroes that relies more on items to make an impact and items in its regular itemset normally scale very well with each other, enabling him to absorb large amount of damage, while he outputs a decent amount of damage through his item and skill set. This is when he truly shines: when he cannot be either ignored nor (easily) killed.
      Moreover, I think that the dependence of carryness or game impact on farm is a bit inflated for most heroes that are normally jungle-reliant, such as doom, lycan, chen and ench. Of these only ench seems to be in her expected position.

  6. Colum Wilson says:

    Remember that doom and alchemist both have abilities that grant them money. This means that they are missing an ability to impact team fights with (a stun/aura/other utility/whatever), so it makes sense that their ability to impact positively in games will be directly correlated with how much farm they get.

    On the opposite end of the spectrum, think of an alchemist or doom that DOESN’T get farm. One of their hero’s abilities is now entirely useless (greed/devour), so they have much less to contribute to the team via pushing/ganking/supporting.

    P.S. I forgot that doom has abilities via the creeps he devours, but their abilities are typically muted/weaker versions of other abilities.

  7. Death Bot says:

    Doom is a character whose gameplay revolves almost entirely around pushing his farm advantage onto the enemy team. The vast majority of good games I see with a doom are ones where, through decent farm ON TOP OF the huge chunk of bonus gold that he gets from devour, he is able to pick up a tier 3 or 4 item around the time most people are able to have something like upgraded boots and part of a tier 2 item.

    And with regards to Slardar, even when he was played as a carry it was never a particularly hard one. He tended to be played as more of an initiator->semicarry kind of character, who would slam down Treads and Blink in order to be landing the stun and get some decent damage afterward, followed by a Basher (back when bashes used to stack) or a Mask of Madness. I understand why he’s listed as a carry, as he does well with some extra farm, but in the public eye he’s seen as more of a beefy semi-carry or even an initiator, just due to the comparative strength of the other carries out there.

  8. Cherry says:

    The high normal win rate of Necrolyte is easily explained. He needs 0 skill to be useful. A good player can use him as a carry, and win in that way. while a weak player on a Necrolyte will benefit the team WAY more than a weak player on lets say Antimage. This of course is due to his aoe heal/nuke that requires you to hit 1 button, and an aura that wrecks no matter how late the game goes.

  9. Jeremy says:

    Hi –

    Very cool work here! I was looking at your google document and I was wondering how you arrive at the final number. The column is labeled “slope” but I’m not sure what exactly you’re regressing.

    This is pure curiosity so feel free to ignore me if it’s any trouble.

    • phantasmal says:

      It’s just a simple linear regression on the heroes’ win rates in each quintile.

      • Jeremy says:

        I see. I guess I mean what are you using for x and y in the regression? Basically I’m trying to figure what exactly the column is labeled “slope.” I guess it’s the regression coefficient? But Then I’m wondering what two variables you’re regressing.

      • phantasmal says:

        For now it’s just the difference between that hero’s win rate in that quintile and their average win rate for x and the number of the quintile for y.

        In the future I’d like to do something more elaborate that avoids quintiles entirely.

  10. Dnothing says:

    No update for Earth and Ember spirits?

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s

%d bloggers like this: