The Best and Worst Aghanim’s Upgrades

January 25, 2018

Came across this thread on the Dota 2 subreddit a few days ago.  The OP is now deleted, but it got me wondering how things had changed since this old spreadsheet from who knows what ancient patch.

For a lot of reasons item win rates kinda suck as a statistic, but one of the biggest is that expensive, late-game items will often have inflated win rates due to being luxury pickups that are primarily bought by the stronger team to close the game (and conversely, cheap early game items will have low win rates simply because they’re most likely to be sold/disassembled/converted).  If we could control for different factors like purchase time, item win rates could be at least a bit more useful.  Unfortunately, that information is not available.

So as a next-best, mediocre workaround, I tried plotting the win rate increase in Aghanim’s games with the purchase rate of Aghanim’s for that hero using Dotabuff’s stat listings.  The idea being here that heroes that pick up the item more often likely tend to, on aggregate, view Aghanim’s as a more core pickup, and therefore should tend to pick it up earlier in their item progression and see less of a luxury inflation to win rates with the item.

As an illustrative example, Invoker’s build Aghanim’s in just under 90% of their games on Dotabuff.  Invoker also receives the second smallest win rate boost from owning an Aghanim’s.  Regardless of how good you view his Aghanim’s upgrade, this makes sense because if Invoker built it in 100% of his games the win rate increase from the item would necessarily be 0%.  In general we would expect the item’s win rate to increase the less often it’s built, and this appears to be the case more or less when you plot out all the heroes.

So the basic idea we’re left with is that the heroes whose win rates beat this trend the most decisively will tend to include the ones with the strongest effects, and the heroes that fall below will tend to include the weakest.

I say “tend” because the title’s a bit of a lie, but you try being sufficiently nuanced under that kind of character limit.  This isn’t a diagnostic that gives us a definitive answer on which ugprades are good or bad, but it does provide us an idea of where to start looking if we decide to investigate further.

For example, Slark scores really highly.  He also has less Aghanim’s purchases in the last week than Tiny, a hero who no longer has an Aghanim’s upgrade.  On top of this, it’s possible that when people do pick up Aghanim’s on Slark, they do so with the intent of making it easier to kill the opposing Ancient and close a game that was already won.  Whatever the cause ends up being, it’s likely that the results from heroes with very small samples are not especially reliable.

Conversely, the Aghanim’s win boost on heroes like Visage and Invoker maybe don’t live up to common expectations.  This doesn’t mean that the expectations are necessarily wrong.  It could be that these upgrades are hard to use effectively and perform better in brackets 4k and above.  Or it’s an upgrade that’s situationally good, but is either built too often or too early by the playerbase at large, and similarly performs better in the higher brackets where people are more aware of why they’re building the item.  Skill level is one of the big things that the Dotabuff data doesn’t include, so keep that caveat in mind as you scroll past this text without reading so you can get to complaining in the comments that I’m telling people to rush Aghanim’s on Slark and never build it on Tinker.

Anyway, here’s the results:

Best Performers: (Aghanim’s Usage rate/increase in win rate (yes, I’m absolutely too lazy to go back and paste in column headers))

Slark 0.25% 22.77%
Naga Siren 5.45% 20.69%
Leshrac 12.19% 19.71%
Anti-Mage 20.57% 18.55%
Bounty Hunter 2.50% 20.38%
Earth Spirit 10.84% 19.34%
Axe 1.23% 19.31%
Razor 19.66% 17.29%
Riki 1.43% 18.70%
Skywrath Mage 25.42% 16.17%

Worst Performers:

Medusa 4.09% 3.32%
Doom 17.20% 5.17%
Luna 11.62% 6.23%
Phantom Lancer 9.33% 6.75%
Meepo 18.47% 6.05%
Sniper 1.47% 8.84%
Juggernaut 10.32% 8.28%
Nature’s Prophet 8.55% 9.15%
Mirana 34.05% 6.86%
Queen of Pain 25.93% 7.86%
Tinker 55.05% 5.39%
Bloodseeker 5.10% 10.71%
Nyx Assassin 26.94% 8.73%
Elder Titan 9.50% 10.71%
Outworld Devourer 1.13% 11.64%

Best Performers with at least 10% Use Rate:

Leshrac 12.19% 19.71%
Anti-Mage 20.57% 18.55%
Earth Spirit 10.84% 19.34%
Razor 19.66% 17.29%
Skywrath Mage 25.42% 16.17%
Rubick 21.75% 16.52%
Legion Commander 10.40% 17.62%
Lion 25.65% 16.00%
Shadow Shaman 37.34% 14.70%
Batrider 12.23% 17.04%

Best Performers with at least 25% Use Rate:

Skywrath Mage 25.42% 16.17%
Rubick 21.75% 16.52%
Lion 25.65% 16.00%
Shadow Shaman 37.34% 14.70%
Disruptor 25.64% 15.62%
Windranger 41.86% 12.62%
Clockwerk 44.59% 12.29%
Witch Doctor 49.87% 11.70%
Visage 35.27% 13.02%
Zeus 71.77% 9.05%

Here’s all the raw data

And here’s the graph that somehow manages to be both comically oversized and unreadable at the same time.


And for what little it’s worth, people might be underestimating Meteor Hammer on at least a situational basis, and Spirit Vessel’s win rate looks crazy good given the purchase price.  Would be completely unsurprised to see an eventual nerf on the latter.


TI5 Retrospective Part 1: The Meta of the Absent

September 24, 2015

TI5 Retrospective Part 2: CDEC, Masters of the Meta now available at

(Part 1 is also available here in what is likely a more mobile-friendly form)

When we talk about the hero meta at the conclusion of a major tournament like TI5, the discussion typically revolves around the heroes that dominated the Picks and Bans (Lina and Leshrac respectively), the heroes that put up absurdly high win rates over a large selection of games (for instance, Bounty Hunter’s 69% win rate over 48 games), or heroes that dominated all three of these metrics (Gyrocopter). But for TI5 a huge part of the story is the heroes that are missing, either as part of the 6.83->6.84 transition or just between the International and all of the other games of the 6.84 patch. This is because there’s a conspicuous absence that helps us to identify the connecting factor that led to the success of heroes like Lina, Leshrac, Bounty Hunter and Gyrocopter. And it also helps inform us on what we might expect from the inevitable 6.85.

With that said, I’d like to draw your attention to an unusual statistic: the ultimate cooldowns of the top 15 picks at TI5.


Without context, this doesn’t mean much, so let’s divide it into two groups: heroes with roughly minute long cooldowns or less, and heroes with nearly two minute cooldowns or more.

On the shorter end of things, we have every hero that has gone from having virtually no draft presence to being some of the hottest picks in 6.84.


The longest ultimate of this group is Undying at 75 seconds, but given that it’s a transformation ultimate with a 30 second duration, you’re looking at an effective downtime of 45 seconds. With that adjustment made, we have a group of composed entirely of heroes with sub 1 minute ults.

(I did, admittedly, leave off two heroes that fit my less than 5% P/B in 6.83 to greater than 30% P/B in 6.84 criteria but did not have as big of an influence on TI5 itself. These heroes are Dragon Knight and Visage. Dragon Knight is another transformation ult with an effective downtime of 55 seconds. Visage is the most complicated case of all given the nature of the familiars, but their expected downtime should be significantly less than their 180 second cooldown at level 6.)

“But what about heroes like Lina, Dazzle, and Storm Spirit?” you might ask. Well, let’s examine a second trend: the heroes whose TI5 selection rate outperformed their 6.84 expectations.


Once again you have a list completely dominated by short cooldown ultimates, albeit with the echoed exceptions from before of Naga Siren, Dark Seer, and Winter Wyvern.

Let’s flip the script now and look at the other end of things, starting with the three heroes I highlighted initially: Queen of Pain, Shadow Fiend, and Winter Wyvern. Of the top 15, they had the three lowest win rates, but you can’t really fault the drafters for expecting a better performance out of these heroes given their solid performances in the months prior to TI5.


But even if these heroes struggled, at least they got TI5 representation. The broader trend is that the 6.83->6.84 transition has coincided with a broad reduction in usage for nearly every hero whose gameplay largely depends on long cooldown ultimates.


If there’s a long ultimate cooldown hero that you’ve noticed missing from this list (not named Naga, Dark Seer, or Earthshaker), then it’s likely they were already a non-factor in 6.83.

Now, it’s certainly true that many of these heroes were already on the decline going into 6.83 or bore the brunt of a specific nerf in 6.84, but let’s examine a few specific cases of conspicuous TI5 absences.


Lion has had a very respectable 6.84 patch period. While he experienced a modest popularity decline in 6.84, from the second highest P/B in 6.83 down to eleventh in 6.84, his 6.84 win rate has been an acceptable .493, down less than a percent from his 6.83 performance.

On top of this, teams were highly valuing Lina’s ability to erase an opponent’s most valuable core through Laguna Blade. While there’s certainly a great degree of difference between the characters, it would be reasonable to expect Lion to see some play in situations where Lina wasn’t available due to their similar ultimates.

Despite all this, Lion’s went largely ignored by TI5 drafters, and when he did play he put up a dismal 7-12 (.368) record.


Due to his 6th highest P+B rate and very strong .567 win rate in 6.83, Juggs saw some nerfs in 6.84. While this definitely drove down his popularity in the current patch, he nonetheless maintained an even better .602 win rate going into TI5. Given the tournament’s emphasis on carries that lane well, come online quickly, and have some sort of escape, there was a decent chance that Juggernaut would see at least some niche selection. Instead, he went 2-4 (.333) and with only a single game outside of group stages.


With Thundergod’s Wrath at only a 90 second cooldown, he pushes the boundaries of my working definition for “long,” but he certainly does qualify as a hero largely defined by his ultimate. Magical burst has been very popular in 6.84, and so it might be reasonable to look at Zeus as a poor man’s replacement for the hugely successful and eternally banned Leshrac. Instead, Zeus’ win rate has cratered from .524 in 6.83 to .393 in 6.84. At TI5, Zeus’ record was an abysmal 3-8 (.273).


Ok, Warlock has been mostly a competitive non-presence in almost every Dota 2 patch. So why include him?

Well, MVP Phoenix’s March has the most recorded games with the hero at 28, a total almost triple the next closest player, and a lifetime record of 20-8 (.714). compLexity gaming also surprised a lot of teams by winning with Warlock twice in the regional TI5 Qualifiers.

Given that both teams had surprisingly successful performances, an unexpected pick like Warlock could theoretically helped them steal an extra series and push their playoff runs even further. However, neither team had confidence in Warlock working, and he saw zero picks and bans throughout the entire tournament.

I’ve thrown a bunch of stats out in an attempt to convince you that for whatever the reason these ultimate-based heroes have struggled or been outright ignored in 6.84 and that, with a handful of exceptions, this trend was amplified dramatically at TI5. The question then remains, what is it about 6.84 has caused this? If you can answer that question, then the inversion likely explains what it is that made Lina, Leshrac, Bounty Hunter and Gyrocopter the central players at TI5.

The first piece of the puzzle is the infamous rubber-band change introduced in 6.82. This change dramatically increased the gold and XP bounty per kill, and while this comeback mechanic has been toned down both shortly after release and in 6.84, it’s likely that there’s more of a networth shift currently at stake during teamfights than there was back in the 6.81 days. It’s a complicated subject and difficult to evaluate statistically, but even the watered down bounties of 6.84 put a greater emphasis on winning teamfights (or at least not losing them).

The second piece is that the 6.84 reduced the value of both lane creeps and many neutrals.

The third and final piece is that 6.84 boosted the “non-net worth portions” of hero kills by 10%, but also made changes so that a greater portion of the typical kill bounty goes to supports.

With all of these factors in play, it’s extremely likely that the ratio of kill bounty income to creep income was higher for this tournament than its been in any recent major. This first leads teams to emphasize heroes that fight well in the early and mid game, which then has a reinforcing feedback effect where teams need to draft lineups that can survive against expected aggression.

For an example in this shift of aggression, look at the approach to the hard lane by teams at the last three majors:

laneUsageTI5 was both the height in popularity for dual hard lanes as well as the most successful tournament for aggressive trilanes out of all Dota 2 majors. The contrast is particularly striking to DAC where both of these laning styles bombed in comparison to the more standard solo hard lane.

The implications of this laning shift were most pronounced for carries. Facing a triple whammy of reduced income rates, elevated safelane pressure, and increased emphasis on early and midgame teamfighting, hard carries died off almost completely, resulting in an approach to core investment similar in some ways to the one that dominated TI4.

One response, particularly in the group stages, was an increased reliance on semi-carries (heroes that trade raw right-click scaling for increased utility or burst damage), often paired up in dual or tri-core lineups. Lina, Queen of Pain, and Storm Spirit were the most popular choices for this, but Dragon Knight, Ember Spirit, and especially Templar Assassin were more niche examples that still saw a good deal of success. Leshrac would be included here had he not achieved what was essentially perma-ban status.

The other response was to build around a better scaling hero that could still somehow survive the pressures of TI5. By far the most popular and successful example of this is Gyrocopter. While Gyro’s scaling barely qualifies as a true carry, in a tournament of the blind, the one-eyed fighter pilot is king. More importantly, he possesses possibly the greatest deterrence to (as well as initiator of) laning aggression in the game in Rocket Barrage, and that ability combined with the low cooldown Call Down made him an early teamfight force in the way very few actual carries could hope to compare to.

Shadowfiend was the second most popular selection for this role, but we’ve already talked about his struggles. The next two most noteworthy selections were Phantom Lancer and Anti-Mage. Mobility in the form of Doppelganger and Blink allowed these heroes to survive early aggression, but of the two, Phantom Lancer was the far more stable pick. Anti-Mage was great at punishing passive teams or teams that bungled their aggression, but he took too long to come online against competent aggressors. Luna also deserves a footnote, but not much more than that and we’ll get to it later.

With all these factors in place, it becomes clear that the defining factor of the TI5 meta was early to midgame fighting. In order to win reliably you either needed to outright win these fights or to draw just long enough for your superior scaling hero to win out, and while many teams preferred that second, investment-driven strategy, the only safe centerpieces to run with it were Gyrocopter and Phantom Lancer.

And in this land of eternal war, ult-centric heroes had a tendency to be a liability. Their laning contribution tends to be weak, or at the very least, only strong in short bursts, and this is a greater than usual liability when aggressive duo and trilanes are at their historic peak. Past the laning phase, the teamfight potential of these heroes is strong, but only while their ultimate is available. When teams were as investment focused as they were at DAC this is not a problem, but maintaining a measured tempo against heroes like Storm Spirit, Tusk, and Undying, heroes that can blow everything in a fight and be ready to go again thirty seconds later…it’s not impossible, but it’s significantly riskier than it used to be, and I’m not surprised that the teams acted in a way consistent with the belief that the risk/reward payoff just wasn’t there for the majority of these heroes.

So what made Earthshaker, Dark Seer, and Naga Siren exceptions to this trend. I suggest three factors:

1. A long cooldown ultimate is less of a detriment if the ultimate is not the centerpiece of your kit.

You can imagine every hero having a ratio in teamfight value between their Ultimate and the rest of their abilities. A hero like Tidehunter, for example, is extremely skewed towards the ultimate side of things. Earthshaker and Dark Seer are likely more balanced with Fissure and Vacuum accounting for a greater proportion of their net utility.

2. A long cooldown ultimate is more prohibitive if it accounts for a larger portion of your net teamfight output. This means that not having a long CD utility ultimates or support ultimate is less of an expected loss than a damage ultimate or core ultimate.

Song of the Siren is a gamechanging ability, but a team with a support Naga Siren is still reasonably capable of winning fights with it on cooldown. Teams that include Juggernaut(Omnislash), Zeus (Thundergod’s Wrath), Chen (Hand of God), or Lion (Finger of Death) are less capable on average of winning fights without the full use of their ultimates.

3. Support Naga is simply back to being good

She was amazing in TI3, and fell out of favor when the post-TI3 changes made it more difficult to run greedy supports. 6.84’s increased emphasis on kill bounties allows you to run greedy supports again and farm them up through teamfights. This is a large part of why aggressive Duo and Trilanes worked as well as they did.

Additionally, Naga has always been very bulky for a support, and Ensnare has always been a great form of reliable CC. This allows her to contribute enough to fights even when Song isn’t available, and get enough experience out of these fights to carry her to the big cooldown reductions she receives at levels 11 and 16.

That covers the major points of the TI5 meta. In part two I’ll look at the impact of the meta on individual teams: how CDEC over-performed by finding their comfort zone, what EG did to take them out of it, and why Secret stumbled in a patch they seemed destined to dominate.

Editor: TheEmulator
Graphics: Ninjan, FO-nTTaX

Is Perfect World Responsible for the Pay-to-Win in New Bloom?

February 19, 2015

As you may have already heard, Dota 2’s recent New Bloom event has not been well received.  The main feature, a 5v5 brawl accentuated with the periodic appearance of semi-controllable ‘Year Beasts,’ has been plagued by a number of questionable design decisions.  The first several days were marred by widespread server issues due queues for the event being restricted to random 10 minute intervals.  This kind of queue restriction is great if you’re trying to boost player density, say to create tightly matched games in the sparsely populated >5000 MMR region.  It’s significantly less productive when you’re dealing with a record high concurrent player count of over one million, most of whom are trying to get inside your shiny new holiday event hat-creation device.

And all this is weird because, as many people have pointed out, none of this seems very like Valve.  I can’t find the video right now, but there was a talk about how Team Fortress 2 patches were structured in a way to periodically reinvigorate player interest.  But for the many casual Dota 2 players out there, New Bloom could easily be an interest killer.  You rush home from school or work, knowing that you’ve already missed a number of opportunities to build up wins for those event couriers.  You find that you can’t queue up for the event, and, assuming you’re outside of the one-hour warning window, you have no idea when the next queue will open.  You can’t practice for the event at all, which makes you nervous because you know random internet strangers can be less-than-cordial when hats are on the line.  Then, once you finally manage to get in the queue, the server crashes, meaning you can wait another ~2 hours to try again, or just do something more productive with your time.

New Bloom feels like a design-by-committee disaster, where requirements and restrictions were pushed with no concern for the viability of the final project.  The F-35 of Holiday Events.  And while there’s been a lot of speculation about Valve’s motives, whether to chalk this up to incompetence or capitalism, I have a theory that they might not have had that much control over the general direction of the event.  I don’t propose this in an attempt to absolve them, but because this alternative explanation is significantly more troubling.  So with the tone properly set, let’s talk about Pay-to-Win (P2W).


Originally I had planned to write a whole spiel about Pay-To-Win: how it’s never literally paying “to win,” how it’s really about paying for an advantage that accelerates a reward cycle, and how in a player-versus-player environment these advantages often need to be subtle in order to avoid frustrating the non-paying players into quitting.  But as people have accumulated larger amounts of ability points, it turns out that the Pay-To-Win in this mode is not subtle in the slightest.

And speaking of a lack of subtlety, take a look at the original page advertising the point system:

NewBloomP2WImage nabbed from’s New Bloom Article

The immediate impression of this points system is that the deck is stacked against a non-paying player.  There are only three ways to get points without paying, all three fall under varying degrees of unreliable, and the points provided by these methods feel utterly inconsequential compared to the 2400 options.  My point here isn’t that this is a bad P2W system because of it’s lack of subtlety; it’s that the system was designed in such a way that makes me believe subtlety was explicitly not a goal.

With that in mind, let’s ask a simple question: why does New Bloom even exist?  As you might remember from the Rekindling Soul Update:

One more thing: we on the Dota 2 team have a number of updates in the works right now that we’re really excited about, some for the rest of this year, and a big update for early next year. But we’re pretty sure we won’t be able to make enough progress on the larger update if we put it down to work on Diretide – so we’ve decided that we’re not going to ship a Diretide event this year. We know that last year we weren’t clear enough in our communication about this, so this year we wanted to be up front about it early. Next year will bring monumental changes to Dota 2, and we’re confident that when you’ve seen what we’ve been working on, you’ll agree it was worth it.

It’s a pretty popular theory that the “big update for early next year” is a conversion to Source 2 based on its inclusion in the Dota 2 Workshop Tools last year.  It’s reasonable that Valve wouldn’t want to put a lot of work into creating new versions of Diretide and Frostivus only to then have to start all over in the new engine.  But New Bloom is apparently important enough to ship regardless, which suggests that the most important holiday on Valve’s development calendar falls in February, and no, I’m not referring to Valentine’s Day.

And why shouldn’t it be?  China is an enormous part of the Dota scene, and Dota 2 is still trying to make inroads there.  The far more concerning possibility is that New Bloom is as much about pleasing Perfect World, the Dota 2 operator in China, as it is pleasing the actual Chinese playerbase.

If it’s true that Perfect World has a great deal of influence over the structure of New Bloom, then that lack of subtlety I mentioned before suddenly makes a lot more sense in a “feature, not bug” sort of way.  Consider this section from an IGN article on Free-to-Play game talks at GDC Europe:

By listening to someone familiar with the Chinese free-to-play browser game market, you’ll get an entirely different perspective. The much maligned ‘pay to win’ label in the West, where users can only really advance by handing over money, is pretty much a standard for distribution and financial gain in China. As described by Jared Psigoda of Reality Squared Games, something like the seemingly omnipresent energy mechanic in Western social games, which caps the amount of things someone can do for free in-game, is the most innocent thing in the world. They make Zynga look almost saintly when it comes to the transparency of their revenue-driven intentions.

“When I talk to Chinese game designers,” said Psigoda, “they say, ‘I just dug this new pit, this monetization pit, that somebody could spend 10,000 dollars or 50,000 dollars in and not reach the highest level.’ It’s all about the monetization of the game.”

Chinese publishers hope to earn back the development budget within two weeks of launch, so everything – armor, pets, equippable angelic wings — can be upgraded, and everything costs money. And users pay it, because they want to be first on the leaderboards. Psigoda provided statistics showing some gamers would spend over 100,000 dollars on a single game. In some games, botting is a built-in feature, though the service will only remain active if you pay a small fee while away from a computer.

(emphasis mine)

And as for Perfect World specifically, I have a little game.  Take the name of any of the games they’ve developed, append ‘p2w’ and search for it in your search engine of choice.  If you’re like me you’ll find an abundance of helpful forum post titles including:

“pay to win?”

“How do you feel about this game being pay to win?”

“This game is absolutely Pay To Win”

“Is PWI the most P2W Game?”

“Why is P2W bad?”

“Paying to Win, why this is actually a positive thing”

I think it’s safe to say that when it comes to monetization design Perfect World doesn’t do subtlety.  And why should they?  The gaming culture they design for does not demand it, and it is not in the interests of anyone making games there to introduce it and ruin the ride for everyone.  Given that environment, New Bloom is a perfect fit.  If anything the pit isn’t deep enough.

All this is predicated on the assumption that Perfect World has the clout to dictate to some extent the terms of this event.  I certainly cannot prove that this is the case, but consider the situation.  China is an undeniably huge part of the Dota community, but in order to get Dota 2 inside of China, Valve needs someone like Perfect World.  Maybe they could shop around for other operators, but this would at the very least be immensely time consuming, possibly a major social faux pas, and even if they somehow managed to find another operator, it’s unlikely that they would find one that would have a more, shall we say, “enlightened” view on P2W mechanics.  These factors give Perfect World a degree of leverage over Valve that could plausibly allow the scenario I’ve laid out in this post.

If you think it’s outrageous that Valve would agree to these concessions, consider it this way: without Perfect World, we would not have had the recent Dota 2 Asian Championships, and we might have even seen greatly reduced Chinese participation at the recent Internationals.  Salt about the TI4 finals notwithstanding, having to put up with a month of a terrible event isn’t an entirely unreasonable price to pay for a healthier international competitive scene.

So from this pragmatic sense, I might be ‘ok’ with New Bloom.  I still find it repulsive, but I can ignore it if that’s what it takes to keep Chinese teams active in the professional scene.  That being said, slippery slopes aren’t always a fallacy.  This year’s New Bloom establishes a troubling precedent, as previously all P2W experiments were done in PvE modes.  P2W in a competitive game is always 100% unacceptable, so hopefully this event is not a harbinger of things to come.

The Big Losers of 6.82’s Rubberbanding Were Not Supports

October 30, 2014

When 6.82 came out, it was a common refrain that this was the 5-carry patch,  as far as pubs were concerned.  And there was at least some supporting evidence of that position.  As I showed last post, 6.82 made pub matches significantly longer, and if you’re expecting long games why not exclusively draft heroes that are strong in the late game?

But while 6.82 might have been the least bad patch for a 5-carry line up in recent history, I didn’t buy into this line of thinking.  You still have to actually win a teamfight or at least create some trades for the kill bounty changes to take effect, and a 4th and 5th carry with no farm and a terrible laning phase is a lot less likely to contribute to those kills than a support.  To me it was more likely that we would see popularity shifts within the support role, but that supports as a whole wouldn’t see a huge hit to their overall success.

Unfortunately, it was difficult to tell much from win rate shifts between patches.  6.82 had a ton of other changes, making it impossible to attribute any particular hero’s success or struggle solely to the new kill bounty formulas.  Fortunately, we have a much more favorable testing environment in the recent 6.82c patch.

As you can see from the patch notes, 6.82c is a pretty simple patch.  We have longer buyback cooldowns and ethereal blade change, neither of which is likely to alter hero balance dramatically.  Twelve heroes received changes, mostly nerfs to the top pub win rate heroes like Omniknight and Spectre or to the top 6.82 professional bans like Brewmaster, Skywrath Mage, Death Prophet, and Terrorblade.  Finally we have yet another scaling back of 6.82’s rubberband mechanic.  So the logic here is pretty simple: these nerfed heroes are going to lose a chunk of win rate, so the biggest recipients of that win rate are likely the heroes that benefit the most from weaker rubberbanding.

To get these pre- and post-patch winrates (Very High only of course) I consulted DotaMax, but they unfortunately do not divide their winrates by mini-patches.  To get around this, I used a Last Week search to create a slice that was purely 6.82c games.  For something to compare that to, I also grabbed the 6.82 results as a whole at that point, and subtracted the 6.82c slice to create what is close enough to those desired pre- and post-patch results.  You can view the raw data here.

As expected, a small group of nerfed heroes represented a lion’s share of the negative momentum:


And who benefited from all this freed up win rate?


With the exceptions of Treant Protector, as well as Chen and Earth Spirit depending on how you define things, supports were pretty much absent from the upper echelon of beneficiaries.  My suspicion based off this is that the heroes most hurt by the kill bounty changes were carries and semi-carries who thrive on early game bullying or scaling and leveraging that into a win but lack the long-term scaling and utility to reliably win fair fights in a late-game scenario, as that’s a pretty fair descriptor for this entire list outside of the three exceptions I already mentioned.  And if you’re wondering about conspicuous absences, Viper was just outside the cutoff point for the graphic at +0.77%.

Another conspicuous absence is Razor, but as you may recall Razor was one of the twelve heroes receiving predominantly nerfs in 6.82c.  His shift was -1.05% which seems relatively minor, but what could be happening is that the win rate loss from his nerfs is being mitigated by 6.82c being a more favorable enviornment for similar reasons to the above heroes.

Incidentally, the same logic could apply to Earth Spirit.  The 6.82 environment as a whole was detrimental to the hero, but this may have disguised the apparent fact that his hero specific changes appear to have been big improvements.  As the rubberbanding is progressively reigned in his win rate is surging, at least in top end pub play.

6.82 Has Resulted in Closer, Longer Matches

October 8, 2014

With its kill bounty adjustments, 6.82 has been a very polarizing patch.  Some people love the way the new kill bounties change the way pub games play out, others feel that the extreme rubberbanding undermines the economic rules of the game, and both sides have cited favorable reddit posts as proof that reddit clearly doesn’t know what its talking about.

End-of-match results aren’t great for evaluating changes like this.  Kill bounties change the dynamics of a game, so you’d ideally want stats that measure the rate of change and not just the end results.  Nevertheless, I’ve been able to find evidence that 6.82 has been successful at creating closer games.  At the same time, there’s also some signs that the patch has had some less desirable side effects.

The Samples

Instead of one sample, I actually have four samples, each of approximately 20-30 thousand games and entirely in the Very High bracket. Each sample corresponds to a different patch period

The first is the initial release patch, and I’ve labeled it 6.82[1].  For the sake of brevity I won’t include the kill changes, but you can find them described here.

6.82[2] is the small patch roughly a day later on the 26th that “Slightly reduced AoE Gold bonus Net Worth Factor for 1 hero kills from 0.5 to 0.38.”

6.82[3] was yet another day later on the 27th.  It changed:

* Kill Streak Bounty from 100->800 to 60->480 (6.81 values are 125->1000)
* Reduced AoE Gold bonus Net Worth Factor for 1/2/3/4/5 hero kills from 0.5/0.35/0.25/0.2/0.15 to 0.26/0.22/0.18/0.14/0.10

And finally there is 6.82b on the 28th.  It also has a long list of relevant changes that you can find here.

I’m not going to go into the mechanical details of each patch, but I think a fair summary is that each subsequent patch is essentially a weakening of 6.82’s kill bonus comeback mechanics.

Closer Games

To measure how close a game was I used the very simple calculations of (Winning Team Average GPM – Losing Team Average GPM) and (Winning Team Average XPM – Losing Team Average XPM).  In the future a more elaborate test might be warranted, but this is good enough for now.  So how did the 6.82 patches compare to 6.81 on this metric?


As you can see, 6.82 corresponded with a huge drop in average GPM and XPM differential, and this stayed true even in the ‘weaker’ versions of the patch.  One interesting quirk though is that 6.82[1] consistently has small differentials despite having the strongest comeback mechanics.  This might indicate that it took players several hours to adjust to and start taking advantage of the new patch.

Game Duration

But as much as closer games is a generally positive development, it’s all for naught if you ruin other aspects of game quality in getting there.  As I said, end-of-game results isn’t a great way for evaluating this, and Valve likely has better approaches.  For example, it’s likely not a coincidence that this patch included the fight recap feature.  But one thing that these API results tell us is that 6.82 has made pub matches take significantly longer.


It varies from patch to patch, but the average 6.82 match is approximately 5 minutes longer than the previous patch period, an increase of over 10%.

Some will say this is an appropriate reaction to the TI4 finals, but this explanation misses the mark.  6.81 was only a fast patch in competitive play.  In public play, it was completely in line with previous patches which had been trending shorter for a long time.  Moreover, the competitive match times were being driven by push strats, which were already directly nerfed in multiple ways in 6.81.  It’s much more likely that the increased match duration is an unintended, though not surprising, consequence of the bounty changes.  It goes against what appears to have been a long-term goal to push Dota towards shorter pub games, but maybe that’s considered an acceptable casualty in the pursuit for closer games.

In any case, this increased duration is just the most obvious example of how wide-reaching the (possibly negative) effects of the bounty changes are.   It’s likely that when it comes to 6.82 reactions both sides were correct and whether you liked it or not just depended on which aspects of the game you were most focused on.  It’s not a surprise that the system saw multiple adjustments the very first weekend it was out, and will likely continue to see changes in future patches based on the feedback of how people react to the patch in the upcoming months.

Radiant vs Dire in 6.82

And in one final note, while it’s still too early to say anything definitive about Roshan balance, 6.82 so far hasn’t disrupted the Radiant/Dire balance very much.  The new Roshan position appears less advantageous for Dire, as evidenced by reduced Dire win rates in longer games, but this has been offset by the overall increase in long games.  This balance could easily shift as players further adapt to the new patch, and it may not even apply to competitive games at all.  Still, it’s interesting that so far the map and bounty changes appear to have offset one another.

The Dota 2 Workshop Tools and the Steam Box

August 8, 2014

There’s been a lot to discuss about the newly released Dota 2 Workshop Tools, ranging from speculation that it’s effectively the launch of Source 2 to its use in the summoning of eldritch horrors, but one thing I haven’t seen brought up is its potential strategic value it offers for the Valve’s Steam Box project.

First, a likely terrible recap on why the Steam Box and SteamOS even exist.  Valve as a company is heavily dependent on the success of Steam, and Steam is in-turn highly dependent on the PC market.  If PC sales suffer a permanent downturn or Microsoft makes Windows a less welcoming platform, Valve’s entire business model is at risk.  A successful Steam Box diversifies Steam’s install base, which leaves Steam less exposed.

But making a successful Steam Box isn’t a simple task, and one of the biggest complications is the Linux-based operating system.  Most games are not going to run natively at the start, and while the SteamOS will support game streaming, it’s hard to see this as anything more than a stopgap solution.  One of Steam’s greatest strengths is convenience, and for casual users the streaming solutions are unlikely to be seen as convenient.

So the SteamOS faces the console paradox: to get popular, the SteamOS needs games, but for game ports to be profitable, the SteamOS needs to be popular.  Traditionally, the most prominent answer to this paradox was the console exclusive.  Super Mario Brothers and the Legend of Zelda not only helped make the original NES a success, but their sequels have played the single largest role in selling new Nintendo consoles for over two decades, the defection of Squaresoft’s Final Fantasy 7 away from the Nintendo 64 to the original Playstation played a huge role in Sony’s takeover of the console market, and Xbox’s eventual challenge to the Playstation was backed by the success of Halo.

Over the years however, the power of the console exclusive has diminished.  For all their influence, games like Super Mario Brothers and Legend of Zelda had comparatively tiny staff sizes and development costs compared to modern AAA mammoths.  A company developing games in the NES day could afford to experiment on multiple titles simultaneously because costs were low enough that a handful of successes could make up for the flops.  In comparison, modern AAA development is sclerotic.  There’s too much at stake financially to take risks, so while you might produce safe top sellers, you’re unlikely to ever get the next Mario.  Even then, to get a top seller you need sales, so committing to an exclusive for an unproven platform is insanity compared to simply going multi-platform.

So to take things back to the Workshop Tools, I’d like to point out that Valve doesn’t really make games, and no, I’m not making a Half-Life 3 joke.  Look at the list yourself.  Half-Life and its sequels are original of course, but besides that you have Counter-Strike, Team Fortress, and Day of Defeat, all of which are like Dota 2 and based on mods.  Left 4 Dead is a bit different in that it was acquired from Turtle Rock Studios, but the difference isn’t a huge one.  The original Portal was based on a student game Narbacular Drop, and if you ever played Portal 2 with commentary on you likely have the words “our student game Tag: The Power of Paint” etched permanently in your brain.

Valve’s MO is finding underrated proofs-of-concept and polishing them up so they can compete with AAA offerings, and in light of that the Dota 2 Workshop Tools become all the more interesting.  If these Workshop Tools become popular, and there’s no reason right now to believe that they won’t, it will create a testbed for game concepts.  The best might get picked up by Valve or they might get developed independently.  Either way their start will be on the Source engine which should eventually make it trivial to run them natively in SteamOS.

But the beautiful thing about all this is it potentially takes us back to oldschool game design where creators are free to experiment in a low-investment environment.  From the modders perspective a lot of your immediate hurdles (engine, art, and especially network code) are addressed for you in the short term, and from Valve’s perspective you get results you can evaluate before committing a huge amount of resources.  It’s theoretically a win-win for everyone involved.

And if you think about it, Warcraft 3’s contribution to gaming is more about serving as the genesis for its descendents (Dota, League of Legends, and World of Warcraft, just to name a few) than it is about its own merits.  It could be the case that 10 years down the line Dota might be overshadowed by a game that’s essentially a mod of a mod.  And to their credit, Valve appears to be pretty well positioned to take advantage to that next step in evolution.

TI4 and 6.82 Part 2: Towers, the Pull Camp, and Blink Dagger

July 23, 2014

I talked about TI4 and why I expect it to foreshadow certain changes in 6.82 yesterday.  I prefer not to get too specific on stuff like this, in Dota there’s hundreds of possible ways to influence the game in a certain direction and I have no confidence that I’ll even come close to the one that goes live, but here’s at least some areas of interest that might come up in the patch.

The obvious change for slowing the game down is towers, but I’d expect any tower changes to be modest like a small boost in armor or bounty.  For one, there’s no desire to kill off push lineups entirely, only tilt the balance away from them a tad.  More importantly I think, the top TI4 strategies featured a broader variety of early aggression than just pushing, and  focusing on towers exclusively would leave those other schemes mostly intact.

What I have heard brought up are nerfs to Smoke of Deceit, which I find intriguing if perhaps a bit off the mark.  Support rotations were a big deal at TI4; fy and Fenrir come to mind, but there was also Liquid who upset a lot of teams largely on the production they were getting out of Bulba and waytosexy’s early roaming.  It might not be the case that Smoke of Deceit is too good right now so much as the opportunity cost of a gank attempt is excessively low.

I talk a lot about 6.79 because I believe it to be a game-changing patch, and I think it needs to come up again.  6.79 effectively nerfed support farm by changing the pull camp to a small camp.  It also reduced the XP bounties of many of the neutral spawns and made it possible for offlaners to steal neutral experience by just maintaining a presence in the area.  Supports can do a lot of things during the laning phase, but the two big ones are gank and neutral farm, and with neutral farming significantly weakened, heavy roaming supports won out pretty hard (6.79 also bumped up passive gold gain, further reducing the farm gap between the two styles of support).  And if ganking supports are decisively more productive that would then favor aggressive lineups that could best take advantage of those early ganks.  I recall TI3 having some crazy support item timings, from Alliance in particular, but I don’t remember anything comparable in TI4 that wasn’t largely a part of intense tower pushing.  I suppose there was LGD’s support Alchemist vs DK in their Saturday night matchup, but that’s a completely different case altogether.

I’m not making the case that farming ought to be favored over ganking, but it should be an actual choice and that maybe it’s not much of one right now.  So with that all being said, I could see 6.82 throwing a bone to more farm-intensive support options.  I don’t know that it will be as extreme as something like reversing the pull camp change, but it’s a possible, indirect way to approach 6.81’s over-reliance on aggression.

Finally, there’s Blink Dagger.  I don’t have anything against it, and I can think of other items that would be more annoying when bought in mass quantities (Shadow Blade, Necrobook, Hand of Midas), but 2014 was definitely the year of the Blink Dagger.

You might remember that 6.80 removed the mana cost from Blink Dagger.  In the preceding patch, Blink Dagger had 1667 purchases in 1341 games, for a rate of 1.2 Blink Daggers per game.  Jump ahead to TI4 and we have 625 Blink Daggers in 166 games, a rate of 3.8 per game or more than three times the purchase rate before its buff.  That’s a pretty crazy surge, mitigated some by the fact that it’s a pretty universally useful item, but if you were looking for factors that might have tilted the game towards early aggression you can’t really ignore it.

So yeah, I wouldn’t be surprised if Blink takes a bit of a hit.  If it does take one, I’d expect it to be smaller than the size of it’s buff, such as regaining a 15 mana cost down from the original 75.  Of course now that I’ve said it, Blink definitely won’t be nerfed in that way.  It might even evade attention entirely, but it still deserves some consideration in a discussion on what made TI4 as aggressive as it was.