Radiant vs Dire Win Rates


Today we’re looking at the Radiant vs Dire win rate in relation to match duration.  It’s speculated that the Dire has an advantage the longer the game goes due to their positional advantage with regards to Roshan, and the stats appear to support this conclusion.  Overall Radiant has a small advantage over Dire (I estimate the overall win rate at 50.83%), but this advantage clearly dissipates the longer the game goes on.  What’s really interesting is that the Radiant advantage seems huge in games that last less than 30 minutes, particularly those in the Very High bracket.

Here’s another way of looking at what’s going on.


If we accept Roshan as the explanation for Dire’s late game success, what drives the early Radiant advantage?  Some explanations that have been thrown around include:

  1. Radiant side is easier to play due to the way the map is oriented and how we visually process information.
  2. Radiant has better pulling opportunities.
  3. Radiant has a safer jungle.
  4. Radiant has a safer midlane.
  5. Radiant’s off-lane tower is the less valuable than Dire’s off-lane tower.
  6. Valve knows about the late-game Dire advantage and gives them a slight handicap during matchmaking.

There’s no clear answer, but I will say that I find 1 and 2 of dubious value.  If 1 were the case I would expect the effect to shrink in Very High where players have simply played more games and have had more time to acclimate.  Instead, Very High is the bracket where the Radiant advantage is the strongest.  As for pull camps, if this were the sole driving factor I would expect the advantage in Normal to be negligible.  I’m almost 100% positive the ability to pull in Radiant’s middle like has had virtually no effect on Normal matchmaking.

This also raises the possibility that perhaps Radiant vs Dire have asymmetrically preferred line-ups.  Radiant could possibly be more effective with a lineup designed for early game dominance and pushing, whereas Dire would be more effective with a lineup designed for counter-push and stalling, while building around some kind of a late game carry.  This is still speculative so don’t yell at your teammates for not going along when you slam down insta-lock Dire Medusa, but it’s a possibility worth considering.

One last curious piece to the puzzle is  that this same effect existed in 6.74 and was actually stronger back then.


This is curious because the talk of 6.75 was speculation that the Aegis duration change would weaken Dire’s Roshan advantage.  I do not have 6.75 or 6.76 data to look at, but at least in 6.77 Dire appears to be stronger than it was before the change.


8 Responses to Radiant vs Dire Win Rates

  1. My guess is that Dire is much more susceptible to ganks. Ganking is most relevant in the early game. After 30 minutes there are no “ganks”, only pick-offs on out of position heroes, which are unlikely to be mitigated by any Radiant vs. Dire asymmetries. Some possible reasons why an asymmetry in ease of ganking may exist:

    1) There are more entrances into the Dire jungle, especially if Radiant’s mid T1 is still standing. Dire’s mid T1 does not prevent access to the jungle from the ramp immediately to the left, and at night, does not provide vision of that entrance.

    2) There are more effective ward spots inside the Dire jungle, namely the hill below the Dire top T2. The comparable ward on the Radiant side only provides vision of one neutral camp and three mostly unused paths. Warding the aforementioned ramp into the Dire jungle is also an effective warding spot, and less likely to be counter-warded than entrances into the Radiant jungle.

    3) Juke spots are not equivalent between the top row of trees on Dire side vs. the bottom row of trees on Radiant. Consider the trees below the Radiant bot T1 – there’s two entrances and a high density of juke paths in multiple directions. Above Dire T1, there’s only one exit, and few obvious quelling blade maneuvers to open up new paths.

    4) Relevant to #1, screen position can play a big role in determining which routes a player considers when attempting to escape. With the standard isometric view, the juke paths into the top row of Dire trees may be less visible and less likely to be directly inside the player’s field of vision.

    Fascinating data. Thanks for all of your hard work.

  2. mattieshoes says:

    I’m not sure I believe the location of Rosh has much impact in normal games either. In a lot of games, nobody even goes in the rosh pit.

    I really wonder if it’s something UI based… Would be nice if there were a bunch of games with the map flipped around the origin, or games where dire’s view is rotated 180 degrees so the suicide lane is “top” for both sides.

    • Brandon Rinebold says:

      It would take a long time for people to adjust. You’d just throw dire into a giant losing streak for a significant time before people got used to the flipped map.

  3. Dragon_clan says:

    It would be nice if there was a little more statistical data like what the standard errors were, and which of these were statistically significant

    • gwern says:

      https://dotametrics.wordpress.com/2012/10/24/hero-usage-and-win-rates-in-the-three-skill-brackets/ implies that the sample size here is 10k for each bracket.

      AFAIK these are binary variables (win/lose) so isn’t your normal standard error not really relevant? Now stop me if I’m getting something wrong, but…

      As a binomial, we can estimate the variance as n*p*(1-p), right? So to take the very first number, a 50.75% win rate: the sqrt of 10000*0.5075*(1-0.5075) is 50. So 2 stddev away from our mean of 10000*0.5075 or 5075 wins, then would be 5075+(2*50) or 5175 or 51.75%, and the other way obviously is 5075-(2*50) or 4975 or 49.75%.

      So if you want a confidence interval, I guess there you have it: 49.75-51.75%.

    • phantasmal says:

      I find it to be a bit of the pain in the ass to calculate and display (WordPress isn’t always cooperative when it comes to efficiently using screen space) the statistical data on everything. I personally follow the two rules of thumb that 1. larger samples will be more trustworthy, and 2. the least trustworthy of the smaller samples are those that disagree with more reliable sources. I do try to offer the sample size of everything so that people can make these judgments for themselves.

      For instance, the chart was just something I quickly threw together so that people could have a more visual representation of the trend. That big dip in High at 15-20 represents less than 1/3 of the High 15-30 sample which was 1763 games. Really small slice (likely under 500 games) that markedly disagrees with the overall trend? Probably unreliable.

      As for the overall trend itself, which is really the important part, look at it this way. To have a 95% confidence level and a 2.5 confidence interval at any really large population size, you’d need a sample of 1537. We have 3 brackets tested at 2 different times, and the smallest bracket/time combination is always 15-30. In 5 out of 6 instances that combo had at least ~1500 matches, and they all agreed that the Radiant win rate is at least 56%. Even if they all happened to be off by the maximum of our confidence interval and in the same direction then we’re still left with a +3% Radiant advantage in short game, and that’s an extremely unlikely worst case scenario when it comes to sample error. So given all that, I’d say that the evidence in favor of the overall trend is pretty strong.

      Edit: Rescued some other comments from the overzealous spam filter.

  4. Decoud says:

    Nice article. The difference in Very High skill games is really surprising. I did a post on the Roshan advantage for pro-games a few months ago (linked in my name). TLDR: Dire gets more Rosh kills than Radiant by about a 2-1 margin, Radiant still has a higher win % (53%), and both sides farm their jungle and neutrals about equally well (pretty much dead on for the first 15 minutes of game time). I could only speculate on why the Radiant wins more.

    If I can get laning data, it’d be a relatively straight-forward analysis to see if the laning discrepancy explains the win % difference. At some point I hope to look into this.

  5. ohne_caps says:

    I will assume that you have your stats from dota2.

    In 6.74 we had significant less heros that can stop pushes so the early game advantage of radiance was stronger and dire could not manage to het into later game stages.

    If you ever try to gank the easy lane it feels a lot easier as radiant imo + your offensive tri lane can pull.
    If you look for a good creep for enchantress/chen you will notice that it is a lot faster + safer in the dire wood so it helps radiance push.

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: