Does the draft decide Dota 2 matches? What our models learned
“We lost the draft” is a favourite explanation after a defeat. To see how much the draft really decides, we compared two of our models: one that knows only the teams and players, and one that also sees all ten heroes. Both were trained on older matches and tested on 22,250 matches from the last 90 days that they had never seen.
The draft helps, but less than people think
Adding the draft makes the model’s favourite win more often:
- in professional leagues - from 66.6% to 68.8% of matches;
- in other leagues - from 58.2% to 59.6%.
That’s a real improvement, but a modest one: about two points in pro play, one and a half elsewhere. Most of what the model knows about a match comes from team and player strength, not from the heroes.
Usually the draft moves the odds by a few points
For a typical match, the draft changes the win probability by about 3 points (say, from 55% to 58%). It moves the odds by 5 points or more in a quarter of matches, and by 10 points or more only in about 3% of matches (5% in pro leagues).
Those rare lopsided drafts do mean something: when the draft moved the odds by 10+ points, the side it favoured won 77% of pro matches (a small sample of 60 games) and 62% in other leagues.
Sometimes the draft flips the favourite
In 9% of pro matches and 17% of matches in other leagues, the team that was the underdog before the picks becomes the favourite after them. When that happens, the new favourite wins 62% of pro matches - the draft was right more often than not - but only 54% in other leagues, barely better than a coin flip.
Heroes alone predict almost nothing
The most interesting result: a model that sees only which heroes were picked, without knowing who plays them, wins about as often as always betting on Radiant - 53%. Heroes by themselves barely matter.
What does matter is who plays what: how comfortable each player is on their hero, how the picks match up against each other and what is strong in the current patch. A model with just that information picks the winner 59% of the time - about as often as a model that knows only team and player strength (59%). Together they reach the 60–69% above.
So, does the draft decide?
Rarely on its own. The draft is a tiebreaker: it adds a few points to what is already known about the teams, and very occasionally it changes the picture completely. The rest is decided in the game itself.
On Swingmeter every live match shows both numbers - the estimate before the game and after the draft - so you can see for yourself how much the picks changed things. For the full test results, see the methodology page.