How much does a good draft actually matter in Dota 2?

November 20, 2026 4 answers
Asked by CRTmenu_4W

I've seen draft models predict short games better than long ones. Does that mean lane matchups matter more than overall hero win rates and team synergy?

Answers

coinopFlipper_3C

One draft model found enemy matchups particularly useful for prediction. The mid matchup and the position 1 versus position 3 matchup were among its highest weighted features. Seven of its ten highest weighted features described enemy matchups.

That suggests looking at who has to play against whom, not just whether each hero has a strong overall win rate. But feature weights describe that particular model. They don't tell us exactly how much a matchup causes a team to win, or prove the same rankings hold for every patch and skill level.

Reply to this answer
DOSbeepMap_7C

The game length result needs a little care. The model reportedly reached about 70% accuracy for games under 30 minutes, but was close to a coin flip beyond 45 minutes.

That doesn't mean you can predict a match at draft time with 70% accuracy. You don't know yet whether it will end that quickly, and grouping games by their eventual duration can select for particular kinds of outcomes.

raidTriage+Doorknob:

So a quick win might be associated with a favorable draft, but that result alone doesn't establish that the draft caused the quick win.

Reply to this answer
lootTable+Button

Don't throw synergy out altogether. In that model, the position 3 and 4 pairing and the position 2 and 5 pairing ranked just outside the top ten features. They contributed, just less than several enemy matchups did.

Same deal with hero win rates. The position 1 hero's win rate was a strong feature, while the other positions ranked lower. That's different from saying individual hero strength doesn't matter.

Reply to this answer
GBCsaveSlot+Thimble

Side advantage could muddy the short game result too. The research reported roughly a 53.5% Radiant win rate in public games since patch 7.38, and over 60% among games under 30 minutes.

If Radiant is winning a lot of the short games, some prediction accuracy may come from knowing the side rather than reading the draft. I'd want to see how the model performs with and without side information before drawing a stronger conclusion.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions