The Mystery 36% Winrate Player
On the worldwide Ability Draft leaderboard there’s one obvious outlier: they are the only player with sub 50% winrate. Very sub-50%. So sub-50% that people have questioned if it’s a bug, or a mistake, and enough to force me to do some investigation into this interesting case.
The mysterious player is called Migo. With a 82-143 record and the last game played nearly 3 weeks ago, they are at ~36.4%, 15.2% lower than the next lowest winrate on the leaderboard. They’re in the top 20 in the Americas leaderboard, and around 75th worldwide.
Warning: mathematics
The algorithm we use for ranking is based on a Trueskill-like model. Trueskill was developed by Microsoft and is a variant on Glicko rating (which is itself a variant on Elo rating). It’s a Bayesian rating system adapted for more complex game-play interactions (not just 1v1 which Elo/Glicko aim to solve). It allows partial play, N vs N and a variety of other extended circumstances — we care mostly about N vs N since Dota is a 5v5 game. We don’t care about margin of victory, just {5 players}, {5 players} => {outcome}.
Investigation begins
My first guess is that our enigmatic friend Migo was primarily being matched up against very high skilled opponents so I dug into that: looking at the average rating of his opponents (who have played ≥ the threshold number of games to have a rank) as well as what % of opponents were actually ranked. My initial sample was the top 100 from Americas.
Migo has, on average, the 8th most skilled opponents in the Top 100 of Americas. 72% of their opponents are ranked — which is not the most in our sample, but is certainly high. Now let’s have a look at Migo’s teammates.
Of the top 100 ranked players in the Americas, Migo has the 4th worst teammates, on average. 66% of their teammates are ranked (once again, this isn’t the highest but is certainly high enough to continue). Combining these two results and his own skill, we see the following Americas leaderboard picture.
Only 4 of the Top 100 players in Americas are, on average, playing against a more skilled enemy team than theirs. Migo is playing at the biggest disadvantage. Casting the net wider to include the top 100 in all regions we see Migo is *still* playing at the biggest disadvantage. This picture doesn’t really change much as you widen the search further and further down the top X for all regions.
Conclusion
So we see no evidence to suggest that the system is somehow super broken. Migo has a 36.44% winrate despite playing against, statistically, much higher rated opponents than his team is rated at. Over the large sample of matches provided, the rating Migo has is reasonable. He has played a large number (~70%) of their games with a low-rated teammate (‘sono’) — but this player performs statistically similarly in games with Migo and games without Migo — they’re not tanking games without him to boost his skill difference difficulty. A few other players have ~20 games with him, but once again nothing significant to suggest something suspicious is happening.
Migo is admittedly a big outlier but the most likely solution is that they just are plainly good. As visible above, most players have a winrate which is highly correlated with the rating skill advantage — and Migo is on that line of correlation, just on a very unexpected extreme. There are further research-related tasks I’m interested in doing for the rating model, but the ideas I want to experiment don’t really aim to ‘fix’ a problem that a low ranked player just so happens to sometimes beat consistently high ranked opponents.
