A Proposal for the definition of “snowball” hero in Dota 2 using Conditional Probability
Something’s that has always puzzled me about Dota 2 is that heroes with less than 50% win rate are consistently picked. While team composition is important i.e. you can’t just put the highest win % heroes on the same team and hope to win, in a vacuum you would still expect heroes that contribute to wins to have at least close to a 50% win rate. In particular, “snowball” heroes are constantly picked yet have some of the lowest win rates in the game; this article below talks more about these “snowball” heroes.
https://dotametrics.wordpress.com/2014/05/07/one-weird-trick-to-raise-mmr-that-raijin-thunderkeg-doesnt-want-you-to-know-about/
One of the main takeaways of this article is that high skill is directly correlated to gold-per-minute (gpm), and that the inherent gold advantage just from being a better player allows a “snowball” hero to snowball more effectively. I think this idea then naturally lends itself to actually being able to define how “snowbally” a hero is in dota, something that was and still is very hand-wavy, using conditional probability as covered in class. What I mean by this is to construct a plot of y = P(Hero wins | x gpm lead at the end) . The importance of the graph itself is important since, as noted, some heroes naturally have higher average win rate than other heroes. Additionally, all heroes will likely do better with a lead. The key then is to look at how quickly win rate increases with increases in gpm lead, as it shows how effectively that hero can translate a lead into a win. In theory, to confirm the article, this steep derivative should be able to make the most “snowball” heroes have well above 50% winrates when given nearly a 100 gpm lead. Unfortunately, though, the most popular Dota 2 match database (www.Dotabuff.com) doesn’t give access to the data itself, so writing a script to query information from it would likely have to be in the form of a webcrawler, which would be tedious work to actual parse through all of the available data. However, if they did give access to the information, or a another, more open, database was created I think it would be interesting to construct these graphs to be able to more qualitatively see who the most “snowbally” heroes are, even among heroes not even considered “snowbally”.