More comprehensive stats
Matchup stats are nice, but I was wondering if you have considered implementing different kinds of stats, such as card level, or turn level stats. For instance, you could analyze the number of times playing innervate + coin into yeti translates to a win, and against which classes/cards it is useful against.
You could even take this a step further! You should be able to determine which cards are in the opponent's hand at all stages of the game (by playing the game in reverse and tracking which card is where and when it is played). That way, you could know what the opponent has in hand for all of the turns in the game (unless if they are never able to play a card). This has to be useful in some way!
Being able to analyze the game from both positions automatically seems like a pretty powerful opportunity.
Marcus Goldman commented
by card level you mean like chillwind yeti: observations of damage taken and dealt before dying, both overflow and non-overflow? so you can assign some power or efficiency level to each minion or spell?
by turn level you mean like board state attack/health in play? so you can build a stochastic volatility model using board state swings and determine which turns are most influential, what cards were played to affect the swings. I guess you can treat each turn as time series lags, and total board state swings are returns.
If you integrate both above, a 7-win bot + machine learning algo for arena draft would finally be possible.