So I just capped it at losing 50 points and set that as 0, since beyond this point the card can just be safely labeled as "bad" and dismissed. I considered throttling low-sample negative scores the same way, but linear scaling wasn't doing the trick and cards still ended up with -300 when they have 0 winrate out of 7 samples in Act 3 or something like that (and it would take forever to get more samples, since nobody picks the card in the first place).
Negative scores are not throttled, but hard-capped at 0 instead. For cards that are picked more often the system gains more confidence in its evaluation and this penalty decreases linearly until it disappears completely for cards in the 40th percentile of popularity and above. Positive scores are "throttled" for cards with low sample size, meaning cards that are rarely picked get only a small boost to their score, otherwise you'd have stuff like Flechettes unreasonably high because the card actually does well when picked in exactly the right deck, but 99% of the time nobody takes it since it's generally really bad. They are derived from winrates of each card compared to the average winrate of all cards, broken down by act - you can see the actual winrates and more info in the detailed tierlist table.Īt the basic level, it's just delta winrate*1000 (to make a nice round number) and then act scores are added up together, but there are some adjustments: If you have no attack cards in your hand, draw 2(3) cards.įilter: Common Uncommon Rare | Low Frequency Card