Game Theory proves that UW Delver deserves a banhammer

On June 20th none of UW Delver cards have been banned. This was a horrible decision, but I'm not going to say one word about specifics of UW Delver in this post, since that has been discussed ad nauseam in other places. The entire post is about DCI's reasoning for lack of bans, and how it shows total lack of understanding of basic statistics and game theory. Let me quote the decision:

The DCI looked at the results of competitive Standard events. We found that while a high percentage of the participants played White-Blue Delver decks, that the win rate of those decks was very close to par. For instance, in a recent MTGO PTQ, the win rate of White-Blue Delver decks against non-Delver decks was a bit under 51%. In general there are decks that the Delver deck is strong against, and decks that it is weak against, but on average the deck tends to get results close to average. 

Additionally the number of people playing high level Standard events is the highest ever. Looking at the Magic 2013 card set, it appears that there may be more tools for other decks than for the White-Blue Delver deck, though time will tell if this bears out. The DCI will continue to observe how this plays out, but is taking no action.

I know nothing of tournament attendance, and as for new M13 cards I'll just quote this as a cautionary tale:

We tried to enable a few specific anti-Jace weapons in Mirrodin Besieged with Phyrexian Revoker, Hero of Oxid Ridge, and Thrun, the Last Troll, but the metagame solved those cards pretty quickly with Squadron Hawks and Swords.

New Phyrexia brought with it Despise and Hex Parasite, but those cards just aren't powerful or versatile enough.

Anyway, the biggest argument is "on average the deck tends to get results close to average". I'll show you in simple mathematical terms why this is not only wrong, it's completely backwards.

Simplifying assumptions

None of these assumptions are essential to the argument, they just make the math simpler. If you have time, you can redo the proof in a more complicated variant.

Let's assume that each player plays to win (more on that later), knows meta, players playing each archetype are equally good on average (this doesn't preclude some players being better, as long as they're either few of such superstars or they're not particularly attached to any archetypes, or both), and has no budget constraints (that's a reasonable assumption in Standard, where competitive deck prices are similar, less so in Vintage).

Let's assume there are N possibly tournament-viable archetypes in a format, and disregard issues like minor variants within each archetype, sideboarding strategies etc.

For each pair of such archetypen I and J, p(I,J) is probability that first player (playing I) wins the match. We don't need pre-sideboard and post-sideboard win probabilities, since if α and β are pre and post sideboard probabilities respectively, then:

p = αβ + α(1 - β)β + (1 - α)ββ

And by similar reasoning all kinds of on-play vs on-draw chances can be folded into a single number. There will be slight inaccuracy if some games are best-of-3 and others are best-of-5, if some games have unusually high chance of being drawn unintentionally due to timeouts (like control mirrors vs aggro mirrors) and if people decide to do intentional draw or not depending on what their opponent is playing, but we'll just fold as much of these as we can into a single "match win %" number, and won't be too concerned with that cannot be simulated this way.

How players decide what to play

There are many ways to simulate this, and they all lead to very similar outcomes. Since we assumed N archetypes, let's start with each of them having 1/N of the field. To make all these long lists of numbers look less dreadful I'll call the decks by randomly assigned 2 and 3 color combinations (but they're actually just labels).

Then people are paired randomly with another player, and each match loss makes a player change their deck with some low probability to a deck randomly chosen from current meta (so if 20% of people are playing archetype X, a player which wants to change their deck will pick X with 20% probability; if they already play X they have 20% chance of picking another decklist within the same archetype for simplicity).

Now this isn't particularly realistic - someone who lost 5/5 games is more than 5x more likely to change their deck than someone who lost just 1/5 on mana screw - but almost any such procedure for evolving meta will lead to very similar outcomes. Got it so far? Now let's create a fresh format with 20 archetypes, generate random matrix of match probabilities within 25%..75% range (except mirror match is always 50% by definition), set decklist-change-after-loss to 1% and simulate some rounds. What will be archetypes' meta shares and their average win percentages? Let's see one such random meta!

Random Fair Meta

It's time to run some simulations. At the beginning meta looks really balanced, no deck has particularly high or particularly low chance against the field (it is honestly just a coincidence that UW got on top):

• UW - 5.0 (55.4)
• Bant - 5.0 (52.7)
• WR - 5.0 (52.5)
• GB - 5.0 (52.5)
• WB - 5.0 (52.3)
• UR - 5.0 (51.9)
• RUG - 5.0 (51.1)
• Grixis - 5.0 (50.8)
• BUG - 5.0 (50.6)
• Kaalia - 5.0 (50.4)
• Jund - 5.0 (49.5)
• Junk - 5.0 (48.9)
• Naya - 5.0 (48.9)
• Esper - 5.0 (48.6)
• GU - 5.0 (48.2)
• GR - 5.0 (48.1)
• BR - 5.0 (47.4)
• UB - 5.0 (47.2)
• WUR - 5.0 (47.1)
• WG - 5.0 (46.0)

After 2000 rounds we start to see meta forming, but no deck is particularly bad:

• UW - 13.7 (54.1)
• Bant - 8.6 (52.5)
• WB - 7.2 (50.6)
• Grixis - 6.9 (52.7)
• GB - 6.6 (50.4)
• WR - 6.3 (49.5)
• UR - 5.5 (48.8)
• RUG - 4.9 (48.4)
• Esper - 4.5 (50.4)
• Kaalia - 4.3 (48.5)
• Jund - 4.1 (49.0)
• GU - 3.7 (49.3)
• BUG - 3.4 (45.5)
• Junk - 3.3 (47.2)
• Naya - 3.3 (46.9)
• UB - 3.1 (48.5)
• GR - 3.1 (47.0)
• BR - 3.0 (47.6)
• WUR - 2.8 (47.4)
• WG - 1.8 (43.7)

4000 rounds, UW is top deck, but Grixis has highest win percentage, so it goes up:

• UW - 21.1 (50.0)
• Grixis - 14.4 (54.0)
• Bant - 11.5 (50.0)
• GB - 7.0 (50.7)
• WB - 6.1 (48.5)
• Esper - 5.0 (49.9)
• WR - 4.4 (47.8)
• GU - 4.0 (51.4)
• Jund - 3.9 (50.7)
• Kaalia - 3.6 (50.5)
• UR - 3.2 (46.1)
• UB - 3.1 (51.7)
• RUG - 2.7 (45.7)
• Junk - 1.9 (47.8)
• BR - 1.9 (47.7)
• WUR - 1.7 (46.8)
• GR - 1.6 (46.9)
• Naya - 1.6 (46.5)
• BUG - 1.0 (42.7)
• WG - 0.4 (41.4)

10000 rounds, a lot of decks are seeing no play, Grixis was even briefly top deck, but then falls down in popularity, a lot of rearrangement in top 5:

• UW - 20.1 (52.2)
• WB - 14.5 (48.7)
• Bant - 12.0 (53.2)
• GB - 11.7 (48.4)
• Kaalia - 10.1 (47.9)
• Grixis - 7.5 (48.8)
• GU - 5.6 (50.1)
• Jund - 4.9 (49.8)
• UB - 3.3 (48.7)
• WR - 2.9 (46.8)
• Esper - 2.3 (52.0)
• Junk - 2.1 (49.6)
• BR - 1.4 (50.6)
• Naya - 0.5 (45.8)
• RUG - 0.5 (49.0)
• UR - 0.4 (46.6)
• GR - 0.1 (42.5)
• WUR - 0.1 (45.8)
• BUG - 0.0 (43.6)
• WG - 0.0 (42.1)

After 20000 rounds:

• Bant - 25.7 (50.7)
• UW - 17.4 (50.0)
• Kaalia - 10.7 (47.7)
• Jund - 9.8 (51.1)
• GB - 7.9 (48.5)
• WB - 7.7 (46.7)
• Grixis - 6.4 (52.6)
• UB - 5.0 (53.2)
• Esper - 2.4 (52.4)
• Junk - 2.1 (49.6)
• GU - 1.7 (48.6)
• BR - 1.7 (50.6)
• RUG - 0.9 (50.7)
• WR - 0.5 (47.1)
• UR - 0.1 (49.1)
• Naya - 0.0 (43.7)
• BUG - 0.0 (45.6)
• WUR - 0.0 (43.9)
• GR - 0.0 (41.9)
• WG - 0.0 (40.6)

50000 rounds:

• UW - 27.3 (51.2)
• Bant - 15.8 (52.2)
• Kaalia - 12.9 (47.1)
• Junk - 11.1 (48.4)
• GB - 9.2 (49.7)
• Grixis - 8.5 (51.9)
• WB - 6.3 (46.8)
• UB - 4.5 (48.9)
• RUG - 2.1 (50.0)
• Jund - 0.7 (49.3)
• BR - 0.6 (49.5)
• GU - 0.3 (49.3)
• UR - 0.3 (48.3)
• Esper - 0.2 (53.1)
• WR - 0.2 (46.2)
• Naya - 0.0 (45.2)
• WUR - 0.0 (47.8)
• BUG - 0.0 (43.2)
• GR - 0.0 (40.7)
• WG - 0.0 (38.3)

75000 rounds:

• Bant - 24.4 (49.9)
• UW - 18.5 (49.2)
• Grixis - 13.5 (52.5)
• Kaalia - 12.0 (47.7)
• WB - 10.9 (47.1)
• UB - 6.5 (54.7)
• GB - 4.5 (48.8)
• Esper - 3.6 (53.3)
• Junk - 2.7 (51.1)
• Jund - 1.3 (51.0)
• BR - 1.1 (52.3)
• RUG - 0.9 (48.0)
• GU - 0.1 (51.3)
• WR - 0.0 (46.2)
• UR - 0.0 (46.1)
• Naya - 0.0 (46.3)
• WUR - 0.0 (43.2)
• BUG - 0.0 (41.9)
• GR - 0.0 (41.9)
• WG - 0.0 (39.8)

100000 rounds:

• UW - 27.1 (48.0)
• Grixis - 21.4 (51.1)
• GB - 13.1 (51.4)
• Kaalia - 11.2 (52.2)
• UB - 7.8 (52.1)
• Bant - 6.9 (48.1)
• RUG - 3.9 (45.7)
• WB - 3.3 (51.8)
• Junk - 2.7 (49.4)
• BR - 1.3 (47.6)
• Esper - 0.7 (47.8)
• Jund - 0.7 (49.8)
• GU - 0.0 (51.1)
• WR - 0.0 (49.4)
• UR - 0.0 (45.1)
• Naya - 0.0 (47.9)
• WUR - 0.0 (47.6)
• BUG - 0.0 (41.6)
• GR - 0.0 (44.3)
• WG - 0.0 (37.6)

OK, let's stop there. This randomly generated meta was actually fairly healthy, with multiple decks being playable, and frequent rearrangement.

But did you notice something interesting about win chances (they're against whole field including mirrors)? At no point did any deck have a chance far higher than 52%, while worst decks were really low like <40% bad. Shouldn't average of win rates be exactly 50%?

Actually, no! Average win rate against random field (round 0) is 50% - but once meta establishes win chances are calculated against other good decks (so they're close to 50%), white bad deck changes are also calculated against good decks (so they're much worse than 50%). It doesn't matter how hard any deck crushes bad decks, since they're not in meta in any appreciable amounts, and so they're not counted.

Let me restate this:

Average win rates of decks in any meta are always lower than 50%, since better decks are played more than bad decks. It is difficult for any deck to have much higher win rate than 50%.

Random Unfair Metas

Now that was a reasonably healthy meta. What about metas which are totally unbalanced and unfair? Let's simulate one.

Now let's pick one deck - let's say UW (for no particular reason, it just happened to be the first in my list of labels, honest...) - for which win probabilities are taken from 0.45..0.75 range, while all other decks have to use 0.25..0.75 range. That's right - this UW doesn't have particularly high chances against any deck, it simply never has particularly low chances against any deck.

Let's give it a try. After 0 rounds:

• UW - 5.0 (all 61.7 / no mirror 62.3 / against UW 50.0 / except UW and mirror 62.3)
• Junk - 5.0 (all 55.9 / no mirror 56.2 / against UW 26.5 / except UW and mirror 57.9)
• BUG - 5.0 (all 53.7 / no mirror 53.9 / against UW 37.4 / except UW and mirror 54.8)
• Grixis - 5.0 (all 52.2 / no mirror 52.3 / against UW 47.3 / except UW and mirror 52.5)
• Naya - 5.0 (all 50.9 / no mirror 51.0 / against UW 30.9 / except UW and mirror 52.1)
• Kaalia - 5.0 (all 50.5 / no mirror 50.5 / against UW 41.0 / except UW and mirror 51.1)
• RUG - 5.0 (all 50.3 / no mirror 50.4 / against UW 46.5 / except UW and mirror 50.6)
• GB - 5.0 (all 50.2 / no mirror 50.2 / against UW 27.3 / except UW and mirror 51.5)
• GU - 5.0 (all 50.1 / no mirror 50.1 / against UW 50.8 / except UW and mirror 50.0)
• WUR - 5.0 (all 49.8 / no mirror 49.8 / against UW 38.4 / except UW and mirror 50.5)
• UB - 5.0 (all 49.7 / no mirror 49.7 / against UW 39.7 / except UW and mirror 50.3)
• BR - 5.0 (all 49.2 / no mirror 49.2 / against UW 31.6 / except UW and mirror 50.2)
• Jund - 5.0 (all 48.8 / no mirror 48.8 / against UW 34.7 / except UW and mirror 49.6)
• Bant - 5.0 (all 48.6 / no mirror 48.5 / against UW 35.5 / except UW and mirror 49.2)
• Esper - 5.0 (all 47.3 / no mirror 47.1 / against UW 51.3 / except UW and mirror 46.9)
• WB - 5.0 (all 47.1 / no mirror 47.0 / against UW 39.6 / except UW and mirror 47.4)
• WG - 5.0 (all 47.1 / no mirror 46.9 / against UW 25.2 / except UW and mirror 48.1)
• WR - 5.0 (all 47.0 / no mirror 46.8 / against UW 37.1 / except UW and mirror 47.4)
• UR - 5.0 (all 46.0 / no mirror 45.8 / against UW 29.1 / except UW and mirror 46.8)
• GR - 5.0 (all 43.9 / no mirror 43.6 / against UW 46.6 / except UW and mirror 43.4)

After 100000 rounds:

• UW - 68.7 (all 49.6 / no mirror 48.7 / against UW 50.0 / except UW and mirror 48.7)
• Esper - 30.8 (all 50.9 / no mirror 51.3 / against UW 51.3 / except UW and mirror 51.1)
• GU - 0.5 (all 50.2 / no mirror 50.2 / against UW 50.8 / except UW and mirror 48.9)
• Grixis - 0.0 (all 49.6 / no mirror 49.6 / against UW 47.3 / except UW and mirror 54.6)
• WUR - 0.0 (all 49.5 / no mirror 49.5 / against UW 38.4 / except UW and mirror 73.7)
• BUG - 0.0 (all 48.2 / no mirror 48.2 / against UW 37.4 / except UW and mirror 71.9)
• WR - 0.0 (all 46.5 / no mirror 46.5 / against UW 37.1 / except UW and mirror 67.3)
• Bant - 0.0 (all 43.9 / no mirror 43.9 / against UW 35.5 / except UW and mirror 62.3)
• RUG - 0.0 (all 43.0 / no mirror 43.0 / against UW 46.5 / except UW and mirror 35.2)
• BR - 0.0 (all 43.9 / no mirror 43.9 / against UW 31.6 / except UW and mirror 70.9)
• Naya - 0.0 (all 42.8 / no mirror 42.8 / against UW 30.9 / except UW and mirror 69.0)
• GR - 0.0 (all 42.5 / no mirror 42.5 / against UW 46.6 / except UW and mirror 33.4)
• Kaalia - 0.0 (all 40.5 / no mirror 40.5 / against UW 41.0 / except UW and mirror 39.5)
• Junk - 0.0 (all 40.3 / no mirror 40.3 / against UW 26.5 / except UW and mirror 70.5)
• WB - 0.0 (all 37.8 / no mirror 37.8 / against UW 39.6 / except UW and mirror 34.0)
• UB - 0.0 (all 38.4 / no mirror 38.4 / against UW 39.7 / except UW and mirror 35.5)
• Jund - 0.0 (all 35.3 / no mirror 35.3 / against UW 34.7 / except UW and mirror 36.6)
• GB - 0.0 (all 36.5 / no mirror 36.5 / against UW 27.3 / except UW and mirror 56.7)
• UR - 0.0 (all 33.1 / no mirror 33.1 / against UW 29.1 / except UW and mirror 41.8)
• WG - 0.0 (all 33.9 / no mirror 33.9 / against UW 25.2 / except UW and mirror 52.8)

This meta is about as degenerate as they ever get. It's 68.7% top deck, 30.8% deck designed to beat the top deck (and it barely does so), and 0.5% third deck. And do you see? Top deck's win percentage is below 50%!

Degenerate meta simply showed in high meta share, and total absence of almost all possible decks (in fair simulation 12 of 20 archetypes had some play, now it's only 3 of 20). What typically shows in simulations is that first all decks that lose to top deck get down to 0%, then top deck establishes ridiculous meta shares (often >90%), then sometimes decks which are good against it get significant share.

Here's second simulation. After 0 rounds:

• UW - 5.0 (all 58.9 / no mirror 59.4 / against UW 50.0 / except UW and mirror 59.4)
• GU - 5.0 (all 55.7 / no mirror 56.0 / against UW 36.8 / except UW and mirror 57.1)
• RUG - 5.0 (all 53.2 / no mirror 53.3 / against UW 43.3 / except UW and mirror 53.9)
• WB - 5.0 (all 53.0 / no mirror 53.2 / against UW 40.5 / except UW and mirror 53.9)
• GR - 5.0 (all 52.6 / no mirror 52.8 / against UW 27.7 / except UW and mirror 54.2)
• Bant - 5.0 (all 52.5 / no mirror 52.7 / against UW 49.0 / except UW and mirror 52.9)
• WG - 5.0 (all 52.5 / no mirror 52.6 / against UW 45.9 / except UW and mirror 53.0)
• Esper - 5.0 (all 51.1 / no mirror 51.1 / against UW 47.5 / except UW and mirror 51.3)
• WUR - 5.0 (all 50.3 / no mirror 50.3 / against UW 43.2 / except UW and mirror 50.7)
• UR - 5.0 (all 49.7 / no mirror 49.7 / against UW 40.7 / except UW and mirror 50.2)
• Grixis - 5.0 (all 49.2 / no mirror 49.2 / against UW 46.1 / except UW and mirror 49.4)
• Junk - 5.0 (all 48.8 / no mirror 48.7 / against UW 51.8 / except UW and mirror 48.6)
• UB - 5.0 (all 48.4 / no mirror 48.3 / against UW 40.3 / except UW and mirror 48.7)
• Naya - 5.0 (all 48.2 / no mirror 48.1 / against UW 28.8 / except UW and mirror 49.2)
• BUG - 5.0 (all 47.5 / no mirror 47.4 / against UW 33.0 / except UW and mirror 48.2)
• GB - 5.0 (all 47.1 / no mirror 47.0 / against UW 52.3 / except UW and mirror 46.7)
• Jund - 5.0 (all 47.1 / no mirror 46.9 / against UW 41.9 / except UW and mirror 47.2)
• WR - 5.0 (all 46.2 / no mirror 46.0 / against UW 33.7 / except UW and mirror 46.6)
• Kaalia - 5.0 (all 45.1 / no mirror 44.8 / against UW 41.1 / except UW and mirror 45.1)
• BR - 5.0 (all 42.8 / no mirror 42.4 / against UW 28.0 / except UW and mirror 43.2)

After 100000 rounds:

• UW - 91.8 (all 50.1 / no mirror 50.7 / against UW 50.0 / except UW and mirror 50.7)
• Esper - 5.1 (all 48.3 / no mirror 48.2 / against UW 47.5 / except UW and mirror 70.0)
• GB - 3.1 (all 51.1 / no mirror 51.1 / against UW 52.3 / except UW and mirror 29.8)
• Grixis - 0.0 (all 47.8 / no mirror 47.8 / against UW 46.1 / except UW and mirror 67.3)
• Junk - 0.0 (all 50.8 / no mirror 50.8 / against UW 51.8 / except UW and mirror 39.6)
• Bant - 0.0 (all 48.7 / no mirror 48.7 / against UW 49.0 / except UW and mirror 45.9)
• RUG - 0.0 (all 45.2 / no mirror 45.2 / against UW 43.3 / except UW and mirror 66.7)
• WG - 0.0 (all 45.7 / no mirror 45.7 / against UW 45.9 / except UW and mirror 43.6)
• WUR - 0.0 (all 44.4 / no mirror 44.4 / against UW 43.2 / except UW and mirror 57.2)
• UR - 0.0 (all 42.9 / no mirror 42.9 / against UW 40.7 / except UW and mirror 67.8)
• Kaalia - 0.0 (all 42.8 / no mirror 42.8 / against UW 41.1 / except UW and mirror 62.2)
• UB - 0.0 (all 40.9 / no mirror 40.9 / against UW 40.3 / except UW and mirror 47.7)
• WB - 0.0 (all 40.6 / no mirror 40.6 / against UW 40.5 / except UW and mirror 42.4)
• Jund - 0.0 (all 42.0 / no mirror 42.0 / against UW 41.9 / except UW and mirror 42.3)
• GU - 0.0 (all 39.0 / no mirror 39.0 / against UW 36.8 / except UW and mirror 63.2)
• WR - 0.0 (all 34.9 / no mirror 34.9 / against UW 33.7 / except UW and mirror 48.1)
• BUG - 0.0 (all 33.5 / no mirror 33.5 / against UW 33.0 / except UW and mirror 39.3)
• Naya - 0.0 (all 30.0 / no mirror 30.0 / against UW 28.8 / except UW and mirror 43.2)
• GR - 0.0 (all 29.2 / no mirror 29.2 / against UW 27.7 / except UW and mirror 45.5)
• BR - 0.0 (all 29.0 / no mirror 29.0 / against UW 28.0 / except UW and mirror 40.5)

This one is even worse - GB has decent matchup against UW, but since Esper is really brutal against GB (70% win rate), it makes GB totally nonviable, and UW has 91.8% meta share. So we have top deck (UW), meta deck (GB), and meta-meta deck (Esper).

Third unfair meta. After 0 rounds:

• BR - 5.0 (all 58.4 / no mirror 58.9 / against UW 50.7 / except UW and mirror 59.3)
• UW - 5.0 (all 56.8 / no mirror 57.2 / against UW 50.0 / except UW and mirror 57.2)
• Esper - 5.0 (all 53.3 / no mirror 53.4 / against UW 39.3 / except UW and mirror 54.2)
• WR - 5.0 (all 52.7 / no mirror 52.9 / against UW 54.5 / except UW and mirror 52.8)
• RUG - 5.0 (all 52.3 / no mirror 52.4 / against UW 53.3 / except UW and mirror 52.4)
• WB - 5.0 (all 52.2 / no mirror 52.3 / against UW 37.9 / except UW and mirror 53.1)
• GR - 5.0 (all 51.8 / no mirror 51.9 / against UW 51.4 / except UW and mirror 51.9)
• UB - 5.0 (all 50.6 / no mirror 50.7 / against UW 48.8 / except UW and mirror 50.8)
• Bant - 5.0 (all 50.5 / no mirror 50.5 / against UW 33.0 / except UW and mirror 51.5)
• Kaalia - 5.0 (all 50.0 / no mirror 50.0 / against UW 41.0 / except UW and mirror 50.5)
• GU - 5.0 (all 48.9 / no mirror 48.9 / against UW 46.2 / except UW and mirror 49.0)
• Junk - 5.0 (all 48.6 / no mirror 48.6 / against UW 29.7 / except UW and mirror 49.6)
• Jund - 5.0 (all 48.4 / no mirror 48.4 / against UW 29.9 / except UW and mirror 49.4)
• BUG - 5.0 (all 48.1 / no mirror 48.0 / against UW 46.7 / except UW and mirror 48.1)
• GB - 5.0 (all 47.6 / no mirror 47.5 / against UW 43.3 / except UW and mirror 47.7)
• Grixis - 5.0 (all 46.7 / no mirror 46.5 / against UW 41.8 / except UW and mirror 46.8)
• UR - 5.0 (all 46.2 / no mirror 46.0 / against UW 26.9 / except UW and mirror 47.1)
• WG - 5.0 (all 46.0 / no mirror 45.8 / against UW 53.6 / except UW and mirror 45.3)
• WUR - 5.0 (all 45.7 / no mirror 45.5 / against UW 40.5 / except UW and mirror 45.7)
• Naya - 5.0 (all 44.9 / no mirror 44.6 / against UW 44.8 / except UW and mirror 44.6)

Now BR is actually better than UW in vacuum! But since UW has no particularly awful matches, and BR presumably does, after 100000 rounds it degenerates to:

• UW - 42.9 (all 50.6 / no mirror 51.0 / against UW 50.0 / except UW and mirror 51.0)
• BR - 26.7 (all 50.6 / no mirror 50.9 / against UW 50.7 / except UW and mirror 51.1)
• RUG - 14.4 (all 51.3 / no mirror 51.6 / against UW 53.3 / except UW and mirror 49.8)
• Kaalia - 10.6 (all 46.0 / no mirror 45.5 / against UW 41.0 / except UW and mirror 49.7)
• GU - 3.2 (all 46.4 / no mirror 46.3 / against UW 46.2 / except UW and mirror 46.4)
• GB - 2.2 (all 47.0 / no mirror 47.0 / against UW 43.3 / except UW and mirror 49.8)
• Naya - 0.0 (all 48.2 / no mirror 48.2 / against UW 44.8 / except UW and mirror 50.7)
• UB - 0.0 (all 51.2 / no mirror 51.2 / against UW 48.8 / except UW and mirror 52.9)
• WR - 0.0 (all 49.8 / no mirror 49.8 / against UW 54.5 / except UW and mirror 46.2)
• Esper - 0.0 (all 44.3 / no mirror 44.3 / against UW 39.3 / except UW and mirror 48.1)
• WB - 0.0 (all 42.1 / no mirror 42.1 / against UW 37.9 / except UW and mirror 45.3)
• GR - 0.0 (all 44.8 / no mirror 44.8 / against UW 51.4 / except UW and mirror 39.9)
• WUR - 0.0 (all 43.3 / no mirror 43.3 / against UW 40.5 / except UW and mirror 45.4)
• Junk - 0.0 (all 40.5 / no mirror 40.5 / against UW 29.7 / except UW and mirror 48.6)
• Bant - 0.0 (all 36.5 / no mirror 36.5 / against UW 33.0 / except UW and mirror 39.2)
• WG - 0.0 (all 44.7 / no mirror 44.7 / against UW 53.6 / except UW and mirror 38.1)
• BUG - 0.0 (all 41.2 / no mirror 41.2 / against UW 46.7 / except UW and mirror 37.0)
• Grixis - 0.0 (all 39.6 / no mirror 39.6 / against UW 41.8 / except UW and mirror 38.0)
• UR - 0.0 (all 37.5 / no mirror 37.5 / against UW 26.9 / except UW and mirror 45.4)
• Jund - 0.0 (all 35.8 / no mirror 35.8 / against UW 29.9 / except UW and mirror 40.3)

Conclusions

I could keep rerunning these simulations, and there are some interesting and nonobvious points there, like having fewer bad matchup being more important than having any amazing matchups. The big thing which should be obvious in retrospect is:

In any stable meta top deck's win percentage will always be close to 50%. If it was much different than that people would switch to the top deck or away from it. Degeneracy can only be seen in meta share %. It never shows in win percentages.

So if win percentage doesn't show which deck is the best, what does it show?

High average win percentage of certain decks means meta didn't adapt to these decks yet. Flat win percentages of top decks near 50% mean meta is stable, not that it's balanced.

That's right - statistics show that current Standard meta is both degenerate (Delver's very high % of the field), and stale (top decks very close to 50% win rate) without much chance of evolving. I thought this should be obvious to anyone with a clue about statistics and game theory (to learn start here) - it's basic result from first chapter of any game theory textbook that results of all strategies will converge to the same average. But apparently from both DCI announcement and ensuing discussion it seems that very few people understand this and the awful argument gets repeated over and over again.

Now this post is concerned only with statistics. Maybe M13 will fix it. Maybe someone will come up with brilliant anti-Delver deck out of nowhere. Maybe people enjoy playing Delver mirrors so much they'll keep coming to tournaments regardless of how many Delver decks are played. It's all possible, and I don't really have any evidence for that. But arguments from win percentages are wrong, and mathematics is very clear about that. It's too late to fix it, but let's hope the next time DCI faces similarly degenerate meta they'll look at numbers that matter (meta share of top deck) not numbers that show something completely different (win rates of top deck).

If you want the simulation script, email me (it's fairly straightforward Ruby code).