How to split teams so the match is actually close

Why two captains picking alternately produces a blowout

Captains pick in the order 1, 2, 1, 2. That looks symmetric and is not, because the gap between the best player and the second best is usually larger than the gap between any other adjacent pair.

Team A takes the best player. Team B takes the second best, who is meaningfully worse. From there the picks are close to equal, so the original gap never closes. The first pick is worth more than every later pick put together.

There is a second problem, and it is the one people actually notice. Being picked last, in front of everybody, every week, is a small humiliation delivered publicly, and it is a common reason somebody quietly stops coming.

Team A picks 10, 8, 6, 4, 2    total 30
Team B picks  9, 7, 5, 3, 1    total 25

Team A is 20% stronger before anybody kicks a ball.

Method one: the snake draft

Sort everybody by strength and pick in the order A, B, B, A, A, B, B, A. The team that picks first in one round picks last in the next, so the first pick advantage is paid back immediately.

On the same ten players a straight draft gives 30 against 25. A snake gives 28 against 27. It costs nothing, it takes the same amount of time, and it is the single largest improvement available to a group that picks by hand.

It still needs somebody to rank the players, which is the part groups avoid, because saying out loud who is seventh best is unpleasant.

Team A picks 10, 7, 6, 3, 2    total 28
Team B picks  9, 8, 5, 4, 1    total 27

A one point gap instead of five, from the pick order alone.

Method two: let the results do the ranking

Nobody has to rank anybody if the table already did it. After a dozen matches, points per game is a defensible ordering, and it has the advantage of being a number rather than an opinion somebody has to defend in the car park.

It is not perfect. A player who has been on strong teams accumulates points that are partly not theirs. But it is unbiased in the way that matters: it does not know who is friends with the captain.

This is the point where recording every result stops being bookkeeping and starts paying for itself.

Method three: account for who plays well together

Two players of equal individual strength are not interchangeable. Some pairs work and some do not, and a group that has played together for two seasons knows which is which without being able to explain it.

The record does explain it. If two players have been on the same side twelve times and won nine, that pair is worth more than the sum of the two players, and splitting them makes both teams weaker in a way a skill ranking cannot see.

The same is true against opponents. Some players consistently do well against one particular other player, and putting the two of them on the same side wastes it.

Method four: what this app does

Sunday Super League gives every player a rating from 0 to 100 and builds the split from it. The rating is in view of the people who organise rather than hidden, because a suggestion nobody can interrogate is a suggestion nobody trusts.

How the candidate splits are generated

Every player has a rating from 0 to 100, and a team's strength is the average rating of the people playing on it. The app looks for the split where the teams come out as close in strength as possible and, inside each team, with a similar mix of strong and weak players.

With a reasonable number of players it tries every possible split and measures each one: with two teams of ten, for example, that is 92,378 splits and the app looks at all of them. With more players it can no longer look at every split, so it starts from many different splits and keeps swapping players between the teams for as long as a swap helps. One rule never bends: each team gets one goalkeeper, when there are goalkeepers. With fewer goalkeepers than teams, some teams go without; with more, the ones with the fewest matches in goal play in the field.

Among the splits within half a point of the best one, by the app's own balance score, it picks one at random. If you ask again, you get another split as fair when there is one; when the best is the only one, the app returns the same split and says so: "There is no other split as fair with these players." When a team has fewer players, the app takes 3.75 points off that team's strength for every player it is missing before it compares the two sides.

The output is a suggestion with the strength gap between the teams in plain view, not an instruction. The admin can regenerate it or ignore it. A machine that insists is worse than a captain who listens.

The uneven teams problem

Nine people turn up. Somebody plays 5v4, and the team of four is not four fifths as strong, it is worse than that, because they cover more ground each and tire sooner.

The app takes 3.75 points off the shorter team's strength for every player they are missing (half a level, on the 0 to 100 scale), so in a 5v4 the four only come out level with a mean 3.75 points higher. That pushes a stronger player onto the short side, which is what a good captain does by instinct.

If you are splitting by hand, the rule of thumb is the same: the short side gets the best available player, not an equal share of them.

What none of this fixes

Balance is a prediction, and predictions are wrong. A perfectly balanced split still ends 7 to 1 sometimes, because somebody had a bad night or a keeper had a great one.

Judge the method over a season, not over a match. If most games finish within two goals, the split is working. If one side wins by five most weeks, the ranking underneath it is wrong, whatever produced it.

Switch from a straight draft to a snake draft and most of the imbalance disappears. Everything after that is refinement.