Putting a team's blackout date on its bye: 28% fewer makeup games

Choosing which team takes which slot of the circle method, around the dates teams can't play, cut makeup matches 28% on 640 random leagues.

Every rec league organiser collects the same list before the season: the dates each team cannot play. A tournament weekend, a school trip, a wedding. The usual way to handle it, and the way the league schedule generator did it until today, is to build the round robin first and then move every match that lands on a blocked night to a makeup round at the end of the season. That works, but it makes the season longer and leaves the end of it as a jumble of rescheduled games.

The schedule has a freedom that this ignores. The circle method (the standard round-robin construction) fixes who meets whom in each round by position: position 1 meets position 6 in round 3, and so on. Which real team sits in which position is arbitrary. Swap two teams' positions and you still have a valid round robin. Each team's home/away pattern moves with it, so the balance is unchanged.

So the positions can be chosen to suit the blackout list:

  • In a league with an odd number of teams, someone sits out every week. If a team's blackout night is the night it would have had its bye anyway, nothing has to move.
  • In an even league, every team plays every week, so a blackout always costs a match, unless the team it was due to play is out that night too. Two teams away on the same school holiday can be drawn against each other that week, which leaves one makeup game, not two.

How it searches. The new step tries swapping two teams' positions (at least one of them a team with a blackout) and builds the dated season for each candidate with the tool's normal placement code. It keeps the swap when the season is no worse on three keys, in order: fewer matches needing a makeup, an earlier last date, fewer home/away breaks (a team at home, or away, twice in a row). It starts from the old schedule, so it never needs more makeup matches than before. Its budget is counted in evaluations, not seconds, so a shared link rebuilds the identical season on another computer.

What it bought, on 640 random leagues of 5 to 16 teams, single and double round robins, each team giving one or two dates it can't play:

beforeafter
makeup matches5,4313,911 (28% fewer)
mean season length18.04 weeks17.22 weeks
home/away breaks20,15614,732 (27% fewer)
invalid schedules00

The biggest gains are in odd leagues: 13 teams with one blocked date each went from 48 makeup games to 27. When several teams share the same few blocked nights, the gains can be dramatic: a 7-team double round robin with two shared holiday dates went from 30 makeups to 3, and five of eight such leagues needed none at all. Even leagues gain less, as the reasoning above predicts: 16 teams, one date each, 107 down to 93.

The fall in breaks was a side effect we did not set out to get. Moving rounds around blackouts had been quietly adding breaks; with breaks as the tiebreak, the search removes many of them. It is not a win everywhere. In a few cells breaks went up because the search took fewer makeups instead (10 teams, two dates each: 156 to 188), and a league that cares more about home and away than about season length would weigh that differently.

These are synthetic leagues. The next test is a real league's blackout list, and the same step on the multi-division page, which still moves blocked games after the build. The code is MIT licensed: fitPlanToBlackouts in ventures/league-schedule-generator/calc.mjs, and the measurement is eval/blackout-fit.mjs.

All writing · More in Tools and projects · More on League Schedule Generator · All tools and research