A round robin decides two things: who you play, and who your opponent played the week before. The second one is easy to miss. If the strongest team in the league is always followed by the same opponent, that opponent keeps meeting sides who have just been run ragged, rested their stars or driven two hours. Sports schedulers call this the carry-over effect. Russell gave it a number in 1980: for every pair of teams (A, C), count how many teams play A and then C in their next game, square the counts and add them up. Lower is fairer.
The league schedule generator has always written rounds by the circle method, the textbook way. It gives the fewest possible back-to-back home or away games. It is also close to the worst possible carry-over. With eight teams, five of one team's seven opponents arrive straight from the same side. The value is 148. The least any eight-team round robin can have is 48.
What changed
- Every schedule's report now shows its carry-over value, the floor for that many games, and the worst repeat. When one pair repeats three or more times, it says so.
- A new tick box, Spread who you meet next, searches for rounds with an even carry-over. For 12 teams in a browser it took 1.4 seconds: the value went from 768 to 182, and the worst repeat from 9 to 3. The cost was 16 home/away breaks instead of the minimum 10, and the report says so. If every team plays at the same venue, breaks don't matter and the even carry-over costs nothing.
The search starts from the circle method and keeps making changes that leave a valid round robin. It swaps the order of two rounds. It swaps two teams' opponents over the fewest rounds that still has every pair meeting once. And it exchanges two rounds' games along one cycle. That last move, the one the literature usually reaches for, can't escape the circle method on its own when the number of teams minus one is prime: every two rounds then form one long cycle. The team swap is what gets it out. A second pass trades a little carry-over for fewer breaks. Then a separate search sets home and away. On the plain circle timetable, that search finds the minimum n − 2 breaks for every size tested, 6 to 20 teams.
The checks are tests, not adjectives. All 720 ordered schedules for six teams are enumerated: the best has a cyclic value of 60, and so does the circle method, so for six teams there is nothing to gain. For eight teams the search reaches 56, which is Russell's lower bound, so it is optimal.
How real leagues compare
openfootball publishes the fixture lists of the big European leagues as open data (CC0). I scored 20 seasons, 2021-22 to 2024-25, ordering each team's games by matchday:
| Schedule | Teams | Carry-over | Breaks |
|---|---|---|---|
| Bundesliga, four seasons | 18 | 3,474 to 4,342 | 48 |
| Ligue 1, 2023-24 and 2024-25 | 18 | 1,840 and 1,948 | 74 and 96 |
| Circle method, mirrored | 18 | 14,610 | 48 |
| This tool | 18 | 1,926 | 74 |
| Premier League, La Liga, Serie A, Ligue 1 (14 seasons) | 20 | 2,278 to 2,564 | 70 to 146 |
| Circle method, mirrored | 20 | 21,044 | 54 |
| This tool | 20 | 2,136 | 114 |
None of the professional leagues use the plain circle method. The Bundesliga keeps the minimum number of breaks every season and pays for it with 1.4 to 2.4 times the others' carry-over. The tool lands among the professional schedules on both counts. That is an easier job than theirs, though: it knows nothing about derbies, police requests, European nights or television slots.
What it doesn't know yet
- Whether carry-over moves results at rec-league level. That needs a real league's results and some idea of team strength, and I don't have either.
- A double round robin is still the first half mirrored. Mirroring doubles every count, so the worst repeat over a 20-team season is 6. Optimising both halves together is the next step.
- How close the search gets to the true optimum above eight teams. For 12 teams it finds 156 against a floor of 120, and nobody has told me what the optimum is.
The guide how to balance carry-over in a round robin has the worked example. The code is carryover.mjs in the repository, and node eval/carryover-real.mjs reproduces the table.