I have 82,043 rated matches going back to 2008, and I had never actually asked the simplest question about them: what does a Tennessee high school soccer result normally look like? The answer turns out to matter, because it is what the projections on this site are quietly assuming.
Both things are true at once, and that is the part worth sitting with. The single most common result is a one goal game. The average margin is far larger than one goal, because the tail is heavy: a big minority of matches are decided by five or more.
A match ends when a side goes nine goals clear in the second half. So there is no such thing as a recorded ten goal win, or rather, there almost is not: every game that was heading for 11-0 or 14-0 is written down as a nine goal win instead. The share of games at each margin falls smoothly all the way up the scale and then steps back up at exactly nine. That step is the rule itself, visible in the data.
It is not a rounding detail. It means the top of the distribution is censored, and any model that treats scorelines as unbounded is fitting a shape the sport does not actually produce.
Goals per game has drifted up modestly over eighteen years. The rate of mercy rule games has roughly doubled. Those two facts together say the field is spreading out rather than simply scoring more: the same number of goals is being distributed less evenly, with more matches that are over as contests well before they are over as games.
This is the table I most wanted, because it is the honest check on every projected scoreline this site publishes. Take every match, look at the ELO gap before kickoff, and ask what actually happened.
That is the finding that changed how the Upcoming widget works. I had assumed the sensible projection was the most likely single scoreline. It is not, because above a 200 point gap the most likely single scoreline is 9-0, and 9-0 is an artifact of the stopping rule rather than a forecast anybody should publish. The projections use the average instead, which tracks what really happens: at a 200 to 250 point gap the favourite genuinely averages a bit over five goals and the underdog a bit over half of one.
The other lesson is about confidence. No single scoreline in any of these buckets is worth more than roughly eight percent. A projected score is a centre of gravity, not a prediction, and it should be read the way you would read an average, not a result.
I expected games to tighten as the season got serious. They do not, and the reason is who you are made to play. District opponents are decided by geography, so a team faces whoever is nearby whether or not they are any kind of match. An in-season tournament is a field somebody chose.
That is larger than it sounds for a sport this low scoring. It is also why the venue on a fixture is not cosmetic, and why the 26-27 schedule work spent as long as it did making sure every match knows whether it is home, away or on neutral ground.
Goals conceded per game mostly rewards being good, since strong teams spend more of the match with the ball. The more interesting question is the second table: when a team does lose, how badly? Across every program the average loss concedes goals. A handful of programs are beaten by one or two and essentially never blown out.
The projected scorelines on this site read a rating gap and turn it into goals. They know nothing about how a particular team scores or concedes. That is a real omission, and it is measurable.
Put two elite defences together and the match produces well over a goal fewer than the rating gap alone would predict. This is the answer to a complaint I could not previously explain. Ravenwood and Franklin are projected at 1-5 in the coming season, yet their last eight meetings finished 2-1, 1-2, 1-2, 0-2, 2-1, 0-2, 2-2 and 1-2, never more than two goals apart. Both concede around a goal a game against a state median near two and a half. The projection cannot see that, so it prices them like any other pair 236 points apart.
The fix is not to lean harder on head to head, which contributes almost nothing here. It is to give the model what a scoreline model normally has and ours does not: a per-team attack and defence strength. That is the next thing I want to test, and it has to be checked against a holdout season rather than simply switched on.