We spent this offseason building a rating system and a research series on top of it. This report asks the question underneath all of it: when TSSE says one team is 100 points better than another, is that claim true, and is it true everywhere? Every game in the database was predicted before it was played, which makes the whole history one long exam. Here are the grades.
An ELO rating is updated only by games already played, so the rating a team carries into any match is a genuine forecast: the engine has never seen that result. That gives us 82,043 graded predictions across eighteen seasons, and two honest ways to score them. Accuracy asks how often the higher-rated team won, among games with a winner. Brier score asks how good the stated probabilities were; it runs from 0 (clairvoyant) to 0.25 (a coin flip on every game), and lower is better.
Four out of five games go to the rated favorite. For high school soccer, with its four-year roster half-life and one-goal margins, that is a strong engine. But accuracy only tests the ordering. The harder question is whether the probabilities mean anything.
Take every game where the model made the favorite 70 to 75 percent. If the probabilities are honest, those favorites should collect about 72 percent of the available points. They collect 85 percent. And that is not a quirk of one bucket; at every confidence level, favorites outperform what the standard ELO formula claims:
Why does this happen? The classic ELO table converts a rating gap to a win probability with a 400-point scale borrowed from chess. Tennessee high school soccer is not chess. Skill gaps here are wider and more decisive: a team that is 100 points better does not win 64 percent of the time, as the chess table says; it wins about three in four. The engine's ordering was right all along. Its translation into percentages was too shy.
We fit the conversion scale directly to all 82,043 outcomes. The answer: Tennessee high school soccer runs on a 226-point scale, not 400. Re-read every prediction through that curve and the calibration snaps into place, within a point of perfect at every confidence level, and the Brier score improves from 0.145 to 0.136:
That gives the ratings an official read-out. This table is the public key for interpreting any two TSSE numbers:
| Rating gap | 25 | 50 | 100 | 150 | 200 | 300 | 400 |
|---|---|---|---|---|---|---|---|
| Favorite's chances | 56% | 62% | 73% | 82% | 88% | 96% | 98% |
The Head-to-Head tool now quotes this calibrated probability for any matchup you build. And to be clear about the curtain: the engine's internal constants stay private, but this read-out curve is deliberately public. It is not a lever inside the machine; it is the label on the gauge.
A single conversion curve is only defensible if the rating behaves the same way everywhere. So we re-graded every slice of the database separately. The test: within each slice, how far does the favorites' actual points share sit from what the raw formula expected? If the rating meant different things in different places, these gaps would scatter. They barely move:
Boys and girls, all three grand divisions, and all three eras sit within about a point and a half of each other. The same 226-point curve serves them all, which is the universality we were after: a rating gap buys the same probability in Memphis as in Mountain City, in 2009 as in 2026, in either pool.
Three slices genuinely bend, and they are worth knowing about:
The engine deliberately models no home advantage, partly because venue data is only about 92 percent reliable and missing entirely for some games. The backtest can now put a number on what that choice costs. Across nearly 40,000 games with a known home side, home teams beat their expectation by about five points of win share; away teams undershoot by a mirror image. Converted to the rating scale, home field in Tennessee high school soccer is worth roughly 30 points of ELO, or about 55-45 between perfectly even teams.
That is real, it is stable, and it is now on the engine's roadmap behind the curtain. Until it ships, know that a small favorite playing away is closer to a coin flip than its rating gap suggests.
Every engine change on this site promised to be judged on prediction, not vibes. The backtest settles it: the current engine's one-step-ahead forecasts beat the original v1 on every window we can score, including the most recent full season. (The numbers below reflect the engine after the state-cup weighting revision — which, notably, improved the backtest rather than trading prediction for trophy respect.)
The margins are modest, which is what recalibration should look like. But they are consistent, they are largest in the roster-registry era where the seasonal model actually has data to read, and they come with no cost anywhere we can find. Weighting the state tournament more heavily could have hurt prediction; instead the state-round segments above show why it did not — those games were being under-counted, and counting them properly made the whole ladder sharper.
The rating is valid. Its ordering is strong (79 percent of decisive games to the favorite), its probabilities are honest once read through the calibrated 226-point curve, and that single curve holds across both pools, all three grand divisions, and all eighteen seasons.
Its edges are mapped. Cross-division and out-of-state numbers deserve skepticism, in-season tournaments are genuinely wilder than the formula knows, and home advantage is a measured, unmodeled 30 points. Those are no longer unknowns; they are documented properties.
That is what "aim for universality" meant. Not that one number explains everything, but that when the number speaks, it means the same thing everywhere, and we know exactly where it whispers.