Rankings · The method

How these rankings are computed

By Joe Spisak, Pennsylvania state champion and 4x NCAA qualifier at Virginia · Published September 3, 2026

A rating computed from every published bout we have been able to read, by the formula on this page, run over rows anyone can check. Nothing is voted on. If two people run the code on the same rows they get the same numbers, to the last decimal.

What these rankings are not

They are not a seeding tool and no coach should use them as one. They are not a projection of who wins in March. They are not complete: they see the bouts that were published in a form we could read, and every set prints how many that was. Where the corpus is thin the match count next to a line says so, and a wrestler under the minimum is not ranked at all.

They are also not a ranking of high school athletes. Every ranked name is a college athlete. Girls' high school rankings wait until this pipeline proves out, and then only state by state where the association publishes complete brackets.

What goes in

Published results only, from the body that ran the event or the school that wrestled in it: a school's own athletics site, the NCAA, NAIA, NJCAA and NWCA, The Guillotine's tournament results posts, and FloArena or TrackWrestling event pages where they are readable. Every row carries the URL it was read from and the date it was read.

Nothing is retyped from another ranking and no aggregator is a source. WrestleStat keeps a women's college section and it was deliberately not used: a ranking whose inputs came out of somebody else's database cannot honestly claim to be computed from published results.

What was not readable, and why that matters

Measured on 2026-09-03 across all 187 programs in the college directory, FloArena returns "There are no results for this event" to a logged-out render and TrackWrestling answers with a 406. The NCAA publishes its championship brackets as a pre-tournament PDF, and a bracket PDF is a drawing: which two wrestlers fed a given match is a fact about where boxes sit on a page, not text anywhere in the file.

Reconstructing bouts from box positions was considered and refused. A wrong pairing published as a result is exactly the failure this site exists not to commit. A missing bout beats a wrong one, so the corpus is what the schools themselves wrote down, and the size of it is printed on every ranking page rather than buried here.

The rating

Every rated bout is one observation in a Bradley-Terry model. Each wrestler has a strength, and the probability that A beats B depends only on the difference between their strengths. The published strengths are the ones that make the season's actual results most likely. This is a maximum likelihood fit, not a running tally.

P(A beats B) = 1 / (1 + exp(-(strength of A - strength of B)))

Why not Elo

Elo updates after each match, so the number it lands on depends on the order the matches arrive in, and an early-season upset leaves a shadow that never quite clears. For an end-of-season ranking every bout is already known, so the fit here uses all of them at once and returns the same answer whichever order they are read in. A test asserts exactly that.

How bonus enters

A fall and a one-point decision are both one win, and a rating that treats them identically throws away the thing every wrestling room argues about. Bonus enters as evidence weight: how much a bout counts in the likelihood.

Major decision is in the table because the code carries it for a folkstyle event. It cannot occur in the sport being ranked: freestyle has no major decision. Over the 1,871 bouts in the first corpus, exactly zero were majors, which is the confirmation rather than a gap.

The weights are deliberately mild. A fall is worth one and a half decisions, not three, because bonus is a signal about the gap and not a second sport. An unopposed win carries no weight at all: a forfeit, an injury default and a disqualification are recorded because they happened and a reader looking up a wrestler should see them, but none of the three is evidence that one wrestler is better than another. They are dropped before the fit and they do not count toward the minimum match requirement.

This is a weighting of the likelihood, not a model of margin. It says a fall is stronger evidence. It does not claim to predict scores.

The virtual opponent

Every wrestler is credited with one win and one loss against a hypothetical average opponent whose strength is fixed at the origin. This does two jobs. It keeps an undefeated record finite, because a wrestler who is 3-0 has a likelihood maximised at infinity without it, and infinity is not a ranking. And it holds the scale still, so 1500 means the same thing in one weight class as in another and in one week as in the next.

The fit

The standard MM iteration for Bradley-Terry strengths, extended with the virtual opponent. Writing pi for strength on the multiplicative scale, W for a wrestler's weighted wins, and n for the weighted number of bouts between two wrestlers:

pi_i <- (W_i + 1) / ( sum_j n_ij / (pi_i + pi_j) + 2 / (pi_i + 1) )

Every strength updates from the previous iteration's values, so the answer does not depend on iteration order. The loop stops when no strength moves by more than 1e-10 on the log scale. The virtual opponent is never updated, which is what pins the scale, so the result is deliberately not normalised afterwards.

The number on the page

rating = 1500 + (400 / ln 10) x strength

A 400 point gap is ten-to-one odds. 1500 is the virtual average opponent. The displayed rating is a display of the model and nothing else: the win probability computed from two ratings is the same number the model gives, and a test asserts that too.

Who gets ranked

4 rated bouts. Below that, the fit is dominated by the one virtual win and virtual loss every wrestler starts with, so the number says more about the prior than about her season. A wrestler under the line is still in the corpus, still counts as an opponent, and still moves everybody else's rating. She is simply not listed with a rank next to her name.

Ratings are fitted one weight class at a time, within a division. A wrestler who moved up in February appears at both weights, rated on the bouts she wrestled at each, which is how wrestling ranks people.

The team rating

Individual ratings are the input. The team number is what those wrestlers would be expected to do in a dual, which is how college wrestling decides team questions.

1. Lineup. At each of the ten weights, a team's slot is its highest-rated ranked wrestler. A team with nobody at a weight forfeits it, which costs five points, exactly as it would in a real dual. Depth is part of the rating because depth is part of the sport. A program needs ranked wrestlers at 3 weights to appear in the table.
2. The bout. The win probability comes from the two ratings. What the win is worth comes from the bonus table below.
3. The dual. The exact distribution of the score difference is computed by convolving the ten weights. Exact, not simulated: a dual decided by random numbers would make the published ranking unreproducible.
4. The rating. Each team plays a round robin of these duals against every other ranked team, and its rating is its mean probability of winning one.

These are freestyle points, not folkstyle ones. College women's wrestling is contested in freestyle, where a fall scores five and a technical superiority four, not the six and five most wrestling fans carry in their heads. Scoring a women's dual on the men's table would inflate every bonus win by a point and change which team the simulation favours. The losing wrestler's single point for scoring in a decision or a technical superiority is modelled from one pooled number, published with the bonus table below.

Tie-breaks, in order: expected dual score margin against the same field, then the number of championship placement finishes in the corpus, then the program name. Actual team scores from a national tournament are deliberately not an input. They are a number somebody else computed, and mixing one into a rating would break the property that makes the whole thing checkable.

The bonus table

How often a win of each kind actually happened, as a function of how big a favourite the winner was. It is estimated from the same corpus the ratings are fitted to rather than assumed, and it is printed here in full because it is the one place where the team ranking could otherwise quietly encode a preference. A bucket holding fewer than 20 bouts uses the whole-corpus frequencies instead of its own, so a thin corner of the season cannot swing a team's expected score.

Nothing is ranked on a first read

Every bout is staged when it is ingested. A second pass fetches every source page again, runs the same reader, and reports any row that is no longer on the page it came from, any bout on the page that is not in the database, and any duplicate. Only then are rows listed, and only listed rows reach the fit. That is the same two-pass rule the athlete profiles run on, and it exists for the same reason: 23 of the first 30 profile drafts had a confirmed wrong fact.

Two schools publish the same bout from opposite corners, and inside one event that is free: the row is keyed on the event, the round, the weight and both names, so the second copy lands on the first row. Across two write-ups of the same tournament the compute step collapses bouts that share a weight class, both names, the round, the method and the score, and reports how many it collapsed.

Known limits

• Coverage. The corpus is the bouts schools published in a readable form, not every bout wrestled. A wrestler whose school posts nothing is underrepresented, and that is a property of the sport's publishing rather than a judgement about her.
• Identity. Wrestlers are counted by name within a weight class and a division. Two different wrestlers with the same name in the same bracket would be merged. None has been found, and one would show up as an impossible match count.
• Unattached and exhibition bouts are counted when a source publishes them as bouts, because they are still one wrestler beating another. The team simulation never reads a dual's team score.
• Context. The rating does not know who was injured, who cut late, or who was wrestling up. It knows results.
• Cross-division opponents. A bout against an NAIA or junior college wrestler at an open tournament stays in the corpus and she is rated as an opponent. The division on a set is the division of the ranked wrestlers, not a claim about every opponent.

How to check us

Every ranked line names the wrestler, her program, her rating, and the number of rated bouts behind it, and carries her record and her wins over other ranked opponents. If a result here is wrong, the fix is a row and not an argument. Mail joe@womenswrestlingclub.com with the URL of the page that published the result and it gets checked against that page.