Every professional MMA contest available, from every promotion, goes into one model. Three things come out of it and all three appear on the site under their own names. The forecaster rates each fighter the way a bettor would have on the night, from the record up to that fight and nothing after. The curve is the site's main object: a fighter's strength through the whole career, drawn with everything known today. The ratings are the numbers both of them produce, on one scale. This page says what the scale means, how the forecaster works, how the curve is built and where it fails, how a fight is scored, and who makes the lists. It also explains how many of the modeling decisions were a result of limitations in the data and would be different with more complete records.
A rating is a number on an Elo-like scale. A gap of 400 points means the model gives the higher fighter ten-to-one odds, a gap of 200 about three-to-one, a gap of 100 about 64 to 36. Every number shown is shifted by one constant that was picked so that the all-time top sits in the 2800s, similar to chess ratings. The shift has no effect on the gaps and win chances. The forecaster, which knows less on any given night, can run higher than the curve: its highest ever is Jon Jones after the Anthony Smith fight, 2921.
The familiar way to rate from results is Elo: one number per fighter, nudged up or down after each contest by a fixed amount scaled to the surprise. It works in chess because a player has hundreds of rated games and the number has time to settle. An MMA fighter has two or three fights a year and a career of ten or twenty. With that few, Elo never arrives. Set the nudge small and a genuinely strong prospect is under-rated for his whole career, still paying off his debut level when he retires. Set it large and every rating jumps with every result and says little more than the last fight did. There is no setting in between that works on fifteen fights. Elo also never revises. A win over an unknown debutant is scored against a nobody, and it stays scored that way when the debutant becomes a champion, because Elo only moves forward. In a sport this thin, the opponent's number on the night is itself a guess from a handful of fights, so the errors compound rather than average out.
So, underneath, both the forecaster and the curve are Whole-History Ratings, WHR (Rémi Coulom, 2008). Each fighter is a line through time, and all the lines are fitted together to every contest at once, so a fight is weighed against what the opponent's whole record says, not against the opponent's number on that night. A win over a strong opponent moves a line more than a win over a weak one, and a loss to someone who later became a champion is less damaging to your rating than one to someone that never made it. The forecaster refits every week on everything before that week and nothing after, so it stays a forecast while still using each of a fighter's few fights in full. WHR is more computationally intensive and trades simplicity for more predictive power. Even on sports like sumo, where the record is dense enough to know the answer, the whole-history fit beat Elo in every test we ran.
The forecaster is a walk-forward run. It sees only the contests up to each night and uses no information from the future: it is a true forecast. It has a handful of constants: how far a rating may drift between fights, how much a debutant is trusted, how fast ratings decline with age, and what each kind of result is worth. None was set by hand. The record was cut at 1 January 2022. The years 2012 to 2021 were the exam: the forecaster walked week by week through those years, each week rating fighters from every fight before it and forecasting the week's fights, and the constants were chosen to make those forecasts as good as possible over the ten years. What a knockout or a split decision is worth was not fixed even then; it is re-learned every week from the fights already seen, so a value is never used on a fight that came before the fights it was learned from. From 2022 onward nothing was tuned. The forecaster walked on with the constants from the exam, and its forecasts on those years, 2022 to today, are the test: they were made before the results, with nothing adjusted afterwards, and they are what the calibration page shows against the market. Before 2012 the forecaster runs the same way, but those years were never scored, so no ± is shown for them.
A fighter enters the forecaster at a level estimated from debutants like them, and the first few contests quickly overrule it. The forecaster does not know a fighter's age until the UFC publishes it and it does not use their amateur record, because it would be a leak of information since it's usually fighters who went on to make it that have a recorded date of birth and/or amateur record. As a result, a young fighter's rating takes a visible step at the UFC debut, when the age correction arrives all at once. That is a price the forecaster pays to stay blind. The curve uses all available information so it doesn't have that issue.
A rating on the night is a guess from a handful of fights. The curve is the opposite kind of statement. It is fitted to the whole career at once, so every fight is read in the light of every other, before and after it. The site is named for it.
The difficulty. A professional MMA career is short. Most fighters have a few fights; a long career is twenty or thirty, spread over a decade, against opponents who are themselves uncertain. That is not enough to learn the shape of a career from MMA alone. Studies of sports with dense records put the number of contests needed to pin down one career's shape at around forty, and few MMA fighters reach it. Our first curves were built from rules instead: how fast a prospect is allowed to rise, how a decline is detected, how far a run of wins over weak fighters may carry someone. Every rule fixed some famous careers and broke others, and tuning them on names is circular, since the names are the only check.
Why sumo. Sumo is the same kind of contest, one man against one man with a win or a loss, with a record dense enough to settle the question. A sumo wrestler fights about ninety bouts a year for fifteen years or more. On 4,385 wrestlers and 582,325 bouts since 1989 the true career line can be fitted with no shape rules at all. We then thinned those records to MMA density, two or three contests a year, and asked which method recovers the true line best. The answer was the plainest one. A smoother that lets a rating drift freely between contests recovered the thinned careers better than our rule-built curve, which placed the peak two to three years off on average against one year for the smoother. Sumo also settled two smaller questions. A surge does not snap back, so the curve has no pull toward a fighter's past level. And a method with no anchor drifts: the shapes come out right while the whole scale slides over the decades, so the curve has to be tied to something. Tennis records (1,182 long careers) were studied alongside sumo for what a career looks like, and the two sports agree: half the rise is done by about three years in, the peak comes seven to eight years in at about 25, and the decline after it runs at about 40 points a year. Those numbers are not inside the curve. They told us what to expect and how to test a method; the curve itself is fitted to MMA.
How it is built. The curve is a smoother with four parts and no rule tuned on any fighter.
Rises are gradual. Between two fights the rating may move, and a move is priced by its square: small moves are nearly free, a jump of 300 points in one gap is very expensive. Skill is built in the gym over months, and the fight only reveals it, so the curve credits a rise across the stretch in which it must have happened. A consequence is that the curve often has a fighter above the man he is about to beat before they meet. McGregor is above Aldo from 2014 and Volkanovski above Holloway from early 2019. We think that is right. He was already the better fighter; the fight just proved it.
Falls can happen in one night. A fall is priced by its size, not its square, so the fit either draws a fall in one gap or draws none. A fighter can get worse in twenty-five minutes, from damage or from being found out, and either way he does not perform at the old level again. The curve does not need to know which. Ferguson drops after Gaethje because he was never the same after the damage he took that night and Rockhold after Bisping because his defensive flaws were exposed at that moment, the result is the same. A man whose record never turned, Ngannou, St-Pierre, Khabib Nurmagomedov, has no fall drawn at all. Aldo keeps his level after the thirteen-second McGregor knockout, because he beat Edgar next. A fall is placed at a fight because that is where the evidence changes, and its size comes from what followed, so a damaging win (Lawler after Condit) is caught the same way, at worst one fight late.
Age. Everyone declines with age at the rate the forecaster estimated from the whole record, from a turning age in the late twenties, heavier fighters a little later. The curve uses each fighter's real date of birth from the first contest, for everyone who has one. Only the forecaster has to wait.
The ends. The start is held near the level the forecaster gives a debutant, adjusted for what the first contests show (a debut against top opposition says the fighter did not start from nothing), so the curve and the forecaster share one scale. The end is today's forecaster rating, so the two agree on where a fighter stands now. One exception: a fighter still on the rise, whose forecaster rating sits at its own high, may end below it. The forecaster credits youth and a UFC debut before the record does, and the curve should not inherit that.
The constants in those four parts, how dear a rise is and how dear a fall, are set by a planted test on real careers, not by eye, and checked against the forecaster across the whole record: the curve and the forecaster should disagree about a fighter's level only by what the later record justifies, and on average they do, to within about fifteen points in every class of career.
What the curve cannot do. A damage night and a run of bad luck look the same until the later fights separate them. On planted careers the curve draws, on average, about four tenths of a 300-point drop within two fights of the night and spreads the rest over the fights around it. A fighter whose last fight was the damaging one is drawn without the damage, since nothing after it exists. A long run of wins over weak opposition reads about a tenth of a fight too strong at the first step up, for the forecaster twice that; we measured it and left it, because it is the same for careers that rose little as for those that rose a lot, and because part of it is matchmaking, not rating.
Women. Women are rated apart from men, in their own model, and nothing on the site compares across the two. The women's curve has the same four parts with its own constants: the record is thinner, so a rating is allowed to move faster between fights and a fall is cheaper, as the women's forecaster already moves faster than the men's.
A fight is not scored as a plain win or loss. Each contest is split as a share between the two fighters, and the forecaster and the curve score it from different evidence.
The forecaster scores it from the result, the method, the round and, in the UFC, the judges' cards: the winner of a clear finish earns close to the whole contest, the winner of a split decision little more than half. What each kind of result is worth is estimated from the record itself and estimated walk-forward, so the forecaster never uses a value learned from fights it had not yet seen. We use judges' cards only on UFC fights because coverage is near 100%.
The curve uses the best record each fight has. A decision with ten or more fans scoring every round is scored on the fans' card alone. Any other decision with three judges' cards, in any promotion, is scored on the judges' card. A finish after the first round, where Sherdog's journalists called every earlier round, is scored as a card: those rounds as called, the finishing round 10–8 to the finisher and every round not fought 10–9. Every other fight keeps the forecaster's value for its kind of result. Each fight's page says which of these it was scored on. The round-by-round statistics shown there are not used by either.
Going into a fight, the two ratings give an expected share. A fighter's page compares the share earned with the share expected, and marks an underperformance where the earned share falls 0.2 or more below the forecaster's expectation. Where another underperformance follows within four years the page rings the first as a night the record may have turned on. The rings are a reader's aid; the curve places its own falls from the results.
Every professional contest in the sources listed on the data page, from every promotion. Amateur bouts stay in a fighter's record, labelled. The forecaster does not rate them. The curve counts them as contests, since a career starts where the fighting starts, but no rating is shown for an amateur night and the record cited anywhere on the site is the professional one. Exhibitions and bouts under other rules are left out.
A result overturned afterwards is scored as it happened in the cage, since the model rates what the fighters did. The one exception is an overturn for a referee's error, which counts as a no contest. No contests and bouts that never started do not count.
A division's list is the top 25 of its rated fighters who have fought within the last 18 months and have not announced retirement; a reigning champion stays on the list, however long the layoff, until the reign ends. The same rule makes the list on every past date. A fighter's division is their home division on the date of the list, from the weight they have been fighting at, so a list from 2012 is the division as it stood then; open weight counts as heavyweight, so the early tournaments belong to the heavyweights. Title reigns were compiled by hand and are shown beside the model's own order, which they need not agree with.
Pound for pound ranks fighters by how many rating points they stand above their own division's benchmark: the average of the division's 10th to 40th best on that date, over divisions with at least 150 active fighters. A heavier division is never measured against a lower benchmark than a lighter one; where its own sits lower, the lighter division's is used, since a thin field rather than size is the likelier reason. Women above 135 lb form one list, open weight; atomweight is measured against strawweight. Every list can be dragged back through time, and the rankings page then shows the list as the model sees it now, for that date.
The calibration page shows the forecaster against what happened, on every contest it priced, and against the betting market wherever a closing line exists. The market has been better every year, and the page shows it. A fighter's page carries the market's closing chance beside the forecaster's on every contest with a line on record (about 23,000 contests in more than a hundred organisations from 2007, a third of them in the UFC; the market is the bookmakers' consensus with their margin removed), and marks the nights the two picked different favourites or stood fifteen points or more apart. Outside the UFC the market is thinner, often one or two bookmakers, and the calibration page shows the two apart. Each decided contest also says which of the two was closer to the result, the one that gave the winner the higher chance: in the UFC the market about 55 times to the forecaster's 45, outside it about 60 to 40. One contest says nothing about skill; the balance does. Where the forecaster rated a fighter higher than the market did, that fighter won about 2 points more often than the market said in the UFC and about 1 point outside it, and far less often than the forecaster said. The forecaster's lean carries a little the market missed; its number is not one to trust over the market's.
The curve cannot be tested that way, since it has seen every fight it draws. It is tested on sumo, where the truth is known, and against the forecaster on MMA, as described above.
New results are added weekly. Between full runs, old results do not move: the forecaster carries on from where it stood and the curves are refitted with the new results in. A full run of the whole history, with every derived input rebuilt, happens roughly monthly.
These are the opinions of a model, not official rankings. The technical report, with the estimates behind every constant, will be linked from this page when it is published.