CageMetrics

The CageMetrics Elo

Every fighter has an Elo rating, and it updates after every UFC fight. The math is the system Arpad Elo designed for chess in the 1960s. Our changes make the rating respond to how a fight was won. A 30-second knockout should move the rating more than a split decision.

Live snapshot

Fighters rated 2781 626 are ranked (3+ UFC fights, active in the last 2 years)
K-factor 64 96 for a fighter's first 5 bouts
Decisions parsed 4052 of 4169 decisions (the rest are draws or have no scorecard on record)
Finish multiplier ×1.80 applied to K for a KO/TKO or submission

What Elo measures

Elo gives every competitor one number that represents their skill relative to everyone else in the pool. Compare two ratings and you get the probability that one beats the other. After each match the winner gains points and the loser drops points. How many points move depends on how surprising the result was. Beating a much higher-rated opponent earns more than beating someone you were already expected to beat.

The math:

expected_a = 1 / (1 + 10^((rating_b - rating_a) / 400))
rating_a  += K × (score_a - expected_a)

Here score_a is 1 for a win, 0.5 for a draw and 0 for a loss. K sets the size of each update: the bigger K is, the more one result moves the rating. As a reference point, a fighter rated 400 points higher is expected to win about 10 times out of 11.

Why classical Elo fits MMA poorly

Chess Elo only cares who won. In MMA that throws information away. A 30-second KO and a razor-thin split decision both count as wins, but they say very different things about the gap between the two fighters. A 30-27 / 30-27 / 30-27 win is more convincing than a 29-28 / 29-28 / 28-29 split, and a rating that treats them the same ignores what happened in the cage.

New fighters are the other problem. UFC careers are short and most fighters compete two or three times a year. A debutant whose real level is around 1,700 should not need 20 fights to climb up from the starting rating of 1,500.

Neither fix removes the randomness of the sport. One punch can end a fight, so no rating system comes close to picking every winner. The aim is to get as much signal as possible out of the results we have.

What we changed

1. Provisional K-factor for newcomers

A fighter's first 5 UFC bouts use K = 96 instead of the standard 64. A newcomer's rating is the least reliable number we have, so bigger updates let it move quickly toward the right level. After 5 fights K drops to the standard value and the rating settles down.

2. Margin-of-victory scaling on K

K is multiplied by a factor that depends on how the fight was won.

We picked these values by replaying every UFC fight in our database under many different settings and keeping the combination that predicted results best. The tuning section below has the details.

3. Decisions scaled by judge rounds won

This is the change that most sets our rating apart from other public MMA Elos.

Three judges score every fight round by round under the 10-point must system, and the fighter ahead on at least two of the three cards wins. ufcstats publishes each judge's totals with every decision. For Dominick Reyes vs Johnny Walker in April 2026 the scorecards read Sal D'Amato 28 - 29. Solimar Miranda 29 - 28. Junichiro Kamijo 28 - 29.

Under the 10-point must system the winner of each round gets 10 points and the loser gets 9 (or 8 or 7 for a lopsided round). So each judge's total tells us how many rounds that judge gave each fighter:

rounds_won = max(0, score − 9 × scored_rounds)

In each pair ufcstats lists the loser's score first, then the winner's. Decoded:

That matches how close the fight was. Reyes won because two of the three judges had him ahead, but across all three cards he took only 5 of the 9 rounds scored. Clear decision wins usually come out at 7 to 9 of 9.

The winner's share of the scored rounds sets the multiplier. It rises in a straight line from 0.50 at a 50% share (the lowest a winner can have) to 1.80 at 100% (30-27 on all three cards):

50% ×0.50 56% ×0.64 78% ×1.22 100% ×1.80 winner's share of judge-rounds

A 30-27 shutout on all three cards (9 of 9 rounds) moves the rating by K × 1.80, the same as a finish. A unanimous 29-28 / 29-28 / 29-28 (6 of 9) gets K × 0.93, and a 29-28 / 29-28 / 28-29 split (5 of 9) gets K × 0.64. Under the old system every unanimous decision got the same ×1.30 and every split or majority decision got ×0.65, whatever the cards said.

Reading the scorecards

ufcstats writes each judge's totals as "loser's score - winner's score", not in the order the fighters are listed. We confirmed this against fights where the winner appears on either side, and we use the recorded winner to assign each score to the right fighter.

How the multipliers were tuned

For each combination of settings we replay every completed UFC fight in date order. Before each fight we note which fighter had the higher rating and what win probability the ratings gave. We then score the combination on fights where both fighters already had at least two UFC bouts, because any system is mostly guessing on debutants. Three measures:

In a November 2026 re-run (9,003 completed fights, 5,061 of them between fighters with two or more prior bouts), the sliding decision scale beat the old flat ×1.30 / ×0.65 decision values by about 0.9 points of accuracy, with better log-loss and Brier as well. We use the 0.50 / 1.80 endpoints because they produced the best probabilities in the search. Stretching them further (0.30 / 2.50) picks about 0.9 points more winners, but its log-loss and Brier are worse, meaning it is overconfident when it is wrong. We kept the setting with the better probabilities.

Vs vanilla Elo

How much do these changes help? We ran the same test on two systems: the CageMetrics settings, and plain Elo with K = 40 for everyone and no adjustment for how the fight was won. Both replayed the same 9,003 UFC fights and were scored on the 5,061 bouts where both fighters had at least two prior UFC fights (data as of November 2026).

MetricCageMetrics EloVanilla Elo (K=40)
Accuracy on established fights58.39%56.06%
Log-loss0.67760.6823
Brier score0.24180.2446

Our version picks about 2.3 more winners per 100 fights (roughly 58 against 56), and its probabilities are better on both measures. Adding the changes one at a time: the larger K values are worth about 0.4 points, the finish bonus with flat decision values adds about 1.0, and scaling decisions by judge rounds adds about 0.9 more.

The full parameter set

ParameterValueRole
Starting rating1500Every new fighter starts here
Base K-factor64Step size for established fighters
Provisional K-factor96Larger step for a fighter's first bouts
Provisional bouts5Number of fights that use the provisional K
Title-fight multiplier1.00Extra weight for title fights (set to 1.00, so no effect; tested and didn't help)
Finish multiplier1.80K scaling on KO/TKO, submission or doctor's stoppage
Decision: close end0.50K scaling when the winner took 50% of judge rounds
Decision: sweep end1.80K scaling when the winner took 100% of judge rounds
Unanimous fallback1.30Unanimous decisions with no scorecard on record
Split/majority fallback0.65Split or majority decisions with no scorecard on record

The current top of the board

For a sense of scale, here are the highest-rated active fighters right now:

#FighterDivisionEloBouts
1 Islam Makhachev Welterweight 2256 19
2 Charles Oliveira Lightweight 2123 37
3 Ciryl Gane Heavyweight 2111 15
4 Alexander Volkanovski Featherweight 2106 19
5 Justin Gaethje Lightweight 2082 16
6 Dricus Du Plessis Middleweight 2074 11
7 Ilia Topuria Lightweight 2069 10
8 Khamzat Chimaev Middleweight 2058 10

The full list, with a leaderboard for each division, is on the rankings page.

What we tried and didn't keep

A title-fight bonus. We tried multiplying K by 1.10, 1.25 and 1.50 for championship fights. None of them improved accuracy or log-loss. The setting is still there at ×1.00 in case we revisit it.

Rating decay for inactive fighters. We tried shrinking a fighter's rating after a year out of the cage. In testing it added about 1.7 points of accuracy, but it came down hard on fighters who returned from long layoffs and won. Jon Jones would have gone into the Gane fight rated 1,544 instead of 1,971, a 427-point drop, and then beaten Gane anyway. Its log-loss was also worse. We removed it.

What it can't see

Styles. Elo boils everything about a fighter down to one number. A wrestler who beats other wrestlers but loses to strikers ends up with one rating that is right on average and wrong for both kinds of matchup.

Weight class changes. A fighter who moves up carries their old rating with them, with no adjustment for the size difference. These moves are uncommon enough that we have not corrected for them.

Draws versus no-contests. Our data cannot tell them apart, so both count as a draw (0.5) for each fighter. Both are rare, so the effect is small.

10-8 and 10-7 rounds. The rounds-won formula assumes every round is scored 10-9. A 10-8 round costs the fighter who lost it an extra point, and the formula reads that point as a round they did not win. A fighter who wins one round on a card but loses another 10-8 gets credited with zero rounds on that card instead of one. The count never goes below zero, and 10-8 rounds are uncommon, so the effect is small.

Our fight predictor feeds Elo into a larger model along with recent form, age and other stats. On recent fights it closes about two-thirds of the gap between Elo and the betting market. It is described on the predictor methodology page.