CageNightly
Fight wire
ResultsUFC 334Gane def. Hokit · DEC (U)13 results · 2 fight clipsResults & highlights
Marking our own homework

Scorecard

A win probability is only worth something if it comes true as often as it says. Here is every UFC fight since 2011 scored against the probability the Cage Nightly Fight Model gave before the fight. Every number is out of sample: each year's model was fitted only on the fights before that year.

62.1%favourites who won
0.227Brier score · a coin flip scores 0.250, lower is better
±1.9points, average gap between what we said and what happened

7,343 fights scored, all out of sample · draws and no contests left out

The honest comparisonBetter than the simple rules?

Bottom lineThe Fight Model scores 0.227, against 0.244 for always-red, 0.244 for the plain Elo and 0.250 for a coin flip, and picks 62.1% of winners. It beat the red-corner rule in 16 of 16 seasons and the Elo in 16 of 16.

How to read this chart
Four ways to put a number on a fight,
scored on the same 7,343 fights since 2011. Each is only allowed what was known before the bell.
The Brier score
is the average squared gap between the probability given and what happened (1 for a win, 0 for a loss). Lower is better: a coin flip scores 0.250, a perfect oracle 0. The lower the score, the longer the bar.
Picked right
is how often the side given more than 50% won. It ignores how confident the number was, which is why the Brier score comes first.
Why the red corner matters:
the UFC usually puts the favourite or the champion in red, so “always red” is a hard rule to beat. Our old Elo-only probability did not beat it. Before 2010 our source lists every winner in red, so those years are left out of every comparison here.
MethodBrier score · lower is betterPicked right
Cage Nightly Fight ModelRating, corner, age, reach, activity and opponent-adjusted stats0.22762.1%
Plain Elo ratingOur old win probability: results only0.24456.8%
Always red at 58%The UFC puts the favourite or champion in red0.24457.8%
Coin flip50% for everyone0.250–

The real benchmark is the betting market. We have logged market prices for 12 finished fights so far: too few to score fairly, so that comparison opens at 200.

Inside the modelWhat moves the number

Bottom lineThe three biggest levers: strikes absorbed vs what opponents land, red corner (the ufc's designated favourite or champion), age.

How to read this chart
One row per ingredient,
ordered by how much it moves the probability. Each bar is the effect of a typical difference between two fighters (one standard deviation), on the log-odds scale the model works in.
Gold pushes towards
the fighter who has more of it, grey against: an older fighter, or one who absorbs more strikes than opponents usually land, is less likely to win.
Opponent-adjusted
means each fight is measured against what that opponent usually allows or lands, so 60 strikes against a great defender count for more than 60 against a poor one.
Strikes absorbed vs what opponents land−0.31
Red corner (the UFC's designated favourite or champion)+0.28
Age−0.28
Striking output vs what opponents allow+0.24
Rating gap (Elo)+0.20
UFC debut−0.17
Takedowns vs what opponents allow+0.16
UFC experience−0.12
Years past 33−0.10
Takedowns conceded vs what opponents land−0.07
Time since last fight−0.07
Control time vs opponents' usual−0.06
Reach+0.06
Last three results−0.05
Knocked out in the last two years−0.05

The full method: every ingredient, how the opponent adjustment works and how the out-of-sample test is run.

Calibration

Does 70% mean 70%?

Bottom lineOn average the model is 1.9 points off what happened; the widest miss is the 75–80% group (said 77%, won 73%, 539 fights).

How to read this chart
Each dot
is a group of fights where the favourite had a similar pre-fight chance. Across = what the model said; up = how often the favourite really won. Bigger dot = more fights.
The gold diagonal
is a perfectly honest number. Above it, favourites won more than we said; below it, less. The short stem shows the size of the miss.
  • 50–55%said 53% · won 49%1,481 fights
  • 55–60%said 57% · won 58%1,569 fights
  • 60–65%said 62% · won 60%1,319 fights
  • 65–70%said 67% · won 67%1,156 fights
  • 70–75%said 72% · won 71%834 fights
  • 75–80%said 77% · won 73%539 fights
  • 80–85%said 82% · won 80%319 fights
  • 85–90%said 88% · won 85%124 fights
30%40%50%60%70%80%90%100%50%60%70%80%90%100%what the model saidhow often it happened49%58%60%67%71%73%80%85%

Year by yearEvery season, out of sample

Bottom lineThe model had the lowest score in 16 of 16 seasons; its best year was 2025 (0.215).

How to read this chart
Three lines, one per method,
left (oldest) to right. Each point is that season's Brier score. Lower is better, so the best line runs along the bottom.
Gold = the Fight Model,
grey = the plain Elo rating, red = always picking the red corner. The dashed line is a coin flip (0.250).
Each year is fresh:
the model used for 2015 had never seen a 2015 fight.
0.210.220.230.240.252011: model 0.235, Elo 0.247, red 0.241 · 295 fights20112012: model 0.228, Elo 0.244, red 0.235 · 333 fights2013: model 0.223, Elo 0.244, red 0.242 · 376 fights2014: model 0.232, Elo 0.240, red 0.233 · 494 fights20142015: model 0.238, Elo 0.243, red 0.247 · 464 fights2016: model 0.227, Elo 0.241, red 0.242 · 483 fights2017: model 0.231, Elo 0.246, red 0.249 · 446 fights20172018: model 0.231, Elo 0.245, red 0.251 · 469 fights2019: model 0.236, Elo 0.250, red 0.245 · 506 fights2020: model 0.224, Elo 0.245, red 0.238 · 444 fights20202021: model 0.229, Elo 0.244, red 0.246 · 497 fights2022: model 0.227, Elo 0.242, red 0.241 · 506 fights2023: model 0.224, Elo 0.243, red 0.245 · 504 fights20232024: model 0.222, Elo 0.241, red 0.251 · 513 fights2025: model 0.215, Elo 0.240, red 0.247 · 513 fights2026: model 0.220, Elo 0.244, red 0.248 · 500 fights2026
Fight ModelPlain EloAlways redCoin flip

Where the model is sharp, where it is notBy situation

Ended inside the distance61.8%3,745 fights · Brier 0.229
Five-round fights62.9%693 fights · Brier 0.224
Non-title fights62.0%6,994 fights · Brier 0.228
Three-round fights62.0%6,650 fights · Brier 0.228
Title fights64.5%349 fights · Brier 0.210
Went to the judges62.5%3,598 fights · Brier 0.226

The bigger the rating gap, the safer the favourite

61%Under 50 pts5,239
64%50-99 pts1,740
68%100-149 pts314
76%150+ pts50

Bottom lineWhen the gap is under 50 points the favourite wins 61% (5,239 fights); at 150+ points, 76% (50 fights, a small sample).

How to read this chart
Each column
groups fights by how far apart the two fighters were on the rating before the bell, in points (left = close, right = a big gap).
Height and %
= how often the higher-rated fighter won; the small number is the fight count.

Division by divisionMost and least predictable weight classes

Bottom lineCatchweight is the easiest to read (67.1% of 76 fights), women's bantamweight the hardest (59.6% of 255). With samples this size, a few points either way is noise.

How to read this chart
One row per weight class
ordered from most to least predictable. The bar and % = how often the model's favourite won there; the small number is the fight count.

How it endsUpsets come by knockout

When the favourite wins
KO/TKO 32%Submission 19%Decision 49%
When the underdog wins
KO/TKO 34%Submission 17%Decision 49%

Bottom lineUnderdogs win by KO/TKO 34% of the time, favourites 32%.

How to read this chart
Two bars
split every fight won by the model's favourite, and every fight won by the underdog, by how it ended: KO/TKO, submission or decision.

The ones we got most wrongBiggest upsets since 2011

The live ledger

12 frozen picks settled

The test above is out of sample year by year, but the list of ingredients was chosen in November 2026 by people who know how MMA history went, so it may still flatter us a little. The real test is the ledger: every probability is logged before the fight, frozen at the first bell and scored here as results come in. So far 58.3% of favourites have won.