Method · cnm-v1 · updated November 16, 2026
The Cage Nightly Fight Model
Every win probability on Cage Nightly, from the fight pages to the prediction ledger, comes from one model. It starts from the Cage Nightly rating (an Elo score built from results) and adds what else was known about both fighters before the opening bell. It never sees anything from the fight it is pricing.
How good it is
Scored on 7,343 UFC fights since 2011, each one out of sample (see below). The Brier score is the average squared gap between the probability and the result: lower is better, a coin flip scores 0.250.
| Method | Brier score | Picked right |
|---|---|---|
| Cage Nightly Fight Model | 0.227 | 62.1% |
| Plain Elo rating (our win probability until 10 November 2026) | 0.244 | 56.8% |
| Always the red corner, at the red corner's win rate | 0.244 | 57.8% |
| Coin flip | 0.250 | – |
Season by season, calibration and the biggest misses are on the scorecard. The comparison that matters most, against the betting market, opens once we have logged market prices for 200 finished fights.
What goes in
Each ingredient is the difference between the red-corner and the blue-corner fighter, measured on the day of the fight from fights that had already happened.
- Corner. The UFC usually puts the favourite or the champion in red. That is information, and the plain Elo ignored it.
- Rating gap. The Cage Nightly Elo rating of both fighters going in.
- Opponent-adjusted striking. For every past fight: significant strikes landed per minute minus what that opponent usually absorbed per minute, and strikes absorbed minus what that opponent usually landed.
- Opponent-adjusted wrestling. The same idea for takedowns (landed against what the opponent usually concedes, and conceded against what the opponent usually lands) and for control time.
- Finishing tools. Submission attempts and knockdowns per 15 minutes.
- The body and the calendar. Age, years past 33, reach, number of UFC fights, a UFC debut, time since the last fight, the last three results and knockout losses in the last two years.
Every per-minute number is pulled towards the league average by 15 imaginary minutes of average fighting, so a fighter with one short fight does not look like a world-beater. The “usually” in the adjustments is itself measured before each fight, with the same pull towards average.
How much each one weighs
Effect of a one-standard-deviation difference between the fighters, on the log-odds scale (the corner is a straight yes or no). Positive favours the fighter with more of it. Ingredients overlap, so read a weight as “on top of everything else”.
| Ingredient | Weight |
|---|---|
| Strikes absorbed vs what opponents land | −0.309 |
| Red corner (the UFC's designated favourite or champion) | +0.285 |
| Age | −0.281 |
| Striking output vs what opponents allow | +0.238 |
| Rating gap (Elo) | +0.198 |
| UFC debut | −0.168 |
| Takedowns vs what opponents allow | +0.156 |
| UFC experience | −0.121 |
| Years past 33 | −0.096 |
| Takedowns conceded vs what opponents land | −0.070 |
| Time since last fight | −0.069 |
| Control time vs opponents' usual | −0.061 |
| Reach | +0.058 |
| Last three results | −0.052 |
| Knocked out in the last two years | −0.046 |
| Submission attempts | +0.010 |
| Knockdowns landed | −0.007 |
How it is tested
A logistic regression with a light penalty against extreme weights. The test is walk-forward: to price the fights of 2016, the model is fitted only on fights up to the end of 2015; then the year is scored, added to the history, and the model is fitted again for the next one. No probability on the scorecard was produced by a model that had seen that fight. For upcoming fights the model is fitted on the full history and updated every hour until the event starts, when the ledger freezes it.
One data trap we had to handle. Our source lists every pre-2010 winner in the red corner (the red corner wins 100% of recorded fights from 1994 to 2009). For those years the corner says nothing, so each pre-2010 fight enters the fit twice, once from each side, with the corner switched off. The corner effect is learned only from 2010 onwards, and every comparison involving the red corner starts in 2011.
What it does not know
- Injuries, camp changes, illness and weight cuts, unless they show up in the results.
- Short-notice replacements, beyond what time since the last fight says.
- Fights outside the UFC: a debut is a debut, whatever the fighter did elsewhere.
- Styles beyond the numbers: it knows a fighter takes people down more than expected, not that the opponent has never been taken down by an Olympian.
The ingredients were chosen in November 2026 by people who know how MMA history went. The walk-forward test keeps the weights honest, but the choice of ingredients can still flatter the past a little. That is why the frozen ledger is the final word.
Free to reuse
The model's probabilities and this report are open data under CC BY 4.0. Credit Cage Nightly with a link.

