CageMetrics

The CageMetrics fight predictor

A machine-learning model gives a win probability for every upcoming UFC bout. It looks at 40 numbers per fight, all known before the fighters walk out. This page covers what goes in, how the model is trained, and how well it does on recent fights it was never shown.

Last retrained Nov 20, 2026

Model accuracy 65.5% on 357 test fights since Nov 14, 2025
Δ vs Elo alone +8.7pp Elo by itself gets 56.9% on the same fights
Trained on 8293 UFC fights before Nov 14, 2025
Features 40 per fight, all known before the bout

What we're predicting

Every bout has a fighter A and a fighter B. For each completed fight the answer is 1 if fighter A won and 0 if fighter B won. The model learns to output a probability between 0 and 1 that fighter A wins. Draws and no-contests are left out of training because they have no winner to learn from.

That probability is what appears as the "CageMetrics pick" badge on upcoming fights on the home and event pages: the favoured fighter and their win chance.

Why prediction in MMA is difficult

A fight can turn on a single punch in the last ten seconds. No amount of data removes that. A model can find the patterns that hold on average, and some share of results will always look random.

The dataset is small. About 9,000 UFC fights have full per-fighter stats. That is enough for a careful model of this kind, but far short of the millions of examples that deep learning usually needs.

The sport changes. The 2008 lightweight division has almost no fighters in common with today's, so the model has to carry lessons across eras. Every fighter in the data was also good enough to get signed by the UFC, and the model never sees the fighters who weren't.

The betting market sets the practical ceiling. It picks the winner around 70% of the time, using everything the public knows about a fight. A model in the mid 60s is doing real work.

How we chose the model

Our data is a table: one row per fight, a few dozen numeric columns on very different scales, and not many rows. We considered four kinds of model.

Logistic regression

Simple, fast and easy to read. It assumes each stat pushes the odds steadily in one direction, and several of ours don't behave that way. Age has a sweet spot around the late 20s rather than "older is always worse". A layoff of 90 days or a year barely matters, but 1,000 days does. A logistic model can only handle shapes like that if we hand-build extra columns for them, which is fragile.

Random forests

Random forests handle those curved relationships without extra work and don't care what scale each stat is on. Their weakness for us is the probabilities. A forest averages many independent trees, which tends to pull its probabilities toward the middle, so a 70% from a forest is less trustworthy than a 70% from gradient boosting. That matters here: when we list a fighter as a 70% favourite, fighters in that spot should win about 70 times in 100.

Neural networks

The usual advice for small tables of data is to skip them. With about 9,000 rows and 40 columns, a neural network would memorise noise before it found anything a tree-based model can't. They are also harder to explain: there is no simple equivalent of the feature-importance chart further down this page.

Gradient-boosted trees

This is what we use, labelled GBT in the tables below. The model is a long chain of small decision trees, each one built to fix the mistakes of the trees before it. It picks up curved relationships and interactions between stats on its own, and it doesn't care that Elo sits around 1,500 while win rate runs from 0 to 1. Its probabilities are also reasonably well calibrated, which the calibration table below lets you check. We use scikit-learn's GradientBoostingClassifier. Training takes well under a minute, so the model is rebuilt from scratch every day rather than saved and reused.

Why not XGBoost / LightGBM / CatBoost

These more specialised boosting libraries might add a little accuracy. They also mean extra native dependencies, separate model files and version pinning. scikit-learn already ships in our container and training is a single fit() call, so the trade isn't worth it at this size.

The features

Each fight is described by 40 numbers. Every one of them is computed only from information that existed before the bout.

Skill (3)

Each fighter's Elo rating going into the fight, and the gap between them.

Recent form, last 5 fights (30)

The same set of numbers for each fighter, taken from their last five UFC bouts (or fewer, for newer fighters): win rate, rate of wins by finish, rate of unanimous decisions, average chaos score, significant strikes landed and thrown per minute and their accuracy, takedowns landed and attempted per minute and their accuracy, share of fight time spent in control, knockdowns per fight, submission attempts per fight, number of prior UFC fights, and days since their last bout.

Physical differences (3)

Reach, height and age, each as fighter A's value minus fighter B's.

Ages (2)

Each fighter's actual age, alongside the age gap. A 10-year gap between a 22-year-old and a 32-year-old plays out differently from the same gap between a 32-year-old and a 42-year-old, and the model needs both numbers to see that.

Context (2)

Whether it is a title fight, and the scheduled number of rounds (5 for a title fight, otherwise 3).

Leak prevention

The easiest way to fool yourself with a model like this is to let the result of a fight sneak into the numbers used to predict it. The recent-form stats come from a fighter's last five UFC bouts. If the fight being predicted were one of those five, the model would already know the answer.

To prevent that, we walk through the fights in date order. For each fight we compute its numbers first, and only then add the fight to each fighter's history. The five-fight window for any prediction therefore never includes the fight itself. Elo follows the same rule: we use each fighter's rating from just before the fight, not the updated one.

Training methodology

The live picks on upcoming fights come from a model trained on every completed UFC fight. The numbers on this page come from a separate test run that hides the most recent year from the model:

The test is repeated after every daily retrain, so the one-year window moves forward as new events happen.

Hyperparameters

The model has 400 trees, each at most 2 levels deep, with a learning rate of 0.05. Each tree is built from a random sample (fraction 0.8) of the training fights, and every final branch of a tree must cover at least 20 fights. The random seed is fixed so the results can be reproduced.

The trees used to be 3 levels deep. When we added each fighter's actual age as a feature, depth 3 started to overfit, finding patterns in the training data that didn't hold up on new fights. Cutting to depth 2 fixed it.

Results: head to head with Elo

Here is the model against Elo alone on the same 357 test fights. Lower is better for log-loss and Brier score (both measure how good the probabilities are; the Elo page explains them).

MetricGBTElo alone
Accuracy 65.55% 56.86%
Log-loss 0.6290 0.6811
Brier score 0.2191 0.2428

Results: vs the betting market

The real test for a fight predictor is the betting market, which sees everything we see and a lot we don't. This comparison uses the 357 test fights where Polymarket had prices on both fighters.

Metric Market GBT Elo alone
Accuracy 71.85% 65.55% 56.86%
Log-loss 0.5950 0.6290 0.6811
Brier score 0.2037 — —

On these fights the market picks 71.8% of winners and the model picks 65.5%, a gap of about 6.3 points. Measured from Elo's 56.9%, the model covers about 58% of the distance between Elo and the market.

What the market sees that we do not

Weigh-ins. Open-workout footage. Camp interviews. Late injury news that never shows up in the stats. Line movement, which reflects what professional bettors think. None of that is in our data. We don't expect to match the market. The aim is to close enough of the gap that the prediction is worth reading on its own.

Calibration

Accuracy doesn't tell you whether the probabilities are honest. A model that called every favourite a 99% lock and got 65% right would pick plenty of winners, but its numbers would be useless. The table below groups the test fights by how confident the model was in its favourite, then shows how often that favourite actually won.

Confidence bin n Model said Favourite won Market avg (n) Market won
0.50–0.60 162 54.9% 58.6% 55.8% (162) 58.6%
0.60–0.70 128 64.9% 67.2% 65.9% (128) 67.2%
0.70–0.80 58 73.3% 77.6% 73.5% (58) 77.6%
0.80–0.90 9 81.7% 88.9% 80.0% (9) 88.9%

A row is well calibrated when "Model said" and "Favourite won" are within a few points of each other. The top rows hold only a handful of fights, so their numbers swing a lot from one retrain to the next.

The two market columns only count fights in that row where Polymarket priced both fighters (the count is in parentheses). "Market avg" is the market's win probability for the same fighter our model picked, so comparing it with "Model said" shows how far the model and the market agree. A value under 50% means the market favoured the other fighter in those fights. "Market won" is how often the model's pick won within that same subset.

What the model actually uses

Feature importance measures how much each feature helped the model split winners from losers during training. The numbers are shares of the total across all 40 features, and each bar is drawn relative to the top one. The top 15:

Elo rating gap (A − B)
0.245
Age difference (years)
0.168
Fighter A — days since last fight
0.067
Fighter B — age
0.051
Fighter B — days since last fight
0.049
Fighter A — age
0.043
Fighter B — sig strikes thrown / min
0.031
Fighter A — sig-strike accuracy
0.030
Fighter B — sig-strike accuracy
0.023
Fighter B — sig strikes landed / min
0.022
Fighter B — Elo rating
0.020
Fighter A — control time %
0.019
Fighter A — Elo rating
0.018
Fighter B — avg chaos
0.018
Fighter B — control time %
0.017

The Elo gap is normally the single most important feature, which is why we think of Elo as the foundation and the rest of the model as a correction on top of it. Age shows up more than once: the age gap ranks high, and so do the two fighters' actual ages, because the same gap means different things at different points in a career.

Things we tried that didn't work

We tested a judge-rounds feature: the share of judges' rounds each fighter won in recent decisions. The same measure drives the decision scaling in our Elo. With one random seed it looked like a gain of about 0.8 points. Averaged over seven seeds it made no difference or made things slightly worse, probably because the decision-rate, average-chaos and win-rate features already carry the same information. It stays in the Elo and stays out of the predictor.

An earlier version of our Elo shrank a fighter's rating after a year out of the cage. It tested about 1.7 points more accurate, but it would have sent Jon Jones into the Gane fight 427 points lower than his actual rating, and it came down hard on every fighter who returned from a long layoff and won. We took it out. The Elo page has more detail.

What the model can't see

Style matchups. The model knows how often a fighter lands takedowns and strikes. It doesn't know that this particular wrestler's takedowns won't work against that particular opponent's defence. Matchup-specific data is thin and hard to model.

Camp and injury news. Sparring footage, weight-cut reports and hidden injuries never reach the public stats. The betting market picks up some of this through line movement.

Fight-week conditions. Altitude, short-notice opponent changes and how well a fighter rehydrates after the weigh-in aren't in the data either.

Adding any of this would mean bringing in sources far less reliable than official fight stats. The gap to the betting market is roughly what all that missing context is worth.