Back to home

Machine Learning / Fight Predictions

How Upset Predicts Fights

An XGBoost model trained on 8,500+ historical fights, powered by Glicko-2 skill ratings and 283 carefully selected features.

65%

Test Accuracy

283

Selected Features

8,500+

Fights Analyzed

4,400+

Fighters Tracked

Performance

How our model compares to other prediction methods.

Random Guess
50%
Majority Pick
51%
Glicko-2 Only
57%
Upset Predictions
65%
Vegas Odds
68%

Vegas closing lines represent the practical ceiling (65–70%). Our model approaches this benchmark using only historical fight data — no odds, no insider info.

What We Analyze

Six categories of fighter data feed into every prediction.

Physical Attributes

  • Reach
  • Height
  • Age
  • Weight class

Fight Record

  • Wins & streaks
  • Experience
  • Win methods
  • Quality win rate

Striking Stats

  • Accuracy & defense
  • Round 1 output
  • Volume & power
  • Head/body/leg mix

Grappling Stats

  • Takedowns
  • Submissions
  • Control time
  • Ground striking

Fighter Ratings

  • Glicko-2 skill
  • Uncertainty
  • Rating trend
  • Opponent quality

Fighting Style

  • Activity & recency
  • Fight duration
  • Finishing rate
  • Cardio profile

All stats computed at point-in-time · No future data leakage · Bayesian priors for debut fighters

Top Factors

The five most predictive features, ranked by importance.

1Age Difference100%

Younger fighters have a measurable edge

2Reach Advantage90%

Physical reach is a key predictor of outcomes

3Glicko-2 Skill Rating85%

Rating gap between fighters predicts winners

4Round 1 Output81%

Early aggression and striking volume matter

5Fighter Activity72%

How recently and frequently a fighter competes

Integrity

Three safeguards that keep our predictions honest.

No Future Data

Every stat is computed using only fights that happened before the predicted bout.

No Corner Bias

Fighter positions are randomly assigned to prevent systematic red/blue advantage.

Real Validation

Tested on 318 fights across 26 cards the model never saw during training.

Documentation

Download our prediction model documentation for a deeper look.