How our UFC prediction model works

Last updated: April 17, 2026

MMA Analytics uses a stacked machine-learning ensemble to predict UFC fights. This page documents the model architecture, training procedure, feature selection, accuracy evaluation, and validation methodology — everything needed to evaluate whether our predictions are trustworthy.

Headline metrics (held-out test set)

• Accuracy: 67.6% (calibrated: 67.7%)
• Log loss: 0.598
• AUC (ROC): 0.727
• Expected Calibration Error (ECE): 0.015
• Symmetry deviation: <0.6% (predictions averaged over both fighter orderings)

1. Model architecture

We train five base learners on 45 features (see Feature Selection below) and combine their predictions with a ridge linear regression meta-learner.

Base learners

• LightGBM (0.22 meta-weight) — gradient-boosted decision trees, fast iteration, strong on tabular features
• XGBoost (0.25 meta-weight) — classic boosting; adds robustness on interaction terms
• CatBoost (0.36 meta-weight — highest) — handles categorical features natively; best single model
• Logistic Regression (0.16 meta-weight) — linear baseline; contributes to calibration
• Siamese Neural Network (0.13 meta-weight) — learns a fighter-embedding space; captures non-linear matchup interactions

Meta-learner

A ridge linear regression takes the five base model outputs and produces a final win probability. Ridge regularization (α tuned via cross-validation) prevents overfitting to the base-model correlations. The weights listed above are the learned meta-weights.

2. Feature selection

We start with 147 candidate features covering striking accuracy and defense, grappling metrics, endurance profiles, opponent quality (ELO-weighted), and recent-form indicators. We then run SHAP feature importance analysis and keep the top 45 features that cover 78.8% of signal.

Top predictive features (by SHAP importance)

1. late_round_finish_rate — finish rate in rounds 3+ (predicts cardio + damage accumulation)
2. damage_ratio — significant strikes landed vs absorbed (offense-defense balance)
3. r1_finish_rate — round-1 finish rate (early aggression signal)
4. avg_opp_elo — average ELO of opponents faced (strength of schedule)
5. standing_time_pct — % of fight time spent standing (grappling exposure)

3. Training procedure

• Data source: historical UFC fight data (fight results, fighter stats, odds movements, event metadata) from 2005–present
• Train/validation/test split: chronological split — model never sees fights from the test window during training
• No future leakage: every feature is computed using only data available before the fight date. A stat snapshot for Fight X uses only stats from fights up to (but not including) X.
• Hyperparameter tuning: Bayesian optimization (Optuna) on the validation set; final hyperparameters frozen before test evaluation
• Calibration: isotonic regression on held-out calibration data; ECE reduced from 0.038 (raw) to 0.015 (calibrated)

4. Symmetry enforcement

A fair prediction for "Fighter A vs Fighter B" should equal 1 minus the prediction for "Fighter B vs Fighter A". Naive models can violate this by up to 8%. We average predictions over both orderings, which reduces our symmetry deviation to under 0.6%. This is a small but important integrity check.

5. What our 67.6% accuracy means

On a held-out test set the model's pick is correct about 2 out of every 3 times. It is wrong 1 in 3 times. Do not treat any single prediction as a guarantee.

Our goal is not always to beat the market — a well-calibrated model can be useful for odds analysis even when public books price things similarly. We surface value opportunities (where our model meaningfully disagrees with market odds) explicitly on fight pages.

6. Limitations and honest caveats

• Short-notice replacement fights degrade accuracy; stats for fighters with only 1–2 UFC bouts are noisy and our model flags these as "high-risk".
• Weight cut failures, injury news, and late-breaking camp changes are not captured in the model — we note these manually in event previews.
• Title-fight five-round format is only weakly represented in training data compared to three-rounders.
• The model is probabilistic, not deterministic. Even a 90/10 favorite loses 10% of the time.

7. Validation vs real cards

We publish real-time predictions before every UFC event and keep a public record of card-by-card hit rates. You can check any past event page to see how we did. We do not hide losses.

8. Data sources and ethics

All training data is public UFC information. No insider information, no scraped private sources. Fighter stats are updated regularly. We prioritize reproducibility: given the same feature inputs, our model will always return the same prediction.

Frequently asked questions

Q: How accurate are MMA Analytics' UFC predictions?

On held-out test data, our ML ensemble is 67.6% accurate with a 0.598 log loss and 0.727 AUC. Calibration error (ECE) is 0.015, meaning our probability estimates match actual win rates closely.

Q: What machine-learning models do you use?

A stacked ensemble of LightGBM, XGBoost, CatBoost, Logistic Regression, and a Siamese Neural Network, combined via a ridge regression meta-learner.

Q: How often is the model retrained?

After every UFC card, we retrain with the latest fight results. Feature computation is recomputed for all active fighters.

Q: Do you have prediction results for a specific past event?

Yes — visit any event page under /events and look for the "Prediction Accuracy" section on the event summary.

Q: How can I use your predictions for betting?

Look for our value opportunities on each fight page — these are fights where the model's fair odds disagree meaningfully with market odds. See our Responsible Gaming page for harm-reduction guidance. Never bet more than you can afford to lose.