What matters
MMA outcomes depend on styles, scheduled length and small samples. A fighter may have only a handful of UFC bouts; a five-round fight is at most five round-level observations. Fight-level winner probabilities are the most reliable output. Method, distance and especially finish timing are progressively harder.
Important data
| Input | Why it matters |
|---|---|
| Fight DNA rates (per 15 minutes) | Comparable striking and grappling rates |
| Sample-size tiers | High / Medium / Low / Insufficient |
| Round-level samples | The underlying observations; thin by nature |
| Scheduled rounds (3 or 5) | Changes distance and finish probabilities |
| Card truth and availability | Only confirmed changes remove a bout |
| Judging and scorecards | Decisions depend on officials |
Common metrics
| Metric | What it measures | How to read it |
|---|---|---|
| Significant strikes landed / absorbed per 15 | Striking volume and defence | Rates, with sample tier |
| Takedowns landed / defended per 15 | Grappling offence and defence | Watch the observed seconds |
| Control time | Time in control positions | Not published as a simulator point estimate |
| Goes the distance | Share reaching the judges | Depends on scheduled rounds |
| Brier vs 0.25 | Model grading | On the model card and track record |
Model concepts
The PBE Fight Model produces pre-fight win probabilities from Fight DNA difference features — never from sportsbook prices. The Fight Simulator runs 10,000 fights per matchup, anchored to the Fight Model so they never disagree about who is favoured, and adds method and distance shares. The PBE Algo locks official calls with confidence tiers (Lean / Medium / High) and no-call rules, and grades them forever.
The simulator's model card lists quantities it deliberately does not show because they are not validated: finish time, round-by-round winners, and point estimates for control time, takedowns landed, knockdowns or submission attempts. A simulated outcome share is a distribution under the model's assumptions — not a forecast of exactly how or when a fight ends.
| Question | Output to read | Reliability |
|---|---|---|
| Who wins? | PBE Fight Model win % | Most validated |
| How does it end? | Simulated method shares | Harder; a distribution, not a pick |
| Does it go the distance? | Goes-the-distance share | Depends on scheduled rounds |
| When does it end? | Not published | Not validated |
Data tiers matter: FULL results are shown normally, LIMITED results with reduced confidence, and matchups with insufficient verified Fight DNA are not simulated. Debut fighters are not called by the PBE Algo at all.
Live intelligence
Live, a fight's state is the round, the clock, the scheduled rounds and the cards so far. After two rounds without a finish the remaining distribution is very different from the pre-fight one — see the pregame vs live lesson.
PropBetEdge UFC tools
Fight DNA ↗
How Fight DNA describes a fighter’s record, with sample-size tiers.
UFC / MMAPBE Fight Model card ↗
Training window, out-of-sample Brier and log loss, baselines, leakage checks.
UFC / MMAFight Simulator ↗
Ten thousand simulated fights per matchup, anchored to the PBE Fight Model.
UFC / MMAPBE Algo ↗
Pre-fight win probabilities, locked before the fight and graded forever.
UFC / MMATrack Record & Past Picks ↗
Every graded call.
UFC / MMAInjuries & Withdrawals ↗
One card per availability episode.
UFC / MMAEvents & cards ↗
Cards, including verified card changes.
UFC / MMARound by round ↗
Round-level fight data.
Worked example
- Fight Model: A 58%. Simulated outcomes: A by decision 26%, A by KO/TKO 24%, A by submission 8%, B 42%.
- "A wins" is a 58% statement. "A by decision" is only 26% — and is still the single most common outcome.
- If the bout is booked for 5 rounds instead of 3, the distance share rises and finish shares shift; the simulator waits for the official round count rather than guessing.
Glossary
Related lessons
Fight DNA
What does a Fight DNA rate mean, and how much should I trust it?
Fight Simulator
The simulator says 58% and 31% by decision. What is each number, and why is it not a finish-time prediction?
Confidence & gates
What do Lean, LIMITED, Insufficient and "gated" each mean?
PBE Picks & track records
How do I know a published record is real, and what should I look at in it?
Brier & log loss
How do you grade a probability, not a pick?
Simulation
What does "10,000 simulations" actually tell me?
Context changes
A player is reported out. How much should that change a number — and when?