- Home
- PBE Algo
PBE Algo
Pre-fight win probabilities. Locked before the fight. Graded forever.
PBE Algo independently evaluates eligible UFC matchups using pre-fight Fight DNA data. Every official call is locked before the bout begins, then graded against the official result. Picks are never rewritten after the fact.
$9.99/month or $3.99/week. No free trial.
pbe-fight-model-v1- Model
- PBE-FIGHT-MODEL-V1
- Official calls
- 32
- Ledger
- Live · scheduler active
- Lock rule
- Before the fight
- Market input
- No
- Grading
- Official result
Live PBE Record
- Official locks
- 32
- Record
- 20-12
- Hit rate
- 62.5%
- Brier
- 0.244 lower is better · coin flip 0.250
- Pending
- 0 locked, awaiting result
Every call counted here was locked before its fight. First lock 2026-09-18, latest 2026-10-09. Individual calls, probabilities and the per-call history are UFC Pro.
| Band | Decided | Claimed | Happened |
|---|---|---|---|
| 55-60% | 9 | 57.1% | 44.4% |
| 60-65% | 10 | 63.0% | 100.0% |
| 65-70% | 4 | 65.8% | 25.0% |
| 70-80% | 8 | 74.4% | 62.5% |
| 80-100% | 1 | 89.6% | 0.0% |
The current calls live in PBE Picks
This page is the method and the public proof. The calls themselves are UFC Pro. For every eligible fight on the next UFC card, PBE Picks shows:
- The pick and its win probability
- Confidence tier and data quality
- Fight-week odds, de-vigged market probability and PBE Edge
- The model's drivers for and against
- Every NO MODEL CALL, with its reason
From fight card to graded call
- 1Fight discovered
- 2Identity verified
- 3Fight DNA features
- 4Model scores
- 5Eligibility gate
- 6Locked
- 7Graded
Read each step in detail
- 1Card discovered
Every UFC card inside a 14-day horizon is picked up hourly by a Cloudflare scheduler. Contender Series and non-UFC events are out of scope.
- 2Identity verified
Both corners must resolve to one reconciled fighter record, with no open same-name identity review.
- 3Features assembled
33 pre-fight difference features are rebuilt from Fight DNA as of the event date: nothing from the fight itself, and never a sportsbook price.
- 4Eligibility decided
A deterministic rule set decides ELIGIBLE or NO MODEL CALL, and a no-call records its exact reasons.
- 5Hourly regeneration
Until the lock, an eligible call is re-scored every hour as the card changes. A provisional call is not part of the record.
- 6Locked on the database clock
The call locks once, on the database's own clock, before the fight. From that moment nothing about the prediction can be edited, and it becomes part of the official PBE record.
- 7Graded after the fight
Results are bound to the stored official result. Corrections are new, dated revisions that must give a reason; the original stays on file.
Locked means locked
Once an official prediction locks, it is part of the record for good.
- The prediction cannot be edited. Pick, side and probability are fixed at the lock.
- The lock time is stored from the database's own clock, before the fight.
- The model and feature version are stored with the call, along with the inputs it was scored on.
- The result is graded later against the stored official result.
- A correction never rewrites the call. It is a new, dated revision that must give a reason; the original stays on file.
No forced picks
Every eligible bout gets a call. A bout that fails any rule gets NO MODEL CALL with the reason on record, and a probability is never nudged to make a bout callable.
What PBE Algo uses
33 pre-fight features, each one the difference between the two corners, rebuilt from Fight DNA as of the event date.
- Age
- Reach and height
- UFC experience, five-round and title-fight experience
- Win rate, last-five form, streak and layoff
- Striking volume, differential and knockdowns
- Striking accuracy and defense
- Durability: knockdowns absorbed, KO and submission losses
- Takedown volume, accuracy and defense
- Control time and submission attempts
- Finish, KO and submission rates
- Pace retention into later rounds
- Stance matchup
- Strength of opposition and quality wins
- How much fight data each corner has
What it does not use
- Sportsbook prices. 0 of 33 model features come from a market.
- Anything from the fight itself. Features are rebuilt as of the event date, and the release passed all 6 leakage checks.
- Information after the lock. A locked call is never re-scored.
- A human override. No editor picks a side or adjusts a probability.
The no-call rules and confidence tiers
- Not a UFC card in PBE Algo scope
- Bout off the card, changed or already fought
- Fighter identity not fully reconciled
- No live model version
- Pre-fight features could not be assembled
- Latest fight not yet in Fight DNA
- A fighter is making their UFC debut
- Fewer than 20 of 33 features available
- Too close to call (below 55%)
- Lock window closed before a call was made
Confidence: Lean below 60%, Medium from 60%, High from 70% only when both fighters have at least 3 prior UFC bouts and at least 31 of 33 features are available. Thin samples are capped at Medium however large the probability.
Market prices are compared after the model has scored the fight. They are never a model input.
Sportsbook prices include a margin, so their implied probabilities add up to more than 100%. During fight week PBE Algo reads the current market, shows the consensus and best available American odds, and removes that margin before comparing. A market older than its fight-week window shows its age and publishes no edge (until 72 hours before lock: current for 12h 10m; until the final day: current for 6h 10m; in the final 24 hours: current for 60 minutes). Brier score, log loss and calibration are tracked by confidence tier and by PBE Edge band as the graded record accumulates. PBE Edge describes a disagreement; it is not betting advice.
Before it made a single live call, the model was tested walk-forward: for each year from 2013, it was trained only on earlier fights and scored on fights it had never seen. These numbers describe what the model would have done. They are kept apart from the live record above and are never added to it.
- Out-of-sample fights
- 7,047 2013–2026, each scored without being trained on.
- Accuracy
- 59.6% How often the side the model favoured went on to win.
- Brier score
- 0.236 Lower is better. 0.250 is coin-flip probability quality.
- Calibration error
- 0.0139 When the model says 70%, how far off is how often it actually happens? Closer to zero is better.
A modest, honest edge over a coin flip across thousands of fights. It does not predict any single fight with certainty, and nothing here is a promise about future results.