An independent probability.A real market price.The difference is PBE Edge.

• American odds are what the books actually offer. PBE Picks shows the consensus price and the best available price with its book.
• Raw implied probability contains vig. Both sides of a fight add up to more than 100%; that excess is the book's margin.
• PBE removes the vig across the current observed market: the median across books, normalised so both sides sum to 100%.
• PBE Edge compares the model probability with that de-vigged market probability, in percentage points. It is never measured against a vigged price.
• Market data is read only after the model has scored the fight. A price never changes a probability, and a stale price never publishes an edge.

The 33 inputs are all pre-fight Fight DNA differences; 0 of them come from a sportsbook. The two lanes meet only at the final subtraction, which is why PBE Edge can disagree with the market at all.

1. 1Discover

   Every UFC card inside the horizon is picked up by the hourly scheduler.

2. 2Verify

   Both fighters resolve to one reconciled identity with current Fight DNA.

3. 3Score

   33 pre-fight features become an independent PBE probability.

4. 4Compare

   The current fight-week market is de-vigged and PBE Edge is measured.

5. 5Regenerate

   Until lock, new pre-fight inputs can change the call, every hour.

6. 6Lock

   The prediction is written on the database clock and becomes immutable.

7. 7Grade

   The official result enters the record; corrections are dated revisions.

This record and the backtest below are never combined. They answer different questions. A backtest says what the model would have done on fights whose results already existed somewhere; a live record says what it did with the future genuinely unknown. They are stored in separate tables, reported by separate views, and there is deliberately no query in this product that adds them together.

7,047 bouts scored between 2013-01-19 and 2026-09-05. Every fold trains only on bouts that had already happened and is scored only on bouts that had not; the model is refit from scratch each calendar year, and the ridge penalty is chosen on an inner chronological split of the training window, never on the year being scored. There is no random train/test split anywhere in this pipeline — with fighters recurring across years, a random split would let the model learn how a career turned out before predicting its middle.

Brier and log loss lead because they are proper scoring rules: only the true probability minimises them, so a model cannot improve them by being confidently wrong. Accuracy is shown because people ask for it, but a model that says 51% and is right is not the same product as one that says 80% and is right, and accuracy cannot tell them apart.

Both corners of a bout are the same prediction stated two ways, so every row below is folded onto the side the model actually picked: confidence is max(p, 1−p) and a hit is that pick winning. The gold line is what happened; the blue block is what was claimed.

Real bouts, real out-of-sample probabilities, real results — and a deliberate mix of right and wrong. These are backtest output, clearly marked as such, and never current PBE Picks; the current calls are in PBE Picks for UFC Pro.

The debut row is the honest headline of this table. When one corner has never fought inside the promotion, the model has almost no signal beyond age, reach and stance, and its AUC there is barely above a coin. It says so by producing probabilities near 50% rather than by manufacturing confidence — which is why the calibration above survives including those fights rather than being computed on a flattering subset. PBE Picks does not call those fights at all.

A fight model that has quietly read the result is easy to build and worthless. Six independent checks run on every build, and the run fails if any one of them does.

Every input is a difference between the two corners, and the model is fitted with no intercept, so swapping the corners negates the vector and the two probabilities are exactly complementary. Sportsbook prices are not an input at any stage; the market appears only after a probability exists.

• age_diff_yearsage · staticAge in years at the event date. Older is normally worse.
• reach_diff_inreach · staticReach in inches.
• height_diff_inreach · staticHeight in inches.
• experience_log_diffexperience · dna snapshotlog1p of completed prior bouts in the snapshot sample.
• five_round_exp_diffexperience · dna snapshotlog1p of prior scheduled-five-round appearances.
• title_exp_diffexperience · dna snapshotlog1p of prior title-bout appearances.
• winrate_diffform · dna snapshotLaplace-smoothed prior win rate, (w+1)/(appearances+2).
• recent5_winrate_diffform · dna bout rowsWin rate over the last five completed bouts before the event date.
• streak_diffform · dna bout rowsSigned current win or loss streak, clamped to plus or minus five.
• layoff_log_diffform · dna bout rowslog1p of days since the previous bout. A long layoff is normally a negative.
• slpm_diffstriking · dna snapshotSignificant strikes landed per minute.
• sapm_diffstriking · dna snapshotSignificant strikes absorbed per minute.
• sig_diff_per_min_diffstriking · dna snapshotSignificant strikes landed minus absorbed, per minute.
• sig_accuracy_diffaccuracy · dna snapshotSignificant strike accuracy.
• sig_defense_diffaccuracy · dna snapshotSignificant strike defence, one minus opponent accuracy.
• kd_per15_diffstriking · dna snapshotKnockdowns scored per fifteen minutes.
• kd_absorbed_per15_diffdurability · dna snapshotKnockdowns conceded per fifteen minutes.
• td_landed_per15_difftakedowns · dna snapshotTakedowns landed per fifteen minutes.
• td_accuracy_difftakedowns · dna snapshotTakedown accuracy.
• td_defense_difftakedowns · dna bout rowsOne minus opponent takedown accuracy across prior stat-covered bouts.
• control_share_diffcontrol · dna snapshotControl time as a share of observed fight time.
• sub_att_per15_diffcontrol · dna snapshotSubmission attempts per fifteen minutes.
• finish_rate_difffinishing · dna snapshotFinishes as a share of prior wins.
• ko_rate_difffinishing · dna snapshotKO/TKO wins as a share of prior wins.
• sub_rate_difffinishing · dna snapshotSubmission wins as a share of prior wins.
• ko_loss_rate_diffdurability · dna snapshotTimes stopped by strikes as a share of prior appearances.
• sub_loss_rate_diffdurability · dna snapshotTimes submitted as a share of prior appearances.
• pace_retention_diffpace · dna snapshotRound-three attempt pace divided by round-one pace.
• champ_round_delta_diffpace · dna snapshotRounds four and five pace minus rounds one to three pace.
• southpaw_edgestance · staticPlus one when this corner is the only southpaw, minus one when the opponent is, otherwise zero. Antisymmetric by construction.
• sos_diffopponent · dna bout rowsMean pre-fight win rate of prior opponents, each measured as of the date they were fought.
• quality_wins_diffopponent · dna bout rowslog1p of prior wins over opponents who were above a .600 pre-fight win rate at the time.
• stat_sample_log_diffconfidence · dna snapshotlog1p of stat-covered prior bouts. Carries how much evidence each corner actually has.

Model pbe-fight-model-v1, features pbe-fight-features-v1, spec hash 75da0a1de14de2e1…. ridge logistic regression, no intercept, antisymmetric differential features, trained on 9,175 graded bouts from 1994-03-11 to 2026-09-05. The full methodology, the leakage argument and the complete backtest tables are in the repository under docs/model/.

The underlying temporal feature source is the repaired historical Fight DNA snapshot layer. How Fight DNA works →