A plain accounting of what the prediction model is, how it was tested, and where it's weaker than we'd like — so the accuracy number below can actually be trusted.
Two different measurements, kept apart on purpose. The live record is the model calling fights before they happened. The backtest is a replay on older fights and is only an estimate of how it should do.
95% range for the hit rate: 54% to 69%. 14 events, Aug 1, 2026 to Nov 8, 2026. Each prediction was recorded once, before its event, and the site never overwrites it.
95% range for the hit rate: 59% to 64%. Trained on the 7,346 oldest fights, then scored on the 1,297 most recent ones it had never seen (early 2024 to July 2026, reconstructed — the original run didn't save exact dates). Fighter stats were rebuilt as they stood before each fight.
With 159 graded fights the live hit rate is still uncertain by about ±7 points. The live (62%) and backtest (61%) ranges overlap, so the gap between them is within normal sampling noise, not evidence the model got better or worse.
| Method | Hit rate | Brier | Measured on |
|---|---|---|---|
| Coin flip (always 50/50) | 50% | 0.250 | any fights |
| Elo rating only | 55% | 0.246 | backtest, 1,297 fights |
| Stat model alone (no Elo blend) | 62% | 0.226 | backtest, 1,297 fights |
| This model (stat model blended with Elo) | 61% | 0.231 | backtest, 1,297 fights |
| Always pick the sportsbook favorite (market probabilities, vig removed) | 68% | 0.223 | live, same 145 fights |
| This model on those same fights | 61% | 0.239 | live, same 145 fights |
Not shown: a “pick the higher-ranked fighter” baseline. Past divisional rankings were not recorded at fight time, and today's rankings would be hindsight, so that comparison cannot be made honestly.
On the fights where we have sportsbook odds, the market's probabilities were better calibrated than the model's. That's common — betting lines are hard to beat — so treat an "edge" on this site as a reason to look closer, not a guaranteed advantage. The market figures use the last odds recorded by fight day, while the model's number was locked in one to two days earlier, so the market had a little more information.
Model feature-v1, trained Jul 20, 2026. The training data and the live record never overlap. See every graded fight.
Accuracy is simply how often the model's favorite actually won. 61.3% beats a coin flip by a real margin, but it's not close to certainty — MMA has real upset variance, and the model isn't pretending otherwise.
Brier score measures calibration, not just win/loss: it penalizes being confidently wrong more than being cautiously wrong. A model that says "51%" for a coin-flip fight and a model that says "90%" for the same fight can have the same accuracy but very different Brier scores — this is the number that tells you whether the percentages themselves are honest.
Why a held-out test set, not just fitting the past: training and testing on the same fights would let the model memorize outcomes it already knows, producing an accuracy number that looks great but tells you nothing about a real upcoming fight. Testing only on fights chronologically after the training cutoff is the only honest way to estimate that.
An Elo rating gives every fighter a baseline from win/loss history alone — simple, always available, even for a fighter with almost no UFC history. A logistic regression model layered on top uses the differential in real fighter stats between the two fighters. The stat model's share of the final number is not fixed: it grows with experience, one fifth per UFC fight, and reaches 100% at 5 fights (a fighter with 4 fights gives an 80% stat / 20% Elo blend; a debut is all Elo). The less-experienced of the two fighters sets the share, since the model is only as reliable as its thinnest input. The explanation on each matchup states the weights actually used.
Ranked by how much each factor actually moved the model's predictions, most to least:
Being the model's favorite is not the same as being a good bet. A pick is made on one fighter when, at the best price any sportsbook is offering: the model gives them at least 60%; the price is no shorter than -200; expected value is at least +3.0% per unit; the model is above the sportsbook consensus (vig removed); and the odds are not stale.
Expected value per unit staked is the model probability × decimal odds − 1. It is separate from the probability edge (model minus market, in percentage points), which says where the model disagrees with the market but not whether the price pays enough. At −200 a bet needs to win 66.7% to break even, so a 61% model number there is about −8.5% and is not a pick. Picks are ranked by expected value, then by edge, then by model probability.
Every number on this site traces back to a real stat or this documented model output — nothing is an invented rating.
| Event | Date | Fights so far | Hit rate so far |
|---|---|---|---|
| UFC Fight Night: Medić vs Rodriguez | 2026-08-01 | 11 | 73% |
| UFC Fight Night: Gamrot vs Salkilld | 2026-08-09 | 22 | 82% |
| UFC 330: Makhachev vs Machado Garry | 2026-08-16 | 34 | 74% |
| UFC Fight Night: Hernandez vs Rodrigues | 2026-08-23 | 47 | 66% |
| UFC Fight Night: Nurmagomedov vs Song | 2026-08-29 | 59 | 69% |
| UFC Fight Night: Hooker vs Parnasse | 2026-09-05 | 72 | 64% |
| Noche UFC: Silva vs Delgado | 2026-09-13 | 84 | 63% |
| Crypto.com UFC 331 | 2026-09-20 | 96 | 63% |
| UFC Fight Night: Rosas Jr. vs Barcelos | 2026-09-26 | 106 | 63% |
| UFC 332: Silva vs Wang | 2026-10-04 | 120 | 63% |
| UFC Fight Night: Allen vs Duncan | 2026-10-11 | 131 | 63% |
| UFC Fight Night: Buckley vs Malott | 2026-10-18 | 144 | 60% |
| UFC 333: Volkanovski vs Evloev | 2026-10-24 | 157 | 62% |
| UFC Fight Night: Cortes Acosta vs Walker | 2026-11-08 | 159 | 62% |