How Vex AI works

Fighter identities, records, rankings, bios and event cards are real data from trusted public sources. From those we build each fighter's statistical profile: significant strikes landed/absorbed per minute, striking accuracy/defense, takedowns per 15 and accuracy/defense, submission attempts, control time and knockdowns, plus engineered indices for cardio, durability and finishing ability. Context inputs include reach, stance, age, real layoff (months since the fighter's last bout) and recent form. Competition (opponent quality) is grounded in real strength-of-schedule — the average win rate of a fighter's recent opponents — not just a rank or a guess.

Where a fighter has real aggregated per-fight statistics, we use them. Where those aren't available, the per-fight stats are transparently estimated from that fighter's real record and finish profile — deterministic and reproducible, never random — and flagged as estimates in the UI. We never present an estimate as if it were measured.

Raw stats are normalized to 0–100 within weight class, then blended into nine interpretable category sub-scores: Striking, Wrestling, Grappling, Submission, Cardio, Durability, Physical, Form and Competition.

The composite rating is a transparent weighted sum of the sub-scores. Win probability uses a logistic (Bradley–Terry-style) function of the rating gap:

R   = Σ (weight_c × subscore_c)
P(A) = 1 / (1 + e^( -k × (R_A − R_B) ))

The rating gap above is only a prior — we do not let one composite number decide the fight. We simulate the bout thousands of times, round by round. Each round the model works out where the fight happens (a fighter's takedown threat against the opponent's takedown defense → how much of the round is on the ground vs standing), then resolves striking in the standing portion and grappling/control on the ground. So a great striker with weak takedown defense correctly fares worse against an elite wrestler.

Fatigue accumulates across rounds (worse for low-cardio fighters, so they fade late) and damage carries over (a hurt fighter is more finishable later). Finishes follow coherent paths — a KO from standing dominance, power and a weakened chin; a submission from control plus submission threat against the opponent's defense. Aggregating the runs yields the method-of-victory split (KO/TKO, submission, decision), the round-by-round breakdown and confidence intervals.

For the final win probability we go one step further: we ensemble this transparent rating-plus-simulation model 50/50 with a gradient-boosted model trained on the historical bouts below. The boosted model captures non-linear matchups a linear rating can't (e.g. an elite wrestler against weak takedown defence), and on a leakage-free backtest it improved the probabilities measurably. The transparent model still drives everything you see — the radar, the category contributions, the method/round breakdown and the plain-language drivers — so the explanation always matches the inputs.

• Uncertainty inflation: probabilities are pulled toward 50% when data is incomplete or the styles are high-variance.
• Hard caps: displayed probabilities are clamped (≈8–92%) — MMA is too chaotic for false certainty.
• Ranges, not points: every prediction ships with a confidence interval and a volatility tag.

Absent stats are shrunk toward the weight-class mean (a Bayesian prior), the data-completeness meter drops accordingly, and imputed fields are flagged. We never silently fabricate a number.

The simulation's finish hazards are validated against real outcomes: we checked the model's predicted finish rate against what actually happened across the historical bouts and corrected it until they matched (about half of fights reach a decision, the rest split between KO/TKO and submission). Those finish rates are also weight-class aware — heavyweights finish far more often by knockout than flyweights — using each division's real historical KO/submission base rates. So the method-of-victory split is calibrated to reality, not invented.

Two honest checks, never one invented number. (1) Going forward: every Vex AI pick is logged before the fight and graded against the real result afterward — never back-dated; the live number is earned only from graded predictions. Our first live card is now graded (and the model started strong), but a single card is a tiny sample — the live record stays deliberately humble and grows honestly, card after card. See the self-building record on the Vex AI accuracy page. (2) Looking back: the engine is also validated on real history — see the leakage-free backtest in the next section.

Beyond the live record, the engine is validated against history. We rebuilt the inputs for 10,206 real UFC bouts (1993–2026, drawn from a dataset of 13,866 fighters) using only each fighter's fights before that date — no hindsight, no leakage — then ran the live engine and compared its pick to the real result.

This is how the engine was tuned, not just described — by walk-forward validation (train on the past, score the future, never the reverse) and bootstrap significance tests. The category weights are data-fitted; the win probability is an ensemble of the transparent model and a gradient-boosted model (see section 4) that, on this set, beat the transparent model alone on both accuracy and log-loss by a statistically significant margin. The result stays well-calibrated: when Vex AI says ~70% it wins ~69–81% of the time. It's sharpest on clear mismatches and honestly near coin-flip on genuine pick'em fights — because those really are. This is a deliberately broad, un-cherry-picked set (33 years, heavy with low-profile, upset-prone bouts), so ~65% winner accuracy is a realistic, conservative figure — the model does better on well-documented, higher-profile matchups.

*Held-out test: the model is tuned only on older fights and scored on the most recent ~30% it never saw. A bout is scored only when both fighters have ≥2 prior fights in the data. This historical backtest is separate from — and complements — the live, pre-registered accuracy record.

For any prediction we show each category's contribution (weight × skill gap mapped to probability points), the top drivers in plain language, and links back to the underlying stats and sources.