Introducing UFC / MMA Predictions + Value Bets

We added a new sport: UFC / MMA. Every upcoming fight now gets a model win probability, a method-of-victory breakdown (KO/TKO, submission, decision), and a 'goes the distance' estimate — plus a full tale-of-the-tape and our fighter Elo ratings, all free for everyone. And we are now placing model-driven value bets on MMA too. This post explains exactly how we bet it, because the *how* is the whole point of how we operate.

What we built

Most public MMA models stop at Elo — a single rating number per fighter. We went further. Our model reads the ufcstats box score for every fight in the modern UFC era: significant strikes landed and absorbed per minute, striking accuracy and defence, takedowns and takedown defence, control time, knockdowns, submission attempts. From those we build a striker-versus-grappler picture of each fighter, point-in-time — only ever using what was known before the fight, never after.

On top of that sit the physicals (reach, height, age, stance), finish tendencies, recent form, and a strength-of-schedule adjustment — because landing 45% of your strikes against elite defenders means more than landing 45% against journeymen. Three heads come out: who wins, how the fight ends, and whether it reaches the judges. Against a pure-Elo baseline the model is a clear improvement, and its win probabilities are honestly calibrated — when it says 60%, that side wins about 60% of the time.

The bug we found (and why we are telling you)

Building this surfaced a subtle, instructive bug — the kind we think is worth being open about. Our raw fight data lists the winner first. That is harmless for a results archive, but when you train a model on it, roughly two-thirds of your training fights have the eventual winner in the 'first fighter' slot. The model quietly learned that being listed first is worth about 22 percentage points — a phantom 'home advantage'.

The fix is called symmetric training: we teach the model every fight in both orderings, so it becomes completely order-independent — and we average both orderings at prediction time to guarantee it. That single correction removed the phantom edge and, as a bonus, fixed a mild over-confidence on close fights. It is a small reminder of a big principle: a model that looks accurate in testing can still be quietly wrong, and finding out requires looking for the artefact, not just admiring the accuracy number.

How we place MMA value bets

We place MMA value bets the same way we do for every sport: when our model's win% for a fighter is meaningfully higher than a bookmaker's price implies, that gap is value — and we take the best price at any book. If our model likes a fighter at 67% and a soft book is paying as though he is a 63% shot, that is a positive-expected-value bet, and we flag it.

With one guardrail we never drop: we anchor to the sharp market. A bet only fires when our model sits within a small margin of Pinnacle's de-vigged line. When the model disagrees *wildly* with Pinnacle, that is almost always an information gap — a camp change, an injury, weight-cut trouble — not an edge, so we defer to the sharp line and skip it. That keeps us betting where the model has a real, modest edge, not where it is guessing against people who know more than a box score can.

How we grade every bet

We grade ourselves in the open. Every MMA bet is logged against the closing line the moment it settles, win or lose, and published on our track record alongside every other sport. Closing-line value is the cleanest early signal that a bet was priced right — so we hold MMA to exactly the same standard as football, tennis and the rest, and we show the numbers as they accumulate, good or bad.

• Win probability — calibrated, order-symmetric, for every upcoming fight.
• Method of victory — KO/TKO, submission, or decision likelihoods.
• Rounds — the chance a fight reaches the judges versus an early finish.
• Tale of the tape + fighter Elo — reach, height, age, stance, record, and our rating.
• Value bets — model-driven, taken at the best price at any book, Pinnacle sanity-checked.

Use the win% as a research tool: a fast, calibrated read on any UFC fight. And when our model finds a real edge at a bookmaker, we flag the value bet — best price at any book, anchored to the sharp line, and settled in public.