🥋 OCTANE ALPHA | UFC Analytics

Base rating: blends Elo (built from recorded fight history, when it exists) with a stats-based rating from career win rate, finish rate, experience, and reach.

Style matchup adjustment ("styles make fights"): takedown accuracy vs. the opponent's takedown defense (or control time percentage, when populated — a fighter who goes 1-for-5 on takedowns but holds 4 minutes of control is more effective than raw accuracy implies), striking accuracy and volume differential (SLpM minus SApM, when populated), and durability all shift the base rating gap.

Age cliff by weight class: the decline threshold varies a lot by division — Flyweight/Bantamweight/Featherweight/Strawweight start seeing a real decline past 35, while Heavyweight and Light Heavyweight often sustain into the late 30s. A flat age penalty across all divisions would be wrong in both directions.

Layoff / ring rust vs. quick return: fighters back after more than a year away take a ring-rust penalty; fighters back too SOON after being finished (under ~6 months) take a separate, opposite-direction penalty.

Missed weight: a documented red flag for a rushed or broken camp — factored in when known, though real weigh-in history isn't fully populated for the roster yet (see limitations).

Method-of-victory props: prior-informed blend — starts at the DIVISIONAL average rate for that method (heavyweight fights have an inherently higher baseline KO/TKO rate than strawweight, which leans toward decisions), shifts toward the fighter's own career tendency (weighted by experience), then further incorporates the specific opponent's demonstrated vulnerability.

Total rounds: blends combined finish rate with first-round finish percentage. Main events get 5-round-calibrated lines (3.5/4.5) instead of 3-round lines (1.5/2.5), since there's simply more scheduled fight to cover.

Line movement: each refresh is compared against saved history to show real price movement and render an actual chart. Honest limitation: this tracks PRICE only, not bet volume or who's behind a move — true "sharp money" detection needs handle/ticket-count data no free source provides, so this is labeled "notable movement," not a sharp-money claim.

Suggested stake: tenth-Kelly, hard-capped at 5% of bankroll regardless of the math. In genuinely efficient sports betting markets, an edge large enough to produce a raw half-Kelly stake above ~10% is itself a warning sign — it usually means the model is overconfident somewhere (a method-of-victory prop resting on a small career sample, a stat that hasn't caught up to recent injury or camp news) rather than a real edge that big. The cap exists to protect against staking real money on the model's own miscalibration, not to suggest the edge itself is wrong.

Closing Line Value (CLV): the moneyline price on the model's favorite the FIRST time it was logged, compared against the last known price before the fight. If the market moved TOWARD the model's side by fight night (the price shortened), that's a "beat" — a real signal the model saw value before the market caught up, independent of whether the pick actually won. This is why CLV is tracked separately from raw accuracy: a model can lose a coinflip fight while still having correctly read where the market was headed, and that skill is what CLV isolates.

Honest data limitations: strike-volume (SLpM/SApM), control-time percentage, and missed-weight history are new schema fields the model is built to use, but they're not populated for the full roster yet — those specific adjustments quietly no-op (rather than guess) until real data backs them. Everything else above is live and active.