CAGEOne brain.
Every fight ever recorded.
Meet Cortex, the engine behind every Cagemath pick: how it works, from raw fight data to the number next to each name. No secrets about the method, only about the weights.
The stack, bottom up
Five layers run unattended every week. Each one exists to remove a human excuse: no vibes, no narratives, no deleting bad weekends.
The data engine
Automated crawlers pull every professional MMA fight in history from public records: over 200,000 fighters and their complete records, plus round by round official UFC statistics. Every significant strike, every takedown, every knockdown, every second of control time, along with historical betting closing lines and judge scorecards. The dataset refreshes itself twice a week, and new results flow in the morning after every event.
The rating system
Every fighter on earth gets a skill rating computed by replaying all of MMA history in chronological order. When a fighter wins, he takes rating points from the loser, and how many depends on who the loser was. Quality of opposition propagates through the whole network: beating a contender who beat a champion moves you more than padding a record against nobodies.
UFC results count more than regional ones, finishes count more than decisions, and ratings decay when a fighter sits out. It is the same family of math used to rank chess players, adapted for MMA's messier reality.
The prediction model
Machine learning on top of the ratings: a two stage model trained on the last six years of fights, because the sport changes too fast to trust older data. Stage one predicts how the fight ends, finish or decision. Stage two predicts who wins each of those scenarios, because the things that win decisions are not the things that win firefights.
It weighs about thirty signals per matchup: the ratings, momentum, age curves, official UFC output stats, and style collisions like a grappler meeting someone with bad takedown defense. Every Sunday it retrains on the newly grown dataset.
The honesty layer
This is what separates us from every tout service. The model is validated by walk forward testing: roll the clock back, train only on fights before a date, predict the cards that came after, and repeat across hundreds of bouts. No hindsight, no cherry picking. Across 897 bouts the model had never seen, it called 65.1% of winners.
The confidence numbers are calibrated. When we say 70%, we win about 70% of those picks, and we cap displayed confidence at what we can actually prove. Every pick is locked into a timestamped ledger before the card and graded in public, wins and losses alike.
The AI analyst (Cortex Sharp)
Before picks are published, a frontier AI model reviews each one with live web search, hunting for what statistics cannot see: injuries, late replacements, weight cut problems, camp turmoil. It can adjust or flip a pick only by naming a specific, sourced reason.
And it does not get a free pass. Cortex Sharp runs head to head against base Cortex on the public ledger, so the data decides whether it earns its keep. The picks page always shows which one is on the board.
The weekly cycle
The whole thing runs unattended, every week, in the open.
A self updating rating system over the entire history of MMA, a machine learned prediction model on top, and a public graded ledger that keeps us honest.
Does the number mean what it says?
Stated confidence versus what actually happened, across 897 fights the model predicted blind.
| When the model says | The pick actually wins |
|---|---|
| 50-59% | 56.9% |
| 60-69% | 68.8% |
| 70-79% | 76.6% |
Get the output
Every pick, every probability, every value flag, for $49/mo ($39/mo billed annually). The record stays public either way.
Get the picks