CAGEDEX

About

How the ratings work · How the predictor works · What's coming next

How ELO Ratings Work

Every fighter in the database starts at 1500 — a neutral baseline meaning "we have no information on you yet." From there, ratings move based on results. Beat a highly-rated opponent and you gain a lot. Grind out a split-decision over a low-rated fighter and you gain much less.

The core idea: your rating reflects the quality of your wins and losses, not just the number of them. A 20-fight veteran feasting on journeymen might be rated lower than a 7-fight prospect who's already beaten two ranked opponents.

[ Finish Type Matters ]

A knockout or submission is worth more than a unanimous decision, and a unanimous decision more than a split. Finishing a fight is a stronger signal of dominance. Fighters who consistently finish get more credit.

[ New Fighters Move Faster ]

Under 5 UFC fights = provisional. Each result moves the rating significantly more until the system has enough data. This lets it calibrate quickly on new talent rather than spending years slowly nudging a number.

[ Title Fights Carry Extra Weight ]

Championship bouts have bigger impact. The stakes are higher, the competition certified elite, and the system treats those results accordingly.

[ Inactivity Decay ]

A rating earned two years ago and undefended isn't worth as much as one earned last month. Fighters inactive for 12+ months see their rating drift back toward 1500. Not a punishment — an acknowledgment of uncertainty about what they currently are.

[ Confidence Scores ]

Every rating has a confidence score. High confidence = recently active with enough fights. Low confidence flags provisional fighters or those deep in a layoff. The number is still informative — treat it with appropriate skepticism.

How the Fight Predictor Works

A machine learning model trained on historical UFC fights. Given two fighters, it outputs a win probability for each — not a gut feeling or a narrative, just what the data says has historically happened when fighters with similar profiles match up.

[ Training Data ]

Trained on UFC fights through end of 2022, tested on 2023 onward — fights it had never seen before. That temporal separation is how you check whether a model learned something real versus memorizing history.

[ 37 Input Signals ]

Striking output (SLpM), striking absorption (SApM), opponent quality, Elo rating, recent form (last 5), finish rate, days inactive, age, reach, stance, and takedown stats. The Vegas opening line gets a seat at the table when available — the market is good, and ignoring it would waste signal.

[ Accuracy ]

On a holdout backtest of 790 fights from 2023 to present, the model picks the right winner 68.4% of the time. Against 370 fights with Vegas odds (Nov 2018 – Mar 2026): model 72.7% vs Vegas 70.3%. It called 49 of 110 upsets — Vegas called zero.

Real numbers, not cherry-picked. The model is solid. It's also not infallible — MMA is chaotic, and no model captures a fighter's chin, their corner, or whether they had a bad camp.

[ What the % Means ]

When Fighter A wins 67%, it means: historically, when fighters with profiles like these matched up, the Fighter A analog won ~⅔ of those fights. A statement about distributions — not a guarantee about Saturday night.

What's Coming

The site is in active development. Post-launch pipeline:

[ Expert Picks Layer ]

Surface picks from top MMA analysts alongside the model probability and Vegas line. Three signals at once: the data, the market, and the minds who watch this sport for a living.

[ Events Tab ]

Upcoming and recent fight cards as prediction panels — fighter vs. fighter, model probability, Vegas line side-by-side.

[ Rematch History Panel ]

When two fighters have history, it should be visible. A fight timeline that gives every rematch the context it deserves.

[ Time-Travel Predictions ]

The model knows every fighter's Elo and stats at any point in history. GSP vs. Khabib at their respective peaks. Fedor vs. Ngannou in 2009. It's a natural extension of the data.

Key Terms

ELO Rating
Starts at 1500. Reflects performance quality weighted by opponent level and finish type. Top P4P contenders typically sit 1650–1750+.
ELO Win Probability
Raw prediction from the Elo gap alone — one of 37 signals in the full model. Layering in striking stats meaningfully improves over Elo alone.
Z-Score (P4P)
Standard deviations above divisional mean. Fixes the cross-division comparison problem — raw Elo ratings can't be compared across weight classes.
Confidence Score
0.0–1.0. High = recently active, enough fights. Low = new, inactive, or provisional. Doesn't affect the math, just tells you how much to trust the number.
Recent Form
Win rate over last 5 fights. Recency matters — a fighter on a streak gets a bump regardless of career Elo.
Finish Rate
% of wins by KO/TKO/sub. High finish rate = more volatility. Factored into matchup dynamics.
Days Inactive
Days since last bout. Triggers inactivity decay in Elo and is a direct feature in the predictor. Ring rust is real.