Skip to content
CageFrame: AI fight predictions for UFC fans← Home

Last updated 2026-10-13

Sources & credits

Every fact on this site comes from a public source and carries the date it was last checked. Fighter stats and records are copied from the pages listed below, never estimated. They were collected with AI-assisted research, then re-fetched and checked by an independent second pass. A few notes (for example, that an average includes Contender Series bouts) are worked out from those pages' own totals. When a value isn't available or doesn't add up, the page shows a dash instead of guessing. Win probabilities are different: they are a statistical model's estimates (see Prediction model below), and where the model lacks an input it uses an average, which the fight page flags.

Prediction model

Win probabilities come from a small statistical model (model-v1@0f0fb55e), not from the AI. It rates every fighter with a Glicko-2 rating built from every UFC result on English Wikipedia since UFC 1 (1993), then a logistic regression weighs that rating with age, reach, time since the last UFC fight, UFC experience, current streak and the strength of recent opponents. Each percentage is the chance of winning if the fight has a winner: draws and no contests are left out. It uses no striking or grappling statistics (there is no licensed fight-by-fight history of them) and no betting odds.

Its inputs are fight results from English Wikipedia, dates of birth and heights from Wikidata (else Wikipedia) and reach from Wikipedia. For fighters on the upcoming cards, when Wikipedia and Wikidata have no date of birth or reach, or disagree with the hand-checked card value (UFC.com, ESPN and Sherdog cross-checks), the model uses the card value when it predicts, and the fight page says “Age from card data” or “Reach from card data” (15 of 21 published numbers). It was trained and tested on Wikipedia and Wikidata only. When an input is missing there and on the card too, the model uses the average of the fighter's division over the previous ten years and the fight page flags it (“Age estimated”, “Reach estimated”) ; none of the 21 published numbers needs one right now. Such numbers are less certain, and in the backtest a missing Wikipedia date of birth tended to flatter that fighter.

It was fitted on 2005–2014 fights, its settings were chosen on 2015–2020 fights, and it was then tested on 2,968 UFC fights with a winner from 2021-01-16 to 2026-10-10, none of which it was trained or tuned on. It picked the winner 60.0% of the time (95% interval 58.2%–61.8%), where a coin flip gets 50%. Its calibration slope there is 1.07 (95% interval 0.93–1.23), where 1 means its probabilities are neither too bold nor too cautious. That test period has been scored three times, so the pass is weaker evidence than a single look. Each scoring is logged: look 1 on 2026-10-13, look 2 on 2026-10-13, look 3 on 2026-10-13 (data corrections). The clean check is the ledger below: every number is frozen before its fight.

Before any number is published the model must pass these gates: leakage, isolation, symmetry and look-logging tests pass; test log loss < 0.693; upper end of its 95% CI < 0.693; 95% CI of the calibration slope contains 1 and 0.8 ≤ slope ≤ 1.25; ECE ≤ 0.03. It passed all of them. If it fails one, or a bout lacks the history for a defensible number (for example when both fighters debut), the bout shows why it has no number instead.

Every published number is frozen in an append-only ledger before the fight, with the time and a hash. It is never edited: a correction is a new entry that supersedes the old one, both stay in the ledger, and the fight page lists the earlier entry, when it was replaced and why, under “Sources and versions”. A data refresh that changes no number adds nothing. The site shows whole percentages that add up to 100; when they read 50/50, neither fighter is highlighted and there is no pick. Right now 21 of the 22 bouts carry a correction.

The written matchup analyses on fight pages are AI text. An AI (an LLM) writes them from an evidence packet of sourced facts and the model's output, and a deterministic validator checks them before publishing: every cited fact must exist, every number must match its source, and any probability must equal the model's. The AI never sets a number. The analyses published now were written by a Claude Code session agent from those evidence packets, fed through the pipeline's replay provider and checked by the same validator; they are not a run of the pipeline's configured API model, and each fight page says so.

Fight histories, the model's data and licenses

UFC results, fight histories, and the ratings and model factors computed from them come from English Wikipedia event and fighter pages (text by Wikipedia contributors), licensed under CC BY-SA 4.0. We parsed, filtered and aggregated them; each result on a fight page links the exact page revision it came from. The Wikipedia-derived material we republish (results, fight histories, ratings, model factors and the facts built from them) is shared under the same license. Dates of birth and heights come from Wikidata, released under CC0. The history behind the model holds 8,943 UFC fights from 1993-11-12 to 2026-10-10 (wikipedia@2026-10-12).

Fighter patterns

The patterns on a fighter's page and in the fight desk's game plans come from pose and strike labels derived from UFC broadcast footage, analysed by CageFrame: an automated pose tracker follows both fighters frame by frame, and the habits are counted from what it measures. It is a pilot: tracking can lose a fighter or mistake one for the other, so every pattern shows how many times it was seen and in how many fights. No footage is shown or stored on this site.

Our Wikipedia collector

CageVizDataBot/0.1 is this site's collector for Wikipedia and Wikidata. It uses only the MediaWiki and Wikidata APIs, one request at a time and at least a second apart, with maxlag=5 and at most 50 titles per request, and identifies itself with a User-Agent that points to this page. It never requests ufc.com, never bypasses bot protection and collects no betting odds.

Matchup breakdown rules

The “Key factor” card in the home page's phone preview is not the model's reasoning. It applies these fixed rules to UFC career averages:

Fight cards and events

Fighter data

Records, date of birth, height, reach and stance come from UFC.com athlete profiles and UFC's published event data, checked against Wikipedia, Sherdog and ESPN. Where a value (for example a stance) could only be confirmed on ESPN, the fighter's profile links to that ESPN page. Career averages (significant strikes landed and absorbed per minute, striking accuracy and defence, takedowns per 15 minutes, takedown defence, submission attempts) are the figures shown on each fighter's UFC.com profile, re-checked independently. For some fighters UFC.com's averages also include Dana White's Contender Series bouts; their profiles say so. We don't show takedown accuracy: UFC.com's published percentage doesn't reconcile with its own takedown counts (for example 5% for a fighter with about 119 UFC takedowns), and we don't patch one source with another. Recent form lists the last five professional results. Each fighter's profile on this site links to its exact sources. These card values are what a fight page's tale of the tape shows. The prediction model reads its own inputs (see Prediction model), and where its input differs from the tape the fight page says so next to the factor.

The full roster

The 4,196 athlete pages under /fighters are built from each athlete's public UFC.com profile, listed in UFC.com's sitemap and fetched at the pace its robots.txt asks for (one athlete, page and photo, then 15 seconds of pause), last checked Oct 13, 2026. Record, division, ranking, bio, wins by method, career averages and the latest fights are copied exactly as UFC.com shows them. Unlike the fighters on the next cards, they are not cross-checked against other sources. For a fighter on the next cards whose UFC.com profile shows a placeholder or contradicting record or bio value (new signees often show 0-0-0 or only their Contender Series bout), the athlete page shows the hand-checked card value instead and says so; when such a fighter has no UFC bout yet, UFC.com's stat block (from Contender Series bouts) is marked as not UFC averages. UFC.com's “Octagon debut” is not shown when its own page contradicts it (a date after the fetch, or after a UFC result it lists), nor for fighters on the next cards, where it was often a Contender Series or cancelled-bout date. Takedown accuracy is left out for the same reason as above, and averages are hidden for athletes UFC.com lists with no fight time. Each athlete page links to its UFC.com source. Photos on those pages are UFC.com's own athlete photos (© UFC, demo only, see below); UFC.com's generic shadow images are never shown as a photo.

Photographs

Photos credited to the UFC are the official athlete photos from each fighter's UFC.com profile, linked below. They are the property of the UFC (Zuffa, LLC), are used on this demo for illustration only and are not licensed for reuse. We cropped and framed them; we never mirror or recolour them.

Other fighter photos are freely licensed images from Wikimedia Commons. We removed the background, cropped them, converted them to black and white, and sometimes mirror them so two fighters face each other. Where a CC BY-SA photo is modified, the modified version is shared under the same license.

2 fighters have no usable photo yet and are shown as a generic dot-pattern silhouette, never with someone else's photo.

Type

Archivo, Geist and Geist Mono, all under the SIL Open Font License.

Independence

CageFrame is independent and is not affiliated with, endorsed by or sponsored by the UFC, Zuffa, TKO or any fighter. Names are used to identify the athletes and events discussed.