Back to the boards

REAL Fighter Rankings: Methodology, version 1.0

Results, Evidence, Analytics, Ledger. Data through 2026-10-31. This document is generated from the live configuration (config/weights.yaml) and the audit trail (outputs/audit_top30.csv) on every rebuild.

1. What the rank means

A resume rank: who has earned the position as of today, updated automatically each week. The single principle behind every rule: a fighter is ranked on how he performed against the fighters he faced, and how good those fighters were. The goal is the most defensible board possible, not agreement with any other board. It is not a prediction of who would win tomorrow; that is a separate model (section 8, and the Prediction page). Champions (C), interim champions (IC) and reserved spots (R: former champions who vacated with an injury and are owed a title shot) sit above the numbered board with their metrics shown. Contenders are numbered 1 to 30.

2. Data and how far to trust it

Known gaps: no pre-UFC records (debutants start at the same rating), no injury data beyond the hand-kept list, no contract or matchmaking information, and statistics before about 2001 are sparse.

3. How A + B = C: the pipeline, in order

1. Resume rating. Every UFC bout since 1994 is processed in date order. Each fighter's rating moves by K x (actual score - expected score), where the expected score comes from the two ratings (Elo). Rules for scoring a result are in section 4.
2. Eligibility. A fighter must have fought within the active window, be on the roster, and be in a ranked division (section 6).
3. Quality ledger. Wins and losses are re-read for resume value (section 5).
4. Score. Score = rating weight x scaled rating + ledger weight x scaled ledger - form penalty. Both inputs are scaled within the division by their interdecile range: (value - median) / (90th percentile - 10th percentile).
5. Head-to-head. The winner of the latest meeting moves above the loser when close enough (section 7).
6. Title cycle. Recent title-fight challengers who lost step back from the top slots (section 7).
7. Weight-sensitivity band. The whole pipeline (score order, head-to-head, title cycle) is rerun with weight vectors drawn near the configured weights; the band is each fighter's 10th to 90th percentile final position. It is a sensitivity interval for the weight choice, not skill uncertainty (see section 10 for the Glicko-style deviation planned for v1.1).
8. Audit. Every contender's placement records where the score put him and which rule moved him.

4. Scoring a single fight (the rating)

K = 60, and 1.5 x K for a fighter's first 5 UFC bouts. There is no rating decay for time off; inactivity is handled in section 6.

Ranks used by the rules below (top 3, top 5, quality-win tiers) are the official ranks on the fight date from 2013 on, and the model's own division position before that.

Other fight-level rules. A no-contest caused by a failed drug test counts as a loss for the fighter who failed (read from the official bout details). A bout taken on short notice (about 3 weeks or less, config/short_notice.yaml) counts 1.2x for a win and 0.5x for a loss.

Loss rules. A loss is dominant when it is a first-round finish, an early finish (first half of the scheduled rounds) while clearly behind on the stats, or a decision with 2 of 3 cards at 3+ points where the stats do not contradict the cards. A late finish while close on the stats (winner ahead by 1.5 or less per minute), a split or majority decision, or a decision scored 0.7 or less is close.

5. The quality ledger

A quality win is valued by the opponent's official rank going into the fight: champion 2.0, ranked 1-5 1.5, 6-10 1.0, 11-15 0.5. An unranked opponent with 8+ UFC wins and a winning UFC record is worth 0.25. Beating four top-10 fighters is worth more than beating eight fighters ranked 11-15. Every ledger item is weighted by its age:

Entrenched (shown as E) = 4+ wins over top-10 opponents, or 5+ over top-15 opponents, in 15 years. Activity alone earns nothing: a fighter who takes many fights and loses to non-elite opponents gives the ledger back.

6. Eligibility, form and division

• Inactivity. No penalty for the first 365 days; up to 50 rating points by 540 days; off the board after that. Documented injury layoffs (config/layoffs.yaml) carry no penalty and stay eligible up to 730 days.
• Form (last five UFC bouts). Subtracted from the score: 2-3: 0.15, 1-4: 0.6, 0-5: 0.8. A 2-3 is a warning; 1-4 counts seriously against a fighter. A fighter with a negative last five also gets no head-to-head lift.
• Division. Two straight bouts in a division settle it. Otherwise the division fought in most over the last 3 years, with ties going to the division of the most recent win. Title holders are ranked in their title's division. Announced moves are in config/division_overrides.yaml, each with its reason.
• Roster. Retirements and releases are removed (config/roster_exclusions.yaml).

7. Matchmaking reality rules

• Head-to-head. If a fighter beat someone in their most recent meeting within 3 years and sits no more than 3 places below him (6 if the fight was in the last 12 months), he moves directly above him.
• Title cycle. A challenger who lost a title fight in the last 270 days is placed no higher than #4: still close, but the champion is fighting someone else next. He earns his way back with 1 win over a top-10 opponent or 2 wins of any kind. A champion who lost the belt is exempt (immediate rematches are common). Anyone that challenger beat in the last year stays below him.
• Head-to-head details. A close win (split or majority decision, or a fight scored as close) more than 12 months old settles nothing, and a fighter with a negative last five gets no head-to-head lift.
• Reserved spots. Former champions who vacated because of injury are listed as R above the numbered board (config/reserved.yaml).

8. Validation

Agreement with the public boards (UFC contenders only, champions removed): our average gap is 1.8 to 2.1 places; the public boards differ from each other by 0.9 to 1.8. Disagreement is expected and reported, not removed: outputs/compare_flags.csv lists every large gap with its cause.

The separate predictive model (performance-adjusted Elo, the specification a 520-point grid selected on 2010-2019) scores 60.6% accuracy and 0.6631 log loss on 3,390 held-out bouts from 2020 on, against 58.1% and 0.674 for results-only Elo (paired log-loss gain 0.011, 95% event-block bootstrap interval 0.006 to 0.016; McNemar exact p = 0.0025). On the 1,445 held-out bouts with closing odds (2020-2023) the de-vigged market scored 67.1% and 0.6088 against the model's 60.0% and 0.669.

Retrospective reconstruction of the resume board (not a forward test). Boards were rebuilt with the v1.0 rules as they would have stood before each of the last 132 events. Across the 940 bouts in which both fighters held a place on that board, the higher-placed fighter won 61.5% (95% Wilson interval 58% to 64%); the frozen score-to-probability map scored 0.6671 log loss against 0.6729 for results-only Elo and 0.6502 for the performance-adjusted model on the same bouts. Restricted to bouts between two top-15 fighters (226 bouts) the figure is 54.4%, against 51.7% for the official board on the 205 bouts it ranked both fighters; ranked-versus-ranked bouts are matched to be close, so every board sits near a coin flip on them and the differences are inside sampling error. Because these boards were reconstructed with today's rules, none of this is evidence of forward validity. The pre-registered prospective test (PREREGISTRATION.md) starts with the first event after 2026-10-01; so far it covers 37 bouts.

9. Worked examples (from this rebuild's audit trail)

10. How REAL compares with published rating systems

11. Why these numbers: what each constant is for and what it costs

Every constant in the engine encodes a stated ranking principle (what a resume should reward or forgive); none was fitted to outcomes. The table shows what each principle costs or buys when the pre-fight resume rating is scored as a forecaster on the same 3,390 held-out bouts (2020-2026) the prediction model is graded on, changing one constant at a time from the v1.0 values (outputs/constants_sensitivity.csv, rebuilt by python -m mmalab.sensitivity). A positive change in log loss means the alternative predicts worse than v1.0; a negative one means it predicts better. The resume board is not graded on prediction (its test is PREREGISTRATION.md), so a small predictive cost is the accepted price of a principle, but the reader can see the price.

v1.0 reference: log loss 0.6757, accuracy 57.9% on 3,390 bouts (results-only Elo 0.674, performance-adjusted Elo 0.663 on the same bouts).

Reading the table: the whole set of loss protections costs about 0.003 log loss out of sample, and no single principle costs more than 0.0015, so the resume rules are cheap in predictive terms. The 50/50 blend of judges' cards and fight statistics is the best of the five blends tried, which supports the Jones-versus-Reyes argument with data. The two constants the data would push are K and the newcomer multiplier (faster ratings predict slightly better); the board keeps them slower on purpose, so that a single fight moves a resume less than it moves a forecast. The board-level constants (70/30 weights, horizons, ledger values, form penalty, title-cycle and head-to-head thresholds) cannot be scored this way because they act on the board, not the rating; the weight band covers the 70/30 choice and the rest are fitted or ablated under the v1.1 plan in section 10.

12. Limits and open decisions

• Pre-UFC records and betting odds after 2023 are not yet loaded (they require a manual Kaggle download); when added, pre-UFC records will set starting ratings and judge debut opponents' quality, never award ranking credit, and odds will define 'heavy favorite' instead of the model's own probability.
• Missed weight is not recorded in the fight data and is not yet used.
• Division overrides, injuries and retirements are hand-kept and need weekly review.
• Weights and thresholds were set by stated judgment and checked against the public boards (which makes those boards an informal tuning target); section 11 reports what each engine constant costs out of sample, and none has been fitted to any outcome.
• Card-quality and matchmaking analyses use the predictive model's positions, not these boards.
• Age. The board carries no age term, on purpose. Measured on 3,748 bouts between two board fighters from 2014 to 2026, the higher-ranked fighter won 45% of bouts when aged 35 or older against 58 to 63% when younger, in every period and division, and a rating discount by age would raise concordance by three to four points (experiments/REGISTER.md, EXP-2026-10-20-09). It was rejected for the board because a resume rank lags ability by construction: a contender earns the higher spot by beating the man who holds it, and an older top-10 fighter loses more often because he is matched against the best. Predictive concordance is therefore reported as a descriptive check, never used as the objective (PREREGISTRATION.md section 8). Age belongs in the prediction model; what the rating misses by age, and each fighter's own trajectory, is measured and shown beside the resume rank, never inside it (section 14).
• Resume-board tests. Because prediction is not the objective, the board's own tests are published every week in outputs/board_quality.json: top-15 turnover, moves of more than three places without a bout, head-to-head results the placement rules leave unresolved (with the reason each was left), published ranks outside their own weight band, roster facts added in the last seven days, and agreement with each public board. None of them is a tuning target.

13. Shadow versions under test

A new board version is never substituted for the registered one during the pre-registered period (PREREGISTRATION.md section 9; GOVERNANCE.md). It is built from its own override file over the frozen configuration, published beside the registered board under its own label, snapshotted to its own directory and graded on the same bouts. The versions currently under test:

• v1.1 (shadow, since 2026-10-20; config/board_v1_1.yaml; page docs/board_v1_1.html; adoption rule in PREREGISTRATION.md appendix C; evidence EXP-2026-10-20-01, EXP-2026-10-20-03, EXP-2026-10-20-08, EXP-2026-10-20-11 in experiments/REGISTER.md). Overrides over the frozen configuration:
• engine: {"protected_loss_mult": 0.3}
• ledger: {"loss_tiers": [[5, 0.25], [10, 0.5], [15, 0.75], [999, 1.0]], "dominant_extra": 0.0}
• form_penalty: {"2-3": 0.1, "1-4": 0.3, "0-5": 0.4}
• form_bad_loss: {"rank_above": 10, "per_loss": 0.1, "cap": 0.3}
• placement: "solver"
• record: prospective 18 bouts, concordance 0.556; retrospective reconstruction (in-sample) 977 bouts, concordance 0.606.

14. Age, decline and rise: what the rating misses, shown beside the rank

Section 12 records why no board carries an age term: a resume rank is earned against the fighters a man was matched with, and an older contender loses more often because he is matched with the best. The question the ranking cannot answer, and that readers ask first about an ageing champion or a rising prospect, is how far the rating now sits from the fighter's current level. The decline and rise index answers it beside the rank (src/mmalab/decline.py, settings in config/decline.yaml, shown in the Trajectory section of every profile page). No board reads it.

Outcome. For every fighter and decided bout, the residual is the engine's performance score minus the pre-fight expectation, r = s - e, where s is the score the resume engine already uses (finishes 1 or 0, decisions a blend of the judges' cards and the fight statistics, section 3) and e comes from the two resume ratings plus a corner credit fitted on the training rows (the first-listed, red-corner fighter is the matchmakers' favourite and wins about 64% of bouts against 53% expected by rating alone; without the credit a fighter who is usually the favourite would look like a riser). The rating update is K (s - e), so a run of same-signed residuals is exactly the drift the rating has not absorbed. One unit of residual sustained over many bouts corresponds to about 695 rating points (the slope of the expectation curve at e = 0.5).

Model. Three parts, every one computed from information available before the bout. A population curve f(age), piecewise linear with knots at 27, 30, 33 and 36, one per sex, fitted on bouts from 2011: the average residual at each age, which is the rating's own age bias. An individual level: the fighter's recency-weighted mean residual around that curve over his earlier bouts, shrunk toward zero by a pseudo-count of bouts. An individual slope: the recency-weighted trend of those residuals against age, shrunk toward zero. The prediction for a fighter is f(age) + level + slope x (age - mean prior age). The three shrinkage settings (level 32, slope 200, half-life 4.0 years) were chosen walk-forward: for each test year from the held-out performance of the years already tested, never from the test year, and they settled on the same values from 2019 on.

The population curve now (fitted on 14,644 fighter-bouts through 2026-11-01, in residual units with the rating-point equivalent):

Read: a 22-year-old on the board typically performs as if his rating were about 61 points higher than it is; a 36-year-old as if it were 53 points lower, a 40-year-old 102 points lower. The women's curve is flatter until the early thirties and then falls at the same rate.

Held-out evidence (walk-forward by calendar year, test years 2016 to 2026, 10,702 fighter-bouts; experiments/REGISTER.md EXP-2026-10-21-01). The population curve explained 2.2% of the variance of the next bout's residual out of sample, positive in every test year; the realized residual by age bucket tracked the curve that predicted it:

The individual terms are a weak, consistent signal: held-out correlation 0.06 between the individual prediction and the realized residual (calibration slope 0.62, where 1 is right), and among the 2,888 held-out bouts of fighters aged 33 or over, the lowest predicted quartile won 38% against 53% expected (-15 per 100), the highest 48% against 52% (-4), all of them -9. Added to the prediction model's shadow feature set they change log loss by -0.0003 (95% event-block interval -0.0014 to 0.0008): nothing. Against the faster prediction Elo (K 130) the individual terms carry no signal at all; the resume rating is slow by design, so it is the one with something left to read. On the tuning panel of 335 boards, converting the population curve to rating points raises board concordance by 1.4 to 2.0 points with no free parameter, the individual terms add nothing on top, and neither is applied to any board (section 12).

Who declines faster (men's bouts from 2011, descriptive, EXP-2026-10-21-02). At the same age, fighters who debuted in the UFC before 25 run further behind their rating from 33 on than those who debuted at 29 or later (about 0.07 of residual, 50 rating points, interval excluding zero at 33, 35 and 37); the lighter divisions decline harder than middleweight and above from 35 on; a knockout loss in the previous three years marks a further 0.05 of lag at 35 to 37; and strikers and wrestlers decline at the same rate, so the common belief that wrestlers age better finds no support here. Fitting the curve within those groups did not beat the pooled curve out of sample, so the index uses one curve per sex and leaves the group differences as findings.

What is published. outputs/decline_index.csv (every fighter with three decided UFC bouts and a bout in the last three years: age component, own-results component, combined, in residual units and rating points, with a plain reading) and outputs/decline_curve.json, rebuilt by the weekly pipeline from the frozen settings. The research runs that chose the settings and measured the evidence are EXP-2026-10-21-01 and EXP-2026-10-21-02 in the register.

15. Changelog

Every change to a ranking rule, a data correction that moves a published number, or a change in what the site displays is logged here with its date, its reason and the evidence behind it. The rules for making a change are in GOVERNANCE.md. Versions follow major.minor: a minor version changes a rule; a patch (1.0.1) changes display, documentation or data handling without changing any rule. The registered method under the pre-registered forward test is always VERSION.

1.0.2 (2026-10-21, display and documentation only)

No rule changes; applies to the site as a whole. Fighter profile pages: every name on the registered and shadow boards opens docs/fighters/<slug>.html (why the fighter is ranked there, built from the score decomposition and placement string; last ten bouts with the opponent's official rank at the time and the predictive model's pre-fight probability; career statistics with division percentiles; board position since Oct 2023 and predictive rating by bout; descriptive context flags), with an index at docs/fighters/index.html. Decline and rise index (src/mmalab/decline.py, config/decline.yaml, METHODOLOGY.md section 14): the population age curve of the rating residual and each fighter's own shrunken level and trend, published as outputs/decline_index.csv and outputs/decline_curve.json and shown in a Trajectory section on every profile page with its held-out evidence beside it (EXP-2026-10-21-01, EXP-2026-10-21-02). No board reads any of it. Pipeline: two steps appended (decline, fighter_pages); the generated pages are committed with the site.

1.1 (shadow board, 2026-10-20)

Status: candidate, published beside the registered v1.0 board at docs/board_v1_1.html, never substituted for it during the registered period (PREREGISTRATION.md section 9 and appendix C). Configuration: config/board_v1_1.yaml, merged over the frozen config/weights.yaml at build time; the frozen file's hash is unchanged.

Reason (audit finding 1): one loss entered three channels with opponent-blind costs, so a losing run against champions and top contenders was punished exactly like one against mid-card opposition. Joaquin Buckley fell from 9 to 21 at welterweight on a 2-3, Israel Adesanya sat 27th at middleweight, Jan Blachowicz 27th at light heavyweight, Marlon Vera 54th at bantamweight.

Rule changes: opponent-aware versions of existing v1.0 rules, and the placement rules solved as one order (added later on Oct 20, before the first snapshot):

Evidence: 43 event-eve boards (979 bouts, Oct 2023 to Oct 17, 2026, rebuilt on the corrected data with every placement rule applied) score concordance 0.606 for v1.1 against 0.607 for v1.0 and log loss 0.665 against 0.669 (paired differences -0.001, 95% interval -0.019 to +0.016, and -0.0045, interval -0.0093 to +0.0001); top-15 bouts 0.549 against 0.537 (EXP-2026-10-20-12). On the 14-day tuning panel (335 boards, 3,748 bouts, 2014 to 2026, score order without placement rules) the same ledger and form rules scored 0.580 concordance and 0.6748 log loss against 0.573 and 0.6772 for v1.0 (EXP-2026-10-20-03). The change is adopted for face validity at equal measured accuracy.

Effect on the Oct 17, 2026 board: 111 of 194 top-15 slots differ in holder or position. Largest moves: Vera 54 to 14, Ortega 40 to 15, Blachowicz 27 to 13, Adesanya 27 to 14, Buckley 21 to 15, Hill 20 to 14; Jim Miller 11 to 19, Jalin Turner 15 to 23, Neil Magny 15 to 22, Bo Nickal 14 to 20. The placement solver puts Topuria first at lightweight (the sequential loops left him third, below the man he knocked out in June 2025). Kamaru Usman moves from 9 to 5 at welterweight: his rating (1,807) is the division's highest below the champion because his four losses in five were to Edwards (twice), Chimaev and Du Plessis, which the rating protects by design; the curator's own view is that 7 is closer to right, and that gap is an open item, not a reason to hand-place him (EXP-2026-10-20-02).

Also in this release, no rule change: provisional label on every board for fighters with three or fewer UFC bouts (finding D3); version, build commit and data date in every page footer (G4); the picks ledger records no pick when the model is within two points of 50% (P2, appendix D) and carries a feed schema check and a cumulative-record chart (E6, G5); pytest suite and pinned dependencies (E1, E2); dated, sourced override files read as of the board date (D2); official-rank name aliases (D1).

Data corrections (logged as facts, not rules)

• 2026-10-20: same-name fighters split (finding D8). UFCStats gives nine display names to two fighter pages each, and the bout files carry only the name, so ingest merged the flyweight Bruno "Bulldog" Silva and the middleweight Bruno "Blindado" Silva into one 23-bout "Bruno Silva". config/fighter_disambiguation.yaml now assigns every bout of a shared name to one fighter page by division (or by explicit bout URL) and names the smaller career with a suffix; ingest stops on any shared name or bout the file does not settle. Effect on the Oct 17, 2026 data: the 11 middleweight bouts become "Bruno Silva (Blindado)" (Middleweight 79, 0-5, inactivity penalty); the flyweight "Bruno Silva" keeps 12 bouts, rating 1,480 to 1,550, ledger -7.05 to -3.20, Flyweight 49 to 47. Direct opponents move by at most 10 rating points (HyunSung Park +9.7, 30 to 28 at flyweight; Naurdiev -8.0, 21 to 23, and Shara Magomedov -3.5, 17 to 20, at middleweight); 13 top-30 placements shift by one to three spots, no top-15 holder changes. fighters.csv gains the second page of each shared name, so Jean Silva, Michael McDonald, Mike Davis, Anthony Romero and Victor Valenzuela now carry their own height, reach and date of birth instead of their namesake's.
• 2026-10-20: eleven official-ranking names that never matched UFCStats spellings now match through config/official_rank_aliases.yaml (Cris Cyborg = Cristiane Justino, Jacare Souza, Costas Philippou = Constantinos Philippou, Tim Johnson = Timothy Johnson, Ulka Sasaki = Yuta Sasaki, Antonio Rogerio Nogueira = Rogerio Nogueira, Francisco Rivera Jr., Luis Henrique Barbosa = Luis Henrique, Magomed Bibulatov, Mara Borella, Bruno Gustavo da Silva = Bruno Silva); one ranked name with no UFC bouts is allow-listed. Effect: 52 bouts change the opponent rank recorded at fight time, 23 change their quality-win flag, one final rating moves more than 10 points (Francis Carmont, +16.8). Under PREREGISTRATION.md section 1 the snapshots already published stand; boards from this date on use the corrected ranks.

1.0.1 (2026-10-01, display and documentation only)

Weight band recomputed after the head-to-head and title-cycle rules so every published rank lies inside its own band; prediction page and card-quality positions switched to the specification the grid selected on 2010-2019; paired bootstrap and McNemar tests, the market comparison on the held-out overlap, the constants sensitivity table (METHODOLOGY.md section 11), PREREGISTRATION.md and DATA_LICENSE.md added; favorites page moved out of the research navigation. 2026-10-05: pre-event picks ledger (appendix A) and the shadow v1.1 prediction model (appendix B). 2026-10-15: AI-assistance disclosure in the README.

1.0 (2026-10-01)

Method frozen for forward grading. Official rank at fight time (2013 on); tiered quality wins; proof-of-concept credit for close losses to top-5 fighters; early-finish and war rules; drug-test overturns count as losses; short-notice credit; last-five form penalty; division, head-to-head, title-cycle and reserved-spot rules; audit trail; prediction model published separately.

0.x (2026-09-29 to 2026-09-30)

Composite of six percentile dimensions, replaced after review because four dimensions carried no ranking signal and losses were counted three times.

Independent fan and research project. Not affiliated with, sponsored or endorsed by UFC, Zuffa, LLC, TKO Group Holdings, or any athlete.

REAL v1.0 (registered method) · build c166ee8 · data through 2026-10-31 · page built 2026-11-06 · changelog. Independent fan research project built on public UFCStats data. Not affiliated with, sponsored or endorsed by UFC, Zuffa, LLC, TKO Group Holdings, or any athlete. All trademarks belong to their owners and are used only to identify athletes and events.