H&G All-Time Index v2.0.0 methodology
How the H&G All-Time Index ranks boxers
The H&G All-Time Index is a career ranking for men's professional boxing. It asks a simple question: based on the record we have, who built the strongest all-time career?
It is not a fantasy-fight predictor and it is not a live record database. The current release was computed on 22 May 2026 from fights up to 10 May 2026.
The Index is the 0-100 number that orders the list. Peak-form Elo is separate: it shows a fighter's best rating level on the rating scale, not their all-time career score.
Active fighters are scored on the career they have completed so far. A fighter still building a resume can rate brilliantly on Peak-form Elo and still sit below where they might finish on the all-time career list.
The Short Version
Results matter first
The rating side comes from the pattern of wins, losses and draws over time. It rewards fighters who reached a high level and stayed there.
Resume still matters
The Index also credits opponent quality, title evidence, class control and sustained top-level career work.
Margins matter
Close ranks should be read as close ranks. Data Confidence and Peak-form Elo bands tell you where to leave more room for uncertainty.
What Goes Into The Score
The Index blends six career dimensions. The weights are fixed for this release and every dimension is put onto a comparable field-relative scale before weighting.
| What readers see | Weight | Plain meaning |
|---|---|---|
| Sustained head-to-head rating | 26% | How high the fighter rated from fight results, and whether that level held up beyond a single hot spell. |
| Quality of opponents beaten | 21% | Credit for wins over title-level opponents using title context, not by looking back at the H&G rank itself. |
| Major title-control signal | 14% | A focused signal for strong control of a weight class, including lineal, undisputed and major multi-belt evidence where the release data supports it. |
| Sustained top-level career | 14% | How long the fighter stayed near the top, capped so simply fighting for longer cannot swamp everything else. |
| Dominance over contemporaries | 13% | How clearly a fighter stood above nearby rivals after the main rating level is already accounted for. |
| Major title evidence | 12% | Major and lineal title context. It helps the career case, but it does not decide the list by itself. |
How To Read A Fighter Page
The expert prior is anchored to a single published source: the International Boxing Research Organization (IBRO) 2019 all-time pound-for-pound ranking. It is one organisation's published list, not an aggregate or averaged consensus of several rankings. The prior acts as a documentation-bias correction for early-era fighters whose surviving records are too sparse for the model to rate on evidence alone. Its weight is small (an equivalent-evidence setting of 4), and only 26 historical anchors carry any weight. Modern fighters are rated on data alone, with no modern anchors. It is not a hand edit, and it does not pick a number one.
Women's boxing is not folded into this release because the historical coverage is materially different. It needs a dedicated model and validation pass rather than an incomplete appendage to the men's list.
Validation Checks
The validation table is a disclosure tool. Passing a check does not make every placement unarguable; failing or incomplete checks show where readers should be more cautious.
Rank movement from the old additive model is deliberately disclosed. The common-scale model line intentionally broke from the older order because the previous model had overlapping rating signals, a scaling defect between dimensions and an opaque additive calibration. v2.0.0 keeps that order: it is rank- and score-neutral against v1.2.3, so rank stability now runs unbroken from v1.2.3 through this release.
With-prior and without-prior top lists are near-identical by design. The prior is there to handle uncertainty at the margins, not to re-order the top of the list by editorial preference.
Some positions diverge from published lists. The ranking is a fitted statistical model, lightly shrunk toward the IBRO 2019 anchor only for sparse early-era records. Where a placement differs from a hand-made list, it reflects the fighter's record as the model reads it. That divergence is intended.
| Check | Status | What it tells you |
|---|---|---|
| Weight sensitivity | Pass | Checks whether a small change to one weight would throw the top of the list around. |
| Leave-one-part-out check | Pass | Shows which fighters depend heavily on one part of the model. |
| Expert prior sensitivity | Pass | Compares the top 25 with and without the expert prior. |
| External-list comparison | Incomplete source coverage | Compares the H&G top 100 with external all-time lists where mapped source data is available. |
| Scale stability | Pass | Checks that the public 0-100 display scale is pinned to fixed anchors. |
| Rank movement from the old model | Needs review | Records where the common-scale model line intentionally changed the older additive order after correcting known model defects. v2.0.0 keeps that order: it is rank-neutral against v1.2.3. |
Known Limitations
Early eras carry wider margins. Some pre-WW1 fighters sit on longer opponent chains and thinner rating pools. v2.0.0 rates them conservatively rather than adding an era-strength correction that would require an editorial judgement about how much of the era gap is real history versus data structure.
Resume-heavy early-era cases need wider reading margins. Titles, class control and longevity are real boxing achievements, but they are harder to compare cleanly across shallow and deep eras. Profile pages should be read with that limitation in mind where the career case leans heavily on those signals.
The external-list comparison is incomplete. IBRO 2019 is mapped locally; Ring, TBRB and ESPN are listed as intended comparators but are not fully wired into the checked-in validation data for this release.
Title and reign facts on profile pages use a separate enrichment layer. The original fight corpus did not capture title context reliably enough for public title-fact claims, so profile highlights use a conservative Wikipedia champion-reign table where identity joins are safe.
For Data Scientists: Reproducibility Appendix
This appendix documents the shipped H&G All-Time Index v2.0.0 public model and its public artifacts. v2.0.0 is a major release: the order-independent bout merge (#47) makes the rating fit fully deterministic, and it folds in the corrections accumulated since v1.2.3. The published rankings stay score- and rank-neutral against v1.2.3 - the new determinism did not move the order.
For the full methods record, the determinism proof and the disclosed trade-offs, read the H&G All-Time Index v2.0.0 dissertation, or the plain-language companion.
Release Inputs
| Model version | v2.0.0 |
|---|---|
| Computed | 22 May 2026 |
| Fight cutoff | 10 May 2026 |
| Corpus range | 1882-01-01 to 2026-05-10 |
| Parsed fighter-page JSON files | 49,788 |
| Rating-eligible bout rows | 459,906 |
| Unique fighters in bout rows | 49,405 |
| Corpus fingerprint | aa24d14c0c14c8dfde4f24ad2c40664e39ca8c0ee43877a7ece95530fbd7330b |
Model Formula
model_score = (six_dimension_score * (1 - expert_prior_weight)) + (expert_prior_mean_score * expert_prior_weight); if no mapped prior exists, expert_prior_weight = 0
fixed linear 0-100 display scale: all_time_hgr_elo_index = round(60.00 + (model_score + 1.45839578) * 9.61599750, 2), clamped to 0-100; 100.00 means the best career on record
The public ordering uses canonical rank or full-precision model score, never the rounded two-decimal Index display value.
| Dimension key | Weight | Fan label |
|---|---|---|
| career_rating_profile | 26% | Sustained head-to-head rating |
| external_elite_wins | 21% | Quality of opponents beaten |
| undisputed_class_count | 14% | Major title-control signal |
| longevity_capped | 14% | Sustained top-level career |
| era_separation_residual | 13% | Dominance over contemporaries |
| lineal_titles | 12% | Major title evidence |
Rating Engine Inputs
Peak-form Elo comes from a Whole-History Rating fit over one global fight graph. The stage-one runner uses whr.whole_history_rating.Base, w2 = 14.0, 50 Newton iterations, a 1500 rating origin and no applied decade offset in the shipped rating layer.
The career rating profile dimension is built from the preserved rating features peak_WHR_3yr, sustained_WHR_5yr and era_dominance_pct, collapsed into one rating-profile signal and then standardised with the other public dimensions before weighting.
Peak-form uncertainty uses the WHR posterior standard deviation at the fighter's peak rating point. Public bands are approximate two-sigma bands on the Peak-form Elo scale, not on the 0-100 Index.
Public rating evolution curves use yearly Peak-form Elo points. They are not fight-by-fight ledgers.
Identity Resolution
Public rows are resolved to the preserved feature rows by normalized identity plus public slug guards where needed. The v2.0.0 pipeline includes explicit guards for Sugar Ray Robinson, Canelo Alvarez and the Roberto Duran public row, plus active-year display overrides for known historical date edge cases.
Public exports omit source IDs, private identifiers, source URLs and complete bout ledgers. The public feature table exposes reduced per-fighter fields: ranks, public slugs, Index, Peak-form Elo, posterior bands, Data Confidence, six weighted contributions and the expert-prior decomposition.
Elite-Opponent Definition
The shipped external_elite_wins signal is title-context based and independent of H&G ranks. It credits wins against opponents in major world or lineal title context at the fight date, or within three years of that opponent's title-context row.
Major markers are WBA, WBC, IBF and WBO world context. Lineal and pre-1962 non-vacant world title context receive full credit; vacant title context receives 0.5 credit. Interim, eliminator, minor, regional, amateur and women's title strings are excluded from this men's model signal.
The raw signal is log1p(title-context elite win credit), z-scored across the preserved feature rows and then common-scaled with the other public dimensions.
Expert Prior
The current prior input contains 26 mapped IBRO 2019 rows plus 6 modern-supplement rows. Ring, TBRB and ESPN lists are validation comparators and intended future prior inputs, but they are not blended into this release's prior.
prior_weight = 4 / (4 + effective_evidence_quality)
career_bout_count * era_evidence_multiplier(first_year) * active_career_multiplier(last_year) * bounded_curve_multiplier(rating_years) * bounded_posterior_multiplier(peak_sigma_elo)
The prior has 26 non-zero public rows. It shrinks lower-effective-evidence mapped careers more than deep, well-documented careers; fighters without a mapped prior have prior weight 0.
With prior
- 1. Naoya Inoue
- 2. Terence Crawford
- 3. Henry Armstrong
- 4. Joe Louis
- 5. Canelo Álvarez
- 6. Sugar Ray Robinson
- 7. Floyd Mayweather Jr
- 8. Roy Jones Jr
- 9. Muhammad Ali
- 10. Oleksandr Usyk
Without prior
- 1. Naoya Inoue
- 2. Terence Crawford
- 3. Henry Armstrong
- 4. Joe Louis
- 5. Canelo Álvarez
- 6. Sugar Ray Robinson
- 7. Floyd Mayweather Jr
- 8. Roy Jones Jr
- 9. Muhammad Ali
- 10. Oleksandr Usyk
Data Confidence And Bands
Data Confidence blends career bout depth, rating-year coverage, the external-elite opposition signal and the Peak-form Elo posterior band. It is a label for reading close ranks, not a posterior probability and not a ranking input.
Current distribution: High 312, Medium 590, Low 98 across the 1,000 ranked fighters.
The WHR posterior source reports 1,000 fighter bands. One-sigma Elo distribution: min 58.94, median 102.03, p90 142.28, max 278.85.
Validation Report
| Check | Status | Threshold or disclosure |
|---|---|---|
| Weight sensitivity | Pass | No top-25 fighter moves more than 10 places under any single +/-25% feature-weight perturbation. |
| Leave-one-part-out check | Pass | Every fighter moving more than 50 places is flagged for review. |
| Expert prior sensitivity | Pass | Standing disclosure, not a pass/fail gate. |
| External-list comparison | Incomplete source coverage | Spearman >= 0.7 against at least three of four named external lists. |
| Scale stability | Pass | Disclosure check; fixed anchors are retained for release stability. |
| Rank movement from the old model | Needs review | Spearman >= 0.97 unless movement is waived and documented. |
Data-Quality Disclosures
Scrape tiering: Parsed opponent pages are deliberately enriched for champions, top-1000 fighters and their opponent chains. Profile relational facts are computed only when every involved opponent has a parsed page or resolved identity; unparsed opponents are excluded from that fact rather than treated as high-value evidence gaps.
Title-context capture gap: The original parsed fight corpus did not capture title facts reliably as structured data. Profile title and reign facts therefore use the dated Wikipedia champion-reign table where a champion page URL joins through the H&G source/Wikidata-backed identity map.
Lineal incompleteness: The lineal/Ring reference is conservative. Current reference summary: 495 rows, 440 with resolved source IDs and 45 unmatched name keys. Missing or ambiguous lineal joins are omitted rather than guessed.
Reign-table coverage: The profile reign table has 876 public profiles with a Wikipedia identity and 687 profiles with joined reign rows. The ID-backed join is intentionally narrower than name-string matching to avoid false public claims.
Colour-line exclusion: Fighters barred from championship opportunity by the colour line are rated on results alone. This understates their standing, because the same system that excluded them also under-recorded them. The results-only treatment is the documented basis here, and the fuller account of their exclusion sits in their profile prose rather than in a number added to the score.
Source data: The ranking derives from a corpus of about 49,000 fighters through an automated pipeline. Known data-quality classes, including identity resolution, title attribution, division labelling and stale display fields, are checked at build time. Underlying source records carry residual error, as every scraped corpus does.
Raw corpus access: End-to-end reproduction from raw fight data requires the private parsed fight corpus, WHR runner, preserved feature artifact and the v2.0.0 export scripts. Public files expose enough to audit the published reduced model, not to reconstruct every raw bout row.
H&G All-Time Index is an Honour & Glory analysis layer built from public boxing record archives. Read the methodology and corrections process.