How the Boxerbook rating works

Every boxer has one rating, whatever the division, computed with Glicko-2: a rating μ and a rating deviation RD, the uncertainty, shown as “± RD”. A new boxer starts at 1500 ± 350 with volatility 0.06; the system constant τ is 0.8.

A rating period is one calendar month: all bouts of a month are applied together, and in months without a bout the uncertainty grows.

A win counts 1, a draw 0.5, a loss 0. No contests and disqualifications without a winner are left out, and so are retracted bouts, bouts dated only by year and bouts the record calls an exhibition. A disqualification with a winner counts as a win.

At most one bout per boxer per day counts: when two sources list the same bout twice and the copies could not be merged, the copy with a linked opponent is kept (then the one with an official result).

Each bout’s contribution is multiplied by a weight, the product of the multipliers below.

Title: world title ×1.5: the title names the WBA, WBC, IBF, WBO, IBO or The Ring (in Spanish records CMB, AMB, FIB, OMB) or is undisputed; interim or franchise title ×1.2; regional, continental, national or silver title, and a bare “world” without a named body, ×1.05; no title ×1.0.

Scheduled distance: 4 rounds or fewer ×0.8, 5–6 rounds ×0.9, 7–8 rounds ×1.0, 9–10 rounds ×1.05, 11 rounds or more ×1.1; unknown distance ×1.0.

Stoppage (KO, TKO, RTD): no bonus, ×1.0. A stoppage bonus of ×1.1 and ×1.2 (and ×0.9) was tested on the 2015–2025 bouts and did not improve prediction, so a stoppage counts exactly like a decision; by rule it never counts less. The share of wins by knockout is shown separately on each boxer’s page.

Bout outside the boxer’s division (more than one division away): ×0.85. The division of a bout is rarely recorded yet, so this applies seldom.

The weight is kept between 0.5 and 2.0.

An opponent without a page in the catalog counts as an unknown boxer rated 1100 ± 350, at weight ×1.0. On average an unlinked opponent is a weak one; this value predicted results best in the backtest.

Each boxer’s page shows how many of the counted bouts were against unknown opponents.

After 12 months without a bout a boxer is inactive: still rated, shown as “inactive since …”, but out of the division lists.

A division list holds active boxers with at least 5 counted bouts and an uncertainty of ± 200 or less, ordered by μ. One correction: when two neighbours met in the last 3 years and the lower one won their latest meeting, they swap places (one pass, no cascade).

A boxer’s division is the more recent of two signals: their latest bout with a known division within 24 months (a linked title reign, or the weight class named in the bout’s title, such as “WBC super lightweight title”), and their latest publisher ranking within 12 months. With neither, it is the division most of their last five linked opponents within 24 months were in when they met, if at least two of them were; a boxer with no division is rated but not listed.

The same rules give a list for the end of every month, so the monthly archive pages show the division as it stood then.

Strength of schedule: the average μ of the opponents at the time of each bout, over the career and over the last 36 months (unknown opponents at their fixed value).

Form: the last five results, weighted 1.0, 0.85, 0.7, 0.55 and 0.4 from the newest, from 0 to 3.5.

Consensus: the average place over the latest list of each of the six publishers (WBC, WBA, IBF, WBO, The Ring, TBRB) in the boxer’s division; a champion counts as 1 and a list without them as 16.

Data confidence: the share of counted bouts confirmed by two independent sources or by an official result.

Honestly: on 2015–2025 (in-sample) the rating picked the winner in 72.67% of 5,012 bouts, against 72.52% for simply picking the boxer with the better win share: level with it within half a point, and below it in 5 of 11 years. Out of sample (holdout): behind it in 2022–2025, 69.43% against 71.06%; behind it in 2026, 61.62% against 65.40%. Its edge is a probability for every bout (log-loss 0.5462), which a win share cannot give, and a scale that weighs who the wins came against. Picking by wins minus losses (52.64%) is near a coin flip: it favours long careers.

Publisher lists cannot be compared yet: Boxerbook’s archive of their snapshots starts in 2026, and only 7 bouts in 2026 had both boxers in a recent list (publishers 100.00%, the rating 71.43% on the same bouts).