Algorithms & Models

Bradley–Terry Model — How Pairwise Comparisons Become Rankings

The Bradley–Terry model turns 'A beat B' into probabilities: each competitor gets a strength, and the odds of any matchup fall out of the ratio.

Reviewed 2026 Updated only when the core methodology changes

What the Bradley–Terry Model Is

The Bradley–Terry model is a statistical model for paired comparisons, introduced by Ralph Bradley and Milton Terry in 1952. It assigns every competitor a positive strength parameter, and declares the probability that i beats j to be i’s strength divided by the sum of both strengths.

It is the mathematical root system beneath much of competitive ranking — Elo is effectively a logistic reparametrization of it — and it has found a second life in machine learning, where it turns human preferences into ranked models.

What the Bradley–Terry Model Ranks

The model ranks any set of items compared pairwise:

  • Players and teams in sports
  • Products in consumer preference studies
  • AI model outputs in human feedback evaluations

Only who-beat-whom data is needed — no margins, no contexts, no continuous scores.

Core Inputs Used by the Bradley–Terry Model

  • Pairwise outcomes — a record of comparisons and their winners
  • Ties (optional) — extensions (e.g., Davidson’s) add a draw parameter

The model deliberately ignores everything else — margin, venue, order — making it robust where data is sparse but blind where context matters.

How the Model Is Fitted (High-Level)

The procedure is maximum-likelihood estimation:

  1. Posit a strength for every competitor (all equal at the start).
  2. The likelihood of the observed results is the product of each matchup’s predicted probability.
  3. Adjust strengths — classically by iterative scaling — to maximize that likelihood: the strengths that make the actual results least surprising.
  4. The fitted strengths induce a total ranking; log-strength differences map to win probabilities on a logistic curve.

Conceptual model: Find the set of strengths under which history was most likely to happen — the ranking is whatever makes the past least surprising.

Key Parameters or Factors

  • Strength parameters — identified only up to a scale constant; ratios carry the meaning
  • Logistic link — Elo’s 400-point scale is a change of variables of the same model
  • Home-advantage extension — a multiplier can be added when venue matters
  • Tie parameter — in extended variants

Update Frequency

The model is typically refit as new comparison data arrives — there is no built-in online update, though incremental and Bayesian variants exist.

Known Limitations and Criticisms

  • Outcome-only — ignores margins and circumstances by design
  • Transitivity assumption — real preferences cycle (rock-paper-scissors effects) that a one-dimensional strength scale cannot express
  • Sparse-comparison instability — unbeaten or winless competitors can push estimates to extremes without regularization
  • Static picture — strengths are assumed constant within the fitting window

Where the Bradley–Terry Model Is Used

The model is used in:

  • Sports rating systems (directly, or via its Elo/logistic kin)
  • Psychometrics and market research on preferences
  • Machine learning — RLHF reward modeling and LLM arena leaderboards fit Bradley–Terry-style models to human pairwise judgments
  • Ranking aggregation research

Summary

Bradley–Terry is the atom of ranking theory: strengths, ratios, likelihood — three ideas from which Elo, countless sports models, and modern AI preference learning all descend. Its limits (one dimension, no margins, static strengths) are precisely the problems its descendants were built to solve.

References and Sources

  • Bradley, R. A., & Terry, M. E. Rank Analysis of Incomplete Block Designs (1952).
  • Wikipedia. Bradley–Terry model.
  • Langville, A. N., & Meyer, C. D. Who’s #1? The Science of Rating and Ranking.