Algorithms & Models

TrueSkill — How Bayesian Skill Ranking Works

TrueSkill treats every player's skill as a probability distribution, not a number — and narrows the uncertainty with every match, for teams and free-for-alls alike.

Reviewed 2026 Updated only when the core methodology changes

What the TrueSkill Ranking System Is

TrueSkill is a Bayesian skill-rating system developed at Microsoft Research (Ralf Herbrich, Tom Minka, and Thore Graepel, published 2006) for Xbox Live matchmaking. It generalizes Elo-style updating to multiplayer games and teams, and it tracks not just each player’s estimated skill but the system’s uncertainty about that estimate.

What TrueSkill Ranks

TrueSkill ranks players in competitive games — including team-based matches (5v5), free-for-all formats (8-player races), and head-to-head games — on a single consistent scale.

Matchmaking uses the rating to assemble games where each side’s predicted win probability is close to even.

Core Inputs Used by TrueSkill

  • Match outcomes — final placements of every player or team (not just win/loss)
  • Team composition — a team’s skill is modeled from its members’ skills
  • Two values per player — mu (estimated skill) and sigma (uncertainty about it)
  • System parameters — the assumed skill spread of the population and performance variability per game

How TrueSkill Is Calculated (High-Level)

The model is Bayesian belief updating:

  1. Every player’s skill is a Gaussian distribution: mu is the best guess, sigma the doubt around it.
  2. Before a match, the model computes expected outcomes from all participants’ distributions.
  3. After the match, beliefs update through factor-graph message passing: players who outperformed expectations gain mu; the surprise size and each player’s sigma scale the shift.
  4. Every game played shrinks sigma — the system grows more confident, so ratings stabilize over time. New or returning players have high sigma and move fast.

Conceptual model: Elo with a confidence interval — the system knows how much it doesn’t know, and bets accordingly.

Key Parameters or Factors

  • Mu (μ) — the skill estimate; the conservative “display” rating is mu minus a multiple of sigma
  • Sigma (σ) — uncertainty; shrinks with games, grows with inactivity in some implementations
  • Beta (β) — expected per-game performance variability, tuning how much a single result can mean
  • TrueSkill 2 — a later version incorporating in-game performance (kills, objectives) and experience, not just outcomes

Update Frequency

TrueSkill updates after every match — ratings are live, continuously recalculated as results arrive.

Known Limitations and Criticisms

  • Gaussian assumptions — real skill isn’t perfectly bell-shaped, and smurfing exploits the new-account uncertainty
  • Outcome focus (v1) — classic TrueSkill ignores how you played, only where you placed
  • Parameter sensitivity — beta and population priors need per-game tuning
  • Opacity to players — sigma is hidden in most deployments, so rating swings can feel arbitrary

Where TrueSkill Is Used

TrueSkill is used in:

  • Xbox Live matchmaking across Halo and other major franchises
  • Other game publishers’ matchmaking (the algorithm is published and widely reimplemented)
  • Chess and board-game platforms seeking uncertainty-aware ratings
  • Machine-learning evaluation of models via head-to-head comparisons

Summary

TrueSkill’s contribution is epistemic honesty at scale: rank players by what the evidence supports, carry the uncertainty openly, and handle teams and free-for-alls in the same Bayesian frame. It is the industrial standard for multiplayer matchmaking — Elo’s logic rebuilt for how modern games are actually played.

References and Sources

  • Herbrich, R., Minka, T., & Graepel, T. TrueSkill™: A Bayesian Skill Rating System (2006).
  • Minka, T., Cleven, R., & Zaykov, Y. TrueSkill 2: An Improved Bayesian Skill Rating System (2018).
  • Wikipedia. TrueSkill.