The gap that uncertainty fills
A simple Elo-style rating gives every player one number. That hides something important: how much evidence sits behind it. A player with five games and a player with five hundred can show the same value, yet the first is far less certain.
Glicko and TrueSkill both address this by attaching a measure of uncertainty to each rating. When a player is poorly known, the system treats a new result as more informative and moves the estimate further. As results accumulate, the uncertainty shrinks and updates become smaller.
Glicko: a rating plus a deviation
Mark Glickman designed the Glicko system as an extension of Elo. Each player has two values:
- a rating, on a scale similar to Elo, with new players commonly starting at 1500
- a rating deviation (RD), which expresses the uncertainty; new players commonly start with a large value, 350 in the standard description
Roughly speaking, the true skill is expected to lie within about two RD of the rating, so a player rated 1500 with an RD of 350 might plausibly be anywhere across a very wide range. After more games the RD falls, and the range tightens.
Two features follow from this design:
- Uncertain players move more. A result against a well-established opponent shifts a high-RD player a lot and a low-RD player only a little.
- Opponent uncertainty matters. Beating an opponent whose own rating is shaky gives a less certain signal than beating one whose rating is firm, and the expected-score calculation takes that into account.
Glicko also lets RD grow again when a player is inactive, reflecting the fact that skill may have changed while they were away. A later version, Glicko-2, adds a volatility value that describes how erratic a player’s results are, so that consistently performing players and unpredictable ones are treated differently.
TrueSkill: skill as a distribution
TrueSkill was developed at Microsoft Research for matchmaking in online console games. Instead of a rating and a deviation, each player has a skill distribution described by a mean (mu) and a standard deviation (sigma). In the commonly cited defaults, mu starts at 25 and sigma at 25/3, about 8.33.
It is built on Bayesian reasoning. The system starts with a prior belief about each player, observes a result, and updates the belief to account for how surprising the result was. This is the same logic as Bayes rule applied in other settings, here used to revise a belief about hidden skill.
TrueSkill is designed for more situations than Glicko:
- Teams. A match result is attributed across the players on each team, so skill estimates for individuals can be derived from team outcomes.
- More than two sides. Free-for-all games and matches with several teams are handled by working with the finishing order.
- Draws. The model includes a way to represent tied results.
A frequently used conservative summary is mu - 3 x sigma. A new player with the defaults would score 25 - 3 x 8.33 = 0, and the value rises as sigma shrinks, even if mu stays the same. Leaderboards sometimes use a measure like this so that a player cannot climb on the strength of a few lucky games.
How they compare
| Feature | Elo | Glicko | TrueSkill |
|---|---|---|---|
| Tracks uncertainty | No | Yes (RD) | Yes (sigma) |
| Natural for 1 v 1 | Yes | Yes | Yes |
| Team and multi-sided games | Awkward | Awkward | Designed for them |
| Complexity | Low | Moderate | Higher |
Glicko is often a good fit for two-player games, where the extra uncertainty term brings most of the benefit with a modest increase in complexity. TrueSkill earns its extra machinery when a result has to be shared among several players, as in squad-based online games.
Neither one is a perfect model. Both assume that skill can be summarised by a single number or distribution per player, and both can be thrown off by smurf accounts, short-term form, or changes to the game itself.
Why uncertainty helps in practice
Tracking uncertainty has two practical benefits. It lets new players find their level quickly instead of needing dozens of games, and it gives matchmaking systems a way to avoid placing very uncertain players in high-stakes matches until the system has seen more. It also gives honest reporting: a rating shown with a margin better reflects the fact that results contain luck.
That connects to the larger theme in competitive data. A number without its uncertainty invites overreading, and the same warning applies to short-run results in player statistics. Elo remains a sensible starting point, and Glicko and TrueSkill build on it rather than replace its logic.