The system uses a custom Elo-based model tuned specifically for EE2. Each game updates the rating of every participating player based on the expected outcome versus the actual result. Team games are handled with a summed-variance team model, so individual contributions are weighted by player uncertainty. Parameters (K-factor, sigma, etc.) were optimised on historical data to maximise predictive accuracy.

There is however a key difference from classic Elo: standard Elo is purely relative to the current pool of players. This means a dominant player from years ago may have reached a very high rating simply because the competition at the time was weaker — their number would be incomparable to a stronger player active today. Here instead, the goal is to estimate absolute skill level as it stands now. Through the monthly re-centring and active-player adjustments, the scale is kept anchored over time so that a player who was dominant years ago but has since been inactive would be given the rating they would realistically deserve if they came back today — not the inflated number they earned against a weaker historical field. In other words, this leaderboard tries to answer the question: if everyone showed up to play right now, where would they rank?

Each player also carries an internal uncertainty value (sigma) that represents how confident the system is about their true skill level. New players start with a high sigma, and players who have been inactive for a while gradually see their sigma increase again. A higher sigma directly scales up the effective K-factor, meaning those players gain and lose points much faster per game. This allows the system to converge quickly toward a new player's real level rather than moving them up or down by tiny increments over hundreds of games.

You need at least 20 counted games in a given mode (R-R or 5-5) to appear on that leaderboard. A game is counted only if it meets the required settings: Conquest game type and a Plains map (Temperate or Tropical Plains). Beyond that, a game can still be excluded if it triggers one of the internal filters — for example, it was too short, all recorded kill ratios were zero, an AI player was present, or a result could not be determined. All other valid games in the right mode accumulate toward your rating and game count.

Win rate tells you how often you win, but it says nothing about who you were winning against. A player who consistently seeks out weaker opponents can maintain a very high win rate while barely gaining any rating — because the system already expected them to win, so each victory is worth very little. Conversely, a player who often faces stronger opponents may have a modest win rate but a high rating, because every win against a stronger player carries a large reward and every loss costs relatively little.

The rating system is designed to give you the deserved amount of points for each game based on how likely your team was expected to win given everyone's current ratings. Quality of opposition matters far more than raw win count.

The numbers below assume both players are veterans with an active play history, so their uncertainty (sigma) is near its minimum value. This is the baseline case — in practice, a new or recently inactive player carries higher uncertainty, which affects not so much the raw win probability (because game-to-game variability already dominates) but rather the rating points gained and lost: a veteran playing against a high-uncertainty opponent will earn and lose fewer points per game than they would against an equally-rated veteran, since a result against an uncertain player carries less information.

R-R — 1v1 win probabilities by rating gap (both players at minimum uncertainty):
 • ~150 pts advantage → ~60% chance of winning
 • ~390 pts advantage → ~75% chance of winning
 • ~740 pts advantage → ~90% chance of winning

5-5 uses the same model structure as R-R but with separately optimised parameter values tuned specifically to 5-5 game data, so the exact thresholds differ slightly.

Team games are calculated similarly. Each team's effective rating is the sum of all its players' individual ratings divided by the team size raised to a small exponent (around 0.26 for R-R). This means it is neither a pure sum nor a plain average — each additional player contributes positively but with some diminishing return as the team grows larger.

Keep in mind that the overall rating is a single number covering all game formats. Some players are genuinely stronger in team games than in 1v1, or vice versa. If you consistently beat a higher-rated opponent in 1v1, it may simply mean they are a better team player than a solo player — or it could reflect limitations of the model itself.

Indirectly, yes. Each month, players who played enough games receive a small rating boost. After that boost is applied, all ratings are globally re-centred so that the average stays at the baseline — this re-centring applies to everyone, active or not. The net effect is that active players gain rating relative to the field, while inactive players slowly fall behind because they miss the boost but still absorb the re-centring. This is intentional: it models the fact that the average skill level of the active player base improves over time.

Feel free to reach out to Matty — he wants to be informed. That said, corrections are handled case by case with fairness in mind. Large, clearly verifiable situations (e.g. "I played with champion handicap for three days") are much more likely to be addressed than isolated or hard-to-verify cases. Applying fixes selectively could itself be unfair to players who never asked.

If you plan to experiment, test strategies, or play non-seriously, consider using an account whose name ends in NotSerious (e.g. Matty_NotSerious) — those accounts are automatically filtered out and won't affect the leaderboard at all.

The leaderboard auto-updates approximately every hour. You can also force an immediate recalculation using the Update Now button in the top-right corner.

No — Matty will keep improving it whenever he has the time and motivation. This means the leaderboard may occasionally shift noticeably after a recalculation driven by a logic improvement, not just new game data. Think of it as the model getting more accurate over time rather than the ratings being unstable.

Additionally, if it becomes clear that certain players are trying to game the rating system by exploiting specific patterns or edge cases, the model will be adapted to counter that — and everything will be recalculated from scratch with the updated logic. The goal is always to make the ratings as fair and meaningful as possible.

But how do you know if a change is actually an improvement? It is not based on personal judgement. Improvements are measured numerically. In fact, the model is already accurate enough that Matty himself is often uncertain whether the ratings of specific players are right or wrong — and it would be unfair to let personal opinions shape the leaderboard anyway.

Instead, every version of the model is evaluated against thousands of historical games using a purely objective metric: for each game, the model's predicted win probability is compared to the actual outcome. For example, if the model says team A has a 65% chance of winning and team A wins, the squared error for that prediction is (1 − 0.65)² = 0.1225. If the model had instead predicted 70%, the error would be (1 − 0.70)² = 0.09 — smaller, meaning that prediction was closer to reality. Averaging this across all games gives a single number called the Mean Squared Error (MSE). A lower MSE means the model's predictions align better with what actually happened. When a logic change reduces the MSE on held-out games the model has never been optimised on, that is a genuine, verifiable improvement — independent of anyone's opinion about individual players.

Almost certainly yes — but in rare cases, a game that appears valid may not be counted. This system fetches game records from an external source, and occasionally a game simply isn't present there even though it was played. This is not something that can be fixed on the rating system's side, as the issue originates further upstream, in the data that gets made available. It is uncommon, but if you notice a specific game is missing from your history despite meeting all the usual criteria, that is likely the cause. Without the data source this project relies on, none of this would exist in the first place — so the occasional gap is a minor trade-off.