How this works

Covering last 1y.

Each tier is worth fixed points set by the committee. A team's strength is the sum of its three players' points, and the only fitted parameter is what one point of advantage is worth:

P(win) = sigmoid(0.47 × (my team's points − theirs))

There is no intercept, by design: two mirrored rosters score exactly 0.5.

The tier letters are not an A-to-E ladder

The strength order is S > E > A > B > C > D. Tier E is the second strongest, not the weakest. Every place that needs "the next tier up" reads it off the points, never the alphabet.

The fit

Fitted on 6394 decided 3v3 matches: 64.4% accuracy, 0.2181 Brier, 0.6254 log loss.

Known limitations

• The approach is partly circular by design: it asks whether a record is consistent with the tier held, not what tier a player deserves in the absolute.
• Tier changes are dated only from the day they started being recorded. A match played after a logged change is scored against the tier held then; anything earlier is scored against the tier held today, so an old promotion can still look like a year of overperformance.
• The alpha side wins about 52% of matches and the model, having no intercept, cannot express that.
• A roster is the three players with the most playtime on each side. A match where four people played a side is still scored as three.
• A player with no committee tier gets one imputed from their shrunken UTRO, capped at A. Imputation error, not luck, is the largest source of false signal here.