How it's calculated

Elo

A sequential rating. Every team starts at 1500. For each match, the two alliances play as the mean of their members' ratings, and the expected result comes from the standard logistic curve:

expected_red = 1 / (1 + 10 ^ ((rating_blue - rating_red) / 400))

actual      = 1 if red won, 0.5 if tied, 0 if red lost
change      = K x (actual - expected)        with K = 32

every red member: rating += change
every blue member: rating -= change

The 400 is what makes a 400-point gap mean roughly a 10:1 expected win ratio — it is the constant that defines the scale, not a tuned parameter. K = 32 is how far one match can move a rating, and it is a choice: higher reacts faster and is noisier, lower is stabler and slower to notice a team improving.

Alliances use the mean rather than the sum so ratings stay on the familiar ~1500 scale and a two-team alliance stays comparable to the one-team entries that appear in some elimination records. Every member takes the same update — the model cannot tell which partner was responsible.

Matches are replayed in order. Where a match has no scheduled time — which is every elimination match, since those are not pre-scheduled — the order falls back to event date plus round, instance and match number, which within an event is the exact order rather than an approximation.

What Elo can and cannot tell you

Unlike OPR it carries across events, which is what makes it the default ordering of the teams table. But it only links events that share teams. Your catalog currently has 27 separate pools; ratings in different pools never exchanged information, so comparing across them compares independent random walks. A team outside the main pool is flagged on its own page.

Elo is zero-sum — every point won is a point lost — so the mean across all 3672 rated teams is 1500 by construction, with a standard deviation of 71. Observed range 1198 to 1883, median 1490.

OPR, DPR and CCWM

These separate a team from its partners, which win rate and average score cannot do. Each alliance in a match is one equation — the sum of its members' ratings should equal its score — and the whole event is solved at once by least squares:

A x = s        A = one row per alliance, 1 where a team played
                x = the per-team rating being solved for
                s = that alliance's score        -> OPR
                s = the opposing alliance's score -> DPR

CCWM = OPR - DPR

OPR is what a team contributed to its own alliance's score. DPR is what it allowed the opposition — lower is better. CCWM is the difference, and it is the one to lead with.

Why DPR earns its place: at your Mall of America event the undefeated winner ranked only 6th on OPR. Their scoring was ordinary; what made them unbeatable was holding opponents near zero. Only DPR sees that, and CCWM put them 1st. OPR and DPR correlate at just −0.28 there, so DPR is not OPR mirrored.

Read them as rankings, not quantities

At that event the fit was R² 0.343 — residual RMS 29.96 points against a score spread of 36.95. So it explains about a third of score variance, and each team's figure rests on roughly five alliances. DPR can come out slightly negative; a team cannot cause negative opponent points, that is the fit overshooting. And a team playing pure defense suppresses rather than scores, so it reads as low OPR and low DPR.

They are only comparable within one event. Teams at different events never shared a match, so the systems are separate and sit on unrelated scales. The teams table shows an average across a team's events as a rough guide; the per-event figures on a team page are the real numbers.

Event importance

A judgement dressed as a number. The weights are a choice, not a measurement, which is why the three parts are shown under every score as level/size/strength instead of being collapsed into one opaque figure.

When field strength is unknown, the weights are renormalised over what is known rather than scoring the event as weak. Otherwise the column would measure how much data we happen to have, and every upcoming event would sink below every past one for reasons unrelated to importance. A · in the strength slot means exactly that.

Known problem with the strength term

The Elo-to-score mapping runs 1400 → 0 and 1800 → 100. Those numbers were chosen, not derived, and measuring shows they are badly calibrated.

Every field mean in your catalog falls between 1438 and 1669 — a 231-point span inside a 400-point scale. So the strength component can only ever produce values in a narrow band, and contributes close to a constant for every event. It is currently a poor discriminator, and that is a calibration fault, not a property of the data.

The cause is that Elo is zero-sum: a large field contains its weak teams alongside its strong ones, so the mean regresses to 1500 no matter who attends. Ranking by the top of a field instead gives 1.4× the spread, and calibrating the endpoints to the observed range rather than to round numbers would use the whole scale.

Smaller rules worth knowing