Rotating is losing, or is it?

Every Rocket League player has been told the same handful of things. Rotate. Shadow, do not dive. Do not be the last man. Learn the speed flip. The advice is everywhere, it is mostly given in good faith, and almost none of it has ever been checked against anything.

Unravel's baseline is 9,000 replays, 3,000 each of 1v1, 2v2 and 3v3, all of them RLCS level or Supersonic Legend. Once you have that sitting on disk an obvious question falls out: for every stat the app tracks, how strongly does it actually move with winning? So I measured it, all 398 of them.

The method is one number per stat, from +1 (always wins) to −1 (always loses). Every value becomes a rate first, so game length cancels, and every team is compared only against the other side of the same match. Without that you are partly measuring which lobbies were sweatier. For scale, 0.20 means the teams best at it win about 35 points more often than the teams worst at it, and 0.05 is noise. (It is a point-biserial correlation, and further down a partial one. I am a developer rather than a statistician, so feel free to fight me on it.)

First, the boring part

Plenty of it lands where you would expect, which is the only reason to trust the rest. Shots is the strongest thing a team does that is not a goal, 0.55 in 1v1 and still 0.46 in 3v3. Possession time, 0.30. It is the control group.

Saves are worth nothing

Saves, the stat on every scoreboard, the thing you get congratulated for, scores 0.020 in 1v1, 0.008 in 2v2 and −0.065 in 3v3. Zero, and in 3v3 leaning the wrong way. Nothing was broken, though. A save is two things at once and they cancel out: making the save is good, needing it is bad, and you only get the chance because the opponent put a shot on your net.

The stat was fine. The question was wrong. Ask a better one, how a team does on saves given how much they were shot at, and saves jump from 0.02 to 0.51, one of the strongest numbers in the whole catalog. Same data, same matches. It only ever needed the denominator nobody puts on a scoreboard.

Rotating more is losing

The most repeated advice in the game is some form of "rotation". Rotation count scores −0.18 in 2v2 and −0.21 in 3v3, one of the strongest team signals there is, pointing down. The more often a team cycled its first man, the more often it lost.

I do not think this means rotation is bad. I think the thing being counted is not rotation, it is churn. Handing the front role over because you touched the ball is a rule, and following a rule produces a high count. Rotating because your team currently lacks a front man produces a low one. Pros do the second thing, and a counter cannot tell the two apart.

Speed flips predict nothing

Speed flip count scores −0.007, −0.003 and −0.042 across the three modes, and execution quality is just as flat. The mechanic people spend weeks in free play on has no relationship with the result at all here.

It does not mean speed flips are useless. It means everyone in these replays already has one. You cannot measure the value of a skill in a group where nobody lacks it — ask whether height predicts basketball ability among NBA centres and you get the same nothing, for the same reason. A zero in a pro dataset means "no information here", not "does not matter".

Shadowing lands in the same place. Time spent shadowing scores −0.139 in 1v1, and corrected for how much the opponent was attacking it settles at +0.04, which is to say nothing in either direction.

The biggest number says the least

The largest score anywhere was goals conceded while you were the last man back, at −0.771 in 1v1. Impressive, until you notice that in 1v1 you are always the last man back. So the stat is really "goals conceded", which is the scoreline, which is the thing being correlated against. It is a circle, and half the work here turned out to be checking for exactly that.

The trap underneath all of it

There is a bigger problem hiding in every number above, and it took my own replays to see it. Average boost held scores 0.22 among pros, which reads as sensible. So I checked it against my own Diamond and Champion replays, expecting to find I was running low.

I was running high. Well above the pro band, and it held across every ranked replay I tried. Ranked games are slower and nobody contests the pads. A stat that predicts winning inside pro play points the opposite way across the gap between pro play and mine.

A number measured inside one group tells you how that group's good days differ from its bad days. It does not tell you how to become that group. Those are different questions, and almost everything that gets called "the stats say" quietly treats them as the same one.

There is no clever fix for that inside a pro dataset. More pro replays only make the same measurement sharper, and sharpness was never the problem. The only thing that can answer "what separates a Gold from a Diamond" is replays from Golds and Diamonds — a second dataset, one that spans the ranks instead of sitting on top of them.

So what is it good for

Everything above is now the measured half of Unravel: each stat's direction and impact tier come from the replays rather than from me, and if a stat is worth nothing the app says so instead of leaving it on a chart next to the ones that matter. But a measured layer cannot be the whole answer, and the speed flip is the proof. Zero correlation, still worth learning. Throw away everything that measures zero and you throw away most of what a coach would tell you.

So it seems it's high time for another dataset, one that contains all eight ranks in each of the three modes. Same method, pointed at a group that still varies on the basics. Boost discipline, shadowing and the speed flip should all have something to say there, because in Gold they are definitely not there. Or are they now?