Click a zone to activate it, and click and drag to move the zone.
It is built from 3,840 FNCS solo games on patch 42.30.
Where players are when a zone closes depends on two things. The first is the direction the zone pulled from: players come in from that side and bunch up along it, leaving the far side of the zone, the deadside, mostly empty. That's partly because rotating players keep the distance they travel into the next zone as short as possible. The deadside is strongest in zones 5 and 6. The second is the map: places with structures, like buildings and POIs, and high ground such as ridges tend to attract players, while open, low ground is avoided. The model learns both from the real games and multiplies them together: the first is a smoothed density of player positions in a frame turned to face the pull, and the second is, for each 10 m patch of the map, how many more or fewer players were there than the first part predicts (a smoothed ratio of observed to expected, pulled toward no effect where there is little data). Zones 7 and 8 are too small for the map to make a difference, so only the pull direction is used there.
Each zone also depends on the ones before it. To capture that, 1,500 simulated players move through your sequence one zone at a time. At each zone, every simulated player copies what a real player in the same spot did in a real game: where they moved, or that they were eliminated. This is a Monte Carlo simulation with nearest-neighbour resampling: each simulated player's next move is drawn at random from the 15 real players nearest its position relative to the zone. The simulated players are smoothed into a density with a Gaussian kernel, and the density shown blends this with the map-based estimate. How much weight the simulation gets, between 10% and 70% depending on the zone, was chosen as the one that best predicts where players really were in games the model wasn't built from.
Rotations come from a second set of 1,500 simulated players that move in 20-second steps, again each copying a real player from the same spot. At each close they are resampled in proportion to the map part of the model, so they carry on from where players gather. Their movement is averaged over part of the zone, so looting back and forth cancels out and only the overall direction remains; those movements are averaged onto a grid and smoothed with a Gaussian kernel into a velocity field, keeping only the places where most of them head the same way. The lines are streamlines traced through that field: wider where more players move together, darker where they move faster.