The receiver hears it about every 6 seconds, so in between we have to guess. The obvious guess is a straight line. Here it is, on a real flight in a holding stack.
The dots are what we heard. Green is what the plane really flew next.
The dashed line, straight on at the last reported speed and heading, ends 4.1 km from the plane.
Can you do better? Try it below, one idea at a time.
Fixes the receiver heard
What it really flew (next 60 s)
Straight on4.1 km off
Every drawing is real: fixes from the receiver, the track flown afterwards, and paths from the code the map and the reel run. The moments are picked to show each idea; the last section tests all of them.
The receiver is in London, so every flight here is near London. The rules are not: they start from what any aircraft shows, wherever it is (height, climb rate, speed, how it is turning), and only at the very end consult a short list of known runways.
Idea 1
Keep going the way it was going
Either wait at the last fix, or carry on straight at its speed and heading. The same flight at two moments, each looking 30 seconds ahead.
Flying straight
Fixes the receiver heard
What it really flew (next 30 s)
Stay at the last fix3.8 km off
Straight on70 m off
Straight flight: straight on is only 70 m off. Waiting at the last fix is 3.8 km behind.
Starting a turn
Fixes the receiver heard
What it really flew (next 30 s)
Stay at the last fix3.8 km off
Straight on2.0 km off
Same plane, a couple of minutes later, turning: straight on is 2.0 km off. It does not know the plane is bending.
30 s ahead
Idea 2
Use more than the last fix
One fix can be a second or two stale. The last 30 seconds of fixes say much more: how fast, how the speed is changing and whether the track is bending.
Fit a smooth curve through the last 30 s. Big dots count more.
The newest fix counts most, so the line starts on the plane.
Speeding up, slowing down and a gentle bend are all in the fit.
The bend fades out over about 15 s, so it never spirals.
Fixes the receiver heard
What it really flew (next 30 s)
Straight on510 m off
Fitted motion120 m off
A second flight, in a gentle curve: straight on 690 m off, fitted 200 m.
60 s ahead
Idea 3
A turn does not stop after 20 seconds
The fit bends a little and then straightens. Real turns last longer, so the heading is planned on its own, from the turn rate the plane reports.
Fading turn: the measured rate dies out, with a time constant of 20 s.
Keep turning: hold the rate until the turn is a semicircle.
A turn that has gone 60° already fades slower the longer it lasts.
Climbing turns and turns that are rolling out fade fast.
Fixes the receiver heard
What it really flew (next 60 s)
Straight on7.2 km off
Fitted motion6.8 km off
Fading turn2.8 km off
Keep turning1.2 km off
60 s ahead
Idea 4
Low and sinking: the turn stops
A plane turning onto its final approach does not turn through the runway line. It lines up and stops turning, and you can tell it is on approach without knowing where the runway is.
Low (under 3,000 ft above the ground), sinking 300 to 1,500 ft/min and slower than 200 kt: that is an approach, anywhere in the world.
The turn then rolls out within about 6 s, instead of fading over 20 s like a turn at height.
Last layer: when a runway on our short list is close, the turn ends exactly on its heading. The dotted line is Gatwick's centreline. Without the list the median error on final, 60 s ahead, is 396 m; with it, 369 m.
Fixes the receiver heard
What it really flew (next 60 s)
Fading turn4.8 km off
Stop on the runway1.1 km off
40 s ahead
Idea 5
Reaching the ground, it stops
A plane about to land does not fly on over the runway. When we lose it low, the map has to decide: has it landed, or did we just stop hearing it?
The feed marks an aircraft on the ground by sending no altitude at all. Straight after a low fix, that is a touchdown, and the roll starts there.
No flag yet? Then low, sinking steadily (at least 150 ft/min), at 45 to 200 kt, and reaching the ground within 45 s at that rate. A cruise, a climb-out, a level pass and a take-off roll all fail one of these.
After touchdown it slows at 1 m/s², the median of 25 real roll-outs at Biggin Hill, and stops. The marker dims and says landed.
Where the ground is: first the altitude just before earlier touchdowns nearby, then a short list of known airfields, then sea level. Guessing the ground too low only makes the call late.
Fixes the receiver heard
What it really flew (next 40 s)
Straight on1.1 km off
Touch down, roll out, stop260 m off
A Biggin Hill arrival at its first fix on the ground, 89 kt. Straight on ends 1.1 km from where it was 40 s later; the landing rule 260 m.
2 min ahead
Idea 6
In a hold, learn the racetrack
Holding planes fly the same oval again and again. After one orbit the pattern is known, so the next turns and legs can be flown ahead of time.
From the last 8 minutes: turn direction, turn rate, leg headings, leg time.
In a turn, finish it on the next leg heading.
On a leg, fly out the leg time, then turn again.
On a leg, straight on gets equal say, because leaving the hold looks the same until the fix.
Fixes the receiver heard
What it really flew (next 120 s)
Straight on9.5 km off
Learned hold1.8 km off
Idea 6, continued
Then notice when it leaves
This is the whole 23-minute track of that holding flight. Orange is where the plan knew the hold. Grey is where it did not.
Known after 10.6 min; dropped for good at 21.6 min.
A leg that runs 25% (+10 s) over the learned time ends the hold. Here, 115 s.
So does flying straight 30° off both legs, or turning the other way.
Or descending faster than 800 ft/min on a leg. Stack step-downs trip this, and it learns the hold again.
Hold known (orange ring: first recognised)
Not known, or dropped (grey ring: dropped for good)
45 s ahead
Idea 7
Let the recent record decide
The flight phase picks a favourite, but it can be wrong. So every candidate is replayed on this plane's own last minute and a half.
The pick starts at weight 1, the others at 0.03.
A candidate that predicted the last few fixes clearly better takes over.
Here the phase picked a fading turn, but keep-turning had been closer, so it takes 99%.
Softer blends were worse: hedging a turn with straight costs more than it saves.
Fixes the receiver heard
What it really flew (next 45 s)
Fading turn3.5 km off
Keep turning1.5 km off
Weight of each candidate: dashed is the start, solid is after scoring
Straight on0%
Fading turn1%
Keep turning99%
Median error, 30 s ahead
Holding (13,742 samples)
1,459 m
555 m
Turning (46,051)
1,271 m
839 m
All traffic (473,000)
258 m
239 m
■ fit only · ■ with the heading plan. Two days from the receiver plus 2 h of London traffic.
Old straight coast vs the fit
30 s ahead
874 m
289 m
60 s
1,914 m
567 m
120 s
4,703 m
1,611 m
■ the old straight-line coast · ■ the fit. First replay: 698 flights, 2 h.
The test
Does it work? Replay recorded flights.
About 5,000 flights, 11 and 12 October 2026. Each test sees only fixes up to a moment, predicts, then checks against what was flown.
Holds at 30 s: 1,459 m down to 555 m. At 60 s: 5,105 m down to 2,126 m.
Cruise, climb and descent stay about the same.
Landing calls: 1,227 fixes were called a landing. 1,207 of them (98%) were followed by the feed's ground state or by the flight ending low, and 17 by a climb (go-arounds, touch-and-goes).
Not better everywhere: the 90th percentile at 60 s is 6% worse in descent and 8% worse in level flight, and one small London stack did worse.
Every setting tried that fixed those tails gave back most of the hold gain.