Demo · the statistical half · Jul–Aug 2026

Departure delay board

Every scheduled departure from IST, scored against its own history: how often has this flight number, this route and this hour run late before? Like an airport board — except the numbers are a prediction, not an announcement.

Flights on the board

Click a row to see where its number came from.

Why does the board stop at 04:00? The free plan returns 100 rows per request, and that page starts at the very beginning of tomorrow's schedule — roughly the first four hours after midnight. It cuts off there because the page ran out, not because these flights were chosen. A whole day (~1,700 departures) would take 17 requests; the free quota is 100 requests a month — and it is currently exhausted.

Does it work?

The same method applied to a day it had never seen: built from everything before that day, then asked to score it, then checked against what actually happened.

The honest part: the model ranks flights well — far more of the ones it called risky really did run late, and where it says 40% it is almost exactly right. Where it misses is the flights it calls safe: tested blind on a busier month it had never seen, the flights it scored at 2% ran late 6% of the time. Chronically late routes stay late; a bad month lands on the ordinarily punctual ones, and nothing in their history warns of it. The missing term is day-level — weather and live airport state.

Method and limits

Delay = aviationstack's gate departure delay, threshold 15+ min. Departures only; cargo and cancellations excluded. Each flight is scored by walking down these rungs: flight number → airline+destination+hour → airline+destination → destination → airport average. Cells with thin history are smoothed toward the rung above, so a route flown three times doesn't shout "0%" or "100%". This is a prediction, not an airline's official estimate.