Phase 2 · Jul–Aug 2026

Delay prediction on the IST dataset

The second stage built on the IST delay dataset: a delay prediction model, and the concrete findings learned along the way. The roadmap and first results are below.

Priority · main quest · ready →

Delay prediction model

Predict the chance a flight leaves 15+ min late — from departure hour, airline, route, weather and the aircraft's previous leg. Trained on May–June, tested blind on July: lookup baseline 0.711, model 0.742 AUC, with 3.44x the delay rate in the riskiest 10%. The difference comes from the aircraft itself — 56.6% of flights whose aircraft had under 60 minutes on the ground leave late, where the lookup table says 20.7%. How it was built, two real cases, and where it fails →

Side quest 1

The airport's pulse

From the 131,334-flight IST dataset: hourly/weekly/monthly traffic patterns and a 3-airport comparison. A snapshot — will refresh as the data grows.

Hourly average: departures vs. landings (Istanbul local time)

Each bar is the average number of flights in that hour, on a typical day (from 85 days of data — not a raw multi-day total). Hover a bar for the exact average.

The overnight-turnaround pattern: departures peak at Istanbul local 07:00 (the morning bank), while landed arrivals peak at 22:00–23:00 (the night bank). That ~8–9 hour gap lines up with how long a plane landing at night needs to sit — crew rest rules — before it's ready to fly again in the morning bank. This is exactly the mechanism the "aircraft rotation" feature in Phase 2 is meant to capture.

3-airport comparison

Toggle airports on/off — the charts update accordingly.

Weekly average: which days are busiest

Each bar is the typical flight count for that day of the week (e.g. the average across every Sunday in the data).

Monthly average: May–July (per day)

Average flights per day WITHIN each month — not a total — so partial July (27 days) and June (27 days, a small collection gap) compare fairly against complete months.