How FlightQueue measures airport wait times

Data sources by tier

How predictions are made

• Rules engine — hour-of-day, day-of-week, airport profile, weather, and FAA delay multipliers produce a baseline wait (rules_pred in snapshots).
• Live overlay — when a fresh official reading exists, it anchors or replaces the model for that checkpoint until the feed goes stale.
• Crowd & feedback — validated traveler reports adjust short-term bias; outcomes land in prediction_accuracy for calibration.
• ML shadow model — an optional gradient-boosted model trains on ml_training_pairs; airports graduate when shadow MAE beats the rules engine in ml_retrain_runs.

Freshness & confidence

Each airport page shows the timestamp of the underlying reading when we have one. Predictions without a live feed display the model run time. Confidence is a 0–1 score from the estimation engine; we map it to user-facing bands:

• High — score > 0.8 (low uncertainty)
• Medium — 0.5–0.8
• Low — below 0.5 or thin sample size at the airport

We never label fallback baselines as "Live · High confidence." If you only see a band (Fast / Moderate / Busy) without minutes, you are on the free tier or the airport lacks a fresh feed.

Accuracy backtest (real production data)

Rolling 90-day window computed 2026-10-29. Ground-truth pairs come from archived snapshot ↔ official feed matches (ml_training_pairs). Traveler feedback uses prediction_accuracy. When sample sizes are small, we show the count honestly rather than extrapolating.