How DelayGuard actually works
No black boxes. Here's exactly what data we use, how the model is built, what the numbers mean — and where its limitations are.
How accurate is AI flight delay prediction?
Here's how we build and test the model so its numbers mean something in the real world.
1.33M
unseen test flights
Tested on October–December 2025 flights the model never saw during training — the way it's used in real life.
3.95M
training flights
Real BTS on-time performance records from 2023 to mid-2025 US domestic flights.
Calibrated
pre-departure signals only
Route and time-of-day patterns, airline and airport history, congestion, weather and seasonality — all known before you fly, with the ML score calibrated on a separate validation period.
Why we don't lead with a single score. A headline accuracy number is easy to inflate — for example by testing on flights from the same period the model trained on, or by using signals that are only known once a flight is already underway. v9 avoids both: it only sees pre-departure information and is tested on a later period. Flight delays stay inherently noisy — crew sick calls, last-minute ATC decisions and pop-up storms can't be foreseen by any model — so treat each prediction as a well-grounded risk estimate, not a guarantee.
Hybrid rule engine + machine learning
DelayGuard combines two complementary approaches so you get both explainability and predictive power.
Rule Engine
65% of final score
A deterministic scoring system that applies known, interpretable factors: airport congestion, slot-control status, airline on-time history, NAS vulnerability, departure time, weather risk, and more. Every factor in the breakdown panel comes from here — fully human-readable.
How it thinks
Each risk factor (weather, congestion, airline history…) gets a score. Those scores are added together, weighted by importance, to produce a single delay probability you can trace back to its causes.
XGBoost ML Model
35% of final score
An ML model trained on 2023 to mid-2025 BTS data (nearly 4 million real US domestic flights) and tested on later flights it never saw. It spots patterns too subtle for a hand-written rule — like how a specific route on a Tuesday evening in March behaves differently from every other combination — and uses them to refine the prediction.
How the two scores are blended
The rule engine's score counts for 65% of the final answer; the ML model contributes the remaining 35%. The ML score is first calibrated on a separate validation period that it was not trained on.
How a prediction is generated
You enter a flight
Route (origin → destination), airline, departure date and time. Optionally a tail number or flight number for live data lookup.
Live data is gathered
We pull current airport weather conditions, active weather advisories, and live aircraft position data from multiple official and industry sources — all in priority order, with automatic fallback if any source is unavailable.
Dozens of signals are assembled
Route history, airline performance, aircraft age, airport congestion, time of day, seasonal trends, and live weather signals are all combined into a unified picture of this specific flight.
Both models score the flight
The rule engine produces an explainable base probability. The XGBoost model produces an independent, calibrated ML probability from pre-departure signals only.
Scores are blended and calibrated
A weighted average (65% rule / 35% ML) gives the final delay probability you see.
How the model has improved
From a simple prototype to today's rebuilt v9 — each version added better data, better signals, or more honest testing.
Rebuilt for real-world reliability
The current model, rebuilt from the ground up. It now uses only information that is available before your flight departs, so test results reflect what you actually get. It trains on 2023 to mid-2025 flights and is tested on a later quarter it never saw. Historical air-traffic-system, weather and airline delay causes from BTS records are built in, along with route patterns by hour, weekday and month. Training and live predictions now share one feature pipeline, and the model output is calibrated on a separate validation period.
What's new in v9
3.95M
training flights
Where the data comes from
DelayGuard draws on official government aviation datasets and industry-standard live flight tracking feeds.
US Flight Performance Records
Training data · Historical
Over 15 million real US domestic flights (2023–2025) sourced from official BTS on-time performance reporting. The current v9 model learns its historical patterns from 2023 to mid-2025, trains on a sample of nearly 4 million of those flights, and is tested on late-2025 flights, as more recent data better reflects post-pandemic airline operations.
Official Aviation Weather Feeds
Live weather · Primary source
Current airport conditions, departure-time forecasts, and active hazardous weather advisories sourced from official government aviation meteorology services. Covers thunderstorms, turbulence, icing, low visibility, and wind.
Live Flight Tracking
Aircraft position · Real-time
Real-time aircraft position data from public ADS-B tracking networks. Used to determine whether the inbound aircraft for your flight is already delayed — one of the strongest predictors of a late departure.
Flight Status Data
Departure & gate status · Real-time
Live departure status and gate information from leading flight data providers, used to cross-check inbound aircraft status and ground delays. Falls back gracefully when unavailable.
Aircraft Registry
Fleet & age · Semi-static
Official aircraft registration data used to identify the age and type of the specific plane assigned to your flight. Aircraft age is a known factor in maintenance-related delays.
Airport Congestion Data
NAS & ground ops · Semi-static
Airport-level congestion patterns, average taxi-out times, runway capacity, and slot-control status derived from official aviation authority records. Helps the model understand systemic ground delay risk at each airport.
Multiple sources, one answer. DelayGuard queries several live data feeds in priority order and blends them automatically. If any source is temporarily unavailable, the model falls back gracefully to the next best signal — so you always get a prediction, even in degraded conditions.
What the confidence score means
The confidence ring tells you how much data was available and how reliable this specific prediction is.
High confidence
Live aircraft position confirmed, live NOAA weather available for both airports, route has strong historical data (500+ comparable flights), departure time falls in a high-coverage window.
Good confidence
Most live data available. Some signals missing (e.g. SIGMET data not covering the route, or tail number unknown). Prediction is reliable but based on partial live data.
Moderate confidence
Live data limited — prediction relies more heavily on historical patterns. This is common for late-night flights, smaller regional airports, or routes with limited historical coverage.
Low confidence
Minimal live data. Prediction is driven almost entirely by statistical averages for this route and airline. Treat with caution — actual conditions may differ significantly.
How fresh is the data?
Current airport weather conditions
Departure-time weather forecasts
Hazardous weather advisories
Turbulence & icing advisories
Pilot weather reports
Live aircraft position
Flight departure status
Prediction cache
What we can't predict
Honest about what the model does and doesn't know.
Crew & operational disruptions
Crew duty time limits, sick calls, and crew positioning issues are not visible in any data feed. These cause roughly 25% of carrier-attributed delays and cannot be predicted.
ATC decisions & ground stops
Real-time ATC flow programs and ground stops are called dynamically by the FAA. NOAA SIGMETs and G-AIRMETs signal elevated risk, but the model can't predict a specific ground stop call.
Unforecast weather events
Rapidly developing convective weather (pop-up thunderstorms) that weren't in the TAF at prediction time will affect accuracy. METAR and PIREP data helps, but nowcasting has inherent limits.
New routes & low-frequency flights
Routes with fewer than 100 comparable historical flights have less reliable priors. The model will signal lower confidence for these. Very new airlines or routes may have minimal training data.
International flights
DelayGuard is currently trained and calibrated on US domestic BTS data only. International flights (even from US airports) have much lower prediction reliability.
Same-day / <2h predictions
Predictions for flights departing within 2 hours are more reliable (live data is fresh) but the model was trained on pre-departure patterns — it does not incorporate real-time gate status.
Important legal notice
Important Notice — Beta Service
Predictions are estimates only. DelayGuard's predictions come from an ML model trained on 2023 to mid-2025 BTS data — nearly 4 million flight records — with live aviation weather feeds (NOAA). All outputs are statistical probabilities — not certainties.
Not a substitute for official information. Always verify your flight status directly with your airline or airport authority. Flight outcomes depend on real-time factors (ATC decisions, crew availability, gate changes) that no prediction tool can fully anticipate.
No liability for travel decisions. DelayGuard accepts no liability for losses, costs, or inconvenience arising from decisions made based on predictions shown in this application. This service is provided "as is" during its public beta phase.
Data accuracy & freshness. Live weather and flight data refreshes frequently — typically every few minutes for current conditions. Training data covers US domestic operations. Prediction accuracy may be lower for unusual weather events, irregular operations, or routes with limited historical data.
Jurisdiction & consumer rights. This notice is provided in accordance with applicable consumer information requirements, including EU Directive 2011/83/EU and US state consumer protection laws. DelayGuard is registered in the Netherlands and serves users globally, including the United States. This service is provided "as is" with no express or implied warranties of merchantability, fitness for a particular purpose, or accuracy, to the fullest extent permitted by applicable law. For questions or concerns, contact [email protected].