Flight Delay Prediction using Machine Learning

Flight delays cost airlines and passengers billions every year, and they follow learnable patterns — monsoon months, evening banks, specific airports and carriers all shift the odds. The US DOT Bureau of Transportation Statistics publishes on-time performance data at massive scale, making this a rare student project with genuinely big, real-world data. Most student builds treat it as a plain classification toy; this project treats it as the dual problem it really is: classify delay vs on-time AND regress expected delay minutes, then decompose the predicted delay into the official BTS cause categories (air carrier, weather, national aviation system, security, late-arriving aircraft). An interactive web demo makes it tangible: enter a flight's details for a delay forecast, or watch a departure board of upcoming flights get scored and ranked. Everything uses standard tools (Python, scikit-learn/XGBoost), so the student can defend every choice in the viva.

1. BTS on-time performance records are loaded (2.1M-record working sample); cancelled and diverted flights are handled per documented rules.
2. Features are engineered: airline, origin/destination, month, day of week, cyclical departure-time encoding, distance, and historical route delay rates.
3. The data is split temporally (train on earlier months, test on later months) so the evaluation mimics forecasting the future, not shuffling it.
4. A gradient-boosted classifier predicts P(delay ≥ 15 min) and a regressor predicts delay minutes; both are tuned with cross-validation.
5. The classifier is evaluated with AUC and a confusion matrix; the regressor with MAE; calibration curves verify the probabilities.
6. The trained models are serialized and wired to the web demo, which scores flights and decomposes delays into BTS cause shares live.

Tech stack:

• Python 3, scikit-learn, XGBoost/LightGBM
• pandas, NumPy, Matplotlib
• Jupyter Notebook (training & evaluation)
• HTML5, CSS, JavaScript (demo)
• US DOT BTS on-time performance data

Dataset & model details

• Dataset: US DOT Bureau of Transportation Statistics — Airline On-Time Performance data; 2.1M-record working sample with scheduled/actual departure and arrival times, carrier, origin/destination, and delay-cause fields (carrier, weather, NAS, security, late aircraft).
• Task: Dual — (a) binary classification: delayed ≥ 15 min or not; (b) regression: expected delay minutes. Input = flight schedule + route features; output = delay probability + expected minutes.
• Model: Gradient-boosted trees (XGBoost/LightGBM) for both heads; cyclical encoding for time features; temporal train/test split (design target).
• Metrics: Classification AUC 0.81, regression MAE 9.7 minutes (design targets for the built-to-order training run). No metric is claimed as measured until the training run is executed for the order.

• [Delay predictor web demo] Enter airline, route, month, day and departure time to get delay probability, expected delay minutes and a likely-cause breakdown.
• [Departure-board mode] Score a list of upcoming flights at once, ranked by delay risk — mimicking an airline operations view.
• [Dual modelling] Classification (delayed ≥ 15 min or not) plus regression (expected delay minutes), each with its own metric.
• [BTS cause taxonomy] Predicted delay decomposed into the five official BTS cause categories with an honest, documented mapping.
• [Seasonality analysis] Delay-rate curves by month showing the monsoon peak — the kind of real insight examiners remember.
• [Calibration analysis] Predicted probabilities checked against observed delay rates so the demo's percentages are trustworthy.
• [Full training notebook] Data loading, feature engineering (cyclical time encoding), training, evaluation and calibration in one reproducible notebook.

• Complete training & evaluation Jupyter notebook
• Trained classifier + regressor model files with feature code
• Interactive delay-predictor and departure-board web demo
• Seasonality, calibration and cause-breakdown plots
• Project report PDF (background, ML theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (AUC, MAE, calibration, temporal splits, gradient boosting)

• 0.81 AUC / 9.7-min MAE are design targets for the training run, stated honestly — the report documents the actual achieved figures after training.
• The demo's cause breakdown is a modelled attribution over BTS categories, not the airline's official delay coding for a specific flight.
• Extreme disruptions (strikes, volcanic ash, system outages) are unpredictable outliers the model cannot foresee; the report says so.
• Trained on US BTS data patterns; Indian domestic delay patterns differ and the report notes the transfer caveat.
• The demo runs representative weights offline; the shipped models are trained on the BTS sample during the build.