Flight Passenger-Count Prediction

Knowing how many passengers a flight will carry sits underneath almost every airline decision: which aircraft to assign, how many seats go to which class, what the price should be. Both directions of a wrong forecast are expensive - a seat flown empty, and demand that couldn't be met.

This forecast has traditionally rested on experience. But on one-stop routes the number of variables goes beyond what anyone holds in their head: carrier, aircraft type, the origin-hub-destination triple, sell and cabin class, departure day and time.

What was needed was a model that weighs those variables together and produces a concrete number per flight.

I built a regression model predicting passengers per flight on worldwide one-stop booking data.

1. I shaped raw operational data for modelling. The data is largely categorical: carrier code, aircraft type, the airport triple, sell and cabin class. These were encoded, and features derived from departure date and time - day of week, month, time band - because demand does not behave the same on a Tuesday morning as a Friday evening.
2. I transformed a skewed distribution. Passenger counts are right-skewed: many low values, a few very high ones. A log transform was applied to the target so the model learned the whole distribution rather than being dragged by a handful of large flights.
3. Eleven models raced under identical conditions. Linear regression, Ridge, Lasso, ElasticNet, KNN, decision tree, random forest, gradient boosting, XGBoost, LightGBM and CatBoost - all measured by RMSE under 10-fold cross-validation. Model choice rested on results, not preference.
4. The winner was tuned. LightGBM, which performed best, was refined through a GridSearch hyperparameter sweep, with final performance re-reported under cross-validation.

The result gives planning teams a numerical demand forecast per flight instead of an experience-based estimate - and, more importantly, makes visible which characteristics actually drive demand.

Stack: Python, pandas, scikit-learn, XGBoost, LightGBM, CatBoost; categorical encoding, date-time feature engineering, log target transform, 10-fold cross-validation, GridSearchCV.