All projects
Machine Learning/MLOps & delivery /Project 03 of 10 M.S. coursework

Flight Delay Prediction Pipeline

A reproducible regression workflow for estimating average departure delay, packaged with versioned data, tracked experiments, and a containerized prediction API.

Python MLOps DVC FastAPI Docker MLflow CI/CD M.S. coursework

Business impact

An airline planner can compare average Atlanta departure delays by destination and scheduled times. The evaluation excluded departures delayed by an hour or more.

8.92 minHeld-out test RMSE
79.65Held-out test MSE
3.8Selected Ridge alpha

Business question

How accurately can a planner estimate the average departure delay for a given Atlanta route and schedule?

Approach

How the work was done

  • Prepared 22,580 cleaned Atlanta flight records. The CSVs are committed directly; DVC pointers record checksums, with no remote configured.
  • Compared Ridge regression settings across 20 alpha values; features include destination airport and scheduled departure and arrival times.
  • Tracked experiments in MLflow and served predictions through FastAPI in Docker, with automated GitHub Actions and unit tests.

Results

What the analysis shows

  • The selected Ridge alpha was 3.8; held-out test RMSE was 8.92 minutes and MSE was 79.65.
  • The pipeline logs a mean predicted delay of 4.48 minutes. A fixed split seed and an explicit refit of the selected Ridge model support reproducibility.
  • The API returns a route-and-schedule average, evaluated on later dates after grouping by destination and scheduled times. Individual flights and major disruptions can depart much later.

From the project

Where the delay estimates miss

Line chart with actual delays in blue and predictions in orange, with actual values varying more widely.

Predicted and observed average delays

The held-out evaluation plot shows predictions concentrated near typical delays while observed averages have larger spikes. The horizontal axis lists evaluated samples. The held-out RMSE is 8.92 minutes.

View the source Open full size

Business relevance

What this makes easier to decide

This project demonstrates how a prediction can move through data preparation, experiment tracking, packaging, and an API interface. The reported test error helps set expectations for how precise a schedule estimate may be.