Free interactive course · real open data

Learn machine learning with real transit data

Nine short chapters take you from a plain average to a neural network. Every model is trained live in your browser, on real reliability figures published by the MTA, Transport for London and SNCF. Each chapter shows how a model works, then why you need the next one.

Pick a network. The course is the same; the data, the numbers and the story change.

New York

The NYC subway, A train

Will the A train be on time?

on-time performance 69.4% on averagelatest 77.9% (2026-08)140 months, 2015-01 to 2026-08Start the course →London

The London Underground, Central line

How much of the Central line's timetable will actually run?

service operated 88.5% on averagelatest 84.0% (2026-27 P6)109 periods, 2018-19 P1 to 2026-27 P6Start the course →Paris

The Paris RER B

Will the RER B get me there on time?

punctuality 85.2% on averagelatest 90.1% (2026-08)159 months, 2013-01 to 2026-08Start the course →

What you will learn

  1. The baseline: why the average is the model to beat, MSE versus MAE.
  2. Linear regression and gradient descent, with a live cost map and learning rate.
  3. Polynomials, bias and variance, overfitting, regularisation and why forecasting is hard.
  4. Logistic regression, the sigmoid, the threshold, precision and recall.
  5. k-nearest neighbours and the curse of dimensionality.
  6. Decision trees, Gini impurity and instability.
  7. Random forests and gradient boosting.
  8. Neural networks, activations and backpropagation.
  9. A showdown of eight models on the same data.

How it works

  • No install, no account: everything runs in the page.
  • Chapters 1 to 3 use each network's real published data.
  • Chapters 4 to 9 use small simulated datasets, so each model's behaviour is easy to see.
  • Each chapter ends with a recap card and a quick quiz.