← All projects

Live demo · City of Melbourne pedestrian sensors

Melbourne foot traffic: the Metro Tunnel, and forecasting a week ahead

Loading the results…

Did the Metro Tunnel change the streets around it?

Each dot is a footpath sensor, coloured by how its weekday 7–10am count changed from to : blue busier, orange quieter, bigger dots for bigger changes. Dashed rings mark m around each new station entrance. Select a dot for details.

  • busier on weekday mornings
  • quieter on weekday mornings
  • not enough working days to compare
  • Metro Tunnel entrance
  • m from an entrance

Median change per site, vs . Near: within m of an entrance. Far: m or more. Permutation p-values over site labels.

Forecast explorer

A gradient-boosting model forecasts each sensor's hourly count up to a week ahead, using only what's known a week in advance. Pick one of the busiest sensors and a week of the test window to compare it with the actual count and with the naive "same hour last week" forecast.

    WAPE = total absolute error divided by total actual count. *Uses observed weather as a stand-in for a weather forecast, so it's an upper bound on what weather adds.

    The model knows a public holiday is coming, but not which kind. Gap = actual vs forecast.

    How it works

    Metro Tunnel comparison

    • Compares with the same dates a year earlier, so seasonality cancels out.
    • Sites within m of a State Library or Town Hall entrance ("near") against sites m or more away ("far"). Sensors under m apart count as one site.
    • A permutation test over site labels asks whether the near/far gap could be chance.

    Week-ahead forecast

    • Gradient boosting (scikit-learn, Poisson loss) trained on , tested on : sensors, sensor-hours.
    • Features: hour, day of week, month, Victorian public holidays, and the same hour weeks earlier.
    • Scored against "same hour last week" and "average of the last 4 weeks".

    Messy sensor data

    • The feed leaves out empty hours, so a missing hour on a working day counts as zero pedestrians.
    • A day missing any hour between , or with a zero total, is a sensor outage and is dropped.
    • Indoor sensors (library and community-hub doors) are left out.

    Caveats: these are footpath sensors, so they measure street activity, not station ridership. There's no measurable overall Metro Tunnel effect yet, and one-off events like the AFL Grand Final break the forecast.

    Files