BART pays for every car it runs, full or empty.

Rushover tells BART how many cars to run each hour. Same crowding standard, 5 to 12% fewer cars than a timetable.

Hourly crowds at all 50 BART stations for next week. Pick a station and an hour to see riders expected and cars needed.


    0 riders entering or leaving this station that hour

    Glow and rings: crowd this hour Trains on BART's timetable

    What it is worth to BART

    BART pays for every car it runs, full or empty. Capacity is planned as expected riders plus a safety margin, and the worse the forecast, the bigger the margin. A better forecast gives the same protection against crowding with fewer cars. Our first try, sizing cars from a forecast trained on average error, only tied a good timetable. Retraining on the cost of running short is what worked.

    The days a timetable cannot know

    Games, concerts and closures next week: what changes, why, and where BART needs more room. Each comes with its source.

    How the plan is made

    This shows how we turn riders expected at each station and hour into a car count. Each part is tested on data it never saw. The language models find and read events. They never write a rider count.

    1. It learns the ordinary day

      A gradient-boosted forecasting model trained on 222 million trips and hourly weather, 2021 to 2025. Walk-forward tested: each year is predicted by a model that never saw it.

      13.1% error in 2025, against 20.1% for repeating last week

    2. It corrects for what repeats

      Self-checking corrections for recurring conditions like Pride, football Saturdays and holidays. 23 were tried; only the 3 that beat the model on a held-out year were kept.

      40.2% → 26.8% error on Pride Sunday

    3. It reads the city

      An agentic web search finds next week's games, concerts and BART service notices near every station. Every claim must carry a source the system has opened; unsourced claims are dropped.

      0 events found for next week, each with a source

    4. It reasons about the exception

      A multi-agent harness, with Claude leading DeepSeek specialists, plans each event's crowd. Code holds the plan to curves measured at real past games and logs every override.

      17.2% → 9.6% error at Embarcadero's busiest Giants-game hour

    Leaving the station Entering Added by events Lightest nearby hour

    The whole week at a glance

    This shows riders expected at every hour this week. Brighter means more riders and less room to run. Pick a square to load it above.

    Games, concerts and closures

    This shows the events next week that call for longer trains, and how many riders each one adds. Riders are added from crowds measured at real past games.

    Tested on a year it never saw

    Error is the share of riders the forecast got wrong, across every station and hour of 2025. The model was trained only on earlier years.

    What this cannot tell BART

    1. There is no hourly data for 2026 yet, so next week's forecast can only be checked once the week has happened.
    2. Event crowds are measured on baseball games. For concerts and other events, the harness reasons from those curves within fixed limits.
    3. Gate counts show people entering and leaving a station, not how full a train car is.
    4. The weather input is itself a forecast, and it gets less certain several days out.
    5. Past game days sit inside the model's own history, which blurs the everyday baseline at stadium stations.