RTA Future Move Hackathon, with Dataiku
The film is played from the presenter's laptop. The day itself starts below.
RTA Future Move Hackathon, with Dataiku
The timetable said 19:23. The bus left at 19:36. Ride down to see why.
01 · The day
card taps on Dubai's buses in one day
Every bus and every stop, on Monday 15 December 2025. This is the day we studied.
Measured on the real day02 · The gap
of buses leave late at 7 in the evening
Same buses, same drivers, same timetable. Late means 5 minutes or more. Only the hour changed.
Measured on the real day03 · One bus, one evening
The trip took 96 minutes. The timetable gave it 67.
Route 310, 6 in the evening. The bus arrives 29 minutes late. Its 16-minute break is gone, so its next trip leaves 13 minutes late.
Illustration built from measured figures04 · The delay travels
late departures: the bus was late back from its last trip
Of 936 late departures we could trace, 870 came after the same bus finished its last trip behind time. By a stricter count, in 589 of 945 the bus was still out when its next trip was due.
Measured on the real day05 · The time that is missing
of driving time the evening needs and the timetable does not give
Added up over every route, from 4 to 7 in the evening. Each pillar is a route; red means too few buses at 6 pm by RTA's own planning rule.
Measured on the real day06 · The forecast
Our forecast of the trip time is closer than the timetable, from 16:00 to 19:00
Made 30 minutes before the bus leaves. Tested on 2,043 evening-peak trips it had never seen. After 19:00 the timetable is as close or closer.
Measured on the real dayThe problem
Someone at the stop planned their evening around 19:23. The driver did nothing wrong. The bus was still finishing its last trip, because that trip took 96 minutes and the timetable gave it 67.
The timetable gives each route the same trip time all day. In the evening the trip takes longer, so buses fall behind, and stay behind.
Illustration: clock times worked out from route 310's measured figures at 18:001 · It is not more passengers
48,724 vs 49,707But late buses go from 19% to 45%.
2 · The trip takes longer
67 → 96Route 310 at 6 pm, a typical trip.
3 · The break is too short
16Not enough to win back 29.
4 · So the next trip starts late
870 of 936In 589 of 945 it was still out when the next trip was due.
Which routes, at which hours, need more time, and how many minutes?
The four figures: measured on the real dayWhy it is not the crowd
The two peaks carry almost the same number of passengers. The morning copes. The evening does not. So the cause is not demand: it is how long the trips take against what the timetable allows.
Read from the Dataiku flow: demand_by_hour and punctuality_by_hour. Hours with fewer than 500 departures are left off the line. Times after midnight belong to the same service day.
Measured on the real dayWhere the timetable is short
For each route and hour we compare two numbers: the trip time in the timetable, and the trip time that would be enough for 85 trips out of 100. Add up the difference and you get the driving time the timetable is missing. Across the whole day it is bus hours.
Read from the Dataiku flow: running_time_ledger_by_hour. A second analysis by a teammate, with other code, finds 569 bus hours for the day.
Measured on the real dayOur answer
A forecast
30 min15% closer than the timetable, 16:00 to 19:00, on an evening it never saw.
A worklist
244Measured from the day, not predicted. The planner accepts or sets aside each row.
A second ranking
695Where lateness reaches the most passengers.
What a planner gets
The routes and hours where the plan falls furthest short, and the minutes each one needs. The planner accepts or sets aside each row. Nothing applies itself.
Read from the Dataiku flow: running_time_bands, rows with 5 or more trips, ranked by extra minutes times trips. In the Dataiku web app this is the "Planner's worklist" tab, with Accept and Set aside on each row.
Measured on the real dayThe forecast
Gradient-boosted trees, trained on the trips before 16:00 and tested on an evening they had never seen. Average miss in minutes; lower is better.
Read from the Dataiku flow: running_time_metrics, trips due from 16:00 to 18:59. MAE is the average miss; RMSE weights large misses more; MAPE is the miss as a share of the trip.
Most of the gain comes from knowing how the route ran earlier that day.
A bus's previous trip has not finished when the forecast is made.
It is an evening-peak tool, and we say so.
Every number comes from Monday 15 December 2025.
One more thing the day shows
Share of departures 5+ minutes late, by type of bus. This is an association, not a cause: the routes each type runs are not the same, and that is not controlled for here.
Read from the Dataiku flow: punctuality_by_bus_type, types with 2,000 or more departures.
Measured on the real dayHow it's built
Eighteen steps in five zones: twelve visual recipes and six short Python recipes. One rebuild from RTA's raw tables takes 131 seconds. The web app's tabs carry the same names as the zones.
Zone 1
Who boards, where, and when.
Zone 2
Does the bus leave on time?
Zone 3
Were enough buses sent?
Zone 4
How long will each trip take?
Zone 5
Where does lateness reach the most passengers?
Who uses it
RTA changes timetables every one to two months and routes four times a year. Each night adds a day of records. Before each change, a planner opens the worklist, sees where the evening runs long, and decides how much time to give it. Boardings, never people "waiting". Nothing about one card, one driver or one vehicle is shown.
About 244 bus hours of extra running time from 16:00 to 18:59 would cover 85 trips in 100. About 630 across the whole day.
Opens the planner's web app on the real day. It asks for a Dataiku sign-in.