Cycle-hire planning research · Python

Cycle Hire Demand and Service Planning

A service planner needs to know which busy station deserves a closer check before the next operating shift. One, three and six-hour forecasts show expected departures and uncertainty; they do not decide where cycles should be moved.

May test resultThe selected model reduced error by 20.8% against the same-hour-last-week baseline across three notice periods.

Its 80% forecast range contained the observed total in 88.8% of final May forecasts. The ranges were deliberately cautious and wider than the observed outcomes required.

Official journeys used: 3,563,266
Journeys used in the study: 356,057
Station-hours modelled: 41,460
May forecasts tested: 26,784

The planning question

Where might departures run ahead of the current service plan?

The operations analyst prepares the forecast for a service planner and station operations lead. A high predicted total is not enough: they also need to know how far ahead it was made, how wide the likely range is and whether the station has a history of large misses.

Hyde Park Corner has the widest mean one-hour range in the May watch list. That makes it a place to check live bike availability and nearby docks. It does not prove a shortage or justify a rebalancing move by itself.

Error by notice

More notice gives the planner more time, but less accurate demand.

The selected model remains ahead of the weekly baseline at every horizon. Mean absolute error rises from 2.14 departures at one hour to 2.32 at six hours, so the page never mixes forecasts made at different times.

1 hour24.7% better than the weekly baseline

2.14

89.0%

3 hours19.1% better than the weekly baseline

2.30

88.6%

6 hours18.5% better than the weekly baseline

2.32

88.7%

Selected modelSame hour last week

Station review field

Wide ranges and repeated misses determine where review starts.

The watch list combines observed May demand, forecast error and mean range width. Amber is reserved for uncertainty that needs attention; it is not a performance rating.

01
Hyde Park Corner, Hyde Park
Wide forecast range; check availability before planning
3.18 MAE
02
Waterloo Station 3, Waterloo
Wide forecast range; check availability before planning
2.98 MAE
03
Waterloo Station 1, Waterloo
Wide forecast range; check availability before planning
2.58 MAE
04
Argyle Street, Kings Cross
High forecast demand; compare with nearby stations
2.09 MAE
05
Hop Exchange, The Borough
High forecast demand; compare with nearby stations
2.11 MAE
06
Wormwood Street, Liverpool Street
High forecast demand; compare with nearby stations
1.70 MAE

One station-day

Follow the Waterloo morning peak from the earlier records to the planning review.

The example follows Waterloo Station 3 at 08:00 Europe/London on 13 May 2026. It records 152 departures. The one-hour forecast was 161.3, compared with 108 departures at the same hour a week earlier. The unusually wide range is shown alongside the point estimate rather than being hidden by it.

Waterloo Station 3 · 13 May 2026Each point shows when the forecast was issued, its expected range and the departures that followed.

00:001

01:000

02:000

03:000

04:000

05:001

06:0020

07:0071

08:00152

09:001

10:001

11:000

12:001

13:001

14:002

15:001

16:003

17:004

18:004

19:001

20:000

21:004

22:000

23:000

80% forecast rangeForecastActual departuresWeekly baseline at 08:00

Same hour one week earlier: 108
One-hour forecast: 161.3
Observed departures: 152
Absolute error: 9.3
80% forecast range: 0.0–169.3

Follow one forecast

Every value can be followed from counted journeys to planner review.

01
station-hour observation
Counted from unique TfL journey numbers
Review role: Operations analyst
152 departures
02
forecast issue
Only earlier observations available
Review role: Operations analyst
2026-05-13 06:00:00+00:00
03
168-hour seasonal lag
Same station and hour one week earlier
Review role: Operations analyst
108
04
selected forecast
80% interval 0.0 to 169.3
Review role: Model reviewer
161.3
05
error review
Flagged for station-day review
Review role: Station operations lead
9.3 hires
06
planner action
Check live availability, nearby docks and operating constraints before intervention
Review role: Service planner
Human review required

The six stages are saved in one-station-day-trace.csvand can be reproduced with npm run verify:cycle-hire-demand. The final step requires the service planner to check live availability, nearby docks and operating constraints before acting.

Time-ordered evaluation

May remains outside training, and every demand feature stops at issue time.

The selected boosted model is compared with a weekly baseline and regularised Poisson model. Selection uses error, peak-hour error, high-volume error and interval coverage rather than a single score.

Ordered splits3 time-ordered tests · May kept for the final test

Training ends 30 April 2026. The final test covers 1 to 31 May, and the records remain in time order.

outputs/model-evaluation.json

Future-data checkLater observations are excluded

Demand, arrivals and net movement are shifted by the forecast horizon; station selection uses January to March only.

tests/test_pipeline.py

Model cardgradient boosted quantile

The model card records the intended decision, error costs, abstention rule, monitoring triggers and retraining conditions.

outputs/model-card.md

Rebuild checkidentical inputs produced identical analytical output files

Artifacts checked: 5. Using the same locked inputs reproduced the Parquet layers, evaluation results and Waterloo predictions byte for byte.

outputs/repeatability-results.txt

Data, period and limits

A forecast shows where to check, not that cycles will be unavailable.

The study uses ten checksum-verified TfL journey extracts from 1 January to 31 May 2026 and a reference list of twelve selected stations.

TfL journey records

Powered by TfL Open Data under the Transport Data Service Licence. The journey extracts contain completed hires; no client or live operating data is used.

What is missing

Completed hires do not show attempted hires, historic bike stock, rebalancing, closures, maintenance or capacity changes. Weather is excluded because equivalent records were not available at the time each forecast was made.

What this does not establish

Twelve busy stations do not represent the whole network. This is a historical test on unseen data. It does not demonstrate live performance, a causal effect or an automatic rebalancing instruction.

See how the analysis was produced

The figures on this page are produced by the Python and DuckDB analysis.

The build checks ten official file hashes, creates typed journeys and complete station-hours, fits three model families across three rolling origins and writes the page figures from the final May forecasts.

Generated analysis report · outputs/native-report/index.html · 7 analytical checks passed

Technical detailsView the data licence, model comparison, rebuild steps and analytical limits

Build

Python 3.12, DuckDB, Parquet and scikit-learn run from a locked environment. Five analytical output files reproduce byte for byte when the same locked environment is rebuilt.

Evaluation

Three expanding time splits compare a weekly baseline, regularised Poisson model and boosted point and quantile forecasts. May 2026 is reserved for the final test and is not used to train or select the model.

Where the data comes from

Powered by TfL Open Data under the Transport Data Service Licence. The study uses ten official extracts from 1 January to 31 May 2026.

Limits

Completed hires do not reveal unmet demand, historic stock or rebalancing. A planner must check live availability and operating constraints before acting.

Discuss a forecasting problem

Start with the planning decision, then test whether the forecast is good enough to support it.

We can begin with the action being planned, the available records, how far ahead the forecast is needed and the cost of getting it wrong.

Discuss a forecasting question