◉ CycleProofA small planning lab.
Demand, with doubt.
Historical demand planning / Washington, D.C.

Don't schedule for
a single number.

The trained models forecast every hour of a day with an uncertainty band. Replay a held-out day, then stress the forecast yourself in the What-if lab.

Offline · No account · No telemetry
Research replay, not a live bike service. Predictions are system-wide hourly rental counts from 2011-2012, not station availability or required fleet size.
ActualPredictionNominal 90% bandWorkload limit

Hourly review

Interactive scenario / live inference

Stress the forecast
yourself.

Pick a real day, then edit what the models see: the count 24 hours before and 168 hours before each target hour. The trained trees and neural network re-forecast every edit, in your browser, right now. This is where you find out how much the plan actually moves.

Lag inputs per target hour 0 edited

lag24: rentals at the same hour yesterday · lag168: same hour last week. Edited cells are highlighted. Guards: whole numbers 0-1000.

What this proves

The models are not a screenshot: change an input and the forecast, the band and the over-limit hours all change, because inference runs here on the exported weights. Nothing is sent anywhere.

The base day supplies the calendar features (hour, weekday, season, working-day and holiday flags). Editing lags explores sensor-style what-ifs; it does not make the model valid for another city or for today.

Your historical scenario / strict schema

Check the input before
you trust the output.

Import up to 744 historical target hours. Use lag counts from exactly the same hour one day and one week before. Import is local and never uploaded.

1. Prepare CSV

Plain unquoted UTF-8 CSV, exact column order. Dates must be in 2011-2012. No target, weather or count-component columns accepted.

2. Review the contract

  • hour: integer 0-23.
  • workingday and holiday: 0 or 1, no weekend or holiday marked working.
  • lag24 and lag168: nonnegative hourly counts, maximum 1,000 as a guardrail.
  • Unique target hours, valid dates, all values required.
  • Numeric checks do not prove lags are truthful or available at forecast time.
This model was trained on one historical city. Passing validation is not proof of suitability for current data, another city or a station.

Import preview

Validated predictions appear here.

Measured, not promised

Let the baseline compete.

A chronological split, no shuffled future rows, one untouched final test block. All values below come from the packaged training run.

Final test / October-December 2012

MAE and RMSE are rental events per hour. Bands are absolute calibration residuals from September, with a finite-sample quantile. Time dependence and drift mean 90% is a target, not a guarantee.

Where error hides

Weather is used only to slice errors after prediction, never as a feature.

Four rolling-origin checks

A fixed depth-3 tree is retrained before each 2012 quarter, compared with the weekly baseline. This is a separate stability check, not the final fixed-model test.

Largest misses / selected model

Neural reproducibility is not one lucky seed

Seed 7 was predeclared for the app. Seeds 19 and 31 are sensitivity runs, not selected after test performance. Two hidden ReLU layers, 48 and 24 units; not a temporal CNN or a pretrained model.

Local browser storage

Keep the decision trail.

Plans stay on this device and browser only. Export them for a durable copy. Clearing browser data deletes saved plans.

Model card / v1.0

Useful because
it shows its limits.

Intended use

Educational day-ahead demand replay, what-if scenario analysis and model comparison for students and community-mobility planners. Explore uncertainty before deciding whether more validation is worth doing.

Not intended for

Live inventory, safety decisions, automatic staffing, station allocation, financial commitments or claims about transport demand today.

Data and credit

UCI Bike Sharing, Hadi Fanaee-T (2013), DOI 10.24432/C5W894. Capital Bikeshare system-wide hourly counts, Washington D.C., 2011-2012. CC BY 4.0 per current UCI page; original README and citation preserved.

Official dataset and license

Fanaee-T, H. and Gama, J. (2013), Event labeling combining ensemble detectors and background knowledge. DOI 10.1007/s13748-013-0040-3.

Known failure modes

  • Growth, holidays and disruptions break yesterday/week patterns.
  • 165 missing timestamps; lag eligibility excludes additional rows. Evaluation is conditional on complete lags, not all hours.
  • October 29-30 has just one eligible row. This dataset cannot support a storm-performance claim.
  • September-calibrated intervals may under-cover the later test months.
  • System demand hides neighborhood inequity and station shortages.
  • Historical rentals are served demand, not unmet demand.

How it was built

Built during November 2026 for Peninsula Innovation Hackathon. AI assistance was used for concept development, code, tests, documentation and narration. Models were actually trained from scratch on the credited dataset using CPU scikit-learn. Browser inference is a transparent JavaScript reimplementation of the exported boosted trees and neural weights. No paid API, fake predictions, external inference service or hidden live data.

Next validation

Month 1: collect a permitted newer city dataset, test missing-data sensitivity, audit demand by neighborhood, use rolling calibration, and compare a calendar-aware seasonal model. Ask planners whether uncertainty changes their decisions before any operational pilot.

Training and evaluation protocol

Privacy and deletion

CSV and saved plans remain in your browser. No analytics, network requests or account. Closing the tab does not delete localStorage. Use Saved work to delete it. Never import private rider-level records.