Exploring over 21 million historical trips, customer behavior, commuting patterns, station demand, and the limits of annual forecasting.
Loading dashboard data…
This interactive report extends the original Divvy analysis with year-specific views, metric selection, station rankings, route relationships, and a transparent forecasting backtest.
Divvy provides shared bicycles across the Chicago area. Historical trip records can reveal when demand concentrates, which stations are heavily used, and how Subscribers and Customers differ. The report analyzes trips rather than unique people because the dataset identifies rides, not distinct riders.
Subscribers use the system more frequently than Customers, particularly during weekday commuting hours. We examine this through trip counts and time-of-day distributions; repeated rides cannot establish individual rider frequency.
Share of recorded trips by user category, based on the selected reporting year.
Customer and Subscriber are trip categories, not unique rider counts.
Explore year, month, weekday, and hour. Switch between demand and trip duration, and compare rider categories without adding multiple static charts.
Use the controls to explore historical activity.
Compare weekdays and start hours to test the commuting hypothesis; evaluate customer and subscriber distributions separately rather than inferring each rider’s individual behavior.
Compare recorded trips and trip duration by gender or age band across reporting years and rider segments.
The demographic categories are those recorded in the historical data; an “Unknown” group is retained where applicable.
Demographic figures describe trips attributed to recorded categories. They should not be interpreted as a census of distinct riders or of all Chicago residents.
Rank departure or arrival stations using station ID as the counting unit, keeping historical name differences from artificially splitting the same ID.
Rankings adjust to the report year and rider segment selected at the top.
Explore the leading ordered origin–destination pairs, then inspect their approximate station-to-station links on a map. Lines show connections, not actual cycling paths.
Top 20 routes are precomputed per selected year and rider category.
Showing approximate links between station coordinates for the selected routes.
Map requires an internet connection to load map tiles.
Ordinary least-squares linear regression trained on annual trip totals from 2015–2018. The 2019 observation provides one backtest; 2020 is a historical forecast, not observed trip data.
Switch between rider segments to inspect model performance and differences in trend.
The original study’s commuting hypothesis motivates targeted operational questions. This expanded dashboard lets visitors test those patterns directly using historical trip data.
Use weekday and hourly demand profiles to schedule bicycle redistribution and assess whether peak-hour supply needs differ for Subscribers and Customers.
Compare origins and destinations before prioritizing station capacity. Review low-usage stations alongside their years of operation rather than judging rankings in isolation.
Use the annual model as a transparent baseline rather than a production demand forecast. Segment-level errors in 2019 indicate that different riding behaviors require more tailored models.
Historical trip-level dataset (2013–2019; 21,234,782 records). Python and DuckDB aggregate trip activity by reporting year, trip start month, weekday, hour, gender, age band, user category, and station ID. The browser consumes small precomputed JSON datasets, not raw trip records. Station map lines connect representative coordinates; annual forecasts use linear regression.
Trip totals are not unique-user totals. Demographic fields may not cover every ride equally. Station names and positions can vary historically. The precomputed routes file contains only the 20 most popular ordered pairs per filter, not the entire network. Monthly, weekday, and hourly groups summarize trips in each selected year; they are not continuous daily time series. A four-year training window limits forecast reliability.
