Capital Bikeshare
MAY 2, 2015
GEORGETOWN UNIVERSITY
SCHOOL OF CONTINUING STUDIES
CERTIFICATE IN DATA ANALYTICS
CAPSTONE PROJECT
SELMA ORR
RYAN DONAHUE
ODETTE RIVERA
NORA GOEBELBECKER
KARINA HIDALGO

Problem Statement
Cities want to build bike systems for economic development and sustainability, but they face
serious fiscal constraints.

Problem Statement
As a result, the public has little tolerance for error – even though usage, and therefore
profitability, is highly variable.
Most popular bikeshare station, 2014:
Union Station – 131,700 trips
Least popular bikeshare station, 2014:
34th St & Minnesota Ave SE – 112 trips
For roughly the same fixed costs, there are more than
1,000 times as may riders each year using the Union
Station location as many others.

Currently, there is no standardized, rigorous methodology for accurately predicting which
stations will be most heavily used.
Problem Statement

Goal
Develop a model, based on Washington DC but applicable to other
U.S. cities, that will predict the popularity of bikeshare stations based
on characteristics of the area surrounding each station.
Such a model could be used to increase the popularity of new and
existing bikeshare systems, making them more financially
sustainable.

Hypotheses
• Bike share station popularity is influenced by station location– specifically, the
economic, demographic, and geographic characteristics of surrounding
neighborhoods.
• Certain determinants of bikeshare station popularity hold true across cities,
allowing for the construction of a model that could accurately predict the
popularity of bikeshare stations in other cities.
•Regression models, such as linear regression, are a good fit to predict station
popularity.

Project Background
This project builds primarily upon two previous analyses of Washington, DC:
1. Maximizing Bicycle Sharing: An Empirical Analysis of Capital Bikeshare Usage
• Multivariate regression to identify five factors that influenced bikeshare station popularity:
population 20-39, non-white population, retail proximity, Metro proximity, and distance from
system center.
2. Predicting the Popularity of Bicycle Sharing Stations: An Accessibility-Based Approach Using
Linear Regression and Random Forests
• Linear regression and random forest analysis to understand how job and residential proximity
influenced station popularity. Although the study attempted to extend the model to San Francisco
and Minneapolis, it found that the model was a poor predictor of station popularity in those cities.
Why revisit this topic?
• Our team identified other characteristics that might drive usage.
• The bikeshare system has expanded considerably since those studies, offering a larger and more
varied sample.

Data Sources and Ingestion
Data Variable Type Source Year Geography Format
Bikeshare trips Dependent Capital Bikeshare 2014 N/A CSV
Bikeshare stations Dependent DC Open Data 2014 Point CSV/shapefile
Population- age Independent U.S. Census/ACS 2013 Block Group CSV
Population- race Independent U.S. Census /ACS 2013 Block Group CSV
DC liquor license locations Independent DC Open Data 2014 Point CSV/shapefile
DC Metro stations (train and
bus)
Independent WMATA 2014 Point CSV/shapefile
Parks (DC, National) Independent DC Open Data 2014 Polygon CSV/shapefile
Campuses (college, university) Independent DC Open Data 2014 Polygon CSV/shapefile
Historic landmarks Independent DC Open Data 2014 Polygon CSV/shapefile
Starbucks and McDonald’s
locations
Independent Various online
sources
2014 Point CSV

Data Sources and Ingestion
Data Variable Type Source Year Geography Format
Total trips per bike per year Dependent Capital Bikeshare 2014 N/A CSV
Bikeshare stations Dependent DC Open Data 2014 Point CSV/shapefile
Population- age Independent U.S. Census/ACS 2013 Block Group CSV
Population- race Independent U.S. Census /ACS 2013 Block Group CSV
DC liquor license locations Independent DC Open Data 2014 Point CSV/shapefile
DC Metro stations (train and
bus)
Independent WMATA 2014 Point CSV/shapefile
Parks (DC, National) Independent DC Open Data 2014 Polygon CSV/shapefile
Campuses (college, university) Independent DC Open Data 2014 Polygon CSV/shapefile
Historic landmarks Independent DC Open Data 2014 Polygon CSV/shapefile
Starbucks and McDonald’s
locations
Independent Various online
sources
2014 Point CSV
Data Variable Type Format
Amenities within walking
distance from bikeshare
stations (sum)
Independent CSV
Distance from bikeshare
station to each type of
amenity (closest amenity
within walking distance of
station-- -.5 miles or less)
Independent CSV
Socioeconomic characteristics
of the population sharing the
same census block group as
the bikeshare station
Independent CSV/shapefile

Capital Bikeshare Data
• Data attributes for bike trip: Trip Duration, Start/End Station (address), Start/End Date Time,
Bike Number, and Member Type
• Divided by time period: A year of trip data downloaded from the Capital Bikeshare website was
divided in four files with each file representing a quarter.
•Separate dataset with coordinates of bikeshare station’s locations was obtained from DC Open
Data website
• How to measure popularity? Trips leaving, trips arriving, total trips? Different capacity at each
station, so popularity = total trips (arrive + depart)/bike/year

Census Data
• Socioeconomic and demographic data collected for all Census blocks within DC, in the form
of CSV files downloaded from American Factfinder
• Census blocks are the smallest geographic area for which sample data is collected (typically
600 to 3,000 residents)
• Challenge 1: how to link stations (discrete points) with block groups (boundaries)?
oSolution: ArcGIS was able to identify block groups by lat/long of bikeshare station
• Challenge 2: how to deal with missing data?
oFor missing rent (4 instances): impute by calculating average rent/income ratio across city
oFor missing income (2 instances): impute by averaging two adjacent block groups
oFor missing population (10 instances, national park areas): leave blank

Nearby Amenity Data
• Selection of amenities based largely on past studies and common-sense drivers
of bikeshare usage: metro and bus stations, college campuses, DC and national
parks, entertainment, restaurants, bars (proxied by liquor licenses)
• Two ways to think about importance of amenities as drivers of usage:
o How close the single closest location of each type of amenity is (likely most
important for metro stations)
o How many locations of each type of amenity there are within a half mile
(likely most important for restaurants, bars)
• One challenge: whenever there wasn’t a single location of an amenity within a
half mile (common result for metro stations), ArcGIS identified distance as “0”

Data Wrangling
The primary challenge in the data
wrangling process was to create an
architecture that links each individual
station with its census block group
(and associated socioeconomic
characteristics) as well as with
distances to surrounding physical
amenities.
Census/Bike
Station.csv
Bike
Station
Address
Census
Block
Group
Amenity
Proximity

Spatial Analysis

Spatial Analysis

Spatial Analysis
Union Station 34th and Minnesota Ave SE

Data Wrangling
All four files with quarterly
ridership data were loaded into a
PostgreSQL table to be merged to
the Census/Bike station.csv file to
add the stationID, latitude and
longitude of the stations.
The resulting file contained all the
trips in 2014 with reference
information about the station and
created a common field uniquely
linking the records between the
files.
BikeStation
_Census.csv
DCBikeTrips2014.csv Census_BikeTrip.csv

Data Exploration
Correlated dependent variables with one another to anticipate collinearity. Shown here are
metro station proximity vs. density, and bar proximity vs. single households.

Data Exploration
A few variables that appear to have little correlation to station popularity: metro station
proximity and % of households in Census block that commute by public transit.

Data Exploration
A few variables that seem to have correlation to station popularity: % of population in Census
block that drives to work, and % of population in Census block with a college degree.

Data Exploration: Correlations with y
-0.600
-0.400
-0.200
0.000
0.200
0.400
0.600
Correlation

2014 Capital Bikeshare Member Survey
“Compared with all commuters in the region, they were, on average, considerably younger,
more likely to be male, Caucasian, and slightly less affluent.”
“Two-thirds (64%) of respondents said that at least one of the Capital Bikeshare trips they made
last month either started or ended at a Metrorail station and 21% had used bikeshare six or
more times for this purpose. About a quarter (24%) of respondents used Capital Bikeshare to
access a bus in the past month.”

Data Exploration: Excerpt from cross-
correlation matrix

Study Methodology: Feature Selection or
better said “Feature Wrangling”
Step 1 : Ran a full Ordinary Least Squares Regression with all thirty independent variables using StatsModels in Python.
◦ R-squared: 0.636
◦ Adjusted R-squared: 0.577
◦ Eight variables with significant p-values included: DRIVE, WHITE, DENSITY, AGE, WALK, BUS_N, TRANSIT, MCDON_N
◦ As expected, very large conditions number, 1.11e+06 indicating strong multicollinearity
◦ F-statistic: 10.73 and Prob(F-statistic): 3.42e-25
◦ Designated this as Model 2
Step 2 : Beginning with the variable DRIVE, the variable with the highest linear correlation to y, sequentially added the
other variables to the OLS regression according to correlation
◦ Any variable that triggered a multicollinearity warning was left out
◦ Any variable without a significant p-value was left out
◦ Seven variables with significant p-values included: DRIVE, WHITE, LIQUOR_N, BUS_N, MCDON_N, SINGLE, CAMPUS_N
◦ Designated this as Model 1
◦ Note: Experimented quite a bit with k features module but this adds features sequentially using descending correlations, but does not
take account of multicollinearity

Definitions of relevant features
Census:
DRIVE: share of population in Census block that drives to work (Model 1 and Model 2)
WALK: share of population in Census block that walks to work (Model 2)
TRANSIT: share of population in Census block that takes transit to work (Model 2)
WHITE: white share of population in Census block (Model 1 and Model 2)
DENSITY: population density in Census block (Model 2)
AGE: median age in Census block (Model 2)
SINGLE: share of households in block group that are single (i.e., non-family) (Model 1)
Amenities:
BUS_N: number of bus stations within half mile (Model 1 and Model 2)
MCDON_N: number of McDonald’s within half mile (Model 1 and Model 2)
LIQUOR_N: distance to nearest establishment with a liquor license (Model 1 and Model 2)
CAMPUS_N: number of college campuses within half a mile (Model 1)

OLS Output for Model 1

Model 1 Results: y -Actual versus y-Pred

OLS Output for Model 2

Model 2 Results: y -Actual versus y-Pred

Study Methodology: Machine Learning
Step 1 : Selected the following regression types for Machine Learning on Model 1 and Model 2
◦ OLS
◦ Ridge
◦ RidgeCV
◦ Lasso
◦ LassoCV
◦ Decision Tree
◦ Random Forest
Step 2 : Prepared the data
◦ Because there were only 201 stations (rows of data) opted against the K-fold cross-validation
◦ Used Repeated Random sub-sampling validation with 20% splits for testing and 80% splits for training
◦ Iterated for each regression type for n=15 times and averaged the results for the 15 trials

Study Methodology: Machine Learning
Model 1 R-SquaredAverages
Ex. 1
R-Squared Averages
Ex. 2
R-Squared Averages
Ex. 3
OLS 0.466372594903 0.472123692839 0.399698146173
Ridge 0.469910211714 0.46817885095 0.406793762824
Ridge CV 0.466525097058 0.472095157445 0.399928079436
Decision Tree (depth = 2) 0.383743276391 0.395203148425 0.359422535555
Decision Tree (depth = 5) 0.396411487916 0.399304877207 0.396627898939
Lasso 0.213675679287 0.192417630297 0.20648584883
Lasso CV 0.466325652215 0.472022317547 0.399672003362
Random Forest 0.513764682165 0.521753949359 0.510745984545

Model 1 Random Forest Results:
y-Actual versus y-Predict

Study Methodology: Machine Learning
Model 2 R-Squared Averages
Ex. 1
R-Squared Averages
Ex. 2
R-Squared Averages
Ex. 3
OLS 0.520828019361 0.540831229205 0.513422473234
Ridge 0.497919255131 0.516490091569 0.506607344645
Ridge CV 0.516873054267 0.54031441364 0.515242580429
Decision Tree (depth = 2) 0.32349467429 0.311292315645 0.285600505202
Decision Tree (depth = 5) 0.353623405764 0.435172086349 0.350751443572
Lasso 0.191216516526 0.217208742758 0.196249828561
Lasso CV 0.52129767449 0.541115629648 0.506104546918
Random Forest 0.468915321109 0.533377083505 0.42975292961

Model 2 Results Ridge CV:
y -Actual versus y-Pred

Model 2 Results Lasso CV:
y -Actual versus y-Pred

Data Product
 Groundwork: By analyzing the correlation between the factors such as bikeshare’s stations location, geographic and
demographic information, we obtained results that allow us to create a data product that predicts the likelihood of
success or failure of a new Bikeshare station prior to implementation in the DC area
 Results: Our models succeed in explaining about half of the variance in the Bikeshare Station popularity as measured by
our utilization factor. This should at least help in predicting the potential popularity of a station based on the
combination of demographic and geographic factors we identified as significant.
 Further applications: In addition to identifying promising locations for new Bikeshare stations in DC, the results may also
generalize to other cities. By using data on the demographic and geographic factors we identified as significant, it could
allow a user to predict promising locations for bike stations during an initial roll-out ,thus enhancing the overall success
of a new project without costly experimentation

What worked?
GIS was critical to creating effective architecture: linked stations to amenities by distance, and
to Census blocks and associated data.
Since the data volume that we handled was small, using local machine rather than a powerful
data base or cloud environment helped to achieve faster results.
Spending time exploring the variables before beginning analysis. This, plus domain knowledge,
allowed team to identify and address data issues manually that the software didn't calculate
accurately from the beginning.
Trying many different types of regressions using different variables (including different forms of
independent variable – log, natural log).

What didn’t work? And lessons learned.
Data sample was relatively small – started at roughly 350 stations, but shrank to roughly 200
once MD and VA locations were removed from the analysis.
Data and feature wrangling take a long time (80% of the process); domain expertise makes this
easier.
Would have been harder to detect and address anomalies and missing values in data had the
sample been larger; familiarity with DC allowed us to understand why data missing for national
parks or institutional land uses.
Couldn’t do k-fold cross validation given small sample size.
Decision tree model didn’t work for our analysis.

Conclusion
Model offers a good starting point for assessing likely popularity of station locations, using data that is readily
available for most major U.S. cities. Some subjective decision-making will still be required around major parks
(i.e., National Mall) or institutional land uses (campuses, hospitals).
Bikeshare continues to gain momentum across the country. Future studies should:
• Use a larger sample
• Idea: use DC as model, test against NYC and Chicago. Instead: use DC, NYC, and Chicago to build model with larger
sample, test against fourth city.
• Categorize stations by function within network
• Stations have different functions: residential feeders to metro stations, tourism. With a larger sample, these station
types could be separated and the drivers of popularity independently determined.
Important to note that there are valid reasons other than current popularity that should determine station
placement (i.e., equity, driving changes in travel behavior). This model helps ensure financial viability so that
these outcomes can be pursued.