Content uploaded by Frederik Schulte
Author content
All content in this area was uploaded by Frederik Schulte on Dec 13, 2023
Content may be subject to copyright.
Research in Transportation Economics 101 (2023) 101340
Available online 26 August 2023
0739-8859/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Research paper
Improving the service of E-bike sharing by demand pattern analysis: A
data-driven approach
Ziru Zhang
a
,
*
, Panchamy Krishnakumari
a
, Frederik Schulte
b
, Niels van Oort
c
a
P.O. Box 5049, 2600 GA, Delft, the Netherlands
b
P.O. Box 5031, 2600 GA, Delft, the Netherlands
c
P.O. Box 2600, 2600 GA, Delft, the Netherlands
ARTICLE INFO
JEL classication:
R41
R48
R49
Keywords:
E-Bike sharing system
Temporal clustering
Demand pattern analysis
Operational strategy
Service level evaluation
ABSTRACT
In recent years, there has been a surge in the popularity of free-oating e-bike sharing. However, the shared
mobility sector is ercely competitive demanding, efcient operations and high-quality service to cater to user
expectations.
We propose several data-driven methods that apply demand pattern analysis. We suggest the use of a new
spatial unit (i.e., overlapping circles) to enhance the cost-efciency and user-friendliness of e-bike sharing.
Moreover, temporal clustering is employed to develop operational strategies that counter the imbalance in
supply and demand in recurrent clusters.
To evaluate the impact of these strategies, we introduce a framework and apply it in a case study of an e-bike
sharing project in The Hague, The Netherlands. We identify 5 hourly clusters which enable reallocation strategies
to alleviate the imbalance among spatial units in these clusters.
The results demonstrate that the derived operational strategies improve the service signicantly, with almost
1.5 times increased ridership, an approximately 20% decrease in vehicle idle time, and a decent monthly net
retention rate of around 60%.
1. Introduction
Shared mobility has become a major trend since 2010, aiming to
improve the sustainability of the transport sector and alleviate trafc
congestion (European Commission, 2011). However, the shared
mobility market is highly competitive, requiring providers to achieve
efcient operations and high service quality (Beirigo, Negenborn,
Alonso-Mora, & Schulte, 2022). Among various shared mobility options,
shared bikes have gained widespread popularity due to their active
mode of transport with the associated health benets (Barbour, Zhang,
& Mannering, 2019; DeMaio, 2009). With the introduction of electri-
cation in the mobility sector, e-bikes, which offer higher travel speeds
and reduce physical efforts, have been gradually incorporated into
bike-sharing schemes (Fishman & Cherry, 2016).
Bike-sharing projects can be classied into two operational types,
viz., station-based and free-oating schemes (see e.g. Ma, Ji, et al.
(2020)). The station-based scheme relies on pre-dened stations for
users to pick up and return bikes, while the free-oating offers
exibility, allowing users to drop bikes at various locations within
designated operational zones (B. Beirigo, Schulte, & Negenborn, 2018).
The latter eliminates the constraints associated with station availability
in station-based systems, contributing to the growing popularity of
free-oating bike sharing in recent years during the past years (Chen,
van Lierop, & Ettema, 2020; Fishman, 2016).
Regardless of the operational type, understanding user travel
behaviour is crucial for matching supply and demand in bike-sharing
systems (Hua, 2020; A. Li, Zhao, Huang, Gao, & Axhausen, 2020).
Extensive research has been conducted on different aspects of
bike-sharing systems, including determinants of bike-sharing demand,
the interaction with public transport (Montes, Gerˇ
zinic, Veeneman, van
Oort, & Hoogendoorn, 2023; van Marsbergen, Ton, Nij¨
enstein, Annema,
& van Oort, 2022; van Mil, Leferink, Annema, & van Oort, 2021), de-
mand pattern analyses, prediction of demand in different time scopes,
and optimization of reallocation of shared bikes (Albuquerque, Sales
Dias, & Bacao, 2021; Eren & Uz, 2020; Fishman, 2016; Fishman,
Washington, & Haworth, 2013; Galatoulas, Genikomsakis, & Ioakimidis,
* Corresponding author.
E-mail addresses: ziruzhang1101@gmail.com (Z. Zhang), N.vanOort@tudelft.nl (P. Krishnakumari), F.Schulte@tudelft.nl (F. Schulte), P.K.Krishnakumari@
tudelft.nl (N. van Oort).
Contents lists available at ScienceDirect
Research in Transportation Economics
journal homepage: www.elsevier.com/locate/retrec
https://doi.org/10.1016/j.retrec.2023.101340
Received 30 November 2022; Received in revised form 23 July 2023; Accepted 29 July 2023
Research in Transportation Economics 101 (2023) 101340
2
2020; Ma, Yuan, Van Oort, & Hoogendoorn, 2020). These studies can be
categorized into three phases: 1) identifying determinants of shared bike
usage; 2) analysing datasets, identifying demand patterns and predicting
future demand; 3) devising optimal strategies to reallocate bikes.
However, most of the research has focused on station-based bike-
sharing systems and limited attention has been paid to free-oating e-
bike sharing. Substantial differences exist between station-based shared
bikes and free-oating shared bikes, as well as between regular bikes
and e-bikes (Chen et al., 2020; Galatoulas et al., 2020; Gu, Kim, &
Currie, 2019). Free-oating bike-sharing systems offer users a higher
degree of freedom by eliminating the need to rent and return bikes at
designated stations. However, this exibility increases the complexity of
modelling, as demand cannot be attributed to specic station units as in
station-based systems. Therefore, spatial analytical units, such as virtual
stations or trafc area zones, need to be dened for free-oating bike--
sharing to model trip generation and attractions (S. Liu, Hou, Liu,
Khadka, & Liu, 2018). Additionally, e-bikes have distinct trip charac-
teristics, such as travel distance, which vary from regular bikes due to
reduced physical effort and the presence of batteries, which exert an
inuence on people’s travel decisions (Galatoulas et al., 2020).
Moreover, the existing studies on bike-sharing operations lack ex-
periments and evaluations of different operational strategies in real-life
contexts. Current approaches often rely on dedicated but complicated
mathematical models to determine the optimal strategies with either
static or dynamic demand input. However, these methods can be quite
time-consuming and unrealistic for small and medium-sized shared
mobility operators, considering the limited resources and uncertainties
of operational actions (Alvarez-Valdes et al., 2016, p.; Angelopoulos,
Gavalas, Konstantopoulos, Kypriadis, & Pantziou, 2018; Chemla, Meu-
nier, & Woler Calvo, 2013; Dell’Amico, Hadjicostantinou, Iori, &
Novellani, 2014; Gavalas, 2016; Raviv, Tzur, & Forma, 2013).
Based on the existing literature, there are 3 scientic gaps: 1) a
spatial analytical unit which is friendly for both the operators and the
users, especially its efciency for operators; 2) studies targeting free-
oating e-bike sharing projects; 3) a well-rounded evaluation
approach of the real-life effects of the proposed operational strategies.
Considering both the scientic gaps and the needs of operators, the
objective of this work is to develop a data-driven approach to derive
benecial operational strategies. Those should then be deployed to
improve the service of e-bike sharing by conducting a data analysis, a
demand pattern analysis and a follow-up examination of the proposed
strategies in reality.
To this end, we 1) introduce an innovative spatial analytical unit,
with the overlapping circles, and prove that the reallocation strategies
derived based on this unit, are more cost-effective, requiring only one
relocation operation per period; 2) we add insights to the eld of e-bike
sharing, taking a different angle than the current studies; 3) we develop
a framework to evaluate the operational strategies and experiments in
real-life settings, considering both operators and users. Results indicate
that the proposed cost-effective operational services contribute to a
positive effect of increasing ridership by roughly 1.5 times.
The remainder of this paper is structured as follows: Section 2 de-
scribes the methodology, which is then applied to a case study presented
in Section 3. The results and corresponding discussion are provided in
Section 4, along with a comparison to parallel work. Finally, Section 5
presents the conclusions, including the main ndings, contributions,
limitations, and recommendations for future research.
Through this study, we introduce an innovative spatial analytical
unit to investigate the demand pattern of e-bike sharing services and
propose cost-effective operational services contributing to a positive
effect of increasing roughly 1.5 times ridership.
2. Methodology
In this work, we propose a four-step approach, as depicted in Fig. 1 to
address the research objective. Firstly, the literature review is conducted
to determine the contributing factors to the bike-sharing demand. The
preliminary data analysis is sequentially done. These results are input to
the demand pattern analysis, which investigates the pattern of bike-
sharing services in depth. The operational strategies are derived based
on these insights accordingly. Finally, the strategies are implemented in
a real-life setting and evaluated systematically.
2.1. Data analysis
The data analysis is the prerequisite of the demand pattern analysis.
This phase involves three steps, which are data description, correlation
analyses between determinants and the demand, and land use pattern
analysis.
2.1.1. Data description
The data description phase encompasses the acquisition of available
data, data cleaning, and data processing/aggregation. The primary
dataset used in this research consists of ride records, as exemplied in
Table 1. Each record contains ride start/end time, ride start/end loca-
tion, trip duration, and trip distance with a unique ride ID, and rider ID.
Additionally, data pertaining to six other groups are included. These
data are obtained from open-source databases, such as Google Maps for
spatial and infrastructure factors, by points of interest, meteorological
institutions for weather-related data and the government for socio-
demographic factors; trip characteristics and temporal data can be
retrieved from ride records and safety factors are usually omitted in the
most studies due to their reliance on deliberative interviews. Following
data obtainment, a data cleaning is employed to eliminate incomplete or
faulty data records and align the data with the research scope. Moreover,
the statistical characteristics of those data are described.
2.1.2. Correlation analysis
Following the data description, linear correlations are conducted
between the determinants and the demand (i.e., the ridership). Only
linear correlations are examined in this study to test the hypothesis of
whether the determinants found in the literature indeed impact the
demand considering the research scope. The methods used in this study
are Pearson’s coefcient and multiple linear regression. They are chosen
Fig. 1. The four-step framework of this study.
Table 1
Ride record sample.
Ride id Rider id Start time End time
2356 3556 16:00:02 July 24,
2023
16:26:24 July 24,
2023
Start location End location Trip duration Trip distance
52.0579935,
4.2638107
52.055401,
4.268105
00:26:22 7.839 km
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
3
because of their simplicity and the power of revealing the correlations.
Pearson’s coefcient is computed for each variable, and it indicates
the normalized covariance between two variables as shown in Equation
(1).
rxQ =n
i=1(xi−x)(Qi−Q)
n
i=1(xi−x)2
n
i=1(Qi−Q)2
Equation 1
Where n is the sample size; x
i
and Q
i
are the individual sample points,
indexed with i; x=1
nn
i=1xi (the sample mean) and analogously for Q; in
this paper, Q always indicates the ridership, aggregated in different time
intervals on different spatial units and x represents different inuential
factors, including weather, number of POIs, and the availability of
public transport.
Multiple linear regression is used when there is more than one var-
iable under a determinant group to see what the essential factor under
the determinant group is.
2.1.3. Land use pattern analysis
In addition to data description and correlation analysis, a land use
pattern analysis is conducted to gain insights into the functional
composition of different locations based on the distribution of POIs. This
analysis is performed using two selected spatial analytical units:
neighbourhood and 400 m overlapping circles.
First, the distributions of POIs are visualized to reveal the general
pattern of different facilities. Then, the function p (i,c) is dened in
Equation (2), representing the POI distribution. If p exceeds 0.5 then the
unit would be dened with the corresponding function. Only workplaces
and recreations (including sustenance and entertainment amenities) are
considered in this study due to their most relevancy to trafc demand.
p(i,c) = |ic|
n
i=1|ic|,∀i∈I,c∈CEquation 2
where p(i,c) presents the proportion of a specic facility in a given
location; i
c
is the type i amenities belonging to unit c, and the denomi-
nator is the total amount of all facilities in this unit, regardless of their
types, where the notation |x| represents the absolute value of x, which is
constantly used in this study; I is the set containing all the types of POIs
and C is the set consisting of all spatial analytical units.
The outputs would be considered in the following demand pattern
analysis, to understand the prediction of ow.
2.2. Demand pattern analysis
The demand pattern analysis serves as the focal point of this
research. First, a descriptive analysis of crucial trip characteristics is
conducted. Second, the demand pattern is performed, incorporating
temporal clustering methods. The insights from demand pattern analysis
support the development of reallocation strategies as they help mitigate
the imbalance between the supply and the demand in different units.
Third, supply efciency is examined by the distribution of vehicle idle
time per unit. Similarly, the average trip time and distance are explored
on the unit level. These analyses, in conjunction with the supply ef-
ciency assessment, provide valuable insights into the operational ef-
ciency and popularity of different spatial units. This information
facilitates prioritizing operational strategies and making necessary ad-
justments to the service area.
A. Spatiotemporal aggregation
After the general demand pattern analysis, the spatial analytical
units are determined, specically the neighbourhood and 400 m over-
lapping circles. Data are aggregated and analysed on these two distinct
levels.
The temporal clustering is then conducted, based on similarities and
dissimilarities in demand of different periods. The purpose of temporal
clustering is to investigate if the demand pattern of different periods
emerges and the insights are studied in the next step, aiding the devel-
opment of operational strategies (T. L. K. Liu, Krishnakumari, & Cats,
2019).
The complete procedures are as follows: rstly, ride records are
aggregated in the spatial units determined in the last step and then OD
(origin-destination) matrix is computed accordingly; it is followed by
the temporal clustering based on OD matrices, gathering different pe-
riods with similar features together. They are used to capture essential
demand peculiarity.
OD matrix presents the ow between different locations with in-
sights on how trips are attracted and generated at zonal levels.
First of all, the ride records are aggregated with the predened
spatial units as the origin and destinations, and the ow q(x, y, t, z) then
corresponds to a specic origin x, a destination y, a ride date z, and a ride
hour, presented by the ride start hour, t.
Secondly, hourly clustering and daily clustering are considered in
this work, and therefore two series of OD matrices are constructed:
hourly OD matrices and daily OD matrices.
For hourly OD matrices, the ow q is aggregated in the increment of
1-h intervals from 0:00 to 24:00, and it is respective to each day.
Therefore, there are 24 D*k*k OD matrices in total where D species the
number of days and k is the number of zones, where each cell corre-
sponds to the ow between the given OD pair during a specic hour for a
given date z, as indicated in Equation (3).
Qt(co,cd,t) =
x∈co
y∈cd
z
q(x,y,t,z)Equation 3
Q
τ
(co,cd,z) =
x∈co
y∈cd
t
q(x,y,t,z)Equation 4
Similarly, daily OD matrices are on a daily basis, generating totally D
24*k*k OD matrices where each cell corresponds to the ow increment
of a 1-h interval between a given OD pair (c
o
, c
d
) for a given date z; c
o
and c
d
represents the origin and destination on the zonal level while x
and y is the exact geolocation of the origin and destination of the ride
records, shown in Equation 4.
Thirdly, temporal clustering is employed with the aggregated OD
matrices as the feature vectors. Each data point is composed of a cor-
responding OD matrix. Agglomerative hierarchical clustering is applied
because of its loose prerequisite of the number of clusters, and its
dendrogram to assist in the determination of the optimal number of
clusters (Rokach & Maimon, 2005).
In this research, Euclidean distance is used to compute the dissimi-
larity metric and the ward method is applied to combining the clusters
by the variance of clusters which is found to be the most suitable method
for quantitative variables (Calinski & Harabasz, 1974).
Agglomerative hierarchical clustering, as a bottom-up algorithm,
starts with the cluster number equal to the number of data points with
zero merging cost since each data point is an individual cluster, and the
successive converging process continues until only one cluster is left.
The number of clusters can then be decided based on the dendrogram
considering the interpretability.
Hourly clustering and daily clustering are applied in this study based
on the OD matrices aggregated at hourly and daily levels as described
before.
2.3. Supply efciency analysis
The supply efciency analysis is performed by examining the vehicle
idle time per spatial unit. Vehicle idle time is the time when the vehicle
is in place while no ride is taken in the vehicle, indicating how long the
vehicle is idle between two rides (Cats, Krishnakumari, Arbez, Chiabaut,
& van Lint, 2020).
Commonly, the vehicle idle time only refers to the time interval
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
4
between two rides while sometimes there are fewer than 2 ride records.
The assignment of the location corresponding to the idle time is tricky.
Thereby, the vehicle idle time is adapted to tackle this problem.
To keep consistency and include all possible idle time records,
vehicle idle time is separated into two types: one corresponding to the
origins of rides as VITv(co,
τ
); the analogous for the destinations as
VITv(cd,
τ
)as Equation (5) and Equation (6).
VITv(co,
τ
) =
t,R
τ
v=0
tsv
r−ts
τ
,R
τ
v=1
tsv
r−ts
τ
,if r is the first ride in R
τ
v
tsv
r+1−tev
r,∀r∈R
τ
v
,R
τ
v>1
Equation 5
VITv(cd,
τ
) =
t,R
τ
v=0
te
τ
−tev
r,R
τ
v=1
tsv
r+1−tev
r,∀r∈R
τ
v
te
τ
−tev
r,if r is the last ride in R
τ
v
,R
τ
v>1
Equation 6
∀v∈V where V is the set of all available vehicles during period
τ
and
t is the total time duration of the period
τ
.
ts indicates the starting time, for both the ride and the period; te is the
ending time for either the ride, r, or the period. ts
τ
is the starting time for
period
τ
and te
τ
is the ending time for period
τ
;
c
o
is the original unit, and c
d
is the destination unit; ts
r+1
v
is the
starting time of the (r+1)
th
ride of vehicle v and te
r
v
is the ending time of
the rth ride of vehicle v; R
v
τ
is the set including all rides for vehicle v,
during time period
τ
;
If there is only 1 ride record belonging to this vehicle during the
given period, the vehicle idle time is separated into two sub-vehicle idle
times, corresponding to the original unit and the destination unit
separately.
For example, if vehicle v only has 1 ride record during time period
τ
,
with the attributes ride start time ts
v
, the original location cv
o, ride end
time te
v
and destination (drop location) cv
d. The rst idle time before
picking up from the starting point of this period would be assigned as the
vehicle idle time for the original unit, while the second idle time from
terminating the ride at the destination to the ending timestamp of this
period is assigned the vehicle time for the destination unit. Analogously,
for more than 1 ride record, the same procedures are applied for the rst
and the last rides: assigning the time between the starting time of this
period, ts
τ
and the starting time of the rst ride, tsv
r; Assigning the time
between the end time of the last ride, tev
r and the end time of this period,
te
τ
.
The time interval t is determined by the usage of the shared service.
For example, if there is 1 ride per vehicle per day on average, the time
interval could be set up as 1 day, preserving the maximal vehicle idle
time of 24 h. Additionally, the frequency of reallocation also exerts an
effect on the determination of the time interval, since the main aim of
this indicator is to assist the reallocation, avoiding low usage in general.
Thereby, this time interval should also be compatible with the frequency
of reallocation.
Unit-based vehicle idle time is calculated based on the formulas. To
reallocation geographically, this vehicle idle time per spatial unit per
period is computed as Equation (7) and Equation (8).
AVITco(Tt) =
i∈ITt
co
VITv(co,
τ
)∀v∈V,
τ
∈Tt
ITt
co
,∀co∈C,Tt∈TEquation 7
AVITcd(Tt) =
i∈ITt
cd
VITv(cd,
τ
)∀v∈V,
τ
∈Tt
ITt
cd
,∀cd∈C,Tt∈TEquation 8
i indicates each vehicle idle time record, and ITt
co is the set of all
vehicle idle time belonging to the unit c
o
and during the period T
t
and
analogously for ITt
cd; C is the set of all spatial units; T
t
is the period, pre-
senting the different operation stages across the whole period, for the
ride records dataset and T is the period of the whole dataset; The de-
nominator, ITt
coand ITt
cd, is the number of all vehicles idle time
belonging to unit c
o
/c
d
during the time period T
t
.
Based on the magnitude of AVITco(Tt)and AVITcd(Tt)in different
units, a heatmap would be visualized, describing which unit(s) vehicles
encounter shorter vehicle idle time and thereby those locations are
appealing to reallocate bikes.
2.4. Average trip time/distance analyses
Similarly, average travel distance and duration are computed on the
spatial unit level. These two metrics are origin-oriented since the
destination is less relevant from the operator’s perspective. The calcu-
lations are conducted as Equation (9) and Equation (10).
ATDc(T) =
i∈RT
c
travel distancei
RT
c
,∀c∈CEquation 9
ATTc(T) =
i∈RT
c
travel timei
RT
c
,∀c∈CEquation 10
where ATDc(T)indicates the average travel distance corresponding to a
unit c and a given period T; R
c
T
is the set including all the ride records
originating at unit c during the period T and |R
c
T
| indicates the size of the
set; Average travel time is computed analogously with the travel time as
the object instead of travel distance.
2.5. Development of the operational strategies
In this study, two primary types of operational strategies are iden-
tied: reallocation strategies and adjustment in the service area.
General demand patterns and temporal clusters provide insights into
the departures, arrivals and total ow of each spatial unit, as well as how
the e-bike sharing trafc ows between different OD pairs. These ana-
lyses serve the purpose of informing the development of reallocation
strategies, which aim to mitigate the imbalance between the supply and
the demand.
Furthermore, the assessment of supply efciency and average trip
distance/duration analyses is performed at the spatial unit level, indi-
cating the popularity and operational efciency geographically. Based
on these ndings, recommendations in the operational areas can be
formulated.
2.6. Operation
During the operational stage, the strategies derived from the previ-
ous phase are implemented in the real-life context and systematically
examined.
To evaluate these strategies, two sets of KPIs are introduced. The rst
group focuses on the operational aspect, while the second category
evaluates user satisfaction from a business perspective. The rst group
includes metrics such as daily ridership, ridership ratio and average
vehicle idle time per vehicle; the second is presented by net retention
ratio and average user expenditure.
➢ Daily ridership
This is represented by the total ridership every single day. The sum is
based on the ride start time. For example, if a ride starts at 23.:59 on 10/
09/21 and ends at 00:15 on 11/09/21, it would be assigned to the rides
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
5
belonging to 10/09/21.
Q
τ
=
x
y
t
q(x,y,t,
τ
),∀
τ
∈TEquation 11
where t is the ride start time. It sums of all the rides in the given day
τ
, if
the origin and destinations are within the operation zones.
➢ Ridership ratio
This indicates the general ridership ratio per day. The supply of this
day is dened as the total available eet size in the whole operational
zone.
r
τ
=Q
τ
supply
τ
Equation 12
➢ Average vehicle idle time
The vehicle idle time is computed in the same way. However, in this
phase, the average vehicle-based vehicle idle time is applied instead of
the unit-based ones, using the same set of vehicle idle time records, in a
different aggregation way, though. The object of this metric corresponds
to each vehicle, and the average idle is computed based on all corre-
sponding vehicle idle time records. Afterwards, the average vehicle idle
time is computed by taking the average of all average idle time corre-
sponding to the available eet during this period. The origin-based and
the destination-based indicators apply in a similar way.
AVITco
v(Tt) =
i∈Ico
v
VITv
i(co,
τ
)∀co∈C,
τ
∈Tt
Ico
v
,∀v∈VTt,Tt∈T
Equation 13
AVITcd
v(Tt) =
i∈Icd
v
VITv
i(cd,
τ
)∀cd∈C,
τ
∈Tt
Icd
v
,∀v∈VTt,Tt∈T
Equation 14
where Ico
v is the set of origin-based vehicle idle records belonging to the
vehicle v, and analogously applies for Id
v.
AVITVco(Tt)and AVITVcd(Tt)are the average of the average origin-
based/destination-based vehicle idle time per vehicle during the
period Tt, the denominator is the eet size belonging to this period
within the operation zone.
AVITVco(Tt) =
v∈V
τ
AVITco
v(Tt)
VTt
,∀Tt∈TEquation 15
AVITVcd(Tt) =
v∈V
τ
AVITcd
v(Tt)
VTt
,∀Tt∈TEquation 16
➢ Net retention rate
Net revenue retention is a metric demonstrating the variations
within the existing revenue base. It is used to describe to what extent the
revenue of the existing customers grows or churns monthly (Guide to
Net Dollar Retention (NDR) - Denition, Calculation, Tips, 2021). It
indicates how much customers spend and their expenditure changes on
the service over time and is a way to understand customers’ satisfaction:
if they are satised with the service, they would keep a subscription and
continue spending money on it.
NRR =Starting MRR +Expansion MRR −Contraction M RR −Churn MRR
Starting MRR
Equation 17
where MRR is the monthly recurring revenue and NRR is computed
based on it.
➢ Average user expenditure
Complimentary to the net retention rate, average user expenditure is
also computed. Three indicators belong to this group, total user average
expenditure, new user average expenditure and retained user average
expenditure.
average user expenditurem=total expenditurem
|usersm|,∀m∈IEquation 18
where M is the set of all months and m refers to a specic month.
This class of indicators provides insights into how much users of
different groups spend on the service monthly.
3. Case study
The case study is conducted based on an e-bike sharing project in The
Hague. The project was launched in mid-June 2021 and the research
focuses on a 4-month duration following the project’s launch. The data
from the rst three months are utilized to develop operational strategies
while the entire period is considered for evaluating the effects of these
strategies. The e-bike sharing project operates with a eet size of a few
hundred e-bikes, with dynamic adjustments made to the eet size based
on operational conditions in the subsequent days.
The Hague, the third-largest city in The Netherlands, serves as the
context for this case study. It has a population of approximately 550,000
inhabitants as of 2021, consists of 8 administrative districts and 44
neighbourhoods, referred to as Wijken in Dutch (The Hague, 2021; The
Hague in Numbers, 2021; Wijken en buurten in Den Haag, 2021). In this
research, the neighbourhood units and the overlapping circles are
treated as the spatial analytical units, as the visualization in Fig. 2.
Neighbourhoods are a convenient unit for research purposes, but
their substantial heterogeneities can pose challenges to operational ef-
ciency. Also, for the users, the operations on the neighbourhood level
are still too coarse and thereby cannot capture their needs in specic
locations. To address this limitation, an innovative unit, overlapping
circles, is introduced. The radius of overlapping circles is set to 400 m:
on one hand, 400 m approximately correspond to 5-min walking based
on the Dutch average walking speed of 4.5 km/h (Waterstaat, 2019). 5-
minute is widely used as the threshold of catchment areas and therefore
this concept is also applied here (‘Basics’, 2011; Sarker et al., 2019); on
Fig. 2. Map of functions on the neighbourhood level.
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
6
the other hand, a sensitivity analysis was conducted to ensure the
robustness of 400 m in terms of revealing demand patterns in hourly
clustering.
The main function of neighbourhoods is illustrated in Fig. 4. It is
evident that recreational areas are predominantly concentrated in the
central and southwestern part of The Hague, particularly in Centrum
and along the beach area. For workplace function, it plots relatively
sparsely. There are quite some ofce-oriented areas along the beach and
in the north-central areas.
Turning to the descriptive statistics of the demand for the rst 1.5
months this timeframe is chosen as it represents the only available data
during that specic research phase. To account for the unstable ridership
caused by the effects of the project’s launch, the ridership is normalized.
The highest ridership is 1 and the rest is the ratio between the ridership
and the highest value. It is observed for the rst 3 days, the ridership was
exceptionally high because the service was free of charge during this
period. Afterwards, the ridership continually declined, with several
uctuations though. The ID verication, aiming to avoid misbehaved
rides, was installed on the 19th day after launch. It has exerted a
negative effect on ridership with a decline to circa 0.15 per day. This also
arises from the bad weather condition. Followingly, reallocation stra-
tegies obtained from the data-driven methods have been implemented,
which would be discussed in detail in the later sections, and they
contribute to the rebounce of ridership to 0.3 per day approximately, as
presented in Fig. 3.
For ridership in terms of the day of the week, there is no obvious
difference between the ridership per day, indicating that the ride day
does not insert a noticeable effect on ridership. Additionally, the de-
mand does not either present the typical morning peak pattern, seen in
Fig. 4, contrary to the previous work (S. Li et al., 2021; Miranda-Moreno
& Nosal, 2011; Tin Tin, Woodward, Robinson, & Ameratunga, 2012;
Xing, Wang, & Lu, 2020). One reason is that people are still unfamiliar
with this service which prevents them from using it during the morning
peak, to avoid being late for work. Similar to the literature, there is an
evening peak between 16:00 to 19:59, which reaches its climate at
around 17:00 to 17:59. It implies the service attracts people after their
work, consistent with the assumption under the fact of no morning peak
in this case study.
The histogram of ride distance shown in Fig. 5, illustrates that the
majority of trips are shorter than 10 km and the peak is between 0 and 2
km. The inspection of trip duration indicates that most rides are shorter
than 50 min and the peak is between 0 and 10 min. The average travel
duration is 18.2 min and the median is 11.5 min, next to the peak range.
4. Results and discussion
This section presents the results of the application of the methods
described in Section 2. First, the general determinants for e-bike sharing
services are introduced with the specication of the determinants in this
case study. Second, the results from the demand pattern are presented,
followed by a summary of the derived operational strategies. Last, the
operational strategies are appraised using the proposed methods.
4.1. Determinants for the demand for e-bike sharing services and
correlation analysis
Based on the literature review, there are 6 types of inuential factors
for e-bike sharing service, viz., spatial and infrastructure factors,
weather-related factors, mobility and trip characteristics, temporal
factors, sociodemographic factors and safety factors (Daddio, 2012;
Fishman, Washington, & Haworth, 2012, 2013; Hampshire, 2012; Ji,
Cherry, Han, & Jordan, 2014). However, only the rst four groups are
applied in this study due to the availability of data.
Correlation analyses are then conducted to see whether these general
determinants do exert effects as expected in this case study. Most of
these determinants are proven to have a Pearson’s coefcient higher
than 0.3 with the demand. Unexpectedly, the precipitation level and
temperature are found to correlate with the demand at a low level,
−0.03 and 0.16 respectively, stemming from their subtle variations of
them under the time scope of this research. Another study also presents a
similar result where precipitation is found to be insignicant to affect
the demand level (He et al., 2019).
In harmony with the previous studies, the number of POIs has pos-
itive effects on the ridership level (Faghih-Imani, Hampshire, Marla, &
Eluru, 2017); the humidity, proved from both Pearson’s coefcient and
the multiple linear regression model, however, impacts the demand
negatively (El-Assi & Mahmoud, 2015).
4.2. Demand pattern analysis and temporal clustering
There are two demand pattern analyses, with the dataset of two
periods. First, general demand pattern analysis is conducted on the
neighbourhood level, using the rides of the rst month; then, temporal
clustering is done on both the neighbourhood and circles level, using the
dataset of the rst 1.5 months.
As seen in Fig. 6, it is found that Centrum is always the hottest spot in
terms of departures and arrivals and the central areas are favoured
compared to other areas. Besides, the beach area catches attention,
standing out as the heat spots. In addition to that, the arrivals outweigh
departures in most neighbourhoods, despite two central ones.
Hence, it is recommended to rebalance the bikes from the beach
areas to Centrum and the south of Centrum. Another operational
Fig. 3. Ridership overview.
Fig. 4. Temporal distribution of the demand by ride start hour of a day.
Fig. 5. Distribution of the distance.
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
7
strategy is to place bikes in batches when rebalancing the bikes, making
the bikes more noticeable.
Afterwards, temporal clustering is conducted to see if different pe-
riods share similar demand patterns. Both spatial units are used in this
analysis Only the clustering on the overlapping circles is present, as
Fig. 7.
As observed from Fig. 8, the peak hours and transition hours
distinguish themselves to be separate clusters, and the rest hours emerge
together, given 5 clusters. Thereby, there are 5 hourly clusters with their
characteristics, shown in Fig. 9 as follows.
i. The rst peak hour (16:00–16:59): for this period, the rides
gather in Centrum and the beachside area. Besides these two
areas, the rides are relatively sparse in the outer units at a low
magnitude. The most prevailing spots are in the centre which
belongs to Centrum and are recreational-oriented areas, located
in the centre of Fig. 9.1. It is noteworthy that the other units of
attention are generally recreational-oriented zones despite the
units located to the east of the centrum with ofce functions.
Taking the function into account, these four ofce-oriented units
present a departure-dominant ow in the rst peak hour. As for
the central area, the arrival/departure pattern is rather hetero-
geneous and varies per unit.
ii. The second peak hour (17:00–17:59): this period demonstrates a
similar pattern as the rst peak hour. The central areas are still
the hottest. However, the most attractive unit shifts to the unit
near the train station and this unit also reveal a slight departure
tendency. The alike phenomenon appears in another unit close to
the central railway station, with a higher departure ratio. It
suggests that travellers prefer to pick up the bikes near the train
station during this period, which is conceivable to serve the last
mile of their trips, compensating for the train trips. The central
units are generally balanced between departures and arrivals,
with slightly more departures. Additionally, the beach units, in
general, attract more trips than the last hour, mainly as the des-
tinations rather than trip origins, while the residential areas are
less appealing.
iii. The rst transition hour (18:00–18:59): Ofce-oriented units
present a predilection for more departures while the recreational
ones tend to attract arrivals instead. In this period, the central
area is quite balanced, with a marginal dominance of arrivals,
especially for the north-western zones.
iv. The second transition hour (19:00–19:59): The unit near the train
station is still the most popular in this period, with almost equal
departures and arrivals, echoing the deduction that these trips
serve the rst/last mile supplementary train trips. Rides are
relatively distributed to other units while the central areas are
still prevalent all the time. In this period, the central units are
generally dominated by more departures while the arrivals are
towards the outer units.
1. The off-peak (20:00–15:59): invariably, the central areas are still
quite essential, followed by units next to the south and east of the
Centrum district. Besides, the outer residential units are either
balanced with similar departure and arrival ratios, or prone to
more arrivals in this period.
Fig. 6. Heatmaps of normalized departures (left) and the difference between arrivals and departures per neighbourhood (right).
Fig. 7. Map of functions on the circle level.
Fig. 8. Dendrogram of hourly clustering on the circle level.
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
8
Correspondingly, the rebalance suggestions are constructed in
Table 2. The repositioning strategies on the neighbourhood levels are
also derived in the same way while the details are omitted in this paper.
The comparisons between these two sets of strategies would be pre-
sented in the evaluation section.
4.3. Supply efciency analysis and average trip duration/distance
analysis
This series of analyses provide an overview of the usage over
different spatial units.
It is found from the supply efciency analysis that the outer units
usually experience a longer vehicle idle time, especially those located in
the southwest, as seen in Fig. 10.
Average travel distance and duration indicate that the southwest part
witnesses a lack of use as presented in Fig. 11, aligned with the obser-
vations in the supply efciency analysis. Moreover, the units located
between the outskirts and central areas, which are usually those ofce-
oriented or residential units, witness a longer trip duration as well as trip
distance.
Thereby, it is suggested to adjust the service area, leaving out the
southeast part as well as placing those bikes in the hotspots of this ser-
vice. By doing so, efforts of battery swaps can also be signicantly
alleviated.
4.4. Evaluation of the operational strategies
There are in total 5 operational strategies with various time scopes. 3
out of 5 are reposition strategies, consecutive in the timeline of two
months from July to September, in which the rst set is based on the
general demand pattern on the neighbourhood level, and the second and
the third are developed from temporal clustering on the neighbourhood
and overlapping circle levels. The details of temporal clustering are
omitted in this study while the derived strategies are still being assessed,
compared with their counterparts on the overlapping circles.
The exact action of temporal-based reallocations is dependent on the
execution time of the action, which is usually between 10:00 to 16:00
during the daytime. It is noted that only one relocation operation is
needed per period, requiring only a few efforts by the operator.
The KPIs from the operational sides are presented in Table 3. It is
obvious that the period of the implementation of ID verication per-
forms the worst out of all the periods. During the periods of rebalancing,
all KPIs are improved to different degrees. Among the three rebalancing
periods, the repositioning strategies based on hourly clustering have
stronger positive effects no matter on ridership or the mitigation of
vehicle idle time. Besides, the reallocation on the circle level has the
foremost advantages in the improvement of the service level.
Though the ridership ratio seems too low, under 1 ride per vehicle
per day, it is acceptable compared to the other schemes with similar
supply levels around the world. For instance, Alacant, Boulder,
Clermont-Ferrand and Perpignan (marked as yellow dots in Fig. 12),
have a similar supply level. The rst two schemes present alike ridership
ratios and the latter two have a lower ridership ratio at around 0.4 and
0.2 respectively (M´
edard de Chardon, Caruso, & Thomas, 2017).
Considering Mobike in Delft, which offered a much higher supply, only
saw the ridership ratio at around 1.5 in 2018 (Ma, Yuan, et al., 2020). All
Fig. 9. Heatmaps of ow (proportional to the maximal ow) and arrival ratio
for 5 hourly clusters.
Table 2
Rebalance suggestions based on temporal clustering in circles.
period From To
1st peak 16:00–16:59 Centrum Station
2nd peak 17:00–17:59 Centrum South of Centrum
1st transition 18:00–18:59 South of Centrum South
2nd transition 19:00–19:59
Off-peak 20:00–15:59 beach Pier of beach
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
9
these cases show the relevance of the results and the efcacy of opera-
tional strategies in this work.
From the users’ perspective, the gures related to the monthly net
retention rate are illustrated in Table 4. The third month from Mid-
August to Mid-September has the highest NRR at 86.87%. However,
NRRs of all three months are below 100%. Taking the average user
expenditure into account, it is easily observed that retained users always
have higher average expenditure than new users, as seen in Table 5.
Fig. 10. Heatmaps of origin-based (left) and destination-based (right) vehicle idle time.
Fig. 11. Heatmaps of average trip time (right) and distance (left).
Table 3
Summary of the operational KPIs.
periods date average
ridership
ratio
average vehicle idle time
per vehicle (h)
origin-
based
destination-
based
free rides Day 1–3 2.11 8.81 17.82
adopting period Day
4–18
0.98 47.28 32.14
ID verication Day
19–28
0.31 59.51 52.66
rst-round rebalance Day
29–43
0.53 51.84 43.25
second-round rebalance
on the neighbourhood
level
Day
44–50
0.59 57.3 41.08
second-round rebalance
on the circle level
Day
51–82
0.67 47.3 33.07
without specic strategy Day
83–97
0.64 50.16 41.08
reduction in the
operational area
Day
98–122
0.52 50.9 43.63
overview 0.66 40.77 33.05
Fig. 12. Ridership ratio overview.
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
10
Based on the results, it is found that the reallocation according to the
hourly clustering demand pattern on the circle level has the most ad-
vantages, contributing to the highest ridership and the lowest vehicle
idle time on the operator’s side.
From the users’ side, the time interval of KPIs is exactly 1 month,
different from the KPIs in the operators’ aspect. Thereby, it is hard to
compare these two sets of KPIs. The time scope of the three reallocation
strategies partially overlapped with the second month while the third
month also accounts for more than half days of the execution period of
the third reallocation strategy. The third month performs the best
considering both NRR and average user expenditure.
Nevertheless, NRRs of all three months are below 100%. It indicates
that customer loyalty is not fully established. It is common for a new
service. Additionally, this level is quite acceptable compared with the
other bike-sharing providers in which the retention rate is also below
100%, ranging from 20% to 70%, even the famous project, such as Lime,
claims a national retention rate of 60% (Shaheen, Martin, Chan, Cohen,
& Pogodzinski, 2014; SmartCitiesWorld, 2017).
The operational strategies are proven to be benecial for improving
the service, observed by an increase in NRRs and average user expen-
diture, with substantial uctuations though.
The fourth month experiences a decline in NRR, while the average
expenditure of new users and all users do not witness a considerable
decrease, even with the shrinkage in the operational areas.
The last strategy, reduction in the operational area, has a lower
positive effect while it decreases the efforts of battery swaps and real-
location to a large degree, not reected in the KPIs used in this evalu-
ation, though.
It is found that rebalancing has better effects although it requires
more effort in operation. However, a reduction in the service area, the
other way, relieves the efforts of operation without harm to the service
level. Thereby, they are complementary to each other and executed at
the same periods.
5. Conclusion
This work addresses several scientic gaps in free-oating e-bike
sharing research by introducing an innovative spatial analytical unit, the
overlapping circles, and demonstrating that the reallocation strategies
derived based on this unit, are more cost-effective because only one
relocation operation per period is needed. Moreover, this work adds new
insights to the scarcely investigated, proposes a framework to evaluate
the operational strategies in a real-life context, while considering the
perspectives of both operators and users.
We introduce data-driven methods such as demand pattern analysis
to design new operational strategies. In addition to that, we conduct a
descriptive analysis, revealing a single peak period of the e-bike sharing
service in The Hague, contrary to the widely observed two peaks in other
studies. Building on this initial analysis, we perform general demand
pattern analyses as well as temporal clustering analyses of the demand
pattern via agglomerative hierarchical clustering. The central units are
found to be the most popular places, in both general and clustering
analyses. Furthermore, 5 periods emerged from the hourly clustering,
which is the rst peak hour, the second peak hour, the rst transition
hour, the second transition hour, and the off-peak period. We observe
that from the second peak hour, people moved towards the recreational-
oriented zones from the ofce-oriented ones and the station unit became
prevalent with the most rides from the 2nd peak hour until the end of the
transition hours. Furthermore, the popularity of the station unit implies
that people have used this service as a supplement to their train trips for
the rst and the last mile. Additionally, the analysis indicates a lack of
use in the outskirts in the southwest of The Hague, based on the supply
efciency and average travel distance/duration analyses. Based on these
ndings, three sets of rebalancing strategies and the reduction in the
service area were proposed.
Subsequently, these strategies were implemented in a real-life
context and their effects are examined using two categories of KPIs
that assesses the perspectives of operators and users. All strategies
demonstrate improvements in service levels. Among these, the third set
of repositions based on hourly clustering on the circle level showed to
improve the service to the largest degree, out of all strategies, with a
ridership ratio of 0.67 and a decrease in the origin-based and
destination-based vehicle idle time at around 12 and 19 h, compared to
the ID verication period. Besides, the adjustment in the operational
zone has a moderate benecial impact on the service, but it decreases
the operational efforts to a substantial extent. After applying the stra-
tegies, the level of ridership ratio is quite decent compared with the
other schemes with a similar supply level around the world. Moreover, it
is also found that users became more satised with the service after the
implementation of the suggested strategies, indicated by an increased
net retention rate ranging from 52.10% to 86.87% and a grown average
user expenditure of around 9 to 15 euros per user. Therefore, we advise
to implement multiple operational strategies in combination to com-
plement each other.
However, there are still some limitations. The explicit supply data is
unavailable thereby the supply level was inferred from the ride records.
The evaluation method does not distinguish the effects from the oper-
ational strategies and other factors, such as the promoting campaigns or
the effects of holiday seasons. Therefore, some future work repeats the
research with better input data, including the precise supply data, and
incorporate the relationship of public transport into analyses, on the
overlapping circle units. Future work should furthermore investigate
predictions of short-term e-bike sharing demand, adapt the evaluation
method of the operational strategies which isolates the effects stemming
from the strategies, by providing better reference cases for the predictive
models.
CRediT authorship contribution statement
Ziru Zhang: Conceptualization, Methodology, Investigation, Formal
analysis, Visualization, Writing – original draft, Writing – review &
editing. Panchamy Krishnakumari: Supervision, Writing – review &
editing. Frederik Schulte: Supervision, Writing – review & editing.
Niels van Oort: Supervision, Writing – review & editing.
Declaration of competing interest
The authors declare the following nancial interests/personal re-
lationships which may be considered as potential competing interests.
Table 4
Summary of NRR.
Period NRR
Mid-June to Mid-July
Mid-July to Mid-August 15.78%
Mid-August to Mid-September 86.87%
Mid-September to Mid-October 52.10%
Table 5
Summary of average user expenditure.
Period New user
average spent
(
€
)
Retained user
average spent (
€
)
total user
average spent
(
€
)
Mid-June to Mid-
July
Mid-July to Mid-
August
7.83 15.78 9.24
Mid-August to Mid-
September
12.80 20.87 15.30
Mid-September to
Mid-October
9.91 19.24 14.45
Z. Zhang et al.
Research in Transportation Economics 101 (2023) 101340
11
During this research, the main author has done an internship at a
company operating e-bike sharing in Europe, who provides the main
data input.
Data availability
The data that has been used is condential.
Acknowledge
This research uses data provided by an anonymous shared-mobility
operator. The authors would like to thank Jam and Mickey Mouse NL
for their help with data acquisition.
References
Albuquerque, V., Sales Dias, M., & Bacao, F. (2021). Machine learning approaches to
bike-sharing systems: A systematic literature review. ISPRS International Journal of
Geo-Information, 10(2), 62. https://doi.org/10.3390/ijgi10020062
Alvarez-Valdes, R., Belenguer, J. M., Benavent, E., Bermudez, J. D., Mu˜
noz, F.,
Vercher, E., et al. (2016). Optimizing the level of service quality of a bike-sharing
system. Omega, 62, 163–175. https://doi.org/10.1016/j.omega.2015.09.007
Angelopoulos, A., Gavalas, D., Konstantopoulos, C., Kypriadis, D., & Pantziou, G. (2018).
Incentivized vehicle relocation in vehicle sharing systems. Transportation Research
Part C: Emerging Technologies, 97, 175–193. https://doi.org/10.1016/j.
trc.2018.10.016
Barbour, N., Zhang, Y., & Mannering, F. (2019). A statistical analysis of bike sharing
usage and its potential as an auto-trip substitute. Journal of Transport & Health, 12,
253–262. https://doi.org/10.1016/j.jth.2019.02.004
Beirigo, B. A., Negenborn, R. R., Alonso-Mora, J., & Schulte, F. (2022). A business class
for autonomous mobility-on-demand: Modeling service quality contracts in dynamic
ridesharing systems. Transportation Research Part C: Emerging Technologies, 136,
Article 103520. https://doi.org/10.1016/j.trc.2021.103520
Beirigo, B., Schulte, F., & Negenborn, R. (2018). Dual-mode vehicle routing in mixed
autonomous and non-autonomous zone networks. 2018 21st international Conference on
intelligent transportation systems (ITSC), 1325–1330. https://doi.org/10.1109/
ITSC.2018.8569344
Calinski, T., & Harabasz, J. (1974). A dendrite method for cluster analysis.
Communications in Statistics - Theory and Methods, 3(1), 1–27. https://doi.org/
10.1080/03610927408827101
Cats, O., Krishnakumari, P., Arbez, N., Chiabaut, N., & van Lint, H. (2020). Empirical
evaluation of the efciency and effectiveness of a ride-sourcing service. In 99th
transportation research board annual meeting, Washington DC.
Chemla, D., Meunier, F., & Woler Calvo, R. (2013). Bike sharing systems: Solving the
static rebalancing problem. Discrete Optimization, 10(2), 120–146. https://doi.org/
10.1016/j.disopt.2012.11.005
Chen, Z., van Lierop, D., & Ettema, D. (2020). Dockless bike-sharing systems: What are
the implications? Transport Reviews, 40(3), 333–353. https://doi.org/10.1080/
01441647.2019.1710306
Daddio, D. (2012). Maximizing bicycle sharing: An empirical analysis of capital bikeshare
usage (Vol. 45).
Dell’Amico, M., Hadjicostantinou, E., Iori, M., & Novellani, S. (2014). The bike sharing
rebalancing problem: Mathematical formulations and benchmark instances. Omega,
45, 7–19. https://doi.org/10.1016/j.omega.2013.12.001
DeMaio, P. (2009). Bike-sharing: History, impacts, models of provision, and future.
Journal of Public Transportation, 12(4), 41–56. https://doi.org/10.5038/2375-
0901.12.4.3
El-Assi, W., & Mahmoud, M. S. (2015). Effects of built environment and weather on bike
sharing demand: A station level analysis of commercial bike sharing in toronto (Vol. 39).
Eren, E., & Uz, V. E. (2020). A review on bike-sharing: The factors affecting bike-sharing
demand. Sustainable Cities and Society, 54, Article 101882. https://doi.org/10.1016/
j.scs.2019.101882
European Commission. (2011). European commission. Transport White Paper (p. 31)
https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:52011DC0144
&from=EN.
Faghih-Imani, A., Hampshire, R., Marla, L., & Eluru, N. (2017). An empirical analysis of
bike sharing usage and rebalancing: Evidence from Barcelona and Seville.
Transportation Research Part A: Policy and Practice, 97, 177–191. https://doi.org/
10.1016/j.tra.2016.12.007
Fishman, E. (2016). Bikeshare: A review of recent literature. Transport Reviews, 36(1),
92–113. https://doi.org/10.1080/01441647.2015.1033036
Fishman, E., & Cherry, C. (2016). E-Bikes in the mainstream: Reviewing a decade of
research. Transport Reviews, 36(1), 72–91. https://doi.org/10.1080/
01441647.2015.1069907
Fishman, E., Washington, S., & Haworth, N. (2012). Understanding the fear of bicycle riding
in Australia (Vol. 23, p. 9).
Fishman, E., Washington, S., & Haworth, N. (2013). Bike share: A synthesis of the
literature. Transport Reviews, 33(2), 148–165. https://doi.org/10.1080/
01441647.2013.775612
Galatoulas, N.-F., Genikomsakis, K. N., & Ioakimidis, C. S. (2020). Spatio-temporal trends
of E-bike sharing system deployment: A review in Europe, north America and asia.
Sustainability, 12(11), 4611. https://doi.org/10.3390/su12114611
Gavalas, D. (2016). Chapter 13—design and management of vehicle-sharing systems: A
survey of algorithmic approaches (Vol. 29).
Guide to Net Dollar Retention (NDR)—Denition, Calculation, Tips. (2021).
UserGuiding. https://userguiding.com/blog/net-dollar-retention-ndr/.
Gu, T., Kim, I., & Currie, G. (2019). To be or not to be dockless: Empirical analysis of
dockless bikeshare development in China. Transportation Research Part A: Policy and
Practice, 119, 122–147. https://doi.org/10.1016/j.tra.2018.11.007
The Hague in numbers. https://www.denhaag.nl/en/municipality-of-the-hague/the-ha
gue-in-numbers.htm, (2021).
Hampshire, R. C. (2012). An analysis of bike sharing usage: Explaining trip generation 2 and
attraction from observed demand (Vol. 17).
Hua, M. (2020). Forecasting usage and bike distribution of dockless bike-sharing using journey
data. https://doi.org/10.1049/iet-its.2020.0305
Ji, S., Cherry, C. R., Han, L. D., & Jordan, D. A. (2014). Electric bike sharing: Simulation
of user demand and system availability. Journal of Cleaner Production, 85, 250–257.
https://doi.org/10.1016/j.jclepro.2013.09.024
Liu, S., Hou, J., Liu, Y., Khadka, A., & Liu, Z. (2018). OD demand forecasting for the
large-scale dockless sharing bike system: A deep learning approach. CICTP, 2018,
1683–1692. https://doi.org/10.1061/9780784481523.167
Liu, T. L. K., Krishnakumari, P., & Cats, O. (2019). Exploring demand patterns of a ride-
sourcing service using spatial and temporal clustering. In 2019 6th international
conference on models and technologies for intelligent transportation systems (MT-ITS)
(pp. 1–9). https://doi.org/10.1109/MTITS.2019.8883312
Li, A., Zhao, P., Huang, Y., Gao, K., & Axhausen, K. W. (2020). An empirical analysis of
dockless bike-sharing utilization and its explanatory factors: Case study from
Shanghai, China. Journal of Transport Geography, 88, Article 102828. https://doi.
org/10.1016/j.jtrangeo.2020.102828
Li, S., Zhuang, C., Tan, Z., Gao, F., Lai, Z., & Wu, Z. (2021). Inferring the trip purposes
and uncovering spatio-temporal activity patterns from dockless shared bike dataset
in Shenzhen, China. Journal of Transport Geography, 91, Article 102974. https://doi.
org/10.1016/j.jtrangeo.2021.102974
Ma, X., Ji, Y., Yuan, Y., Van Oort, N., Jin, Y., & Hoogendoorn, S. (2020). A comparison in
travel patterns and determinants of user demand between docked and dockless bike-
sharing systems using multi-sourced data. Transportation Research Part A: Policy and
Practice, 139, 148–173. https://doi.org/10.1016/j.tra.2020.06.022
van Marsbergen, A., Ton, D., Nij¨
enstein, S., Annema, J. A., & van Oort, N. (2022).
Exploring the role of bicycle sharing programs in relation to urban transit. Case
Studies on Transport Policy, 10(1), 529–538. https://doi.org/10.1016/j.
cstp.2022.01.013
Ma, X., Yuan, Y., Van Oort, N., & Hoogendoorn, S. (2020). Bike-sharing systems’ impact
on modal shift: A case study in Delft, The Netherlands. Journal of Cleaner Production,
259, Article 120846. https://doi.org/10.1016/j.jclepro.2020.120846
M´
edard de Chardon, C., Caruso, G., & Thomas, I. (2017). Bicycle sharing system ‘success’
determinants. Transportation Research Part A: Policy and Practice, 100, 202–214.
https://doi.org/10.1016/j.tra.2017.04.020
van Mil, J. F. P., Leferink, T. S., Annema, J. A., & van Oort, N. (2021). Insights into
factors affecting the combined bicycle-transit mode. Public Transport, 13(3),
649–673. https://doi.org/10.1007/s12469-020-00240-2
Miranda-Moreno, L. F., & Nosal, T. (2011). Weather or not to cycle: Temporal trends and
impact of weather on cycling in an urban environment. Transportation Research
Record: Journal of the Transportation Research Board, 2247(1), 42–52. https://doi.
org/10.3141/2247-06
Montes, A., Gerˇ
zinic, N., Veeneman, W., van Oort, N., & Hoogendoorn, S. (2023). Shared
micromobility and public transport integration - a mode choice study using stated
preference data. Research in Transportation Economics, 99, Article 101302. https://
doi.org/10.1016/j.retrec.2023.101302
Raviv, T., Tzur, M., & Forma, I. A. (2013). Static repositioning in a bike-sharing system:
Models and solution approaches. EURO Journal on Transportation and Logistics, 2(3),
187–229. https://doi.org/10.1007/s13676-012-0017-6
Rokach, L., & Maimon, O. (2005). Clustering methods. In O. Maimon, & L. Rokach (Eds.),
Data mining and knowledge discovery handbook (pp. 321–352). Springer-Verlag.
https://doi.org/10.1007/0-387-25465-X_15.
Shaheen, S. A., Martin, E. W., Chan, N. D., Cohen, A. P., & Pogodzinski, M. (2014). Public
bikesharing in north America during a period of rapid expansion: Understanding business
models, industry trends and user impacts (Vol. 234).
SmartCitiesWorld. (2017). The benets of smart bike-sharing. Smart Cities World. https://
www.smartcitiesworld.net/news/the-benets-of-smart-bike-sharing-2420.
The Hague. (2021). In Wikipedia. https://en.wikipedia.org/w/index.php?title=The_Hagu
e&oldid=1048037264.
Tin Tin, S., Woodward, A., Robinson, E., & Ameratunga, S. (2012). Temporal, seasonal
and weather effects on cycle volume: An ecological study. Environmental Health, 11
(1), 12. https://doi.org/10.1186/1476-069X-11-12
Wijken en buurten in Den Haag. (2021). In Wikipedia. https://nl.wikipedia.org/w/index.
php?title=Wijken_en_buurten_in_Den_Haag&oldid=60056431.
Xing, Y., Wang, K., & Lu, J. J. (2020). Exploring travel patterns and trip purposes of
dockless bike-sharing by analyzing massive bike-sharing data in Shanghai, China.
Journal of Transport Geography, 87, Article 102787. https://doi.org/10.1016/j.
jtrangeo.2020.102787
Z. Zhang et al.
