International Journal of Forecasting ( ) –

Contents lists available at ScienceDirect

International Journal of Forecasting

journal homepage: www.elsevier.com/locate/ijforecast

Spatio-temporal modeling of yellow taxi demands in New

York City using generalized STAR models
Abolfazl Safikhani a, *, Camille Kamga b , Sandeep Mudigonda c ,
Sabiheh Sadat Faghih b , Bahman Moghimi b
a
Columbia University, Department of Statistics, 1255 Amsterdam Avenue, New York, NY 10027-6938, United States
b
City University of New York, New York, NY, United States
c
City College of New York, New York, NY, United States

article info a b s t r a c t
Keywords: The spatio-temporal variation in the demand for transportation, particularly taxis, in the
STARMA highly dynamic urban space of a metropolis such as New York City is impacted by various
Spatio-temporal factors such as commuting, weather, road work and closures, disruptions in transit services,
Time series
etc. This study endeavors to explain the user demand for taxis through space and time by
Taxi demand prediction
proposing a generalized spatio-temporal autoregressive (STAR) model. It deals with the
high dimensionality of the model by proposing the use of LASSO-type penalized methods
for tackling parameter estimation. The forecasting performance of the proposed models
is measured using the out-of-sample mean squared prediction error (MSPE), and the
proposed models are found to outperform other alternative models such as vector au-
toregressive (VAR) models. The proposed modeling framework has an easily interpretable
parameter structure and is suitable for practical application by taxi operators. The efficiency
of the proposed model also helps with model estimation in real-time applications.
© 2018 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

1. Introduction to a much greater efficiency and more nuanced economic

Taxi services have long been an important part of ur-
ban transportation. Traditionally, ride hailing has been
performed by customers from the curbside of streets;
however, such street-hail taxis could be inefficient at ad-
dressing spatio-temporal variations in demand. The highly
dynamic urban space in a metropolis such as New York City
means that the spatio-temporal variation in the demand
for taxis is impacted by various factors, such as commuting,
weather, special events, parades, road work and closures,
disruptions in transit services, etc. However, the rise of
transportation network companies (TNCs) is making ride
hailing more on-demand. TNCs use economic means such
as surge pricing to address some of the inefficiencies, but
accurate predictions of the demand for taxis could lead
* Corresponding author.
E-mail address: as5012@columbia.edu (A. Safikhani).
measures. It has been shown that providing more accu-
rate predictions of the demand for taxis can improve the
efficiency of the transportation system (Moreira-Matias,
Gama, Ferreira, Mendes-Moreira, & Damas, 2013), since a
better knowledge of the demand in various parts of the city
at different times allows cab drivers to reduce their ride-
seeking trips, resulting in a reduction in traffic congestion
and pollution. On the other hand, inaccurate taxi demand
forecasts could leave passengers in some parts of the city
with long wait times. Hence, accurate short-term taxi de-
mand prediction enables the TNCs to dynamically reroute,
schedule and optimize operations.
The demand for ride hailing taxis in New York City
(NYC) is highly variable, with between about 150,000 and
600,000 trips per day being provided by 21,263 street hail
taxis in 2015 (New York City Taxi & Limousine Commission,
2016; see also Fig. 1). This demand also has a high spatial
variability, with about 383,000 pickups in Manhattan and

https://doi.org/10.1016/j.ijforecast.2018.10.001
0169-2070/© 2018 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
2 A. Safikhani et al. / International Journal of Forecasting ( ) –

Fig. 1. Temporal and spatial variation in taxi demand.

Fig. 2. Spatial variation of taxi demand aggregated by zip code in Manhattan.

only 3,150 pickups in the Bronx on an average day. The use
of GPS enabled the spatio-temporal historical demand for
taxis in the year of 2015 to be disaggregated to several sub-
regions within the city (see Fig. 2).
These taxis travelled approximately 460 million miles
in 2015 (New York City Taxi & Limousine Commission,
2016), the distribution of which can be seen in Fig. 3. Due
to the myriad of factors that can impact demand, which
may or may not be known in advance, there is scope for
taxis to drive around seeking rides – some of which could
be included in the number of trips under one mile in Fig. 3.
This study models the demand for taxis as a dy-
namic spatio-temporal process. The historical GPS-enabled
spatio-temporal demand for taxis in the year 2015 (pro-
vided by the Taxi and Limousine Commission of New York
City) is used and aggregated to several sub-regions within
the city.
We examine the demand for taxis in NYC through
space and time using a spatio-temporal autoregressive
model. The spatio-temporal autoregressive moving aver-
age (STARMA) model is a well-established spatio-temporal
process that was first introduced by Pfeifer and Deutsch
(1980, 1981) and has been applied in many different dis-
ciplines since, such as economics and finance (Giacomini
& Granger, 2004; Hernández-Murillo & Owyang, 2006;
Shoesmith, 2013), social science (Pfeifer & Deutsch, 1980;
Sartoris, 2005), transportation (Cheng, Wang, Harworth,
Heydecker, & Chow, 2011; Duan, Mao, Zhang, & Wang,
2016; Kamarianakis & Prastacos, 2003), climatology (Kyr-
iakidis & Journel, 1999), health sciences (Baklanov et al.,
2007), etc. Modeling the demand through time in all sub-
regions simultaneously – a vector autoregressive model
(VAR) – is a high-dimensional problem, since the number of
parameters in the model is proportional to the square of the
number of sub-regions. A STAR-type modeling approach
reduces the number of parameters dramatically by gov-
erning the neighborhood structure between the regions.
This structure is also useful for capturing the spatial depen-
dence of the demand between the regions, and makes the
results more interpretable.
In this research, we use the spatio-temporal structure
for predicting the yellow taxi demand for different zip
codes of Manhattan, a borough of NYC. For this purpose,
GPS taxi demand data for two ordinary days in October

Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
 A. Safikhani et al. / International Journal of Forecasting (
Fig. 3. Distribution of distance travelled by taxis in 2015.
2015 are disaggregated based on the zip code. Such an
approach requires the use of a zoning system for studying
the spatial characteristics of the data. The system should
divide the study area into smaller districts in order to assess
the existence of spatial correlations among them. However,
districts cannot be too small (like census tracks), since the
GPS devices are not 100% accurate and can cause pick-up
counts in each district to be less accurate.
We also ensure temporal variation in the demand by
disaggregating the GPS data into 15 min intervals, result-
ing in 96 time-points per day. Other studies (e.g. Qian,
Ukkusuri, Yang, & Yan, 2017) have also decomposed such
data by time (15-min intervals) and space (zip code basis).
In addition to the use of spatio-temporal structure for taxi
demand prediction, the main contribution of this paper
is the introduction of new penalty functions in the STAR
model. Various different penalty functions are used in this
paper for parameter estimation in the proposed general-
ized STAR model, and we show that they improve the accu-
racy of the taxi demand prediction. The double hierarchical
group LASSO (DHGLASSO) is the new penalty scheme that
is introduced in this paper for penalizing parameters in the
hierarchical structure of modeling in both the space and
time dimensions. We evaluate the forecasting performance
of the proposed method using the out-of-sample mean
squared prediction error (MSPE), and the results show that
the proposed model outperforms some of the alternative
algorithms such as ARMA and VAR models.
Given that there are about 12 million taxi trips a month,
amounting to 2 GB of data, a demand forecasting model
with a good spatial and temporal predictive accuracy is
very useful. In particular, the proposed model has the abil-
ity to forecast the taxi demand a few steps into the future
at various locations in NYC, and this enables the agency’s
framework to plan taxi positioning and provide demand-
sensitive taxi dispatching for various locations and specific
times of the day over the year. When using a model for
placement of taxis, the yellow taxis may not operate as yel-
low taxis per se. In other words, they could serve demand-
based rides as well as curb-side pickups. In fact, some
Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
) – 3
yellow taxis are already using app-based ride providers
such as Curb. The objective of the proposed modeling
framework would be to direct taxis to certain zones at cer-
tain times. Actual dispatch to pick up a certain passenger
is not a part of the proposed framework, though it could
be done using the actual location (within the zone) of a
passenger requesting a ride through an app-based ride-
providing service. The main benefit of this approach is that
it avoids taxis cruising for rides, especially across different
zones, which would decrease the unnecessary distance
travelled, with the associated fuel costs and pollution. In
addition, from a policy standpoint, the spatio-temporal
structure inferred from the demand data provides a basis
to allow regulating agencies to explore cordon pricing ini-
tiatives.
This paper is organized as follows. Section 2 discusses
the literature on time series modeling in transportation
and short-term taxi demand prediction. Section 3 provides
a detailed description of the spatio-temporal modeling and
formulation of taxi demand using the STARMA approach.
Section 4 presents findings for various types of STARMA
models and prediction errors, and compares them with
those from other time series models. Finally, we present
our conclusions and future research directions.
2. Literature review
With the rise of intelligent systems and increases in
the availability of data, taxi pick-up demand prediction
has recently come to the attention of many scholars. It is
obvious that the taxi demand in any given zone changes
from one time interval to another in real time. Hence,
time series models can form a strong statistical tool for
capturing this time variation in taxi demand, figuring out
the correlations within the taxi data, and producing real
time predictions. For univariate data, a well-known family
of time series models called ARIMA (autoregressive in-
tegrated moving average) models can be beneficial, and
have been applied to many transportation-related prob-
lems (Moghimi, Safikhani, Kamga, & Hao, 2017). However,
4 A. Safikhani et al. / International Journal of Forecasting (
in the dense urban transportation network that we con-
sider, with many different areas or zip codes that each
have their own demand dynamics and may be correlated
with one another, the taxi demand variation of a given zip
code is not related only to its own values, but also to the
demand in the neighboring zip codes. Since there are too
many parameters to estimate due to multiple zip codes,
VAR models, as the most common multivariate time series
model, will not be able to perform well and will do a poor
job of forecasting the taxi demand. This study attempts to
mitigate this problem by applying a multivariate spatial–
temporal time series model which is discussed in details in
Section 3
Some of the primary research about taxi demand has
aimed to find the factors that influence taxi demand.
Schaller (1999) developed a citywide empirical time series
regression model of NYC taxis in an attempt to understand
the relationship between the taxicab revenue per mile and
economic activity in the city, taxi supply, taxi fares, and
bus fares. Later, Schaller (2005) tried to figure out the rela-
tionships between taxi demand and other factors including
the city size, the availability and cost of privately owned
autos, the use of complements to taxicabs, the cost of taxi
usage, the taxi service quality, the presence of competing
modes, and the presence of a senior or disabled popula-
tion. With the emergence of GPS technology, subsequent
extensive research into the use of spatial information has
been applied in the context of transportation-related prob-
lems. GPS based systems are also used to track the taxis
of New York City and to analyze the taxi ridership. Yang
and Gonzales (2017) processed the New York City GPS taxi
data and used the negative binomial method to capture
the variation in the taxi pick-up demand. Their study used
six explanatory variables, namely population, education,
median age, median income per capita, employment by
industry sector, and transit accessibility. Correa, Xie, and
Ozbay (2017) performed an empirical analysis in order
to explore the spatial dependence between Uber and taxi
pick-up data. The results from Moran’s I tests confirmed
the significant spatial correlation in both taxi and Uber
demand.
Several studies have considered prepositioning taxis so
as to reduce wait times (Chang, Tai, & Hsu, 2010; Yuan,
Zheng, Zhang, Xie, & Sun, 2011) using spatio-temporal
clustering. Time series models such as ARIMA have also
been tested for taxi demand prediction (Moreira-Matias
et al., 2013; Qian et al., 2017; Sayarshad & Chow, 2016).
Moreira-Matias et al. (2013) proposed a methodology for
the prediction of short-term taxi demand at 30-min time
intervals. Their methodology is an ensemble of three pre-
dictive models, namely a time-varying Poisson model,
a weighted time-varying Poisson model, and an ARIMA
model. They found that their proposed model outper-
formed all three models run individually. The recent study
by Qian et al. (2017) also used artificial neural networks
to combat nonlinearities in the tax demand. Furthermore,
they attempted to capture spatio-temporal variations us-
ing conditional random fields. The proposed model and
two other algorithms (ARIMA and ANN) were run in four
different scenarios and their performances evaluated. The
results reported that the proposed model outperformed
Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
) –
both of the other algorithms, with mean absolute percent-
age errors (MAPE) close to 0.1. The case study used in the
current paper is the same as that used by Qian et al. (2017).
It is well-known that spatial information can increase
the accuracy of prediction, especially for traffic congestion
and at longer horizons. The idea of capturing spatial in-
formation in time series studies of transportation-related
problems was first introduced in the study by Okutani and
Stephanedes (1984) for the prediction of traffic flow. Later,
the spatial concept was deployed in the study by Kamar-
ianakis and Prastacos (2003) for forecasting the relative
velocity on major roads in Athens, Greece, in a method re-
ferred to as the space–time autoregressive integrated mov-
ing average (STARIMA) model. The model is quite different
from traditional ARIMA models due to the inclusion of
spatial information regarding neighboring links for traffic
forecasting. They compared the forecasting performances
of four models, namely the historical average, ARIMA,
VARMA, and STARIMA. The results demonstrated that there
are no significant differences among the last three models,
although the last three models all performed better than
the historical average one. Spatial–temporal modeling is
also used in various other areas of transportation; for ex-
ample, the traffic condition of the downstream section of
a road is highly correlated with that observed upstream.
Stathopoulos and Karlaftis (2003) considered the spatial
information from four consecutive loop detectors in the
area upstream of the study section for predicting the traffic
flow in the downstream of an urban corridor, while the
same idea was used in the studies by Cheng et al. (2011)
and Duan et al. (2016) for predicting the traffic speed of
the downstream link.
One of the most important parts of STARIMA modeling
is the spatial weighting matrix, which indicates the spa-
tial dependency between multiple time series. Thus, the
optimal spatial weighting matrix varies with the nature of
the problem, and determining it requires some engineer-
ing judgment. In general, two approaches have been used
for selecting the neighboring dependence: (a) correlation-
coefficient assessment and (b) distance adjustment. The
values in STARIMA’s weighting matrix can vary by time and
location. In one method that has been developed, called
General STARIMA, the spatial parameters are designed to
vary by location, instead of having fixed values over all
locations (Min, Hu, & Zhang, 2010). Another approach that
is associated with the weighting matrix is to consider only
the link/zone that is adjacent to the target link/zone. It
can be elaborated by a ring of dependency, defined by the
‘‘order’’. For instance, a first-order adjacent matrix repre-
sents the dependency between the study link/zone and
its immediate adjacent link/zone (first-order link/zone). A
second-order adjacent matrix shows the dependency of
zone that is not directly adjacent to the study zone, but it
is an immediate adjacent to the link/zone defined as first-
order stated earlier. It can also be expanded to a third-
order adjacent matrix, and so forth. First- and second-order
adjacency-weighting matrices were used in the study by
Kamarianakis, Prastacos, and Kotzinos (2004). On the other
hand, it is more practical to use the distance between the
two links/zones, where the value of the dependency is
reduced by increasing the distance.
 A. Safikhani et al. / International Journal of Forecasting (
Fig. 4. Sample ACFs of the first 5 components.
3. Methodology
This section introduces the proposed spatio-temporal
model and briefly discusses the implementation of the
model. Suppose that k different time series are observed
over a period of duration T . If one chooses vector autore-
gression (VAR) models with a maximum time lag of p to
fit the data, a total of k2 × p parameters need to estimated
using the k × T observed data points. Now, if k is relatively
large compared to T , then the number of parameters in
the model will be more than the number of data points
observed. This is called a high-dimensional problem. The
typical least square methods cannot be used, as the design
matrix will not be invertible. The data set that we explore
in this paper has features similar to high-dimensional time
series. More specifically, we consider the yellow taxi de-
mand in NYC for the day October 6th, 2015, a date chosen
because it is a typical day without any holidays or special
events nearby. The demand is aggregated spatially over
the zip codes, and temporally every 15 min, making it a
multivariate time series with more than 100 components.
However, only 39 of the zip codes have enough non-zero
counts to keep them in the model. It is worth noting that
the zip code zoning system includes some small zones
(even as small as a block/building) that should be removed
from the inputs, as the demand in these small zones is not
of interest. Thus, the dataset ultimately consists of k = 39
locations and T = 96 time points. Fig. 4 shows the sample
ACFs of the first five components of the data, which imply
the existence of a strong temporal dependence. Hence, a
multivariate time series model is chosen for analyzing this
dataset.
Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
) – 5
Due to the high dimensionality of the data, simple VAR
models are not appropriate here. Instead, this section de-
velops a generalized version of the STARMA model, which
takes into account the topology of the locations at which
the data is observed, with the aim of increasing the predic-
tion performance efficiency. STARMA models, introduced
by Pfeifer and Deutsch (1980, 1981), are spatio-temporal
models that reduce the number of parameters in a typical
VAR model by introducing neighborhood structures. Here,
we focus only on the autoregressive (AR) part of this model,
since it is more interpretable. A multivariate time series
Y (t ) = (Y1 (t ) , . . . , Yk (t )) , t = 1, 2, . . . , T , is called a
generalized STAR of order p (see Di Giacinto, 1994, and
Terzi, 1995, for an introduction, and Di Giacinto, 2006, for
its application to regional unemployment analysis) if for
each t = 1, 2, . . . , T and i = 1, 2, . . . , k,
P ηj −1
φi(j,l) Wi(l) Y (t − j) + εi (t ) ,
∑ ∑
Yi (t ) = (1)
j=1 l=0
where εi (t ) = (ε1 (t ) , . . . , εk (t )) is a k-variate normal
variable with mean zero and
{
σ 2 Ik , s = 0
E ε (t ) ε(t + s)′ =
( )
0, ∧other w ise.
Also, the W (l) s are k × k weighting matrices that govern
(0)
the lth neighborhood location, with Wi = Ik . Denote
(l)
the ith row of W by Wi . One possible choice for W (l)
(l)
is to set W (l) (i, j) = 1 if the ith and jth locations are
lth level neighbors, and W (l) (i, j) = 0 otherwise. These
matrices are then normalized in such a way that the sum
of each row is 1. Finally, for each i = 1, 2, . . . , k, and
6 A. Safikhani et al. / International Journal of Forecasting ( ) –

(j,0:ηj −1) (j,η −1)

( )
j = 1, 2, . . . , p, φi = φi(j,0) , φi(j,1) , . . . , φi j is • HGLASSO (hierarchical group LASSO): similar to the
a vector of coefficients of size ηj that relates the current ob- HVAR method introduced by Nicholson, Bien, and Matte-
servation at location i , Yi (t ), to all weighted observations son (2014) and Nicholson, Matteson, and Bien (2017) for
in ηj different neighborhoods j time lags in the past. With- sparse VAR models. The coefficients for each time lag are
out loss of generality, it is assumed that η1 = ·(· · = ηp = η) grouped together, and they are penalized more if the time
(if they are different, one can choose η = max η1 , . . . , ηp lags are higher by means of a time-lag hierarchical group
(j,l) structure. More specifically, denoting
and set some (of the φi coefficients to)zero). Furthermore, (
(j,0:η−1) (p,0:η−1)
)
(1,0:η−1) (p,0:η−1) (j:p)
denote Φi = φi , . . . , φi . In the generalized Φi = φi , . . . , φi for j = 1, 2, . . . , p,
STAR model in Eq. (1), the current value Yi (t ) depends on
the previous values observed at nearby locations through p  
∑  (j:p) 
both the φ parameters and the weighting matrices W. Ω (Φi ) = Φi  . (6)
2
The difference between VAR and STAR models is that VAR j=1
models have the current value depending on the previ-
• DHGLASSO (double hierarchical group LASSO): similar to
ous values of all other time series components, whether
HGLASSO, but with an additional neighborhood-lag hierar-
they are nearby observations or not. Thus, VAR models
chical group structure penalty term. Denoting
don’t utilize the spatial structure, meaning that they may
perform poorly for prediction, especially when the spatial (j:p,l:η−1)
(
(j,l:η−1) (j+1,0:η−1) (p,0:η−1)
)
Φi = φi , φi , . . . , φi ,
correlations are high. It is worth noting that there are
also various other models that use the spatial structure, j = 1, 2, . . . , p, l = 0, 2, . . . , η − 1, (7)
including the conditional autoregressive model (CAR), spa-
one can write the penalty function as
tial autoregressive models, and the simultaneously autore-
gressive model (SAR; see Cressie, 2015, and LeSage, 1997, p η−1 
∑ ∑  (j:p,l:η−1) 
for more details). However, none of these have any notion Ω (Φi ) = Φi  . (8)

2
of time, meaning that they are not applicable for our time j=1 l=0
series data, especially given that our objective is to predict
the values of demand in the future.
3.1. Implementation
It would be more convenient to write Eq. (1) in a com-
pact matrix form. For that, let Yi = Yi (1) , . . . , Yi (T ) ∧
εi = (εi (1) , . . . , εi (T )) , and define Zi as T × ηp with The solving of optimization problems of the type in
(l) Eq. (4) has been studied extensively under the penalty
Zi (t , (j − 1) η + l) = Wi Y (t − j) for t = 1, 2, . . . , T , j =
terms introduced previously (see Tibshirani, 1996, and ref-
1, 2, . . . , p, and l = 0, 2, . . . , η − 1. Now, one can write the
erences therein). Due to the hierarchical structure of the
data equation for the ith time series component as
group penalties in HGLASSO and DHGLASSO, here we ap-
Yi = Zi Φi + εi , (2) ply the proximal gradient method introduced by Jenatton,
Mairal, Obozinski, and Bach (2011). Furthermore, the con-
2
This model reduces the number of parameters from k × p vergence rate of the proximal gradient method has been
in the VAR model to k × η × p, assuming η ≪ k. Least improved by Beck and Teboulle (2009) by introducing the
squares estimation can be implemented for parameter es- fast iterative soft-thresholding algorithm (FISTA), where a
timation; i.e., for i = 1, 2, . . . , k, sequence of matrix coefficients Φ̂i [r] , r = 1, 2, . . ., are
1 introduced iteratively through
Φ̂i = argminΦi ∥Yi − Zi Φi ∥22 , (3)
2 r −2 ( )
with ∥.∥2 being the Euclidean norm. However, for the cases φ̂ = Φ̂i [r − 1] + Φ̂i [r − 1] − Φ̂i [r − 2]
r +1
where T is small compared to k, it might still be beneficial
to reduce the number of parameters in the model with the ( ( ))
aim of improving the forecast performance. For that pur- Φ̂i [r] = ProxsλΩ φ̂ − s∇ fi φ̂ , (9)
pose, a penalty function Ω (Φ ) will be added to Eq. (3) with
the aim of setting some of the small parameters to zero so where fi (Φi ) = 21 ∥Yi − Zi Φi ∥22 ∇ fi (Φi ) = −Zi′ (Yi − Zi Φi )
as to increase the forecast efficiency. More specifically, is the vector of derivatives of fi (Φi ) , s is the step size (here,
we choose s to be 1/σ1 (Zi )2 , where σ1 (Zi ) is the largest
1
Φ̂i = argminΦi ∥Yi − Zi Φi ∥22 + λΩ (Φi ) , (4) singular value of Zi ), and
2 ( )
1
where λ is the tuning parameter to be selected by cross ProxsλΩ (u) = argminυ ∥u − υ∥2 + sλΩ (υ) . (10)
validation techniques. We define several penalty functions 2
and evaluate their performances on the yellow taxi de- The proximal function may not have a closed form in
mand data. More specifically, we consider the following general, and in that case, it must itself be approximated
penalty functions. numerically. However, in the case of a hierarchical group
• LASSO: a simple element-wise L1 penalty on all the com- penalty, this function does in fact have a simple closed form
ponents of Φi ; i.e., for i = 1, 2, . . . , k, (see for example algorithm 2 of Nicholson et al. (2014)),
p η−1 ⏐ which makes the whole optimization efficient. The tuning
∑ ⏐ (j,l) ⏐⏐
parameter λ is selected based on a rolling scheme cross-
∑
Ω (Φi ) = ⏐φi ⏐ . (5)
j=1 l=0 validation procedure that was used also by Nicholson et al.

Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
 A. Safikhani et al. / International Journal of Forecasting ( ) – 7

Table 1
Results for October 6th data with η = 1.
Model MSPE MRPE AIC BIC
VAR 1.7153 4.8259 216.4933 257.1222
STAR (Univariate AR) 0.2815 2.4463 176.9854 178.0271
LASSO 0.2977 1.8467 173.8735 174.9153
HGLASSO 0.2977 1.8467 173.8735 174.9153
DHGLASSO 0.2977 1.8467 173.8735 174.9153

(2014, 2017) and Song and Bickel (2011). For this purpose, Table 2
the time points are divided into three parts (usually at Results for October 6th data with η = 2.
equal distances), 0 < T1 < T2 < T . The estimation Model MSPE MRPE AIC BIC
procedure for fixed values of λ is applied for the first part, STAR 0.2707 1.9913 177.3313 179.4148
i.e., t = 1, 2, . . . , T1 ; then the mean squared prediction LASSO 0.2728 1.9616 177.0052 179.0353
error (MSPE) for predicting one step ahead is calculated HGLASSO 0.2909 1.8942 176.0614 178.1449
DHGLASSO 0.2907 1.9543 178.6925 180.6425
over all k time series components on the time interval
[T1 + 1, T2 ] :
Table 3
k T2 Results for October 6th data with η = 3.
1 ∑ ∑ )2
Yi (t ) − PT1 Yi (t ) ,
(
MSPE = (11) Model MSPE MRPE AIC BIC
k (T2 − T1 )
i=1 t =T1 +1 STAR 0.2932 2.1346 178.7531 181.8784
k T2 LASSO 0.3254 2.1413 177.3218 179.1115
1 ∑ ∑
HGLASSO 0.301 2.114 175.1811 178.3064
MRPE =
k (T2 − T1 ) DHGLASSO 0.2821 1.9472 176.3991 179.3107
i=1 t =T1 +1

× |Yi (t ) − PT1 Yi (t )| |Yi (t )|, (12) Table 4

Results for October 6th data with η = 4.
where PT1 Yi (t ) is the best linear predictor of Yi (t ) based
Model MSPE MRPE AIC BIC
on the first T1 observations. The mean of the relative pre-
diction error (MRPE) is also shown in Eq. (12). Now, we STAR 0.3582 2.4261 182.2474 186.3877
select the tuning parameter λ that minimizes this MSPE, LASSO 0.3506 2.2577 177.8353 180.3196
HGLASSO 0.3412 2.2928 177.0926 181.2329
and quantify the model performance based on the MSPE DHGLASSO 0.2968 1.882 176.5145 180.0939
on the last part of the data, which covers the time interval
[T | |2 + 1, T ].

4. Results 4.1. Case study using data from October 6th only

This section applies the proposed methods to the yellow
taxi demand data on different days, and calculates their
prediction performances under different scenarios. Based
on the sample ACFs of the data, p is chosen to be 1, and the
calculation of the AIC/BIC also supports this selection. How-
ever, before applying different methods to this dataset, it
needs to be scaled properly. For this purpose, the sample
mean is subtracted from each time series that corresponds
to a zip code, and the resulting series are divided by the
sample standard deviation, so that all time series have the
same scales. Also, the weighting matrices W are chosen for
five different neighborhood levels based on the authors’
judgment, or, more specifically, by counting the numbers of
boundaries between the target zip code and its neighbors.
For example, a zip code that is adjacent to the target zip
code is considered as the first-order neighborhood; zip
codes adjacent to the first-order neighborhood are the
second-order neighborhood for the target zip code, and so
on. This study extends the neighborhood up to five levels by
means of an eyeballing procedure. October 6th and 7th are
chosen for this research because they are typical weekdays,
being away from both weekends and days with special
events. Two different approaches have been considered
for evaluating the performance of the developed model.
First, we consider time points from October 6th only; and,
second, the two days of October 6th and 7th are merged to
give a longer range of time points.
Considering data from October 6th, only T = 96 time
points are available. A rolling window scheme is used to
divide T into three parts, with T1 being set to ⌊T /3⌋ and
T2 to ⌊2T /3⌋. Different orders of neighborhood (η) are
chosen, and the MSPE, mean squared relative prediction
error (MRPE), AIC and BIC (see Lutkepohl, 2007, for the
definitions and formulas) values are reported in each case.
Tables 1–4 show the results for η = 1, 2, 3, 4, respec-
tively. In simple words, η = 1 only considers the previous
time data of the study zone which does not consider any
neighborhood information, η = 2 considers only the in-
formation of the first-order neighborhood, and it continues
up to 3rd-order neighbor which is η = 4. Obviously,
the VAR model does not perform well relative to STAR-
based models due to the huge number of parameters in-
volved. Based on the MSPE, the STAR and LASSO models
for η = 2 outperform the rest. The difference between
the prediction performances of these two methods, STAR
and LASSO, and those of the other methods is statistically
significant, with p -values of less than 0.01 based on the
Diebold–Mariano test statistics developed by Diebold and
Mariano (1995). This means that the inclusion of the first
neighborhood structure improves the forecasting perfor-
mance of the STAR model. Meanwhile, the proposed spatio-
temporal structure using the topology and zip code based
disaggregation of Manhattan with a first-order neighbor-
hood (η = 2) performs the best in this case study. Also, it

Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
8 A. Safikhani et al. / International Journal of Forecasting ( ) –

Table 5 Table 8
Results for October 6th and 7th data combined, with η = 1. Results for October 6th and 7th data combined, with η = 4.
Model MSPE MRPE AIC BIC Model MSPE MRPE AIC BIC
VAR 0.7103 14.544 204.3445 230.15 STAR 0.2247 4.4178 178.271 180.9008
STAR (univariate AR) 0.253 3.9068 182.6419 183.3035 LASSO 0.2244 4.3892 178.0977 180.6765
LASSO 0.2527 3.8983 182.5923 183.254 HGLASSO 0.2247 4.419 178.2357 180.8654
HGLASSO 0.2527 3.8983 182.5923 183.254 DHGLASSO 0.2224 4.162 177.6367 180.2156
DHGLASSO 0.2527 3.8983 182.5923 183.254
Table 9
Table 6 Results for October 6th and 7th data combined, with η = 5.
Results for October 6th and 7th data combined, with η = 2.
Model MSPE MRPE AIC BIC
Model MSPE MRPE AIC BIC STAR 0.2279 3.5851 178.7162 182.0077
STAR 0.2273 4.0633 178.3005 179.6239 LASSO 0.2257 3.5265 178.0827 181.1196
LASSO 0.2273 4.0633 178.3005 179.6239 HGLASSO 0.2277 3.5857 178.6503 181.9418
HGLASSO 0.2273 4.0633 178.3005 179.6239 DHGLASSO 0.2212 3.835 177.8113 180.95
DHGLASSO 0.2273 4.0633 178.3003 179.6237
Table 10
Table 7 Results for October 6th and 7th data combined, with η = 6.
Results for October 6th and 7th data combined, with η = 3.
Model MSPE MRPE AIC BIC
Model MSPE MRPE AIC BIC STAR 0.2405 3.5611 178.4606 182.4137
STAR 0.2249 4.1741 177.9721 179.9571 LASSO 0.238 3.4261 177.7624 181.2913
LASSO 0.2248 4.1703 177.9496 179.9347 HGLASSO 0.238 3.5376 178.2116 182.1647
HGLASSO 0.2249 4.1742 177.957 179.9421 DHGLASSO 0.2291 3.9304 178.8057 182.4703
DHGLASSO 0.2238 4.1062 177.6838 179.6519

Another benefit of the use of STAR-based models is

is worth mentioning that the DHGLASSO penalty function
provides a consistent model performance overall, since
its MSPE and MRPE values do not increase dramatically
with η. In other words, the DHGLASSO penalty structure is
better at correcting for the increase in the parameter space
dimension. Also, increasing the number of neighborhood
levels η, and thus the number of parameters in Φ , allows
the DHGLASSO method to reduce the MSPE by around 3%
relative to the STAR model when η = 3, and by around 17%
when η = 4. If the MRPE is selected as the measure of the
forecasting performance, the DHGLASSO is comparable to
the other leading models when η = 4.
4.2. Case study using data for October 6th and 7th combined
The same models and methods that were applied in
the previous case study are applied again here to the taxi
demand for two days, October 6th and 7th. This makes
the total number of time points 192, instead of 96 as in
the previous case study. Increasing T while keeping k fixed
reduces the effect of the penalization on parameter esti-
mation, and hence on the forecasting performance, as can
be seen from Tables 5–10, which show the performances
of the methods when η = 1, 2, 3, 4, 5, 6, respectively.
In this scenario, DHGLASSO for η = 5 outperforms the
other methods in terms of MSPE. The DHGLASSO with
η = 5 reduces the MSPE significantly relative to other
methods with different values of η. The Diebold–Mariano
test statistics developed by Diebold and Mariano (1995)
are applied to the different components of the data, and
the test shows that the improvement in the performance
of DHGLASSO relative to those of the other methods is sta-
tistically significant, with p-values of less than 0.01 in most
cases. Fig. 5 summarizes the MSPE results of all methods for
the six levels of η. Again, DHGLASSO is the most consistent
penalty function with respect to the increase in η.
that one can infer the neighborhood influences of other
zip codes’ demands on a given target zip code. Figs. 6, 7,
and 8 show the inferred neighborhood correlations among
the η = 5 different neighborhood orders for lower, mid-
town, and upper Manhattan, respectively. The colors on
these plots are basically |Φi | for different components of
i based on the DHGLASSO method. It is clear from these
plots that the correlation/influence between neighboring
zip codes decreases as they get farther apart. This corre-
lation structure is reasonable and well-aligned with the
assumptions involved in using a spatio-temporal model
such as the STARMA model for predicting taxi demand in
Manhattan, New York. In other words, a knowledge of the
short-term demand histories of neighboring zip codes in
lower Manhattan will be more informative for predicting
the taxi demand over the next 15 min for a zip code in lower
Manhattan than a knowledge of the short-term demand
histories of zip codes in upper Manhattan. The proposed
DHGLASSO model is able to capture this decreasing trend
accurately within the STARMA structure, reaching the low-
est prediction error among all methods considered.
Another notable feature that can be highlighted using
the proposed generalized STAR model with DHGLASSO is
the variation in the spatial differences in the dependence of
demand of neighboring zip codes. It can be seen from Fig. 6
that the values of the coefficients of the second and third
levels of neighboring zip codes are not the same among the
zip codes, even in lower Manhattan. More specifically, for
zip code 10280, the coefficient for the second-level neigh-
bors’ demand is less than that for the third-level neighbors.
However, for zip code 10002, the coefficients for first-,
second- and third-level neighbors’ zip codes demands all
decrease with the level of neighborhood. This non-linear
trend in the coefficients for neighboring zip codes could be
due to the smaller areas of the zip codes – particularly for
zip codes 10004 and 10280.

Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
 A. Safikhani et al. / International Journal of Forecasting ( ) – 9

Fig. 5. MSPE results for October 6th and 7th data combined, with η = 1, 2, . . . , 6.

Fig. 6. Neighborhood-level estimated coefficients for lower Manhattan (zip codes: 10004, 10002, 10280).

Fig. 7. Neighborhood-level estimated coefficients for midtown Manhattan (zip codes: 10019, 10022, 10128).

Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
10 A. Safikhani et al. / International Journal of Forecasting (
Fig. 8. Neighborhood-level estimated coefficients for upper Manhattan (zip codes: 10021, 10028, 10027).
5. Conclusion
The accurate prediction of yellow taxi demand in large,
populous, and dense areas of cities like New York is hard
to achieve, since there are numerous parameters that af-
fect the demand. Moreover, the demand for taxis in such
densely populated areas is highly variable in different parts
of the city, depending on the time of the day. In this
study, taxi demand data obtained from the historical us-
age of individual (GPS-enabled) taxis, obtained from NYC
TLC, are aggregated both spatially by zip code and tempo-
rally by 15-min time intervals. We propose a multivariate
spatio-temporal method called STARMA that reduces the
number of parameters dramatically compared to typical
multivariate time series models such as VAR, by means
of the neighborhood structure between the regions. This
structure is useful for capturing the spatial dependence of
the demand between the regions, and makes the results
more interpretable. We also present a new method for
penalizing prediction parameters, called the double hier-
archical group LASSO (DHGLASSO). DHGLASSO penalizes
parameters that are farther away not only temporally but
also spatially to a larger extent, thus establishing a ‘double’
hierarchy.
The proposed model and several other comparable time
series models and penalty functions are applied to the
yellow taxi demand of Manhattan for a typical day of the
week. The results reveal that the proposed model captures
the structure of the data well, with lower prediction errors
than the other time series models considered, such as
VAR, STAR with and without LASSO, etc. Using data from
both a single day and two consecutive days, the proposed
generalized STAR model with DHGLASSO performed the
best in terms of predictive performances. For the model
using data from two consecutive days, a maximum level of
five neighborhood orders performed the best. Additionally,
DHGLASSO is shown to be most consistent and stable when
dealing with increasing parameter dimensions.
The proposed generalized STAR model and penalty
function is able to capture the spatial variation in the de-
mand for taxis among zip codes very well. The effect of the
Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
) –
neighborhood structure changes depending on the location
of interest. The influence of neighborhood taxi demand lev-
els is easy to interpret, especially for agencies that manage
taxi operations and other TNCs. Taxi companies and TNCs
can easily use the neighborhood taxi demand dependence
to direct taxi drivers to remain in certain areas, depending
on the time of day and location. This helps to reduce the
lengths of empty taxi trips seeking new rides, thus reducing
emissions, improving air quality and reducing fuel costs for
the operators. The increased computational efficiency due
to the DHGLASSO penalization structure also helps with
estimating the model in real-time.
The proposed modeling approach can be used as a de-
mand prediction tool in its current format for planning
the positioning of taxis. However, in addition, one could
also include several other covariates that may potentially
represent or act as surrogates for the demand being served
by other providers for a real-time implementation. Also,
ideally, the parameters may be re-estimated periodically
to account for changes in the demand for yellow taxis,
which would account indirectly for the passenger de-
mand being served by other providers and help to predict
the demand for more current conditions. Even with such
modifications, the computational efficiency offered by the
double-hierarchical LASSO structure means that the pro-
posed modeling approach could still be useful for a real-
time implementation.
As part of ongoing and future work, the modeling frame-
work is being extended using other forms of disaggrega-
tion. The utilization of additional travel demand-related
information such as subway and bus ridership, bike de-
mand, weather, etc., will also be being considered through
the addition of exogenous variables to the time series
regime.
Acknowledgment
This work was supported in part by NSF, United States,
Grant IIS-1302423.
 A. Safikhani et al. / International Journal of Forecasting (
References
Baklanov, A., Hänninen, O., Slørdal, L. H., Kukkonen, J., Bjergene, N., Fay, B.,
et al. (2007). Integrated systems for forecasting urban meteorology,
air pollution and population exposure. Atmospheric Chemistry and
Physics, 7, 855–874.
Beck, A., & Teboulle, M. (2009). A fast iterative shrinkage-thresholding al-
gorithm for linear inverse problems. SIAM Journal on Imaging Sciences,
2(1), 183–202.
Chang, H., Tai, Y., & Hsu, J. Y. (2010). Context-aware taxi demand hotspots
prediction. International Journal of Business Intelligence and Data Min-
ing, 5(1), 3–18.
Cheng, T., Wang, J., Harworth, J., Heydecker, B. G., & Chow, A. H. F. (2011).
Modeling dynamic space–time autocorrelations of urban transport
network. GeoComputation, Session 5A: Network Complexity, 215–220.
Correa, D., Xie, K., & Ozbay, K. (2017). Exploring the taxi and uber demands
in New York City: an empirical analysis and spatial modeling. In
Transportation research board’s 96th annual meeting, Washington, D.C.
Cressie, N. (2015). Statistics for spatial data. John Wiley & Sons.
Di Giacinto, V. (1994). Su una generalizzazione dei modelIi spazio-
temporali autoregressivi media mobile (STARMAG). In Atti della
XXXVII riunione scienti_ca SIS, Sanremo, Aprile 1994, vol. H.
Di Giacinto, V. (2006). A generalized space–time ARMA model with an
application to regional unemployment analysis in Italy. International
Regional Science Review, 29(2), 159–198.
Diebold, F. X., & Mariano, R. S. (1995). Comparing predictive accuracy.
Journal of Business & Economic Statistics, 13, 134–145.
Duan, P., Mao, G., Zhang, C., & Wang, S. (2016). STARIMA-based traffic
prediction with time-varying lags. In IEEE 19th international conference
on intelligent transportation system, Rio, Brazil.
Giacomini, R., & Granger, C. W. J. (2004). Aggregation of space–time
processes. Journal of Econometrics, 118, 7–26.
Hernández-Murillo, R., & Owyang, M. T. (2006). The information content
of regional employment data for forecasting aggregate conditions.
Economics Letters, 90(3), 335–339.
Jenatton, R., Mairal, J., Obozinski, G., & Bach, F. (2011). Proximal methods
for hierarchical sparse coding. Journal of Machine Learning Research
(JMLR), 12, 2297–2334.
Kamarianakis, Y., & Prastacos, P. (2003). Forecasting traffic flow conditions
in an urban network: comparison of multivariate and univariate ap-
proaches. Transportation Research Record: Journal of the Transportation
Research Board, 1857, 74–84.
Kamarianakis, Y., Prastacos, P., & Kotzinos, D. (2004). Bivariate traffic
relations: A space–time modeling approach. In AGILE proceedings
(pp. 465–474).
Kyriakidis, P. C., & Journel, A. G. (1999). Geostatistical space–time models:
a review. Mathematical Geology, 31, 651–683.
LeSage, J. P. (1997). Bayesian estimation of spatial autoregressive models.
International Regional Science Review, 20(1–2), 113–129.
Lutkepohl, H. (2007). New introduction to multiple time series analysis.
Springer.
Min, X., Hu, J., & Zhang, Z. (2010). Urban traffic network modeling and
short-term traffic flow forecasting based on GSTARIMA model. In
13th international IEEE annual conference on intelligent transportation
systems, September 19-22, Madeira Island, Portugal.
Please cite this article in press as: Safikhani, A., et al., Spatio-temporal modeling of yellow taxi demands in New York City using generalized STAR models.
International Journal of Forecasting (2018), https://doi.org/10.1016/j.ijforecast.2018.10.001.
) – 11
Moghimi, B., Safikhani, A., Kamga, C., & Hao, W. (2017). Cycle-length pre-
diction in actuated traffic-signal control using ARIMA model. Journal
of Computing in Civil Engineering, 32(2), 04017083.
Moreira-Matias, L., Gama, J., Ferreira, M., Mendes-Moreira, J., & Damas, L.
(2013). Predicting taxi-passenger demand using streaming data. IEEE
Transactions on Intelligent Transportation Systems, 14(3), 1393–1402.
New York City Taxi & Limousine Commission, (2016). TLC factbook. http:
//www.nyc.gov/html/tlc/downloads/pdf/2016_tlc_factbook.pdf.
Nicholson, W. B., Bien, J., & Matteson, D. S. (2014). Hierarchical vector
autoregression. arXiv preprint, arXiv:1412.5250.
Nicholson, W. B., Matteson, D. S., & Bien, J. (2017). VARX-L: structured reg-
ularization for large vector autoregressions with exogenous variables.
International Journal of Forecasting, 33(3), 627–651.
Okutani, I., & Stephanedes, Y. J. (1984). Dynamic prediction of traffic
volume through Kalman filtering theory. Transportation Research, Part
B (Methodological), 18(1), 1–11.
Pfeifer, P. E., & Deutsch, S. J. (1980). A three-stage iterative procedure for
space–time modeling phillip. Technometrics, 22(1), 35–47.
Pfeifer, P. E., & Deutsch, S. J. (1981). Variance of the sample space–time
autocorrelation function. Journal of the Royal Statistical Society. Series
B. Statistical Methodology, 43, 28–33.
Qian, X., Ukkusuri, S. V., Yang, C., & Yan, F. (2017). A model for short-term
taxi demand forecasting accounting for spatio-temporal correlations.
In Transportation research board annual 2017, Washington D.C.
Sartoris, A. (2005). A STARMA model for homicides in the city of Sao Paulo.
In Proceedings of the spatial economics workshop, kiel institute for world
economics, 8–9 April, 2005, Kiel, Germany.
Sayarshad, H. R., & Chow, J. Y. J. (2016). Survey and empirical evaluation
of nonhomogeneous arrival process models with taxi data. Journal of
Advanced Transportation, 50, 1275–1294.
Schaller, B. (1999). Elasticities for taxicab fares and service availability.
Transportation, 26, 283–297.
Schaller, B. A. (2005). Regression model of the number of taxicabs in U.S.
cities. Journal of Public Transportation, 8, 63–78.
Shoesmith, G. L. (2013). Space–time autoregressive models and forecast-
ing national, regional and state crime rates. International Journal of
Forecasting, 29(1), 191–201.
Song, S., & Bickel, P. (2011). Large vector auto regressions. arXiv preprint
arXiv:1106.3915.
Stathopoulos, A., & Karlaftis, M. G. (2003). A multivariate state space ap-
proach for urban traffic flow modeling and prediction. Transportation
Research Part C: Emerging Technologies, 11(2), 121–135.
Terzi, S. (1995). Maximum likelihood estimation of a generalized STAR(p,
lp) model. Journal of the Italian Statistical Society, 4(3), 377–393.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso.
Journal of the Royal Statistical Society. Series B. Statistical Methodology,
58, 267–288.
Yang, C., & Gonzales, E. (2017). Modeling taxi demand and supply in New
York City using large-scale taxi GPS data. In P. Thakuriah, N. Tilhun, &
M. Zellner (Eds.), Seeing cities through big data–research, methods and
applications in urban informatics (pp. 405–425). Springer.
Yuan, J., Zheng, Y., Zhang, L., Xie, X., & Sun, G. (2011). Where to find my
next passenger. In Proceedings of the 13th international conference on
ubiquitous computing, Beijing, China —September 17–21 (pp. 109–118).
New York, NY: ACM.