Journal of Data, Information and Management (2019) 1:45–64

https://doi.org/10.1007/s42488-019-00004-z

ORIGINAL ARTICLE

Retail supply chain management: a review of theories and practices

Deng Ge1 · Yi Pan1 · Zuo-Jun (Max) Shen1,2 · Di Wu1 · Rong Yuan1 · Chao Zhang1

Received: 23 March 2019 / Accepted: 19 June 2019 / Published online: 18 July 2019
© Springer Nature Switzerland AG 2019

Abstract
Retail business has been rapidly evolving in the past decades with the boom of internet, mobile technologies and most
importantly e-commerce. Supply chain management, as a core part of retail business, has also gone through significant
changes with new business scenarios and more advanced technologies in both algorithm design and computation power. In
this review, we focus on several core components of supply chain management, i.e. vendor management, demand forecasting,
inventory management and order fulfillment. We will discuss the key innovations from both academia and industry and
highlight the current trend and future challenges.

Keywords Supply chain management · Vendor management · Demand forecast · Inventory management · Order fulfillment

1 Introduction number for the manufactures. As a result, retail supply chain

has to be able to scale more efficiently and focus more
From the first department store in 1852, the retail revo- on supply chain coordination. Secondly, retailers face end
lution has never stopped. Retail business has been rapidly consumers in the number of hundreds of millions compared
evolving in the past decades with the boom of internet, to a limited number of wholesalers for manufacturers.
mobile technologies and most importantly e-commerce. Retailers need to invest more on understanding customer
More recently, new technologies are making it possible for demand in order to plan and adjust accordingly. Thirdly, the
even more forms of retail business. Although the fundamen- cost structure of a retailer is different from a manufacturer.
tal problems of supply chain management for retail business
The majority of retailers cost is inventory cost while
remains the same as what was seen decades ago as Forrester
manufacturer spends significantly more on equipment and
(1958) illustrated, the tools and methodologies have dramat-
product lines. As a result, inventory management is more
ically changed over the years. In this paper, we will high-
important for retailers. Manufacturers tend to focus more
light several key components in supply chain management
on resource scheduling and planning. Lastly, with the ever-
that is particularly relevant to retail business, especially
growing customer needs in service quality and delivery
technology-driven retail business, and provide an overview
speed, retailers have to take fulfillment into consideration
of the evolution of supply chain management methodologies
when building their supply chain networks. Although some
and practices.
methodologies can also be applied to other supply chain
There are several key factors that differentiate a retail
supply chain from a manufacturer supply chain. Firstly, systems, in this review, we will primarily focus on the
retailers generally face significantly more partners. A large- components that are more relevant to retail supply chain
scale retailer could have tens of thousands of vendors to systems.
supply its inventory compared to a much more manageable In retail, supply chain systems normally consist of
many different components and have different set ups.
In this paper, we only consider the related components
 Di Wu that the retailer as a decision maker has direct impact
di.wu@jd.com on. Traditionally, a typical supply chain in retail consists
of 4 main components: wholesalers (vendors), warehouses
1 JD.com American Technologies Corporation, 675 E (distribution centers), stores and customers as shown in
Middlefield Rd, Mountain View, CA 94043, USA Fig. 1. Other parties such as manufacturers and their
2 College of Engineering, University of California, Berkeley, suppliers that are further upstream of the supply chain
4143 Etcheverry Hall, Berkeley, CA 94709, USA system are normally beyond the scope from a retailer

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

46 J. of Data, Inf. and Manag. (2019) 1:45–64

Fig. 1 Traditional retail supply

chain

perspective. In a more modern setup, however, the supply over time. Machine learning based inventory algorithms
chain system visible to the retailer becomes much more (Shi et al. 2018; O’Neil et al. 2016) can directly take in
complex (as shown in Fig. 2). In additional to the increasing the historical sales and purchase order (PO) information
size of the warehouse and store network, retailers now have to produce optimal decisions without placing assumptions
more options to satisfy their customers’ demand (e.g. from on demand estimation. Embedding technique (Shi 2018;
stores, from warehouses, or directly from the suppliers, Barkan and Koenigstein 2016) can also effectively learn the
etc.). Further, more and more retailers are forming a deeper product similarities through historical customer order data
integration with the manufactures and even raw material without consulting to domain knowledge.
suppliers through the introduction of store brands or private The growing interests in data-driven algorithms are
labels (Chen 2018a, b). Retailers today still face the same universal across both industry and academia. At the same
or similar supply chain management challenges such as the time, we do observe promising outcomes from newer
bullwhip effect (Lee et al. 1997). But the approaches and approaches showing their ability to solve practical problems
methodologies used by retailers are continuously evolving with better empirical performance. With the ever-growing
in order to further drive for lower cost, higher efficiency and complexity of modern supply chain networks, data-driven
ultimately better customer experience. algorithms are playing more and more important roles.
One of the most distinguishable changes over the In this paper, we will primarily discuss several key
time in supply chain management methodologies is the components of retail supply chain management systems
shift from “experience-driven” to “data-driven” approaches. and present their corresponding evolution to the data era.
An “experience-driven” supply chain management system Besides research literature, the review will also focus
heavily relies on human experience to make the critical on industrial applications and demonstrate how theories
decisions such as inventory placement, fulfillment network are applied in the real-world. The structure of the paper
design, etc. Although there will still be usage of data is organized as follows: Section 2 discusses retailers’
in the process, assumptions based on experience and coordination and interaction with vendors; Section 3
human judgment are usually pivotal to the success of such discusses topics related to demand forecasting, one of the
models. For example, traditional inventory replenishment most critical components in supply chain management;
model relies on the assumption of the shape of demand Section 4 reviews inventory management systems; Section 5
distribution as well as vendor lead times in order to make follows on the fulfillment perspective of the supply chain
an optimal decision (Covert and Philip 1973; Brahimi system; Section 6 provides a summary for the future
et al. 2006); Product assortment solutions (Kök et al. research directions.
2008) also heavily depend on product similarities which
are mostly rated by experienced category managers. On
the other hand, “data-driven” models rely more on the 2 Vendor management
data itself rather than human experience. Due to the
greatly increased data availability and the advancement in In a supply chain, vendor management is critical for
data mining and machine learning technologies, it is now a firm since it has a direct impact on product quality,
possible for supply chain management systems to make service level and company profits. An effective vendor
decisions on available data alone. For example, learning management requires information sharing between the
based replenishment algorithms (Zhang et al. 2017) utilize retailer and vendors, such as sharing of demand forecast and
observed demand data to help improve inventory decisions inventory information through a vendor managed inventory

Fig. 2 Modern retail supply

chain

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 47

(VMI) system in a marketplace with stationary and non- By ensuring end-to-end information sharing, the occur-
stationary demand. In this section, we will review the recent rence of the bullwhip effect is reduced thus lowering inven-
development of vendor management from both theoretical tory levels across the chain. It also allows the partners in
research and industrial practice perspectives. a supply chain to visualize a bigger picture in terms of the
entire supply chain rather than their individual enterprise.
2.1 Introduction As partner collaboration is initiated from the planning to
the replenishment stage, the supply chain can better respond
Recent advances in information technology, particularly in to the exceptional circumstances so as to make it a more
the e-business arena, are enabling firms to rethink their proactive system rather than a reactive one. On a more
vendor management strategies and explore new avenues for abstract level, CPFR aims at creating an environment of
the cooperation with vendors through information sharing. trust between trading partners where the benefits of sharing
Sharing demand related information among supply chain information are realized. The role of CPFR in various stages
members has achieved huge impact in practice. As pointed of supply chain activity is aptly represented in Fig. 3.
out by Stein and Sweat (1998), by exchanging information, Within an efficient and integrated CPFR environment where
such as Point of Sales (POS), forecasting data, inventory firms share promotion plan, sales data and retail analytics
level and sales trends, many companies are reducing their and vendors share inventory and shipment information, the
cycle times, fulfilling orders more quickly, cutting out supply chain is more responsive to the external business
excess inventory, and improving customer service. change and all the partners can benefit from the information
Many companies not only share information with their sharing.
supply chain partners, but also jointly make decisions to In the vendor management, however, an incomplete
improve supply chain performance by using Collaborative understanding of the value of information sharing and phys-
Planning, Forecasting and Replenishment (CPFR). Accord- ical flow coordination may hinder the efforts that promote
ing to Panahifar et al. (2015), CPFR is a technological the efficiency and responsiveness of a supply chain. We
innovation tool that was first registered as a trademark by attempts to better understand the information sharing and
the Voluntary Inter-industry Commerce Standards (VICS) flow coordination by reviewing and categorizing the recent
in 1998 and is defined by VICS as a collection of new busi- research and practice in this area.
ness practices that leverage the Internet and EDI (electronic
data interchange) in order to achieve two goals: radically 2.2 Literature review
reduce inventories and expenses while improving customer
service. Relevant literature consistently recognizes that inventory
Verity (1996) reported in Business Week that Wal-Mart reduction can be achieved by implementing initiatives such
and Warner Lambert attained significant improvements of as information sharing, continuous replenishment program
in-stock positions while reducing inventory through CPFR. and VMI. In the field of these initiatives, there are several
CPFR is one of a series of supply chain initiatives like streams of the related literature: literature on (1) inter-
JIT (Just-In-Time), ECR (Efficient Customer Response) and organizational systems (IOS); (2) quantitative models in
VMI (Sheffi 2002) driven by organizations to make their information sharing; and (3) implementation of information
supply chains more responsive and keep all the supply chain sharing through VMI and CPFR (Yao and Dresner 2008).
members in tune with the end customer demand, both in The first stream of research has revealed the business
terms of the product and its volumes. value of IOS by studying the supply chain management

Fig. 3 The CPFR in a supply

chain

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

48 J. of Data, Inf. and Manag. (2019) 1:45–64

initiatives, such as information sharing, continuous replen- management, improved product offering, operational effi-
ishment programs and VMI. Research has shown that IOS, ciency, product availability assurance, decreased replenish-
as a link between suppliers and buyers, can improve a firm’s ment lead time, increased customization capability. The
performance and bring it competitive advantages (Sethi financial dimension is the most important objective for
et al. 1993; Palmer and Markus 2000; Premkumar 2000; firms implementing CPFR. This covers several criteria
Srinivasan et al. 1994). reported in the previous studies such as increased rev-
The second stream of research has quantitatively enues and earnings, increased margins, increasing EVA
examined the value of information sharing in supply chains. (Economic Value Added), increasing shareholder wealth,
The consequences of the bullwhip effect, for example, can decreasing cost of production, planning and deployment,
be minimized through information sharing (Lee et al. 1997; maximum efficiency of members, a reduction of inventory
Lee and Whang 1999; Lee et al. 2000; Cachon and Fisher in the supply chain, decreasing working capital, reduction
2000; Chen et al. 2000). in production and inventory costs, reduced overall costs,
Some research has found that policies such as VMI increasing the sales of products and reduction in stock-outs.
can decrease the bullwhip effect, thereby improving supply
chain efficiency, such as by lowering inventory levels and 2.3 Industry practice of information sharing
reducing cycle time (Cachon and Zipkin 1999; Kulp et al.
2004; Mishra and Raghunathan 2004). Angulo et al. (2004) Potential economic benefits of information sharing through
use simulation to find demand information sharing is a CPFR are well-recognized and have been publicized in
significant part of VMI implementation and can improve practice by successful retail businesses such as Wal-Mart.
the fill rate by up to 42%. Çetinkaya and Lee (2000) In its CPFR partnership with P&G, Wal-Mart’s market-
develop an analytical model for coordinating inventory ing information is integrated with P&G’s manufacturing
and transportation decisions with VMI systems. Lee and systems to make better consumer-based decisions across
Whang (1999) present empirical evidence to confirm the their firm-level boundaries. For example, Wal-Mart’s POS
value of supply chain coordination. These authors show that data show the transaction-level information about con-
inventory turns and stock-outs have been improved after sumer’s choices, thus providing the actual demand informa-
the implementation of continuous replenishment programs, tion on what is selling and the selling price. P&G’s products
using data collected from 31 grocery retail chains. are then developed, manufactured and delivered to meet
The third stream of research has widely discussed the those customer needs in a timely manner. CPFR pilot with
collaboration and integration in the context of supply chains P&G provided a structured contractual platform for joint
in recent years, especially in the area of CPFR. Panahifar forecasting and planning activities between Wal-Mart and
et al. (2015) review the scope and value of CPFR using a its vendors that ultimately drive the replenishment process
devised state-of-the-art taxonomy for the classification of through the entire supply chain. As pointed out by Andraski
selected references related to CPFR. In the paper, based and Haedicke (2003), by 2003 Wal-Mart has established
on a total of 93 papers published from 1998 to 2013 on over 600 trading partners through CPFR to reduce its oper-
CPFR, the authors attempt to seek answers to the question ating expenses to the lowest in the industry. Successful col-
of what are the main constructs and efficient framework laboration with CPFR partners allowed Wal-Mart to price
for successful implementation of CPFR. The key findings its products 10% below most of the competitors.
of the paper is that four main constructs for successful Kim and Mahoney (2010) provided a detailed case study
implementation of CPFR have been identified: 1. CPFR of the CPFR arrangement between Wal-Mart and P&G. The
enablers; 2. CPFR barriers; 3. trading partner selection; case study reveals that the successful implementation of
4. incentive alignment. The findings indicate that there is CPFR depends not only on extensive information sharing
a need for better understanding of the amount and level but also on mutual learning as well as commitments to
of information sharing as an important function of CPFR the dedicated partners from the repeated interactions. It
implementation. grows out of first gaining an awareness of its contractual
This paper also categorizes the CPFR implementation partners’ business needs by asking: 1. what is competitive
benefits for companies, which consists of three main dimen- advantage of your partners; 2. what is the competitive
sions: Information, Service and Finance. The informa- advantage to you if you combine them with yours; 3.
tion dimension encompasses improvement of forecasting what kind of business relationship does that create. Thus,
accuracy, reducing the amount of exchanged information successful implementation of CPFR requires higher levels
and reducing the bullwhip effect. The second dimension of communication including the exchange of strategies and
involves more criteria including increased responsiveness, objectives between partners at the beginning of a planning
enhanced customer service quality, improved inventory period.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 49

As e-commerce retail industry in China has been replenishment to JD.com warehouses; 3. Using Vendor B’s
booming recently, the information sharing through CPFR warehouses to fulfill JD.com customer orders. In 2016,
has been widely implemented by Chinese retailers. To with the CPFR practice and information sharing efforts, the
build an efficient and responsive supply chain that provides supply lead time from Vendor B has decreased from 8 days
fast delivery to customers, JD.com, China’s largest retailer, to 3 days. The in-stock rate increases from 73% to 96%, and
online or offline, has invested in the collaboration with GMV increased by 5%.
its vendors through CPFR by continuously applying
the advanced information technology in supply chain Higher in-stock rate for peak promotion with Vendor C In
innovation. Below we will illustrate several successful cases the important promotional and holiday events, such as
from JD.com, where the value of information sharing in the Thanksgiving Black Friday event in US, 618 and Single Day
vendor management is fully realized. events in China, the sales could account for 30% or higher
of the annual sales, and retailers usually need to prepare
More efficient vendor management with Vendor A One inventory 1-2 months in advance to meet the peak demand,
practical example of information sharing is the EDI system especially for the top-selling products. Through the CPFR,
coordination between JD.com and Vendor A, who is a Chi- firms can book the vendors inventory and improve the in-
nese leading electrical appliance manufacturer. Vendor A is stock rate during the promotional events. Another success
one of the top vendors for JD.com by providing electrical case of information sharing is the implementation of CPFR
appliances, such as small kitchen appliances, laundry, large between JD.com and Vendor C. During the 618 promotional
cooking appliances, and refrigeration appliances. event in 2017, Vendor C fulfilled 99% of JD.com’s
The original information exchange between JD.com and replenishment orders. The promotional information sharing
Vendor A such as replenishment purchase orders was simply resulted in Vendor C’s products in-stock rate improved by
by emails and telephone, which resulted in information 3%, and lead time reduced by 20%.
delays and inaccuracies, and orders being intractable. To
overcome this challenge, JD.com and the vendor cooperated
in building an EDI system, allowing the information 3 Demand forecasting
exchange in real-time, being transparent and traceable.
The information shared between JD.com and Vendor A Demand forecasting is one of the most important compo-
includes the JD.com’s sales plan, monthly demand forecast nents in supply chain management. In fact, demand fore-
by region, purchase orders, and vendor confirmation casting results are key inputs for many decision-making
responding to the orders. The business benefits of this processes in retail such as inventory management, network
information sharing through EDI and the integrated demand planning, pricing and revenue management, marketing, etc.
forecasting is significant: 1. JD.com has better estimation Over the recent years, many retail companies, especially
for the Vendor A’s inventory and production capacity, e-commerce companies have significantly increased their
increasing the prediction time window from one month to investments in improving demand forecasting performance.
three months. 2. With JD.com’s sharing demand prediction Further, with the increased data availability and the devel-
for longer periods, Vendor A can make a better arrangement opment of more sophisticated machine learning algorithms,
of raw material procurement, production capacity and we observe many new developments in this space in recent
workers schedule. The production plan of Vendor A is 95% years. We will review both theoretical research and indus-
alignment with JD.com’s purchase plan. 3. The time delay trial practices of demand forecasting in this section.
between JD.com’s purchase orders and Vendor A’s response
and confirmation is significantly reduced by 70%. 3.1 Literature review

Supply lead time reduction with Vendor B Vendor B is a Demand forecasting is critical to the success of a retail
major beverage and food supplier for JD.com. Since 2014, company. Particularly in supply chain management, an
JD.com set a goal to provide fresh products and high- accurate and practical demand forecasting system can be a
quality customer service in the beverage market. Knowing significant source of competitive advantage by improving
the potential benefits of supply chain coordination, the customer service levels and by reducing costs related to
executive managements of both JD.com and Vendor B supply-demand mismatches (Snyder and Shen 2011). As
reached a strategic agreement that would strengthen the a result, the forecasting problem has a long research
collaboration in information sharing. Since then, three history in the field of statistics and recently in machine
key steps have been implemented: 1. Demand forecast learning. Demand forecasting is a practical domain of time
and promotion plan sharing; 2. Vendor B direct inventory series modeling and forecasting (Hamilton 1995). Many

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

50 J. of Data, Inf. and Manag. (2019) 1:45–64

important models have been proposed in the literature machine learning methods are more data-driven; in general,
for improving the accuracy and efficiency of time series these machine learning methods are exploited to improve
modeling and forecasting. In the section below, we describe time series predictions by minimizing a loss function.
three important classes of time series models that are Ahmed et al. (2010) compares the accuracy and time
commonly used in the practice of demand forecasting, consumption of these machine learning models.
i.e. the stochastic time series, machine learning and deep Zheng and Su (2014) proposed a two-step enhanced
learning models, together with their inherent forecasting KNN method, and the method consistently improved the
strengths and weaknesses. Further, we will also point forecasting accuracy in short-term forecasting. Recurrent
out several specific forecasting scenarios that are closely least-square SVR (Suykens and Vandewalle 2000) and
associated with retail industry. dynamic least-square SVR (Fan et al. 2006) are two popular
SVR models for time series forecasting. Girard et al.
3.1.1 Stochastic time series models (2003) proposed to use the non-parametric Gaussian process
model for multi-step ahead time series prediction, so that
Stochastic time series models are the most popular the uncertainty about intermediate regressor values can be
methods used in demand forecasting. There are two incorporated, thus the uncertainty on the current prediction
main classes of stochastic time series models: linear can be updated. Decision Tree provides a foundation
and non-linear models. Linear model includes the two for various tree-based ensemble algorithms including the
most widely used stochastic time series approaches, i.e. two most widely used machine leaning techniques, i.e.,
the Holt-Winters method (Holt 2004; Winters 1960) and Random Forest (RF) (Kam 1995) and Gradient Boosting
the Autoregressive Integrated Moving Average (ARIMA) Decision Trees (GBDT) (Chen and Guestrin 2016; Ke
model (Box et al. 2015). ARIMA model has subclasses et al. 2017; Prokhorenkova et al. 2017). RF has been
of other models, such as the Autoregressive (AR), used for electricity load forecasting (Nedellec et al. 2014;
Moving Average (MA), Autoregressive Moving Average Dudek 2015), and it is shown that the RF model provides
(ARMA), and Seasonal ARIMA (SARIMA) (Hamzaçebi as accurate forecasts as artificial neural networks (ANN)
2008) models. Although these models can capture trend and outperformed the ARIMA and Decision Tree models
and seasonality, they are ineffective in modeling highly (Dudek 2015). RF has been proven to be competent in one-
nonlinear time series. To overcome this drawback, various step-ahead time series forecasting, and it is also shown that
non-linear stochastic models have been proposed in a low number of recent lagged variables performs better,
the literature, such as the Autoregressive Conditional highlighting the importance of the training set’s length
Heteroskedasticity (ARCH) (Engle 1982), and Generalized (Tyralis and Papacharalampous 2017). Similar to RF, GBDT
ARCH (GARCH) (Bollerslev 1986) models; however, the has exhibited competent performance comparing to other
implementations are not as straight-forward and simple as stochastic time series and machine learning methods on time
the linear models. series forecasting (Kusiak et al. 2009). However, there is
The strengths of stochastic time series models are the some limitation as tree-based models cannot extrapolate, i.e.
model interpretability and computational efficiency. It is cannot predict value bigger or smaller than the value in the
well received by practitioners due to the easiness to training set, so they are not suitable for data with a trend
implement and the intuitive results. However, the mod- in time series. The solution is feature engineering, which is
els have relatively strong assumptions in terms of the to remove the trend first by constructing many time series
underlying stochastic processes, which are normally over features (e.g. day of week, holiday, season, moving average,
simplifying the practical situation. Recently, with the and lag) (Kusiak et al. 2009). Another weakness of machine
advancements in big data and machine learning, researchers learning models is that they require more computation
have found Machine Learning and Deep Learning algo- comparing to stochastic time series models, which increases
rithms to be able to provide better forecasting performance. the time complexity of the solutions.

3.1.2 Forecasting with machine learning techniques 3.1.3 Forecasting with deep learning techniques

In the last few decades, machine learning techniques are Recently, artificial neural networks (ANNs) have attracted
widely used in the field of forecasting, such as Decision Tree increasing attentions in the domain of time series forecast-
(Breiman 2017), K-Nearest Neighbor Regression (KNN) ing (Zhang 2003; Kihoro et al. 2004; Kamruzzaman and
(Hastie and Tibshirani 1996), Support Vector Regression Sarker 2006). The excellent feature of ANNs, when applied
(SVR) (Drucker et al. 1997), and Gaussian Processes (GP) to time series forecasting problems is their inherent capabil-
(Williams and Rasmussen 1996). Comparing to stochastic ity of non-linear modeling, without any presumption about
time-series models which are more model-driven, the the statistical distribution followed by the observations.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 51

Recurrent Neural Networks (RNNs) are a special type of inventory (Jain et al. 2014). Empirically, out of stock has a
ANNs which contain self-connections. Unlike feedforward high correlation with promotion activities in retail business.
NNs, the hidden states of RNNs serve as memory to help So, the forecasting algorithm needs to deal with the missing
map both current time inputs and previous time internal data very carefully in order to produce accurate results.
states to new desired outputs. This allows RNNs to capture The expectation-maximization (EM) algorithm (Dempster
temporal information in sequence data. Long Short-Term et al. 1977) is one of the approaches that widely adopted
Memory (LSTM) (Hochreiter and Schmidhuber 1997) is for this need (Anupindi et al. 1998; Talluri and Van Ryzin
an important class of RNN variants, which has additional 2004; Vulcano et al. 2010; Conlon and Mortimer 2013).
memory-control gates and memory cells to selectively store Secondly, with the ever-increasing product selection, new
historical information and keep long term information. product forecast is becoming a more and more important
RNNs and LSTMs have been successfully applied to time topic. Since the publication of the Bass model in 1969
series data forecasting problems. Längkvist et al. (2014) (Bass 1969), research has been made to make diffusion
reviewed recent research in unsupervised feature learning models theoretically sounder and practically more effective
and sequence modeling with deep learning methods for for new product forecast (Ismail and Abu 2013; Lee et al.
various time series data such as video frames, speech 2014; Kahn 2014). Thirdly, product complementarity and
signals, and stock market prices. Bianchi et al. (2017) substitutability plays an important role in the actual sales of
compared different RNN variants and showed that LSTMs products, and the demand for a product can depend directly
outperformed others on highly non-linear sequences with and indirectly on other products in different categories
sharp spikes thanks to the quick memory cell modification (Shocker et al. 2004; Duan et al. 2015). Fourthly, a retailer,
mechanism. Cinar et al. (2017) proposed using an LSTM especially e-commerce retailer generally has a large product
encoder-decoder with position-based attention model to selection with highly diversified products and with different
capture patterns of pseudo-periods in sequence data. They product characteristics, such as fashion, fresh groceries, etc.
applied the attention mechanism (Bahdanau et al. 2014) to For fashion products, demand uncertainty, lack of historical
explore similar local patterns in historical data for future data and seasonal trends usually coexist and make demand
prediction. However, it is impractical to look into the full forecasting more challenging (Nenni et al. 2013). For fresh
history of time series and the selection of which part of the groceries, successful demand forecasting is very critical
history to attend to relies on human knowledge. Taieb and because of the short shelf-life of fresh products and the
Atiya (2016) analyzed the performance of different multi- importance of the product quality (Doganis et al. 2006;
horizon forecasting strategies on synthetic datasets with Shukla and Jharkharia 2013). For long-tail products, the
different factors, such as length of time series and number problem has proved to be more challenging due to the
of horizons. Flunkert et al. (2017) proposed a technique sparsity of the sales data that limits the degree to which
called DeepAR to make probabilistic forecasts by assuming traditional analytics can be deployed. Pitkin et al. (2018)
an underlying distribution for time series data. DeepAR developed a Bayesian hierarchical model to forecast this
could produce the probability density functions for target type of demands.
variables by estimating the distribution parameters on each
time point with multi-layer perceptions (MLPs). However, 3.2 Industry practice
the distributional assumption is often too strong to apply to
real-world datasets. As mentioned in the earlier section, demand forecasting
Comparing to stochastic time series and machine is a heavily invested domain in the industry, especially in
learning models, the strengths of deep learning models are retail. There are many existing commercial solutions on the
the model capability to incorporate a variety of information, market. At the same time, many companies have established
no need for feature engineering, and generally higher dedicated teams on demand forecasting to keep pushing for
forecast accuracy. The weakness of deep learning models is the boundary of the forecasting accuracy.
that when the training size is large, it requires an enormous
amount of computation, and the model interpretability is the 3.2.1 Commercial forecasting tools
lowest among the three classes.
Demand forecasting is a key part of traditional supply
3.1.4 Retail specific forecasting scenarios chain management software. Leading companies in this
space include JDA, IBM, SAS, Oracle and more. The
Retail demand forecast has its own unique challenges that prediction methods they use are mostly stochastic time
makes the problem more complex. Firstly, retail demand series algorithms that emphasize the need for planning
forecasting normally have to deal with censored demand in software functions, such as assuming simulations and
where observed demand data are affected by out of stock artificial adjustments to requirements. Although these

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

52 J. of Data, Inf. and Manag. (2019) 1:45–64

companies lack sophisticated predictive algorithms, they Many researches focus on generating quantile estima-
have a high market share in large and medium-sized tions for target variables by formulating forecasting prob-
enterprises because of their long history, wide product lems as quantile regressions (Koenker and Bassett 1978).
coverage, and good integration with ERP (Enterprise Zheng (2010) proposed to minimize objective functions for
Resource Planning) systems. quantile regressions with high dimensional predictors by
With the popularity of machine learning in time series gradient boosting (Friedman 2001). Xu et al. (2016) pro-
prediction, in recent years, there have been many products posed a quantile autoregressive model which can output
with machine learning and prediction accuracy as the multi-horizon quantile predictions by sequentially feeding
main selling points. Such products can use machine predictions of previous steps into the same NN for current
learning algorithms to make more accurate forecasting of prediction. Their NN-based autoregressive model is differ-
demand than stochastic time series methods, leveraging ent from RNN as prediction values, instead of hidden states,
internal and external data such as product features, store are fed recursively. For the above methods, separate models
locations, weather, geography and economic indicators. have to be trained for different quantiles for multi-quantile
Representative companies in this field include LLamasoft forecasting tasks, which is inefficient in practice.
and Blue Yonder. It is worth noting that JDA acquired Recent forecasting researches from e-commerce com-
Blue Yonder in 2018, and then launched the machine- panies show strong interests for probabilistic forecasting,
learning-based cloud prediction product Luminate Demand such as DeepAR (Flunkert et al. 2017) and MQ-RNN (Wen
Edge in October of that year, marking JDA’s transition et al. 2017) from Amazon and the multi-horizon fore-
to cloud services and artificial intelligence (JDA Software casting algorithm (Fan et al. 2019) developed by JD.com.
2018). Besides, Amazon recently launched its time series MQ-RNN (Wen et al. 2017) generates multiple quantile
forecasting cloud service “Amazon Forecast” based on forecasts for multiple time horizons. It uses an LSTM to
machine learning and deep learning techniques (e.g. encode the history of time series into one hidden vector, and
DeepAR (Flunkert et al. 2017) and MQ-RNN (Wen et al. uses two MLPs to summarize this hidden vector, together
2017)), and no machine learning experience is required to with all future inputs, into one global context feature and
use this service. horizon-specific context features for all horizons. However,
In the last two years, a number of common machine this global context feature may be too general to capture
learning platforms have emerged, such as H2O, DataRobot short-term patterns.
and Ali PAI. This type of product service is targeted at Fan et al. (2019) presented an end-to-end deep learning
data scientists, allowing them to perform modeling through framework for multi-horizon probabilistic demand forecast-
visual operations without programming. In this trend, there ing. This LSTM based system improves the consideration
is a branch that is a general predictive software that does not for temporal relationship within the forecast result by allow-
need to be modeled by itself. The most representative one ing propagation of information both forward and backward
is H2O Driverless AI, which has many advanced algorithms in the LSTM decoder (shown in Fig. 4). This approach out-
built in, and has the ability to automatically target any set of performs current state-of-the-art models on the GOC20181
data by the machine to find the optimal algorithm. Sales Forecasting dataset (Yuan and Jing 2018).

3.2.2 Probabilistic demand forecasting system

4 Inventory management
Many practical decision-making scenarios require richer
information provided by probabilistic forecasting that Inventory management is to make informed decisions about
returns the full conditional distribution, rather than point the quantity and placement of stocked goods. As a com-
forecasting that predicts the conditional mean. For real- plex multi-layer system, it is required at various loca-
valued time series, this is traditionally achieved by assum- tions/facilities across the inventory network. It covers the
ing an error distribution on the residual series. However, entire inventory control process including the monitoring of
an exact parametric distribution is often not directly rel- goods moved into and out of warehouses/distribution cen-
evant in applications as mentioned above. Instead, par- ters and the reconciling of the inventory balances. In this
ticular quantiles of the forecast distribution are useful in section, we discuss the state-of-the-art regarding three key
making optimal decisions, both to quantify risks and mini- aspects of inventory management, namely, replenishment,
mize losses (e.g. risk management, operation optimization). transshipment, and placement.
Probabilistic forecasting is a key enabler for forecasting
demand and optimizing business processes in retail busi-
1 https://jdata.joybuy.com/en/html/detail.html?id=4
nesses.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 53

Fig. 4 Structure of deep

learning probabilistic
forecasting framework

4.1 Literature review policy (R, s, S) or (R, S) (Snyder and Shen 2011). The
former uses a replenishment quantity of Q which is usually
Inventory management is an extensively studied field in determined using the idea of economic order quantity
the literature. Due to space limitation, we only cover (EOQ) (Harris 1915), while the latter typically assumes a
main methodologies and invite readers to refer to the more realistic stochastic demand and takes a similar form as
cited reference for more comprehensive introductions and the safety stock derived above.
reviews. In the following subsections, three key aspects The strategic safety stocks as described above can
of inventory management are covered, i.e. replenishment, help to manage the risk of stock-out, which usually
transshipment and placement. These are the foundations for outweighs other relevant costs. A comprehensive review
many practical inventory management systems in practice of safety stock techniques is available in Guide and
and still see many new developments recently. Srivastava (2000). Various safety stock strategies have been
developed, for instance, Salameh (1997) and Brandolese
4.1.1 Replenishment and Cigolini (1999). Estes (1973) developed a re-ordering
point inventory model that accounts for demand and
If the customer demand and the vendor lead-times are leadtime variability by assuming that the demand during
appropriately predicted, one would be able to vary stocks the leadtime follows a normal distribution. Later on, Ruiz-
accordingly to accommodate their needs. However, both Torres and Mahmoodi (2010) determined safety stock by
demand and lead-times are difficult to forecast to a focusing on historical data without making any particular
satisfying accuracy level due to the stochastic nature of distributional assumptions of demand and leadtime. The
these quantities (e.g., variability in demand and lead-time). optimal placement of safety stocks also has been extensively
In addition, items may be required faster than the supply can studied with seminal work including Clark and Scarf (1960)
provide it. Among all different types of incurred inventory and Lagodimos and Anderson (1993). Graves and Willems
costs, stock-out cost is usually the most significant part, (2000) formulated the safety stock placement problems as
which may happen in cases such as fluctuating customer a network optimization problem and developed a dynamic
demand, forecast inaccuracy, and variability in lead-times, programming algorithm as solution procedure. Lesnaia
etc. To compensate for these, one can otherwise hold (2004) formulated the safety stock problem in supply chains
sufficient stock to cope with unexpected or excess demand as a general network problem, which is shown to be NP-
so as to prevent stock-outs. The minimum stock level hard.
maintained is referred to as safety stock and this inventory The inventory system in practice may consist of multiple
management strategy is called safety stock policy. Besides, layers of warehouses such as regional distribution center
safety stock can also compensate for the uncertainty in the (RDC) and forward distribution center (FDC). At each level
vendor lead-time prediction. Arguably, safety stock is the of the supply chain, stock levels of various SKUs need
basis for the broader problem: inventory replenishment. to be determined. Multi-echelon inventory management
The safety stock idea has been used to devise a few deals with multiple layers of inventory simultaneously.
classic replenishment models such as the continuous review Specifically, a multi-echelon system can be categorized
inventory policy (s, Q) and the periodic review inventory into a few types: a serial chain, a divergent system and a

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

54 J. of Data, Inf. and Manag. (2019) 1:45–64

convergent system. A divergent system consists of a single i.e. proactive and reactive transshipment, distinguished by
central stage and several successors. A convergent system the timing of the decision-making process with respect
has a one-end stage with several predecessors. A serial chain to demand realization. Proactive transshipment redistribute
is a special case of the divergent systems that have one inventory between warehouses at predetermined moments
single successor. in advance of any potential customer orders while reactive
Many real-world supply chains are multi-echelon sys- transshipment, as it is suggested by the name, reacts to real
tems consisting of several stages of procurement, manu- demands and act accordingly. Proactive transshipment mod-
facturing, and transportation. The theory of multi-echelon els, e.g. Allen (1958) and Karmarkar (1981), range from a
inventory management was initialized by Clark and Scarf simple single-period, single-transshipment with no network
(1960), in which a basic model for a supply chain consist- inventory replenishment model to more complex model
ing of multiple stages with a serial structure is considered. considering multi-period, multi-transshipment with network
The authors proved the optimal base stock levels can be inventory replenishment. Research on reactive transship-
obtained by the minimization of one-dimensional convex ment can be split between periodic review (Krishnan and
cost functions. Another seminal work on this topic is Sher- Rao 1965; Tagaras 1989) versus continuous review (Lee
brooke (1968). Later on, the work of Clark and Scarf 1987; Wong et al. 2006) models, which can be further
(1960) was extended in terms of many aspects, such as detailed into single-echelon, multi-echelon and decentral-
systems with a pure assembly/convergent structure, fixed ized systems. A comprehensive review of lateral transship-
batch sizes or fixed replenishment intervals, and advanced ment models can be found in Paterson et al. (2011). It
demand information, etc. For instance, the optimality of is worth noting that with e-commerce retailers, there is a
base stock policies and the decomposition result (Clark growing emphasize on reactive transshipment not only to
and Scarf 1960) was based on a stochastic dynamic pro- improve product availability but also to reduce order splits
gram in a finite-horizon setting, which was then extended (Zhang et al. 2018). As last mile delivery is one of the largest
to the infinite-horizon case (Federgruen and Zipkin 1984). components of fulfilment cost (Xu et al. 2009), an efficient
The optimal inventory policy for a serial system assuming reactive transshipment algorithm can significantly reduce
Markov-modulated demand was derived by Chen and Song overall fulfillment cost.
(2001). Recent research further increases the intricacy of the
There are various solution procedures proposed in system but primarily in a focused area. For instance, a
literature for the inventory problem of multi-echelon two-item, two-warehouse periodic-review inventory model
systems, including both heuristics, approximate methods, with transshipment was examined in Ramakrishna et al.
and exact methods. The optimal policy and optimal costs (2015). Noham and Tzur (2014) implemented a fixed
of a multi-echelon, serial system have been computed by transshipment cost to improve on previous methods of
both the approximate method based on two-moment fits and modeling single and multi-item transshipment problems.
exact method in van Houtum and Zijm (1991). Shang and Torabi et al. (2015) modeled the transshipment problem in
Song (2003) proposed a simple heuristic-based inventory e-commerce as a mixed integer linear programming (MILP)
policy for the multi-echelon, serial system with linear costs and solved optimally to minimize the cost of logistics, while
and stationary random demands through the lower and not consider the effects of shipment and leadtime.
upper bound subsystems.
4.1.3 Inventory placement
4.1.2 Transshipment
Inventory placement is the problem of determining the
Transshipment problem deals with the problem of managing choice of fulfillment centers in which to place each SKU.
goods between distribution centers, facilitates inventory The inventory placement problem is specifically relevant
re-balancing, bulk inventory receiving, and customer in the era of e-commerce due to the greatly increased
order fulfillment (Herer et al. 2002). Different from item selection available to the retailers. The retailer has
replenishment problem, transshipment problem handles the to strategically store the items in order to better balance
necessary movements of goods between warehouses within shipping speed and inventory cost. As an optimization
an echelon, e.g. within the retailer’s network. Through problem (e.g., mixed integer linear programming), it aims
transshipment, the overall inventory level can be reduced to find a placement plan for a given planning horizon
whilst maintaining the required service levels. that minimizes the total costs (i.e., shipping, overhead)
This topic has been studied extensively since the late and satisfies the expected geographical demand and
1950’s (Allen 1958, 1961) focusing on varying number of respects fulfillment center capacity constraints. It covers
items, warehouse echelons, warehouses, ordering policies, topics on facility (e.g., warehouses) location and product
and cost analysis. Two main types of transshipment exist, placement.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 55

Inventory placement problem can be viewed as an multinomial logit model, exogenous demand, or locational
extension of the facility location problem and a special case choice models, is utilized to capture consumer behaviors.
of the multi-commodity capacitated fixed-charge network For uncapacitated assortment problems, algorithms ranking
flow problem. The problem varies by the number of products by margin under the multinomial logit (MNL)
commodities (i.e., single-commodity, multi-commodity), discrete choice model are available in Talluri and Van
capacity (i.e., uncapacitated, node-capacitated or arc- Ryzin (2004), Gallego et al. (2004), and Liu and Van
capacitated), fixed charge (i.e., charge on nodes, edges Ryzin (2008). Later, Schön (2010) and Wang (2012) studied
or both), and network structure (e.g., two-echelon, multi- the assortment problem with capacity constraint imposed.
echelon). In contrast to the static view of the assortment planning
There has been extensive research on topics related problem, Rusmevichientong et al. (2010) and Caro and
to inventory placement. Earlier work includes the single- Gallien (2007) addressed the dynamic assortment problem
commodity uncapacitated facility location problem (single that revises or changes assortment selection as time elapses.
plant location problem or SPLP) (Jakob and Pruzan 1983;
Cornnejols et al. 1977), and the multi-commodity location 4.2 Industry practice
problem (Warszawski and Peer 1973; Karkazis and Boffey
1981; Revelle and Laporte 1996) which was extended based Inventory management has long been the center piece for
on SPLP. The capacity constraints, on nodes, edges or both, retailers. In this section, we present industry approaches to
impose additional challenges for the solution procedures. the inventory management problem regarding the following
Various techniques have been used for solving the problem aspects. As will be seen, these companies take significantly
such as greedy heuristics, dual ascent method, and Benders different approaches and emphasize different aspects of the
decomposition. Sridharan (1995) provides a survey on these inventory management problem.
solution techniques. Melkote and Daskin (2001) proposed
the capacitated facility location/network design problem to 4.2.1 Replenishment
bridge the gap between the problems of facility location
and network design. The multi-commodity capacitated The idea of safety stock has been popular for decades in
fixed-charge network flow problem is a more generalized industry as inventory replenishment method. Starting in
formulation compared to the aforementioned approaches. 1994, IBM started to develop the Asset Management Tool
Gendron (2011), Gendron and Larose (2014), and Melo (AMT) aiming to achieve quick responsiveness to customers
et al. (2009) provide comprehensive reviews on the exact with minimal inventory (Lin et al. 2000). In addition, IBM
solution approaches, including cutting plane methods, has developed a software tool named the IBM Supply Chain
Benders decomposition, and Lagrangian relaxation, and Simulator (SCS) to help with making strategic business
recent advances to this problem. To capture other practical decisions about the design and operation of its supply chain,
considerations, researchers have been making efforts in which include a function for determining safety stocks for
considering multi-period dynamics (Melo et al. 2006), the modeling of distribution (Bagchi et al. 1998).
stochasticity of demand or costs (Snyder 2006), and Intel uses a periodical review model in its multi-echelon
nonlinear cost structures. inventory optimization (MEIO) system, where inventory
Another related topic on inventory placement of broad targets are reviewed weekly with up-to-date sales and
and current interest focuses on the placement of specific inventory information. The inventory targets are set based
types of goods named product placement. It is to determine on a weeks-of-inventory (WOI) policy by product family.
the set of SKUs (i.e., products) to store in a specific location The inventory target is then fed into an Intel-developed
(i.e., distribution center) at a time point so as to better advanced-planning and scheduling (APS) optimizer that
service customers. It can also be viewed as an assortment minimizes production costs, lost-sales costs, and costs for
problem. The assortment planning problem has been deviating from the inventory targets (Manary and Willems
extensively studied and is still an emerging research area. 2008).
Reviews of the state-of-the-art on assortment planning are Walmart has been collaborating with P&G to deploy a
provided in Mahajan and van Ryzin (1999) and Kök et al. continuous review replenishment policy in which they con-
(2008). The latter also summarizes the current approaches tinuously monitor the inventory levels by RFID technology
used in the industry practice for inventory placement. A and automatically replenish their inventory when levels go
typical approach in literature to the assortment problem is below the safety stock level (Kosasi et al. 2014). Similarly,
to formulate it as an optimization problem with which to Amazon.com tracks its inventory position in real-time based
select a subset of products maximizing the revenue captured on warehouse receipts and shipments and places purchase
at a single store/location. To take the interaction between orders to vendors based on the forecasted amount needed
assortment and demand, a demand model, such as the subtracting the current on-hand inventory in the warehouse

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

56 J. of Data, Inf. and Manag. (2019) 1:45–64

(Zeppieri 2004). Chiles and Dau (2005) provide a detailed JD.com is well known for its super-fast delivery thanks to
description of the replenishment process at Walmart and the smart inventory allocation mechanism. It features same
Amazon. and next day delivery as a standard, allowing customers
Multi-echelon inventory management is increasingly to receive same-day delivery for orders placed before
incorporated into the practice of supply chain management. 11am and next-day delivery by 3pm for those placed
For instance, ideas along this line have been successfully before 11pm. More than 90% of orders on JD.com are
applied in practice such as IBM (Lin et al. 2000), Eastman delivered same-day or next-day. Due to the fast delivery
Kodak (Graves and Willems 2000), and Philips Electronics speed requirement, reactive transshipment, although still
(de Kok et al. 2005). exists in JD.com network, takes a smaller portion of the
At JD.com, a sophisticated inventory control system overall transshipment volume. In contrast to Amazon’s
is implemented with a system of inventory replenishment lateral transshipment, JD.com owns a more complex
models. The system selects the appropriated models based distribution network and uses a multi-step inventory
on product characteristics as well as business requirements. allocation algorithm. Specifically, they calculate the target
For example, products with low daily sales may be treated inventory level for each product and then select a subset
using a dedicated model to account for the higher relative of products to transship via optimization. This innovative
variability while high volume items are generally driven allocation strategy has led visible improvement regarding
by continuous or periodical review models. In general, stock-out rate at the distribution centers and the proportion
the replenishment models are formulated as optimization of orders meeting their same- and next-day delivery
problem minimizing total cost including inventory holding standards.
cost, and stock-out cost.
JD.com utilizes a fulfillment network structure that 4.2.3 Inventory placement
resembles multi-echelon inventory system with a RDC-
FDC design. Recently, approximation algorithms have been Compared to the more operational and detail-oriented
proposed for solving the replenishment (to RDCs) and approaches in academia, industry uses a more strategic and
allocation problems (from RDCs to FDCs) jointly (Qi et al. holistic approach for assortment planning.
2018). This work is extended based on the seminal work As reviewed by Kök et al. (2008), in retailing industry,
of Levi et al. (2007), which proposed the dual-balancing one common approach to assortment planning is for
policies. The algorithm is validated with actual JD.com corporate headquarters to decide on a single common
operation data and provides high computational efficiency assortment that is carried by all stores of the chain (e.g.,
by leveraging from two novel techniques: marginal cost Best Buy), except that in smaller stores, the breadth of
accounting and cost balancing. the assortment may be reduced by removing some of the
least important SKUs. On the contrary, a few retailers (e.g.,
4.2.2 Transshipment Bed Bath & Beyond) allow their store managers to manage
the assortment on their own. One example that follows
Transshipment is a vital and growing part in the inventory a hierarchical approach to assortment planning is Albert
management in practice. Transshipment allows for more Heijn, BV, a leading supermarket chain in the Netherlands.
efficient use of network resources and reduces costs Specifically, they first divide SKUs into categories, and
of fulfillment and replenishment throughout the entire store space is allocated to each category accordingly.
inventory network. Amazon uses a transshipment method Subsequently, they carry out product selection and facing
called lateral transshipment which deals with the movement allocation to products as detailed in Kök et al. (2008).
of goods within the same echelon of the supply chain. Three In contrast to academic studies, product placement for
transshipment types exist in Amazon, i.e. Customer Order industry practitioners is an even more intricate decision that
(reactive), Network Inventory Re-balancing (proactive) and requires many more practical issues to be considered and/or
Domestic Cross-dock (proactive). Reactive transshipment is customized to their specific business needs. For instance,
primarily used to reduce fulfillment cost while proactive as a strategic necessity, Best Buy would carry a rarely
transshipment is used for cost savings on replenishment demanded product just to maintain consumer perception
to fulfillment centers. Young (2016) details the key of it as offering the latest lines of technological products.
steps of transshipment process in Amazon, which include Retailers are also known to frequently carry unprofitable
inventory, transshipment pick, merge/palletize, outbound products in order to strategically drive for growth or overall
dock, transportation, inbound dock, receive, and stow, profitability of a whole category. Besides, Best Buy also
etc. The same process covers both reactive and transfer considers various other supply chain considerations, such
(proactive) transshipment. as vendor relations, vendor performance and the scope of

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 57

products offered by a vendor, to develop the assortment 5 Order fulfillment

plans (Kök et al. 2008).
In JD.com, there are two types of warehouses or Order fulfillment is the process of accepting, processing,
distribution centers, namely regional distribution centers and delivering customer orders. Order fulfillment is a
(RDC) and forward distribution centers (FDC). The former crucial component contributing to the customer experience,
have larger capacities and tend to store comprehensive as well as an important aspect of cost control in supply
SKUs, including slow moving products, and also serve as chain. For a large retailer like JD.com, the order fulfillment
the suppliers for the latter. On the other hand, the latter have decision is among the most difficult problems given that
limited capacities, are geographically closer to customers, there are billions of products in its large-scale and multi-
and usually store mostly high-selling SKUs in order to stage inventory network.
fulfill orders in a timely manner. As a primary feature that
customers pay attention to, delivery time is an important 5.1 Literature review
metric for e-commerce and e-retailers are thus seeking
to cut their delivery time in order to be more attractive Order fulfillment has become a popular research topic since
to customers. As a result, JD.com need to maximize the e-commerce entered the main stage of retail business. In
total number of orders that are shipped from FDCs so the era of e-commerce, the packages placed online can be
as to benefit from fast shipping. In case an order cannot delivered to the customers within a few days, sometimes
be satisfied by an FDC (e.g., one product is missing in even the same day. The scale and complexity of the online
the FDC), it may be routed to the closest RDC. As a order fulfillment system gives rise to a broad range of
consequence, the delivery is subject to delay. To make research opportunities. In this section, we review three
things more complicated, it is normally preferable to have main technical aspects of the order fulfillment process,
an order with multiple items delivered together. As a result, namely Fulfillment Optimization, Order Prioritization, and
simply placing the best sellers in FDC may not be able Fulfillment Flexibility.
to best utilize the capacities. Since a significant portion
of orders contain more than 2 products, it is necessary Fulfillment Optimization is the problem in the center of
to group items with similar attributes that sell together the order fulfillment. When a customer placed an order
in the same FDC to avoid splitting orders into multiple online, the retailer needs to decide where to ship the
packages. products to the customer destination. Different from the
Shi (2018) presented one example of the solutions used traditional retailing, customers do not select where the
for inventory replacement in JD.com. To fully leverage products should be shipped from as the logistic process of
the fast delivery from FDCs, they aim to maximize the an order fulfillment is hidden from the customer after the
number of orders that can be fulfilled entirely from the order has been placed. This allows for greater flexibility
FDCs and minimize the occurrences of order splits. As in selecting the best location and transportation method
a large-scale optimization problem, a practical solution to ship the products to meet the customer requirements
procedure to this is to use heuristics that rank orders by while minimizing the fulfillment cost. Acimovic and
their popularity. Specifically, they use SKU2Vec algorithm, Graves (2017) consider the fulfillment problem for a
which is motivated by Google’s Word2Vec (Mikolov et al. large distribution network of minimizing shipping costs
2013) to compute a latent vector for each SKU which is then for a single SKU, and propose an LP-based policy that
used to model the closeness to another SKU. An end-to-end incorporates the forecast of future orders. Xu et al. (2009)
neural network framework is then used to make inventory consider a batch optimization policy where each single
assortment decisions by directly capturing the co-purchase order fulfillment decision is delayed until solving the
relationship between products reducing order split by about fulfillment for a batch of orders together using an integer
2% (i.e., 2 million less split packages per year) compared to programming. Jasin and Sinha (2015) consider a problem
a benchmark Greedy Ranked algorithm. similar to Xu et al. (2009) that incorporates different costs
Alternatively, they also propose a data-driven graph- of bundled shipments.
based algorithm at JD.com that outputs subsets of products Besides solving the fulfillment optimization problem
to be placed at FDCs. This is achieved by sampling for a given level of customer requirement, another stream
batches of orders and aggregating the solutions of a of research works focuses on investigating policies that
selection problem solved with parametric cut minimization. can improve fulfillment performance by differentiating
This innovative approach is demonstrated to have superior customer requirements, or Order Prioritization. The idea
performance than the former inventory placement strategy of this can be traced back to the inventory ration policies
(Jehl et al. 2018). that use limited inventory to fulfill multiple classes of

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

58 J. of Data, Inf. and Manag. (2019) 1:45–64

customer demand discussed in Topkis (1968), Kaplan 5.2 Industry practice

(1969), Veinott (1965), and Nahmias and Demmy (1981).
In this stream of literature, customers are prioritized, and With a fast growth in the past ten years, large e-commerce
the inventory system is allowed to either back-order or retailers like Amazon.com and JD.com have established
lose demand for low priority customers to fulfill future highly efficient order fulfillment processes. We introduce
demand for high priority customers. In the context of the framework of a general order fulfillment engine and the
e-commerce, it is not easy to differentiate customers network structure it has been built upon in this section.
directly as usually all customers purchase products from
a single website. However, differentiating customer orders 5.2.1 Order fulfillment engine
by different levels of desired delivery-time standard is
feasible. Cattani and Souza (2002) investigate rationing The order fulfillment process usually starts with Order
in a direct marketing environment (similar to online Approval. A fraud detection procedure is conducted to
retailing), where customers may pay a higher fulfillment identify potential risks. This is typically done with
fee to reduce their delivery times. The integration of e- a classification algorithm. In the case of a high-risk
commerce and omni-channel retailing also brings new transaction, a human verification procedure may be
research opportunities. As offline customers usually have involved to further analyze the risk. After the order is
a priority in purchasing the products than the online verified, the order will be dispatched to the warehouses
customers, Karp (2017) discusses a protection level policy for processing. This step is usually controlled by an Order
to ensure the right amount of inventory is reserved for Management System (OMS). The fulfillment optimization
the offline customers. Govindarajan et al. (2018) discusses algorithms play a key role in this step. The algorithm
a joint decision for the inventory and fulfillment policies needs to take in all the realistic constraints such as
that optimize the total costs. Agatz et al. (2008) provide inventory availability, delivery cut-off times, warehouse
an excellent review on the other related omni-channel processing capacity, outbound transportation capacity, and
researches. the specific features of the order and then minimizes
the overall fulfillment cost. For various reasons, an order
Fulfillment Flexibility strategies approach the order fulfill- consisting with multiple products may be split into multiple
ment from a network design point of view. Inventories of warehouses for processing. As a result, the customer
the same product are stored in multiple storage locations in may receive multiple packages for an order placed. This
the network so as to provide “flexibility” for order fulfill- Order Dispatching step usually takes a short time and is
ment. Higher degree of flexibility allows for more options in processed without human intervention. The next key process
satisfying customer orders which eventually leads to lower is Picking, Assembling, and Packing at the warehouses.
order fulfillment costs due to less order split and inventory This step is usually controlled by a Warehouse Management
stock-out at nearby storage locations. Designing a flexible System (WMS) which ensures the orders dispatched to the
fulfillment network is related to the concept of process flex- warehouse can be completed before the outbound due times
ibility, which is developed by Jordan and Graves (1995) in and at the same time maximizes the operational efficiency.
1995. DeValve et al. (2018) extend the flexibility concept The packed orders are then loaded onto trucks, waiting to
to online retailing by investigating a threshold policy with be shipped. Depending on the distance from the warehouse
the structure which is motivated by the protection levels to the customer, different transportation methods may be
in revenue management originated from Littlewood (1972). involved in the shipment, ranging from air to motorcycles.
Acimovic and Graves (2017) show how traditional decen- The order fulfillment completes with the success of the final
tralized allocation policies may perform sub-optimally and mile delivery of the package to the customer.
induce dynamics (whiplash) that result in costly spillover.
The literature of fulfillment flexibility also shares some Product Return is an important aspect of customer experi-
similar features with models from inventory transshipment ence. Most companies provide a convenient return policy
literature. Paterson et al. (2011) show flexibility can be with no or small additional return shipment fees. To provide
created both by transshipping inventory among the facili- more flexibility to the customer, many companies also apply
ties proactively and reactively. Axsäter (2003) develops a a free Cancellation Window policy which allows customers
decision rule dictating whether to transship, or whether to to cancel the purchase before a cut-off time. After the can-
incur the back-order costs. Yang and Qin (2007) discuss cellation window, customers who want to cancel orders can
a model that utilizes virtual lateral transshipment between only do so by returning goods after they have arrived and
two factories. Archibald et al. (2009) develop a trans- by mailing the package back. To further increase customer
shipment heuristic for a realistic multi-location inventory experience, e-commerce retailers are trying to extend the
system. cancellation windows until the delivery of a package. For

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 59

example, JD.com uses a concept called Cool-Down Period Different from a fully-connected distribution network, a
(Wu 2018) to reduce the cost for order cancellation. For each hierarchical distribution network is composed of multiple
order, a cancellation probability is calculated based on the levels of inventory storage facilities. When a product is
nature of the order and who is buying. An order may be put out-of-stock at the lower level storage facility, the demand
on hold for a cool-down period before dispatching orders is fulfilled by a higher-level storage facility, which is
to the warehouses for picking. The cool-down period will usually much farther from the customer. This hierarchical
be longer if the order is associated with a high cancellation distribution network is typified by JD.com. Products need
probability. to be carefully placed (Shi 2018) at each tier of the network
The new trend of the integration of online and omni- in order to maximize the utilization of the storage capacity.
channel retailing raises new opportunities to the order The risk of inventory stock-out at the lower level storage
fulfillment. On one hand, with the omni-channel stores facility needs to be carefully managed by frequent inventory
having the inventory of the same ideal product when the transshipment (Ma et al. 2018) from the upper level storage
customer placed the order online, the order may be shipped facility. The advantage of a hierarchical distribution network
to customer directly from the store which reduces the is its high efficiency. As the lower-level storage facility is
delivery time. On the other hand, with multiple omni- usually close to the customers, it enables ultra-fast delivery
channel stores as potential sources for order fulfillment, to the customers. A hierarchical system is also easier to
it increases the number of choices for order dispatching manage as the inventory transshipment flow among the
and thus increases the complexity of the fulfillment storage facilities is usually unidirectional.
algorithm. For example, at JD.com, with its expansion
to the omni-channel retailing including JD Home for 3C
and home appliance, 7Fresh for food and groceries, and 6 Future challenges
the collaboration with Walmart Stores (Shan 2018), it has
built a fulfillment system to select the most cost-efficient In this section, we discuss several challenges that retailers
fulfillment methods to meet customer requirements by are still facing today in the area of supply chain
leveraging the power of online and offline. management. As the retail business evolves overtime, we do
see new challenges as well as new methodologies emerging
5.2.2 Order fulfillment network in practice which could provide some insights to the future
supply chain management research directions.
The design of the order fulfillment network can be generally
categorized as Fully-Connected Distribution Network and 6.1 From supply chain to supply network
Hierarchical Distribution Network. A fully-connected
distribution network refers to a fulfillment system where Supply chain is originally named to describe the chain of
each inventory storage facility (a warehouse or a depot) can activities involved in moving a product or service from
be used to fulfill the order from any place in the network. supplier to customer. However, as technology and business
This distribution network is exemplified by Amazon.com scenario evolves, the supply “chain” is also evolving to
and Walmart.com, the two largest e-commerce retailers a supply “network” where the different parties within the
in the US. With a fully-connected distribution network, system forms a tighter interconnection.
as the inventory of a product is stored in many storage As described in Section 5, the real fulfillment networks
facilities, the order fulfillment decision can be very flexible used by large scale retailers have grown to very complex
as there will be many ways to fulfill an order from structures (JD operates more than 550 warehouses nation-
different storage facilities. While this flexibility guarantees wide in China (JD.com 2018b); Amazon is operating 75 ful-
the required product can be fulfilled as long as there is fillment centers within North America (Amazon.com 2018);
inventory somewhere in the entire network, it increases the Walmart and Sam’s Club combined have 175 fulfillment
transportation cost as orders are more likely to be shipped centers in United States (MWPVL International Inc. 2018)).
in longer distance and even with air methods in order to A modern fulfillment network may include tiered network
meet the delivery promise. To avoid high transportation structure similar to the classic multi-echelon design. How-
cost, the fulfillment decision should not only look at the ever, the network topology could also be subject to addi-
direct fulfillment cost (usually composed of transportation tional constraints such as availability of transship links or
and processing cost) but also consider the future demand compatibility of product types. For example, JD.com’s ful-
from the customers in order to fulfill orders strategically. fillment network is a superset of 6 networks, e.g. small-to-
Acimovic and Graves (2017) provide a good review of medium sized warehousing, oversized warehousing, cross
the current industry practice from an academic point of border, cold chain delivery, frozen and chilled warehousing,
view. B2B and crowdsourcing logistics (JD.com 2018a), each has

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

60 J. of Data, Inf. and Manag. (2019) 1:45–64

its own capability and capacity. These networks can share But the gap between revenue management and inventory
the same transportation or fulfillment facilities in certain management still exists despite efforts by researchers
sections of the supply chain process while requires special- (Elmaghraby and Keskinocak 2003).
ized channel for some other sections. For example, some The rise of group purchase (Liang 2018) in China
products in the cold chain delivery can be delivered using recently also creates new challenges in supply chain
the regular delivery channel (shared with other networks) management where massive demand is created within a
using a specialized packaging, but these products must be very short period, creating pressure on both inventory
stored in refrigerated warehouses. At the same time, over- management and logistics. With the help of social networks,
sized products have their own last mile delivery mechanism the group purchase frequently establish highly localized
but can share the same or similar storage facility as small-to- pattern creating geographic hot spots for the product
medium sized items. How to efficiently utilize the complex demand. As the group purchase products are normally
networks by efficiently sharing the facilities and capacities limited time promotions, these hot spots exist in both spatial
still requires extensive research. and temporal space, creating ever increasing difficulty in
Crowd sourcing has also grown to be a viable approach managing inventory and fulfillment.
to supplement or even act as the backbone of the trans-
portation and fulfillment system (Rougès and Montreuil 6.3 Big data, small data
2014). Companies such as Amazon.com, Walmart has all
established certain type of crowd-sourced delivery solu- With the growth of online marketplace as well as
tions within the last several years. JD.com and Walmart digitization in offline stores, retailers now have extensive
recently invested $500 Million into Dada-JD Daojia, which data to work with in order to figure out the best management
is a crowd-sourcing delivery platform in China with oper- strategy. With the rise of data availability, there is a
ations in more than 400 major cities (Choudhury 2018). clear trend toward data-driven approaches in supply chain
Using such delivery approach provides better scalability to management research. However, how to effectively utilize
the supply chain system, but also brings in additional uncer- such data in providing better supply chain management
tainty in both reliability and cost implication. There are still decision is still not clear to many retailers. Further, the
a lot of questions to be answered, especially when in-house data size available to supply chain management is still
fulfillment capacity is used alongside the crowd-sourcing considerably smaller than other more popular AI areas
abilities. such as image or natural language processing, etc. For
Furthermore, following recent pushes of omni-channel example, current effort on reinforcement learning based
retailing (Verhoef et al. 2015), it is more and more com- replenishment algorithm (Sui et al. 2010; Chaharsooghi
mon for retailers to use physical store’s shelf inventory et al. 2008) still need a relatively large number of iterations
to fulfill online orders as well. Some issues and exist- for the algorithm to converge to optimal which may be
ing research are summarized by Hübner et al. (2016) and unrealistic for practical implementation. Although there
Melacini et al. (2018). With the recent collaboration with are studies (Shi et al. 2018; Zhang and Gao 2017;
Walmart (Shan 2018), JD.com gains the ability to access Oroojlooyjadid et al. 2016) on using deep learning to solve
Walmart’s local stores’ inventory. This provides better cus- inventory problems in supply chain, few has been able to
tomer experience for both better selection and faster deliv- demonstrate performance on practical set ups.
ery. However, this creates complexities in inventory man- Besides the “big data” issue, e-commerce marketplaces
agement and fulfillment decisions as well for both JD.com have greatly improved the number of small players in the
and Walmart. How to efficiently manage inventory across retail industry, who only manufacturer or manage a small
the fulfillment network is still a challenging problem to all set of SKUs or tailored to a particular customer base. Such
practitioners. retailers face a huge disadvantage in the era of big data.
How to effectively utilize the “small data” available to the
6.2 From managing supply to managing demand small players to help them make effective decisions is also
an exciting problem to solve.
Supply Chain management is now not just about managing
supply but also on managing demand. Recent advances
in search and recommendation algorithm has made it References
possible for managing demand and supply simultaneously
to maximize retailer’s revenue and/or profit (Bernstein et al. Acimovic J, Graves SC (2017) Mitigating spillover in online retailing
via replenishment. Manuf Serv Oper Manag 19(3):419–436
2015; Chen et al. 2016). Inventory markdowns (Lazear Agatz NA, Fleischmann M, Van Nunen JA (2008) E-fulfillment and
1986) has long been a common practice for managing multi-channel distribution–a review. Eur J Oper Res 187(2):339–
products by retailers, especially for seasonal products. 356

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 61

Ahmed NK, Atiya AF, Gayar NE, El-Shishiny H (2010) An Chen C (2018a) Amazon now sells 76 of its own private-label brands
empirical comparison of machine learning models for time series — from clothes to baby wipes. https://www.businessinsider.com/
forecasting. Econ Rev 29(5-6):594–621 amazon-private-label-brands-list-2018-4 (accessed December 3,
Allen SG (1958) Redistribution of total stock over several user 2018)
locations. Naval Research Logistics Quarterly 5(4):337–345 Chen C (2018b) Online retailer JD.com follows Amazon’s footsteps
Allen SG (1961) A redistribution model with set-up charge. Manag Sci with own brand selling household basics. https://www.scmp.com/
8(1):99–108 tech/china-tech/article/2128481/online-retailer-jdcom-follows-ama
Amazon.com (2018) Amazon’s Fulfillment Network. https://www.ab zons-footsteps-own-brand-selling (accessed December 3, 2018)
outamazon.com/working-at-amazon/amazons-fulfillment-network Chen F, Song JS (2001) Optimal policies for multiechelon inventory
(accessed November 19, 2018) problems with markov-modulated demand. Oper Res 49(2):226–
Andraski JC, Haedicke J (2003) CPFR: time For the breakthrough? 234
Supply Chain Management Review 7(3):54–60 Chen F, Drezner Z, Ryan JK, Simchi-Levi D (2000) Quantifying the
Angulo A, Nachtmann H, Waller MA (2004) Supply chain information bullwhip effect in a simple supply chain: the impact of forecasting,
sharing in a vendor managed inventory partnership. J Bus Logist lead times, and information. Manag Sci 46(3):436–443
25(1):101–120 Chen T, Guestrin C (2016) Xgboost: A scalable tree boosting
Anupindi R, Dada M, Gupta S (1998) Estimation of consumer demand system. In: Proceedings of the 22nd ACM SIGKDD International
with stock-out based substitution: an application to vending Conference on Knowledge Discovery and Data Mining, ACM, pp
machine products. Mark Sci 17(4):406–423 785–794
Archibald TW, Black D, Glazebrook KD (2009) An index heuristic for Chen X, Ma W, Simchi-Levi D, Xin L (2016) Assortment planning
transshipment decisions in multi-location inventory systems based for recommendations at checkout under inventory constraints,
on a pairwise decomposition. Eur J Oper Res 192(1):69–78 available at SSRN: https://ssrn.com/abstract=2853093
Axsäter S (2003) A new decision rule for lateral transshipments in Chiles CR, Dau MT (2005) An analysis of current supply chain
inventory systems. Manag Sci 49(9):1168–1179 best practices in the retail industry with case studies of wal-
Bagchi S, Buckley SJ, Ettl M, Lin GY (1998) Experience using mart and amazon.com. Master’s thesis, Massachusetts Institute of
the IBM supply chain simulator. In: Simulation Conference Technology
Proceedings, 1998. Winter, vol 2. IEEE, pp 1387–1394 Choudhury SR (2018) Walmart and JD.com invest $500 million in a
Bahdanau D, Cho K, Bengio Y (2014) Neural machine translation by Chinese online delivery company. https://www.cnbc.com/2018/
jointly learning to align and translate. arXiv:14090473 08/09/jdcom-walmart-invest-500-million-in-chinese-company-dad
Barkan O, Koenigstein N (2016) Item2vec: neural item embedding for a-jd-daojia.html (accessed November 30, 2018)
collaborative filtering. In: 2016 IEEE 26th International Workshop Cinar YG, Mirisaee H, Goswami P, Gaussier E, Aı̈t-Bachir A,
on Machine Learning for Signal Processing (MLSP), IEEE, pp 1–6 Strijov V (2017) Position-based content attention for time series
Bass FM (1969) A new product growth for model consumer durables. forecasting with sequence-to-sequence RNNs. In: International
Manag Sci 15(5):215–227 Conference on Neural Information Processing, Springer, pp 533–
Bernstein F, Kök AG, Xie L (2015) Dynamic assortment customiza- 544
tion with limited inventories. Manuf Serv Oper Manag 17(4):538– Clark AJ, Scarf H (1960) Optimal policies for a multi-echelon
553 inventory problem. Manag Sci 6(4):475–490
Bianchi FM, Maiorino E, Kampffmeyer MC, Rizzi A, Jenssen R Conlon CT, Mortimer JH (2013) Demand estimation under incom-
(2017) An overview and comparative analysis of recurrent neural plete product availability. American Economic Journal:, Microe-
networks for short term load forecasting. arXiv:170504378 conomics 5(4):1–30
Bollerslev T (1986) Generalized autoregressive conditional het- Cornnejols G, Fisher M, Nemhauser G (1977) Location of bank
eroskedasticity. J Econ 31(3):307–327 accounts of optimize float: an analytic study of exact and
Box GE, Jenkins GM, Reinsel GC, Ljung GM (2015) Time Series approximate algorithm. Manag Sci 23(8):789–810
Analysis: Forecasting and Control. Wiley, Hoboken Covert RP, Philip GC (1973) An EOQ model for items with Weibull
Brahimi N, Dauzere-Peres S, Najid NM, Nordli A (2006) Single item distribution deterioration. AIIE Transactions 5(4):323–326
lot sizing problems. Eur J Oper Res 168(1):1–16 Dempster AP, Laird NM, Rubin DB (1977) Maximum likelihood from
Brandolese A, Cigolini R (1999) A new model for the strategic incomplete data via the em algorithm. J R Stat Soc Ser B Methodol
management of inventories subject to peaks in market demand. Int 39(1):1–38
J Prod Res 37(8):1859–1880 DeValve L, Wei Y, Wu D, Yuan R (2018) Understanding the value of
Breiman L (2017) Classification and regression trees. Routledge fulfillment flexibility in an online retailing environment, available
Cachon GP, Fisher M (2000) Supply chain inventory management and at SSRN: http://ssrn.com/abstract=3265838
the value of shared information. Manag Sci 46(8):1032–1048 Doganis P, Alexandridis A, Patrinos P, Sarimveis H (2006) Time
Cachon GP, Zipkin PH (1999) Competitive and cooperative inventory series sales forecasting for short shelf-life food products based on
policies in a two-stage supply chain. Manag Sci 45(7):936–953 artificial neural networks and evolutionary computing. J Food Eng
Caro F, Gallien J (2007) Dynamic assortment with demand learning 75(2):196–204
for seasonal consumer goods. Manag Sci 53(2):276–292 Drucker H, Burges CJ, Kaufman L, Smola AJ, Vapnik V (1997)
Cattani KD, Souza GC (2002) Inventory rationing and shipment Support vector regression machines. In: Advances in Neural
flexibility alternatives for direct market firms. Prod Oper Manag Information Processing Systems, pp 155–161
11(4):441–457 Duan Y, Yao Y, Huo J (2015) Bullwhip effect under substitute
Çetinkaya S, Lee CY (2000) Stock replenishment and shipment products. J Oper Manag 36:75–89
scheduling for vendor-managed inventory systems. Manag Sci Dudek G (2015) Short-term load forecasting using random forests. In:
46(2):217–232 Intelligent Systems’ 2014, Springer, pp 821–828
Chaharsooghi SK, Heydari J, Zegordi SH (2008) A reinforcement Elmaghraby W, Keskinocak P (2003) Dynamic pricing in the presence
learning model for supply chain ordering management: an of inventory considerations: research overview, current practices,
application to the beer game. Decis Support Syst 45(4):949–959 and future directions. Manag Sci 49(10):1287–1309

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

62 J. of Data, Inf. and Manag. (2019) 1:45–64

Engle RF (1982) Autoregressive conditional heteroscedasticity with Jain A, Rudi N, Wang T (2014) Demand estimation and ordering under
estimates of the variance of United Kingdom inflation. Economet- censoring: stock-out timing is (almost) all you need. Oper Res
rica:, Journal of the Econometric Society pp 987–1007 63(1):134–150
Estes R (1973) The joint probability approach and reorder point Jakob K, Pruzan PM (1983) The simple plant location problem: survey
determination. Journal of Production and Inventory Management and synthesis. Eur J Oper Res 12:36–81
14(2):50–56 Jasin S, Sinha A (2015) An LP-based correlated rounding scheme for
Fan C, Zhang Y, Pan Y, Li X, Zhang C, Yuan R, Wu D, multi-item ecommerce order fulfillment. Oper Res 63(6):1336–
Wang W, Pei J, Huang H (2019) Multi-horizon time series 1351
forecasting with temporal attention learning. In: 25th ACM JDA Software (2018) JDA Introduces Luminate Demand Edge to
SIGKDD Conference on Knowledge Discovery and Data Mining Power Self-Learning, Autonomous Demand Forecasts. https://jda.
(KDD 2019), Anchorage, Alaska, USA com/knowledge-center/press-release/jda-introduces-luminate-dem
Fan Y, Li P, Song Z (2006) Dynamic least squares support vector and-edge-to-power-selflearning-autonomous-demand-forecasts
machine. In: 2006 6Th World Congress on Intelligent Control and (accessed May 15, 2019)
Automation, vol 1. IEEE, pp 4886–4889 JD.com (2018a) JD.com Business. https://corporate.jd.com/
Federgruen A, Zipkin P (1984) Computational issues in an infinite- ourBusiness (accessed November 19, 2018)
horizon, multiechelon inventory model. Oper Res 32(4):818–836 JD.com (2018b) JD.com Inc 3Q2018 Financial and Operational High-
Flunkert V, Salinas D, Gasthaus J (2017) DeepAR: Probabilistic fore- lights. http://ir.jd.com/static-files/e7ae12fc-41b9-43c9-9ddb-e6afe
casting with autoregressive recurrent networks. arXiv:170404110 1155f11 (accessed November 19, 2018)
Forrester JW (1958) Industrial dynamics. a major breakthrough for Jehl T, Shi Y, Wu D, Shen ZJM (2018) Product placement
decision makers. Harv Bus Rev 36(4):37–66 optimization using parametric cuts, working paper
Friedman JH (2001) Greedy function approximation: a gradient Jordan WC, Graves SC (1995) Principles on the benefits of
boosting machine. Ann Stat pp 1189–1232 manufacturing process flexibility. Manag Sci 41(4):577–594
Gallego G, Iyengar G, Phillips R, Dubey A (2004) CORC Technical Kahn KB (2014) New product forecasting: an applied approach.
Report TR-2004-01. Managing flexible products on a network Routledge
Gendron B (2011) Decomposition methods for network design. Kam HT (1995) Random decision forest. In: Proc of the 3rd int’l conf
Procedia Soc Behav Sci 20:31–37 on document analysis and recognition, Montreal, Canada, August,
Gendron B, Larose M (2014) Branch-and-price-and-cut for large-scale pp 14–18
multicommodity capacitated fixed-charge network design. EURO Kamruzzaman J, Sarker RA (2006) Artificial neural networks:
Journal on Computational Optimization 2(1-2):55–75 applications in finance and manufacturing. In: Artificial Neural
Girard A, Rasmussen CE, Candela JQ, Murray-Smith R (2003) Networks in Finance and Manufacturing, IGI Global, pp 1–27
Gaussian process priors with uncertain inputs application to Kaplan A (1969) Stock rationing. Manag Sci 15(5):260–267
multiple-step ahead time series forecasting. In: Advances in Karkazis J, Boffey T (1981) The multi-commodity facilities location
Neural Information Processing Systems, pp 545–552 problem. J Oper Res Soc 32(9):803–814
Govindarajan A, Sinha A, Uichanco J (2018) Joint inventory and ful Karmarkar US (1981) The multiperiod multilocation inventory
fillment decisions for omnichannel retail networks, available at problem. Oper Res 29(2):215–228
SSRN: http://ssrn.com/abstract=2924850 Karp J (2017) Models and methods for omni-channel fulfillment. PhD
Graves SC, Willems SP (2000) Optimizing strategic safety stock thesis, Carnegie Mellon University
placement in supply chains. Manuf Serv Oper Manag 2(1):68–83 Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu TY
Guide VDR Jr, Srivastava R (2000) A review of techniques for (2017) Lightgbm: A highly efficient gradient boosting decision
buffering against uncertainty with MRP systems. Prod Plan tree. In: Advances in Neural Information Processing Systems, pp
Control 11(3):223–233 3146–3154
Hamilton JD (1995) Time Series Analysis. Economic Theory II. Kihoro JM, Otieno RO, Wafula C (2004) Seasonal time series
Princeton University Press, USA, pp 625–630 forecasting: a comparative study of ARIMA and ANN models. Afr
Hamzaçebi C (2008) Improving artificial neural networks’ perfor- J Sci Technol 5(2):41–49
mance in seasonal time series forecasting. Inf Sci 178(23):4550– Kim SM, Mahoney JT (2010) Collaborative planning, forecasting
4559 and replenishment (CPFR) as a relational contract: an incomplete
Harris FW (1915) Operations and cost. Factory Management Series pp contracting perspective. Int J Learn Intellect Cap 7(3-4):403–428
48–52 Koenker R, Bassett G Jr (1978) Regression quantiles. Econometrica:
Hastie T, Tibshirani R (1996) Discriminant adaptive nearest neighbor Journal of the Econometric Society 46:33–50
classification and regression. In: Advances in Neural Information Kök AG, Fisher ML, Vaidyanathan R (2008) Assortment planning:
Processing Systems, pp 409–415 Review of literature and industry practice. In: Retail Supply Chain
Herer YT, Tzur M, Yücesan E (2002) Transshipments: an emerging Management, Springer, pp 99–153
inventory recourse to achieve supply chain leagility. Int J Prod de Kok T, Janssen F, Van Doremalen J, Van Wachem E, Clerkx M,
Econ 80(3):201–212 Peeters W (2005) Philips electronics synchronizes its supply chain
Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural to end the bullwhip effect. Interfaces 35(1):37–48
Comput 9(8):1735–1780 Kosasi S, Saragih H, Kom M (2014) How RFID technology boosts
Holt CC (2004) Forecasting seasonals and trends by exponentially Walmart’s supply chain management. International Journal of
weighted moving averages. Int J Forecast 20(1):5–10 Information Technology and Business Management 24(1):1–9
van Houtum GJ, Zijm WHM (1991) Computational procedures for Krishnan KS, Rao VRK (1965) Inventory control in N warehouses. J
stochastic multi-echelon production systems. Int J Prod Econ Ind Eng 16(3):212–215
23(1-3):223–237 Kulp SC, Lee HL, Ofek E (2004) Manufacturer benefits from informa-
Hübner A, Holzapfel A, Kuhn H (2016) Distribution systems in tion integration with retail customers. Manag Sci 50(4):431–444
omni-channel retailing. Bus Res 9(2):255–296 Kusiak A, Zheng H, Song Z (2009) Short-term prediction of wind
Ismail Z, Abu N (2013) A study on new product demand forecasting farm power: a data mining approach. IEEE Trans Energy Convers
based on bass diffusion model. J Math Stat 9(2):84 24(1):125–136

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J. of Data, Inf. and Manag. (2019) 1:45–64 63

Lagodimos AG, Anderson EJ (1993) Optimal positioning of safety Mishra BK, Raghunathan S (2004) Retailer- vs. vendor-managed
stocks in MRP. The International Journal Of Production Research inventory and brand competition. Manag Sci 50(4):445–457
31(8):1797–1813 MWPVL International Inc. (2018) The Walmart Distribution Center
Längkvist M, Karlsson L, Loutfi A (2014) A review of unsupervised Network in the United States. http://www.mwpvl.com/html/
feature learning and deep learning for time-series modeling. walmart.html (accessed November 19, 2018)
Pattern Recogn Lett 42:11–24 Nahmias S, Demmy WS (1981) Operating characteristics of an
Lazear EP (1986) Retail pricing and clearance sales. Am. Econ Rev inventory system with rationing. Manag Sci 27(11):1236–1245
76(1):14–32 Nedellec R, Cugliari J, Goude Y (2014) GEFCom2012 electric load
Lee H, Whang S (1999) Decentralized multi-echelon supply chains: forecasting and backcasting with semi-parametric models. Int J
incentives and information. Manag Sci 45(5):633–640 Forecast 30(2):375–381
Lee H, Kim SG, Hw Park, Kang P (2014) Pre-launch new product Nenni ME, Giustiniano L, Pirolo L (2013) Demand forecasting in the
demand forecasting using the Bass model: A statistical and fashion industry: a review. International Journal of Engineering
machine learning-based approach. Technol Forecast Soc Chang Business Management 5:5–36
86:49–64 Noham R, Tzur M (2014) The single and multi-item transshipment
Lee HL (1987) A multi-echelon inventory model for repairable items problem with fixed transshipment costs. Nav Res Logist (NRL)
with emergency lateral transshipments. Manag Sci 33(10):1302– 61(8):637–664
1316 O’Neil S, Zhao X, Sun D, Wei JC (2016) Newsvendor problems with
Lee HL, Padmanabhan V, Whang S (1997) Information distortion in a demand shocks and unknown demand distributions. Decis Sci
supply chain: the bullwhip effect. Manag Sci 43(4):546–558 47(1):125–156
Lee HL, So KC, Tang CS (2000) The value of information sharing in Oroojlooyjadid A, Snyder L, Takáč M (2016) Applying deep learning
a two-level supply chain. Manag Sci 46(5):626–643 to the newsvendor problem. arXiv:160702177
Lesnaia E (2004) Optimizing safety stock placement in general Palmer JW, Markus ML (2000) The performance impacts of quick
network supply chains. PhD thesis, Massachusetts Institute of response and strategic alignment in specialty retailing. Inf Syst Res
Technology 11(3):241–259
Levi R, Pál M, Roundy RO, Shmoys DB (2007) Approximation Panahifar F, Heavey C, Byrne PJ, Fazlollahtabar H (2015) A frame-
algorithms for stochastic inventory control models. Math Oper Res work for collaborative planning, forecasting and replenishment
32(2):284–302 (CPFR) state of the art. J Enterp Inf Manag 28(6):838–871
Liang C (2018) New Group-Buying Services Have Pinduoduo in Paterson C, Kiesmüller G, Teunter R, Glazebrook K (2011) Inventory
Their Crosshairs. https://www.sixthtone.com/news/1002740/ models with lateral transshipments: a review. Eur J Oper Res
new-group-buying-services-have-pinduoduo-in-their-crosshairs 210(2):125–136
(accessed December 2, 2018) Pitkin J, Manolopoulou I, Ross G (2018) Bayesian hierarchi-
Lin G, Ettl M, Buckley S, Bagchi S, Yao DD, Naccarato BL, Allan cal modelling of sparse count processes in retail analytics.
R, Kim K, Koenig L (2000) Extended-enterprise supply-chain arXiv:180505657
management at IBM personal systems group and other divisions. Premkumar GP (2000) Interorganization systems and supply chain
Interfaces 30(1):7–25 management. Inf Syst Manag 17(3):1–14
Prokhorenkova L, Gusev G, Vorobev A, Dorogush AV, Gulin A
Littlewood K (1972) Forecasting and control of passenger bookings.
(2017) Catboost: unbiased boosting with categorical features.
Airline Group International Federation of Operational Research
arXiv:170609516
Societies Proceedings 12:95–117
Qi M, Yuan R, Wu D, Shen ZJM (2018) An approximation algorithm
Liu Q, Van Ryzin G (2008) On the choice-based linear programming
for joint replenishment and allocation problem, working paper
model for network revenue management. Manuf Serv Oper Manag
Ramakrishna KS, Sharafali M, Lim YF (2015) A two-item two-
10(2):288–310
warehouse periodic review inventory model with transshipment.
Ma C, Lu J, Yuan R (2018) The Secret Behind JD.com’s
Ann Oper Res 233(1):365–381
Super Fast Delivery. https://medium.com/jd-technology-blog/ Revelle CS, Laporte G (1996) The plant location problem: new models
the-secret-behind-jd-coms-super-fast-delivery-71a6c0792405 and research prospects. Oper Res 44(6):864–874
(accessed February 26, 2019) Rougès JF, Montreuil B (2014) Crowdsourcing delivery: New
Mahajan S, van Ryzin GJ (1999) Retail inventories and consumer interconnected business models to reinvent delivery. In: 1st
choice. In: Quantitative Models for Supply Chain Management, International Physical Internet Conference, pp 1–19
Springer, pp 491–551 Ruiz-Torres AJ, Mahmoodi F (2010) Safety stock determination based
Manary MP, Willems SP (2008) Setting safety-stock targets at Intel in on parametric lead time and demand information. Int J Prod Res
the presence of forecast bias. Interfaces 38(2):112–122 48(10):2841–2857
Melacini M, Perotti S, Rasini M, Tappia E (2018) E-fulfilment Rusmevichientong P, Shen ZJM, Shmoys DB (2010) Dynamic
and distribution in omni-channel retailing: a systematic literature assortment optimization with a multinomial logit choice model
review. Int J Phys Distrib Logist Manag 48(4):391–414 and capacity constraint. Oper Res 58(6):1666–1680
Melkote S, Daskin MS (2001) Capacitated facility location/network Salameh MK (1997) Buffer stock under the effect of fluctuating
design problems. Eur J Oper Res 129(3):481–495 demand. Prod Plan Control 8(1):37–40
Melo MT, Nickel S, Da Gama FS (2006) Dynamic multi-commodity Schön C (2010) On the optimal product line selection problem with
capacitated facility location: a mathematical modeling framework price discrimination. Manag Sci 56(5):896–902
for strategic supply chain planning. Comput Oper Res 33(1):181– Sethi V, Hwang KT, Pegels C (1993) Information technology and
208 organizational performance: a critical evaluation of computer-
Melo MT, Nickel S, Saldanha-Da-Gama F (2009) Facility location and world’s index of information systems effectiveness. Inf Manag
supply chain management–a review. Eur J Oper Res 196(2):401– 25(4):193–205
412 Shan X (2018) JD to share inventory with Wal-Mart. http://en.people.
Mikolov T, Chen K, Corrado G, Dean J (2013) Efficient estimation of cn/n3/2017/0726/c90000-9246773.html (accessed December 2,
word representations in vector space. arXiv:13013781 2018)

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

64 J. of Data, Inf. and Manag. (2019) 1:45–64

Shang KH, Song JS (2003) Newsvendor bounds and heuristic for Wang R (2012) Capacitated assortment and price optimization
optimal policies in serial supply chains. Manag Sci 49(5):618–638 under the multinomial logit model. Oper Res Lett 40(6):492–
Sheffi Y (2002) The value of CPFR. In: Proceedings of RIRL 497
Conference, Lisboa Warszawski A, Peer S (1973) Optimizing the location of facilities on
Sherbrooke CC (1968) METRIC: A multi-echelon technique for a building site. J Oper Res Soc 24(1):35–44
recoverable item control. Oper Res 16(1):122–141 Wen R, Torkkola K, Narayanaswamy B (2017) A multi-horizon
Shi Y (2018) Solving Inventory Assortment Problem by Utilizing quantile recurrent forecaster. arXiv:171111053
Product Embedding. https://medium.com/jd-technology-blog/ Williams CK, Rasmussen CE (1996) Gaussian processes for regres-
solving-inventory-assortment-problem-by-utilizing-product-embe sion. In: Advances in Neural Information Processing Systems, pp
dding-807f992cb35a (accessed December 3, 2018) 514–520
Shi Y, Qi M, Ma C, Yuan R, Wu D, Shen ZJ (2018) A practical end- Winters PR (1960) Forecasting sales by exponentially weighted
to-end deep learning inventory model for e-commerce. Working moving averages. Manag Sci 6(3):324–342
Paper Wong H, van Houtum GJ, Cattrysse D, Van Oudheusden D (2006)
Shocker AD, Bayus BL, Kim N (2004) Product complements and Multi-item spare parts systems with lateral transshipments and
substitutes in the real world: the relevance of “other products”. J waiting time constraints. Eur J Oper Res 171(3):1071–1093
Mark 68(1):28–40 Wu D (2018) Going the Extra Mile In Customer Service: How Does
Shukla M, Jharkharia S (2013) Agri-fresh produce supply chain JD.com Make Order Cancellation So Easy? https://medium.
management: a state-of-the-art literature review. Int J Oper Prod com/jd-technology-blog/going-the-extra-mile-in-customer-service-
Manag 33(2):114–158 how-does-jd-com-make-order-cancellation-so-easy-a00ebc7fd1ee
Snyder LV (2006) Facility location under uncertainty: a review. IIE (accessed February 26, 2019)
Trans 38(7):547–564 Xu PJ, Allgor R, Graves SC (2009) Benefits of reevaluating real-time
Snyder LV, Shen ZJM (2011) Fundamentals of Supply Chain Theory. order fulfillment decisions. Manuf Serv Oper Manag 11(2):340–
Wiley, Hoboken 355
Sridharan R (1995) The capacitated plant location problem. Eur J Oper Xu Q, Liu X, Jiang C, Yu K (2016) Quantile autoregression neural
Res 87(2):203–213 network model with applications to evaluating value at risk. Appl
Srinivasan K, Kekre S, Mukhopadhyay T (1994) Impact of electronic Soft Comput 49:1–12
data interchange technology on JIT shipments. Manag Sci Yang J, Qin Z (2007) Capacitated production control with virtual
40(10):1291–1304 lateral transshipments. Oper Res 55(6):1104–1119
Stein T, Sweat J (1998) Killer supply chains. Information Week Yao Y, Dresner M (2008) The inventory value of information sharing,
708(9):36–46 continuous replenishment, and vendor-managed inventory. Trans-
Sui Z, Gosavi A, Lin L (2010) A reinforcement learning approach portation Research Part E:, Logistics and Transportation Review
for inventory replenishment in vendor-managed inventory systems 44(3):361–378
with consignment inventory. Eng Manag J 22(4):44–53 Young E (2016) Amazon UK transshipment stow process analysis
Suykens JAK, Vandewalle J (2000) Recurrent least squares support through discrete event simulation. Master’s thesis, Massachusetts
vector machines. IEEE Transactions on Circuits and Systems Part Institute of Technology
1 Fundamental Theory and Applications 47(7):1109–1114 Yuan R, Jing L (2018) Highlights from the Global Optimization Chal-
Tagaras G (1989) Effects of pooling on the optimization and service lenge. https://medium.com/jd-technology-blog/highlights-from-th
levels of two-location inventory systems. IIE Trans 21(3):250–257 e-global-optimization-challenge-17f4ad357337 (accessed Decem-
Taieb SB, Atiya AF (2016) A bias and variance analysis for multistep- ber 12, 2018)
ahead time series forecasting. IEEE Transactions on Neural Zeppieri MA (2004) Predictive capacity planning modeling with
Networks and Learning Systems 27(1):62–76 tactical and strategic applications. Master’s thesis, Massachusetts
Talluri K, Van Ryzin G (2004) Revenue management under a general Institute of Technology
discrete choice model of consumer behavior. Manag Sci 50(1):15– Zhang GP (2003) Time series forecasting using a hybrid ARIMA and
33 neural network model. Neurocomputing 50:159–175
Topkis DM (1968) Optimal ordering and rationing policies in a Zhang H, Chao X, Shi C (2017) Closing the gap: A learning algorithm
nonstationary dynamic inventory model with n demand classes. for the lost-sales inventory system with lead times. Working Paper
Manag Sci 15(3):160–176 Zhang Y, Gao J (2017) Assessing the performance of deep learning
Torabi SA, Hassini E, Jeihoonian M (2015) Fulfillment source algorithms for newsvendor problem. In: International Conference
allocation, inventory transshipment, and customer order transfer on Neural Information Processing, Springer, pp 912–921
in e-tailing. Transportation Research Part E:, Logistics and Zhang Y, Huang M, Hu X, Sun L (2018) Package consolidation
Transportation Review 79:128–144 approach to the split-order fulfillment problem of online
Tyralis H, Papacharalampous G (2017) Variable selection in time supermarkets. J Oper Res Soc 69(1):127–141
series forecasting using random forests. Algorithms 10(4):114 Zheng S (2010) Boosting based conditional quantile estimation for
Veinott AF Jr (1965) Optimal policy for a multi-product, dynamic, regression and binary classification. In: Mexican International
nonstationary inventory problem. Manag Sci 12(3):206–222 Conference on Artificial Intelligence, Springer, pp 67–79
Verhoef PC, Kannan PK, Inman JJ (2015) From multi-channel Zheng Z, Su D (2014) Short-term traffic volume forecasting: a
retailing to omni-channel retailing: introduction to the special k-nearest neighbor approach enhanced by constrained linearly
issue on multi-channel retailing. J Retail 91(2):174–181 sewing principle component algorithm. Transportation Research
Verity JW (1996) Clearing the cobwebs from the stockroom. Bus Week Part C:, Emerging Technologies 43:143–157
3498:140
Vulcano G, Van Ryzin G, Chaar W (2010) Choice-based revenue
management: an empirical study of estimation and optimization. Publisher’s note Springer Nature remains neutral with regard to
Manuf Serv Oper Manag 12(3):371–392 jurisdictional claims in published maps and institutional affiliations.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

onlineservice@springernature.com