International Journal of Forecasting 35 (2019) 170–180

Contents lists available at ScienceDirect

International Journal of Forecasting

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

Forecasting sales in the supply chain: Consumer analytics in

the big data era
∗
Tonya Boone a , Ram Ganeshan a , , Aditya Jain b , Nada R. Sanders c
a
Raymond A. Mason Mason School of Business, The College of William and Mary, United States
b
Zicklin School of Business, City University of New York, United States
c
D’Amor-McKim School of Business, Northeastern University, United States

article info a b s t r a c t

Keywords: Forecasts have traditionally served as the basis for planning and executing supply chain
Sales forecasting activities. Forecasts drive supply chain decisions, and they have become critically important
Predictive analytics due to increasing customer expectations, shortening lead times, and the need to manage
Literature review
scarce resources. Over the last ten years, advances in technology and data collection
Big data
systems have resulted in the generation of huge volumes of data on a wide variety of topics
and at great speed. This paper reviews the impact that this explosion of data is having on
product forecasting and how it is improving it. While much of this review will focus on
time series data, we will also explore how such data can be used to obtain insights into
consumer behavior, and the impact of such data on organizational forecasting.
© 2018 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

1. Introduction improve demand prediction and better execute supply

chain management. The intent of this paper is to provide a
Traditionally, forecasts have served as the basis for plan- review of the way in which this explosion of data is impact-
ning and executing supply chain activities such as sourcing, ing forecasting. For the purposes of this review, we focus
making, and distributing products and services to cus- on ‘‘consumer analytics’’ from a forecasting perspective,
tomers. Over the years, trends such as the globalization of encompassing a set of data-driven techniques that provide
supply chains, the explosion of product variety, the short- insights into consumer buying behaviors. The better a firm
ening of product life cycles and increasingly competitive understands its customers’ buying behaviors, the more
markets have made forecasting more complex, yet its role accurate its demand forecasts will be, which in turn helps
more critical. it to plan and execute supply chain operations more effi-
Digital technologies such as advanced Point-of-Sale ciently. The context of this review will be that of retailing,
(POS) systems, the ‘‘Internet of Things" (IoT), user- and we focus specifically on time series data. However,
generated content from social media, and cloud computing the ideas described (new data streams, personalization,
have enabled firms to collect vast amounts of data in real- new variables in forecasting, etc.) are not limited to the
time. We use the term ‘‘big data’’ to refer to data sets customer side of the supply chain, but can also be extended
that are large (large‘‘volume’’); collected in near real-time to inter- and intra-company interfaces.
(high ‘‘velocity’’); and present in a myriad of unstructured This paper is organized as follows. Section 2 offers an
forms (great ‘‘variety’’). The vast amounts of data that are overview of the sources from which vast quantities and
collected and analyzed in near real-time have the poten- new types of data are becoming available. Section 3 dis-
tial to improve our understanding of customer behavior, cusses existing research and potential future opportunities
in consumer analytics, specifically demand forecasting, us-
∗ Corresponding author. ing the data from these new sources. Section 4 then pro-
E-mail address: [email protected] (R. Ganeshan). vides an overview of the challenges of integrating big data

https://doi.org/10.1016/j.ijforecast.2018.09.003
0169-2070/© 2018 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180 171

into the sales and operations planning process, and finally, both the retail worker and the customer access to product
Section 5 presents a summary and our major conclusions. availability information and the ability to purchase in the
broader network of physical and virtual stores. While this
2. Sources of big data may make it difficult to forecast the demand at any partic-
ular store, ‘‘system-wide" service levels will improve.
For the purpose of exposition, we identify a three-stage
path that is followed by a customer when making a pur- 2.2. In-store path data
chase. In the first stage, the customer realizes their need
for the product and gathers information about the product. New in-store technologies are blurring the differences
In the second stage, the customer interacts with the firm between brick-and-mortar and online browsing and pur-
to gather information, and processes it in order to make chasing experiences. Data on customers’ browsing (‘‘path
a purchase decision. Then, in the third and final stage, the data’’) and purchase (intent to buy by adding to cart, aban-
customer makes a purchase. In the ensuing discussion, we doning the cart, etc.) behaviors that were once available
identify new sources of data that are now available on each only to online retailers are now being integrated into phys-
of these stages, after which Section 3 reviews the literature ical stores. Beacons that connect to customer smartphones
and identifies opportunities for research in each of these via bluetooth technology identify when a customer has
stages of the customer decision journey. entered the store. For example, Macy’s Shopkick App is
an example of one such proximity-based engagement. It
2.1. Point-of-sale data is triggered by beacons, and when combined with online
browsing experience and the customer’s position in the
Traditionally, POS systems have been used to collect
store, the store is able to engage customers by pushing con-
sales data which are then used as a baseline for forecasting
tent, providing ‘‘offers’’, and awarding prizes.4 Kroger, one
future demand. Gone are the days of independent ‘‘cash
of the largest grocers in the USA, is deploying digital shelf-
registers’’ and ‘‘credit card’’ readers: POS systems today
edges, where personalized information, including pricing,
provide a so-called ‘‘frictionless’’ checkout experience for
can be triggered as a customer walks by.5 Retailers such as
the customer and a connected data gathering system for
Marks and Spencer are experimenting with ‘Virtual Rails’,
the merchant. Self-checkout and minimally-assisted POS
where the store’s entire catalog is displayed electroni-
kiosks reduce lines and increase the checkout speed. POS
cally on large TV screens. They play related content, and
devices are now moving to mobile devices (either NFC-
also, as eCommerce customers are accustomed to, suggest
based payments or hardware-assisted such as Square1 ),
additions to the outfit when a customer scans a chosen
which is enabling a wider array of retailers to engage with
item. Augmented reality applications are used for virtual
their customers in multiple ways. For example, a customer
dressing rooms,6 where customers can ‘try on’ different
in an Apple, Inc. retail store can seek help from any re-
products and sizes. Traffic counters, infrared sensors and
tail staff member, and checkout right next to the product
video cameras can now track customer traffic and paths
using ApplePay, Apple’s payment app on a mobile device
through the store (Section 3.2 explores research in this
or a watch. Interestingly, even the POS device is simply a
area). They are used most commonly to reduce checkout
specialized iPhone. In the most extreme case, Amazon Go2
lines and schedule and deploy the workforce.
promises ‘‘No Lines. No Checkout" — payment is handled
by the Amazon Go App on the customer’s phone and sen-
sors detect the contents of the shopping cart. Finally, POS 2.3. User-generated content
systems are integrated across multiple selling channels,
making it possible to blur the differences between the User-generated content (UGC) commonly refers to data
physical and Internet stores. that is created by unpaid contributors. In our context, it can
The advantage of such modern systems is that they now be internet searches such as a Google search, which is avail-
connect the customer (via software applications or through able only on an aggregated basis; or it can refer to uploaded
loyalty initiatives) and the firm’s customer relationship Facebook posts, pictures, videos, testimonials, tweets, blog
management (CRM) system directly. Vast amounts of gran- posts, etc., where the identity of the contributor is explicitly
ular data indexed by customer are now readily available.3 available.
Such data include demographic profiles, order histories, UGC has become a critical factor in the product dis-
loyalty card information, coupon redemption rates, etc. A covery, research, and buy phases of consumer activity.
second advantage is that modern POS systems are also Increasingly, consumers are discovering new products and
connected to the inventory and warehouse systems, giving trends online, trusting and using online reviews and ratings

1 See squareup.com. 4 See http://time.com/money/3432693/macys-shopkick-ibeacon/.

2 See https://www.amazon.com/b?node=16008589011. 5 See for example https://www.wsj.com/articles/at-kroger-technolog
3 For example, Target, a large US retailer, uses what they call a ‘‘Guest y-is-changing-the-grocery-store-shopping-experience-1487646362, ac-
ID’’ to track individual customers. This ID is linked to the customer’s cessed December 7, 2017. See also the related YouTube video: https:
browsing and purchase behavior, as well as to their demographic infor- //www.youtube.com/watch?v=w2vvcz-fki4, accessed December 7, 2017.
mation. See: http://www.nytimes.com/2012/02/19/magazine/shopping- 6 Gap Inc., is experimenting with the DressingRoom by Gap
habits.html?pagewanted=1&_r=1&hp. Also see Feng and Shanthikumar augmented reality-based virtual dressing room. See https://adressed.
(2018) for examples of ways in which such real time data can be used gapinc.com/blog/gap-ces-announcement-2017-dressingroom-app,
to build forecasting models. accessed July 20, 2017.
172 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180

of products and services from peers, and engaging more complex for practical implementation. In recent years, re-
with brands in the presence of UGC. searchers have proposed various data-driven approaches
Firms are taking ‘‘digital media" strategies that make in order to counter the problem of demand censoring and
use of UGC — search, social media, mobile, and email mar- thus improve forecasting. These approaches are based on
keting — seriously. According to one estimate,7 the digital the higher level of granularity in sales and inventory data,
marketing ‘spend’ for US companies will be around $120 and new forecasting models that can be applied to such
billion by 2021, representing 46% of all marketing expen- data. Modern Point-of-Sales systems have the ability to
ditures. Firms are spending significant sums of money to capture detailed data on both the timings of sales and the
position themselves in search engine searches; and are availability of various products in inventory at different
investing in exploring how best to use social media plat- points in time. Combining this data allows the firm to see
forms in conjunction with their most valued consumers sales as a function of the product availability over time,
(fans, influencers, etc.) to ‘spread the word’, not only to thus enabling us to obtain a better picture of demand. In a
inform but to persuade potential consumers. There are recent paper, Jain, Rudi, and Wang (2015) offered a simple
several case studies in the literature on such efforts that illustration of improvements in forecasting from the use
have been successful (for example see Goh, Heng, & Lin, of such granular data. They show that firms can improve
2013 and Kumar, Bhaskaran, Mirchandani, & Shah, 2013); their forecasting and inventory decisions significantly by
however, many of these studies do not focus on SKU-level using the data on the timing of sales. They also show that,
forecasting, creating a need for research in this area. in many cases, it is sufficient for the firm to capture only
the time at which a stock-out takes place. Using the timing
3. Opportunities for consumer analytics and forecasting of stock-outs allows better estimates of actual sales to be
made, improving forecasts. Queenan, Ferguson, Higbie, and
This section discusses existing and potential advances Kapoor (2007) offer more practical applications of this idea
in consumer analytics and demand forecasting that are to the forecasting of demand in the context of revenue
achieved using the data collected through new sources. management.
This idea of using more granular data on sales and
3.1. Point-of-sales data inventory over time has even stronger implications for
the development of multi-product demand forecasts. Most
The data used most often in time series based fore- firms worry about having both the right assortment and
casting are historical data on sales. The obvious attraction the appropriate inventory for each of the products in the
of such data is that they are available readily from the assortment. Solving this problem requires a knowledge of
most rudimentary of point-of-sales systems. Time series the demands for different products as a function of the
methods are typically used for estimating patterns in past assortment offered — effectively, an understanding of cus-
sales data, which are then extrapolated for forecasting tomers’ inherent demands for different products and their
future demand. One of the key issues with this method of willingness to substitute between them. The availability
forecasting is the bias that results from demand censoring: of more granular data can lead to a great improvement
firms do not record information on demand during stock- in the way in which sales data can be used for learning
out, meaning that information on the demand in excess of about customers’ choices between multiple products in an
inventory carried is simply lost. Such demand censoring assortment. Using inventory data permits one to deduce
leads to forecasts which are biased downwards, resulting the availability of the product assortment over time, and
in poorer inventory decisions, which further increase the marrying this with the sales transaction data gives a more
extent of censoring. The problem is exacerbated further complete picture of a customer’s choice process. The multi-
when the firm is selling multiple products and stock-outs product forecasting method proposed by Karabati, Tan, and
may result in substitutions. In such cases, the firm may end Öztürk (2009) uses this idea: in their proposed scheme,
up underestimating the demand for fast-selling products the sales data are split into different time intervals, each
which stock-out often, and overestimating the demand for of which corresponds to the availability to customers of
slow-selling products which do not sell out, but may have a different product assortment. These disaggregated data
higher sales due to substitution. are then used to estimate the demand rates for individual
A number of research papers have highlighted the prob- products, as well as the substitution probabilities between
lem of demand censoring and its adverse effects not only them. More recent papers have applied this idea of using
on demand forecasting (Wecker, 1978), but also on the granular sales data to different types of customer choice
estimation of important consumer characteristics such as models in an attempt to develop a better understanding of
price elasticity (Bruno & Vilcassim, 2008), as well as on the customer choice process; for example, Musalem, Oli-
making important inventory decisions (Tan & Karabati, vares, Bradlow, Terwiesch, and Corsten (2010) consider a
2004). Researchers in the area of operations management multinomial logit (MNL) choice model, van Ryzin and Vul-
have proposed a number of advanced econometric meth- cano (2014) consider a more generalized rank-based choice
ods for correcting this problem, but most have been too model, and Chen, Mersereau, and Wang (2017) show how
granular sales data can improve pre-launch merchandise
7 See the report from Forrester, a leading research and advisory testing significantly.
firm: http://blogs.forrester.com/shar_vanboskirk/17-01-24-us_digital_
In addition to increasing the granularity of data over
marketing_spend_will_near_120_billion_by_2021, accessed July 16, time, the modern Point-of-Sales and data storage systems
2017. also allow firms to observe the contents of a customer’s
 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180 173

basket; in other words, which products were bought to- elements and product demand/sales should be explored in
gether and how often. In many contexts (for example gro- future research. Third, the existing research streams are
cery retail), customers generally buy multiple products in based mainly on the number of customers entering the
a single visit, and the availability (or not) of one product store, with no attention being paid to either the amount of
may influence the purchase decisions of other products in time customers spend in the store or the number of people
the customers’ baskets. Researchers in marketing have rec- in the store. It is well known that the number of people in a
ognized this in their work on the development of models store affects the purchase decision – crowds hinder prod-
of customer choice (Chung & Rao, 2003; Hruschka, 2017; uct browsing and access to sales persons, and increase the
Russell & Petersen, 2000), but the application of such anal- ‘friction’ in completing sales transactions. Future research
yses to the forecasting of product-level demand remains an should consider these additional operational dimensions of
open avenue for future research. customer flow.

3.2. In-store data 3.2.2. Path data

Traffic data offer information only on how many cus-
Next, we consider other sources of in-store data related tomers enter the store. After entering the store, customers
to customer behavior. These are data related to customers’ engage with various elements in the store, such as products
entry into the store and subsequent interactions with vari- on shelves and sales personnel. The ways in which cus-
ous store elements before making their purchase decision. tomers go about interacting with these store elements pro-
vide information about their decision making processes,
3.2.1. Traffic counter data
and hence, the demand. Traditionally, such data have been
In recent years, it has become much easier for firms
difficult to gather; however, technologies such as RFID
to gather traffic data, i.e., information on how many cus-
tags, digital video capture, and nimble database solutions
tomers visited their stores at different times. First, many re-
have made collecting, storing, coding and accessing such
tailers have ventured into electronic retailing, where such
data in a brick-and-mortal retailing context easier. In the
data are a lot easier to collect and analyze. Second, elec-
context of electronic retailing, such data are available read-
tronic traffic counters have become more pervasive and
ily, since store websites usually track each visitor’s digital
accurate. While many of the applications of traffic data re-
clickstream. Such data can be used to assess customer
late to tactical decisions (e.g. store-level staffing decisions),
intent, which can then be used for forecasting demand. For
firms are beginning to realize the value of this informa-
example, an electronic retail store can use customer path
tion for other operational decisions (out-of-store and in-
information to detect customer interest, and to position
store promotions, salesperson training, store layout design
inventory in (warehouse) locations closer to the customer
etc.). Traffic data inform the retailer as to how effective
for fast delivery when such orders materialize.
their store front and other out-of-store promotions are in
Over the last decade, a substantial amount of research
attracting customers to step into their store. When com-
bined with sales records, the potential of this data grows has utilized path data collected using these novel meth-
many-fold. By combining traffic data with sales data, a ods and sources. The main focus of such papers has usu-
firm can also learn about trends in conversion (i.e., what ally been on obtaining a better understanding of customer
fraction of visiting customers choose to purchase), which behavior; specifically, the ways in which customers in-
in turn can inform it about the effectiveness of different teract with various elements on the physical or digital
in-store elements in driving conversion. Research papers store. For example, Hui, Bradlow, and Fader (2009) and Hui,
on this subject (see for example Lam, Vandenbosch, & Fader, and Bradlow (2009) use data collected using RFID
Pearce, 1998; Mani, Kesavan, & Swaminathan, 2015 and tags installed on shopping carts/baskets to verify behav-
Perdikaki, Kesavan, & Swaminathan, 2012) have shown ioral hypotheses on customers’ purchase processes. They
how sales can be predicted as a function of traffic and show that customers are more likely to make a purchase
store staffing levels. These models can then be used not when they spend a longer time shopping. In another ef-
only for the improved forecasting of aggregate sales (given fort, Lu, Musalem, Olivares, and Schilkrut (2013) measure
some information on future traffic), but also for optimizing customers’ sensitivity to waiting in a line. Jain, Misra, and
staffing levels. Another application of traffic data is offered Rudi (2016) use video data to quantify the impact of sales
by Lam, Vandenbosch, Hulland, and Pearce (2001), who assistance on purchase decisions. However, the use of data
show the usefulness of such data for evaluating the impacts from such sources for improving demand forecasting has
of different types of promotions. received only a limited amount of attention. Path data are
There are several possible avenues for future research in ‘‘short-term", and typically are not suitable for planning
this area, especially relating to understanding the demand typical supply chain procurement or replenishment, ac-
at brick-and-mortar retailer stores. First, while tivities with lead times that can range from a few days
researchers have constructed models that relate sales to to several months. In practice, retailers need to combine
traffic, it is not clear how much improvement in accuracy path data with data from other sources in order to predict
can be achieved by incorporating traffic data. Second, the demand (say through econometric models). However, such
existing research is also based on highly aggregate (store- sophisticated and time-consuming tools may not be acces-
level) data, and does little to differentiate between differ- sible to smaller retailers. Thus, future work could focus on
ent types of customers entering the store or different types the best way of integrating such data into existing easy-to-
of products being sold. The relationships between these employ forecasting methods.
174 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180

3.3. Internet/user-generated content Trend data have also been shown to be valuable in
industries where having real-time data is critical. In finan-
Last but not least, we consider the impact of data on cial markets, search query data are used for measuring
a customer’s behavior prior to a store visit. The key idea retail investor attention (Bank, Larch, & Peter, 2011; Da,
here is that such customer activities are indicative of an Engelberg, & Gao, 2011) and market volatility (Dimpfl &
intent to purchase a given product, and can be used to Jank, 2016), and for predicting earnings (Da, Engelberg, &
improve the forecasting process. Traditionally, such data Gao, 2014; Drake, Roulstone, & Thornock, 2012). Trend-
have been difficult to obtain, barring expensive market enhanced models have been used in estimating customer
surveys. However, the explosion of online activities over ‘‘traffic’’, for example in tourism for predicting arrivals
the last decade, especially the emergence of ‘‘social media’’, (Bangwayo-Skeete & Skeete, 2015; Choi & Varian, 2009),
and firms’ abilities track them, has changed this. Nowa- hotel bookings (Pan, Chenguang Wu, & Song, 2012; Rivera,
days, firms have access to many tools that allow them not 2016), and movie admissions (Hand & Judge, 2012). The
only to see the latest trends, but also to shape them in MAPE improvements in many of these industry-specific
order to influence customer behavior. Most firms maintain models are between 10% and 40%, raising the possibility
a presence on social media platforms such as Facebook and that trend data can enhance prediction.
Twitter, and can use them to interact with their customers However, research on the application of Google trends
directly. data to the prediction of product sales is very limited.
The few studies that have tackled this challenge have
focused on the forecasting of aggregate-level sales (Choi
3.3.1. Google trends
& Varian, 2009; Fantazzini & Toktamysova, 2015; Nun-
Even without an explicit social media presence, though,
nari & Nunnari, 2017; Schaer, Kourentzes, & Fildes, 2019),
there are technologies available today that can track cus-
and have shown modest improvements in MAPE (2%–
tomer search behaviors, i.e., patterns of what customers
15%) as a result of using Google trends data for enhanc-
search for on the web, and how often they do it. There
ing forecasting. We found two studies that investigated
is a growing stream of literature on the use of Google whether Google trends data can improve SKU-level fore-
Trends, a free tool available from Google, Inc., that tracks casts. Through a case study of a food retailer, Boone, Gane-
customer search data. The trends are reported on a 1–100 shan, and Hicks (2015) show that Google Trends improve
scale, where a value of 100 indicates peak popularity of a in-sample performances by 6%–8%. Later, using five years of
search term whereas 50 indicates that it is half of its peak data over five SKUs in multiple categories from a specialty
value. For example, the popularity of the search term ‘‘L. L. food and cookware retailer, (Boone et al., 2017) show out-
Bean’’ on January 11, 2017, was at a value of 9; but when of-sample MAPE improvements of 2%–8%, depending on the
President Trump tweeted ‘‘Buy L. L. Bean’’ on January 12, SKU category.8
the popularity of ‘‘L. L. Bean’’ surged to 100, its peak. While there is considerable enthusiasm regarding the
A growing stream of research is finding evidence that potential for Google trends data to improve forecasting,
the use of Google trends indexes (and those of competing a lot of work is needed in order to realize this poten-
search engines like Baidu) reduces forecast errors, both in- tial. Following the example of Boone et al. (2017), there
and out-of-sample. The premise is that smaller forecast need to be more studies testing the out-of-sample fore-
errors lead to better supply chain performance (see Boone, cast improvements that can result from the use of search
Ganeshan, Hicks, & Sanders, 2017; Cui, Gallino, Moreno, data. Second, there is a need for systematic protocols for
& Zhang, 2017). Researchers typically use traditional time generating relevant search terms for a product. There are
series forecasts, most often with some form of autoregres- potentially thousands of terms that may be correlated with
sive model as a baseline. They then include the search target variables — how does one choose which ones to use?
index for specified search terms as a predictor variable In existing studies, researchers have relied on manager
when constructing the trend-enhanced forecasting model. intuition (Boone et al., 2017) or used part of a portfolio of
The procedure for testing the value of trend variables terms that are likely to be related to the product. How-
is to compare the out-of-sample forecast errors of these ever, these are not scalable for firms that sell thousands
two models using a rolling window forecast, most com- of products. Third, Google only indexes terms that meet
monly with one-step-ahead forecasts. Existing research a certain threshold of traffic, meaning that index values
has shown that Google trends can be used to improve the for relevant search terms may simply not be available. In
forecasting of economic indicators such as unemployment such situations, the researcher has to rely on proxy terms,
rates and benefit claims (Choi & Varian, 2009, 2012; Smith, which adds another source of uncertainty. Fourth, the way
2016), GDP (Castle, Fawcett, & Hendry, 2009), and auto- in which Google or other search engines calculate the index
mobile, home and retail sales (Choi & Varian, 2009; Wu is a black box. According to Google, the index is based on
& Brynjolfsson, 2015). The Google trends-enhanced eco- sampled data, meaning that they are not entirely consistent
nomic models typically have out-of-sample MAEs that are over time, calling into question the long-term reliability of
anywhere from 4% to 25% smaller than those of the baseline the use of trend variables. Fifth, search terms may vary with
models. Google trend-enhanced models are also common the product lifecycle or with time even for the same target
in public health, often for tracking incidences or the spread variables; future studies can perhaps address and gain
of disease outbreaks (for example, see Ginsberg et al., 2009,
for predicting flu outbreaks and Seifter, Schwarzwalder, 8 While only five SKUs were reported, the authors found similar results
Geis, & Aucott, 2010, for tracking Lyme disease). for over 15 SKUs.
 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180 175

insights from the ways in which changes in search terms errors can be improved by adding social media informa-
impact demand forecasting. Finally, as was also pointed tion, but we have not found any studies that investigate
out by Schaer et al. (2019), such models perform poorly how social media platforms impact SKU-level forecasts.
in forecasting demand over the longer time frames (even Using data from an online retailer, Cui et al. (2017) show
a few days) that are required for supply chain planning. how the number and relevance of Facebook entries reduce
Since Google trends use contemporaneous terms, dynamic the out-of-sample MAPEs of aggregate daily sales for an
forecasting quickly gets very noisy, even over only a few online fashion retailer by 7%–23%. However, they concede
days, which limits these models to very short forecasting that their forecasts hold only for short timeframes. In a
horizons. Addressing the ways in which these models can controlled study with a premium ice cream brand, Kumar
be used strategically with the sales and operations plan- et al. (2013) were able to show that a marketing campaign
ning process is another a viable area for future research. that is constructed carefully to make use of Facebook can
grow sales. However, their study focused on how to spread
3.3.2. Social media the ‘‘word of mouth" about the product, not specifically on
There is a growing stream of literature on ways of har- forecasting SKU demand.
While models that incorporate user-generated content
nessing the data created on social media (Facebook, Twit-
show promise for generating sales and revenue, forecasting
ter, YouTube, etc.) in order to provide forecasting insights.
models specifically have a number of practical limitations
The premise is that as more people share their experiences
in regard to supply chain planning. First, the forecast hori-
on these platforms, often in real time, it can not only inform
zons of models that use such data are often short. In a sup-
the consumers of this data, but also shape how they make
ply chain context, where lead times tend to be longer than
decisions. From a technical perspective, the primary focus
just a few days, short-term search or social media data in-
of most of the academic studies in this arena has been on formation cannot be used for demand and inventory plan-
establishing how such social media entries can be captured ning. However, these models can be useful for the staging
and processed into quantitative measures or indices (that of inventory as it is consumed. For example, as a trend or a
capture the ‘‘mood" or ‘‘sentiment" of these posts). There ‘buzz’ is detected for a certain SKU in a certain area, contin-
is a wide range of views as to how these indices should gency plans can be made to move the inventory to that area
be constructed, including ‘‘volume" measures such as the in order to maximize profits. Second, such models can also
number of posts,9 the fraction of positive posts, who is be used to make replenishment decisions for consumable
posting, the posters’ ratings, the number of upvotes, etc.; goods. For example, as the interest in a particular SKU (as
and ‘‘valence" measures, which relate to the context and opposed to actual inventory) grows, actions to replenish
quality of posts, such as the subject of the post, the ‘‘bag of can be taken earlier than normal in anticipation of that
words" that are used in the post, the credibility and influ- demand. In our opinion, these Internet search and social
ence of the poster, the context in which the post originated, media variables have their highest impact on short-term
etc. (see Goh et al., 2013, for a review). The aim is to capture ‘yield’ management scenarios. Connected stores can be fed
the mood related to the topic; or, in a marketing campaign, digital signage based on current sales and trends. These can
to determine how product information can be diffused work in conjunction with beacons and customer devices
effectively using social media. For example, a potential to identify customers and offer them promotions such as
index for stock market predictions could be the number coupons to stimulate sales of selected SKUs or trending
of positive tweets about Dow or Nasdaq. The creation of items. Flexible digital price tags can change based on the
these indices is in itself a vibrant area of research. While evolution of these contemporaneous Internet variables,
reviewing such indices is outside the scope of this paper, it giving customers a price break based on their browsing and
provides a rich area of study for the forecasting community. buying patterns. Customers can also be given the choice of
The second step in assessing user-generated content is procuring products through other channels, such as other
to augment forecasting models with these social-media store locations by means of eCommerce, for products that
measures or indices in order to test whether they improve are unavailable in the store they visit. However, making use
the forecasts. In addition to linear models, a significant of contemporaneous Internet and user-generated content
number of research studies that use social media use non- depends on how well these variables are integrated into
linear models, typically machine learning methods (with the sales and operations planning and execution processes.
support vector machines and neural networks being those
4. Organizational challenges of big data forecasting
most commonly used).
Twitter data have been used widely for predicting flu 4.1. Integrating big data into sales and operations processes
outbreaks (Broniatowski, Paul, & Dredze, 2013; Lamb, Paul,
& Dredze, 2013; Lampos, De Bie, & Cristianini, 2010), stock On a strategic level, every firm has to decide whether
prices (Bollen, Mao, & Zeng, 2011; Rao & Srivastava, 2012; and how much to incorporate big data technologies into
Zhang, Fuehres, & Gloor, 2011), box office revenue (Liu, their planning process. This depends on the relative ben-
Ding, Chen, Chen, & Guo, 2016; Mishne, Glance, et al., efits the firm can potentially accrue against the cost of
2006; Sanguinet, 2016), and TV ratings (Wakamiya, Lee, collecting and analyzing such data. Our intent is not to
& Sumiya, 2011). Most studies have shown that forecast address this strategic question of whether and how much
to integrate big data into organizational processes; rather,
9 We use the word ‘‘post" in its very generic form. This could include a we want to highlight the changing supply chain landscape
tweet or a retweet, or a Facebook entry or a like, or a comment on a public and the challenges of integrating big data specifically for
forum, etc. demand and supply chain planning.
176 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180

4.1.1. Capturing big data and connecting it to traditional SOP 4.1.3. Changing customer experiences
processes The widespread use of connected devices such as smart-
Big data brings with it the potential to improve product phones or wearable technologies are changing the in-store
forecasts and give interesting insights into customer be- customer experience. Beacons in stores (or online) can
haviors. However, these potential benefits come with enor- detect a customer entering a store, thus providing signif-
mous practical challenges for demand planners. First, the icant opportunities for customer engagement, gathering
sheer volume of data can be overwhelming. For example, data and making links to historical buying patterns. As a
Walmart collects more than 2.5 petabytes (1 petabyte = 1 result, demand is not just forecast, but can be influenced
million gigabytes) of data every hour from one million cus- and modified directly, as consumer buying can now be
tomer transactions.10 However, only about 0.5% of all data directed (see ; Cohen, 2018; Feng & Shanthikumar, 2018;
collected is analyzed.11 Thus, a practical question is, what and the various references within). Content such as product
data should be stored, and for how long? Second, Feng and information, pricing, and promotions such as e-coupons
Shanthikumar (2018) point out that, while ‘‘theoretically can be personalized to the customer (and sent to their
more information leads to better forecasts, the challenge, phones) based on their past browsing experience (Gane-
however, comes from dealing with the increased number
shan, 2014). Products or substitutes can be suggested, and
of variables and their ambiguous relationships.’’ Since big
in some cases, customers can be led to product locations.
datasets, especially those used in forecasting, tend to be
This is an area of active research in the marketing and
sparse and non-repetitive, they posit that semi- or non-
operations management communities (see for example Ca-
parametric methods such as machine learning (see for ex-
chon, Daniels, & Lobel, 2017; Chen, Ma, Simchi-Levi, & Xin,
ample Cui et al., 2017) are better suited for analyzing them
than traditional time series forecasting methods. Demand 2016, for pricing; van Heerde & Neslin, 2017, for promo-
planners need to adjust so as to accommodate a wider tions; Bobadilla, Ortega, Hernando, & Gutiérrez, 2013, for
range of methodologies into the planning process. recommendation systems; and Lau, Zhang, & Xu, 2017, for
sentiment analysis).
4.1.2. From human judgement to data-driven decisions From a customer’s perspective, smart devices are help-
Studies of industry practices repeatedly show that judg- ful for researching a product (for example, looking for
mental forecasts are the norm and are based on ‘‘gut feel- reviews), looking up a competitor’s price or even ordering a
ings.’’ Here, statistical forecasts are adjusted to account competing product right from the store. In many ways, the
for numerous factors that demand planners perceive as customer experience has become a very personalized one.
being difficult to measure, including promotional activ- Demand planners, and indeed the research community,
ities, seasonal activities, demand risk, demand and sup- need to address this apparent contradiction of disaggregate
ply chain disruptions (see for example Fildes & Goodwin, planning for the unique customer at significantly shorter
2007). While human judgement has the potential to im- time horizons while at the same time planning for aggre-
prove forecasts, experts often introduce their own biases gate needs over the longer-term. Second, customers are
in the process (see for example Hypotheses H4 and H5 now demanding the flexibility of being able to order the
of Fildes, Goodwin, Lawrence, & Nikolopoulos, 2009). Of- product via multiple channels (the ‘‘omni-channel’’ expe-
ten, such judgmental forecasts have an adverse impact on rience), and want product delivery (and potential returns)
the forecast accuracy, since they weight their contribution in a wide variety of locations, including their home, stores,
disproportionately (Franses & Legerstee, 2010). At least in or third-party locations. While big data technologies have
theory, newly available data have the potential to reduce enabled and enhanced the planning between channels, it
the negative effects of ‘‘adjustments’’ to the forecast (see still requires a significant amount of coordination between
the prototypical models of Feng & Shanthikumar, 2018), channels (eCommerce, retail, etc.), in terms not only of
and there have been efforts to integrate such new data
tactical planning, but also of shared strategic goals, perfor-
streams (Sagaert, Aghezzaf, Kourentzes, & Desmet, 2018)
mance measures, and incentives. Indeed, forecasting omni-
into forecasts in a systemic manner. However, significant
channel demand remains a viable area of enquiry.
practical challenges remain. First, recent surveys (Weller &
Crone, 2012) indicate that such new data streams may not
4.1.4. Integrating the connected supply chain
be available readily, or be used by a majority of demand
planners. Second, significant hardware, software, and an- Big data technologies have enabled the echelons of the
alytical support is required in order to integrate the data supply chain to be ‘‘connected’’ and to ‘‘communicate’’
into the ERP systems that planners typically use. Third, a with one another, creating the ‘‘intelligent’’ supply chain
significant learning curve is required in order to interpret Sanders (2016). At the customer level, there has been
the results from the newer machine learning algorithms. a profusion of ‘‘connected" devices. This has enabled a
Another area of research for the forecasting community is more active engagement with the customer. Targeted mes-
to investigate whether big data streams can substitute for sages/promotions can reach the customer’s living room;
expert judgement. While we do not believe that big data consumption can be tracked to actuate replenishment;
can ever supplant expert opinion, it can alleviate some of problems can be diagnosed online; and finally, specialized
the issues with judgmental forecasts. products and services can be tailored to the customer.
Such connectivity provides a treasure trove of data that
10 see https://www.dezyre.com/article/how-big-data-analysis-
the demand planner can potentially use to capture the cus-
helped-increase-walmarts-sales-turnover/109. tomers’ intent, to shape demand, and to increase loyalty.
11 https://www.technologyreview.com/s/514346/the-data-made-me- It can also provide a greater supply chain responsiveness
do-it/. to forecast errors. However, in our experience, demand
 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180 177

planners seldom use this data for the S & OP process. The customer records — names, addresses, and social security
integration of customer engagement provides a viable field numbers — were hacked in this breach. At best, the cost
of enquiry for the forecasting community. to the consumer is the cost of protecting their identity; at
Initiatives on the supply side, such as collaborative plan- worst, it is the cost of reversing the consequences of their
ning, forecasting and replenishment (CPFR) and efficient identity being stolen.
customer response (ECR), are setting up processes where Big data are also not immune to the typical problems
demand planning and replenishment is done in collabo- faced by traditional datasets, such as that of bias and rep-
ration with key suppliers. Events such as store openings, resentation in datasets. The device or sensor collecting the
promotions, and supply shortages are shared between data can introduce bias into the data. For example, the
trading partners, improving visibility and forecast accu- City of Boston used the ‘‘Street Bump’’ App15 to collect
racy. In many cases, firms also extend this visibility to the data from citizens on potholes. The user’s smartphone’s
customer (Boone & Ganeshan, 2008; Panahifar, Heavey, accelerometer would record ‘‘bumps’’ and their locations
Byrne, & Fazlollahtabar, 2015), especially in regard to in- as they drove along on the road, a potential sign of a
ventory availability. RFID tags help to improve inventory pothole that the city could then fix. The app reported
tracking and maintain consistent service levels (Bertolini, disproportionate numbers of potholes in wealthier parts
Bottani, Rizzi, & Volpi, 2010). Emerging technologies such of town where more residents owned smartphones and
as Blockchains promise secure, accurate transactions and were digitally engaged. Thus, even such well-intentioned
easy reporting. They also make ‘‘track and trace" in the programs have the ability to have a negative impact on
supply chain easier, providing the demand planner with members of our society.
valuable information on planning replenishment and Another famous example is the overestimation of flu
order-promise dates. Finally, instrumented supply chains incidence in algorithms that used Google Trends search
have improved yield management (for example by using terms, due to people beginning searching for flu symptoms
dynamic price tags) and supply chain efficiency. The chal- when they heard of flu incidences on the news. The CDC
lenge for demand planners is how these new technologies data were field-based and tended to estimate flu incidences
should be integrated into the planning process. For exam- better. These are examples of so-called ‘‘big data hubris’’,
ple, should Blockchain technology be rolled out to all prod- the implicit assumption that big data collected via auto-
ucts? Should they embrace smart contracts? How should mated or volunteered methods can replace traditional data
they establish the chain of command? Answering these collection and analysis (Lazer, Kennedy, King, & Vespig-
questions successfully will enable an improved visibility nani, 2014).
and potentially better forecasts.
4.2.1. Algorithmic ethics and injustice
4.2. Privacy, bias, and discrimination For a recent job advertisement in finance located in
Washington DC, Verizon, a US telecom company, targeted
Connected devices and sensors are constantly collecting their promotion at ‘‘the Facebook feeds of users 25 to
significant amounts of data on customers. Often, such data 36 years old who lived in the nation’s capital, or had re-
(such as mobile phone pings, for example) do not have any cently visited there, and had demonstrated an interest in
specific purpose. In addition, the data are not anonymous: finance.’’16 Such targeted ads raise questions of fairness
often the identity of the consumer is linked to the behavior. to older workers, and many critics suggested that the ad
Since the forecasting of consumer behavior often warrants violated the federal Age Discrimination in Employment Act
the use of these private datasets, it raises significant issues of 1967. Such biases by algorithms based on demographic
of privacy and data integrity, as well as bias and discrim- features (race17 and disability18 are also cited commonly)
ination, as the ensuing discussion will show. These issues have come to be known as ‘‘machine bias’’. O’Neil’s (2017)
are increasing in prominence,12 and the forecasting com- best-selling book Weapons of Math Destruction provides
munity needs to be actively engaged in finding solutions. multiple examples of such algorithmic injustice — poor
The Identity Theft Resource Center estimates13 that evaluation procedures for high school teachers, race, gen-
there have been 8190 data breaches since 2005, with over der and economic biases in product offerings, etc. — and
one billion customer records having been exposed to theft. these algorithmic decisions often affect the most vulnera-
ble populations.
These breaches can have disastrous consequences, both
for the firms whose data were hacked and potentially for
the consumers whose identities were exposed.14 In 2017, 15 John Podesta, President Barack Obama’s senior counselor, references

the largest breach was Equifax, a credit monitoring agency this bias in an interview. See https://www.reuters.com/article/us-usa-
obama-privacy/white-house-looks-at-how-big-data-can-discriminate-
in the US. Over 140 million US and half a million British
idUSBREA3Q00M20140427.
16 https://www.propublica.org/article/facebook-ads-age-
12 Not only are companies, especially retailers, reacting to many discrimination-targeting.
breaches and the algorithmic bias discussed in this section, policy mak- 17 Propublica reported that Facebook was allowing housing advertisers
ers are also taking steps to legislate against privacy- and algorithmic- to exclude viewers by race. See: https://www.propublica.org/article/
injustice. See: https://www.cnet.com/news/elizabeth-warren-equifax- facebook-advertising-discrimination-housing-race-sex-national-origin.
mark-warner-credit-reporting-agencies-data-breach-bill-fines/. Amazon also came under criticism when it rolled out Amazon same
13 https://www.idtheftcenter.org/Data-Breaches/data-breaches. day deliveries in cities, as their algorithms inadvertently excluded black
14 See for example the story of Drew Armstrong, whose iden- neighborhoods.
tity was stolen: https://www.bloomberg.com/news/articles/2017-09-13/ 18 O’Neil (2017) describes how job applicants were weeded out based
my-three-years-in-identity-theft-hell. on mental health.
178 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180

Several research streams tackle the issue of machine a security risk (Lu, Zhu, Liu, Liu, & Shao, 2014). Any failure
bias (see for example Calders & Verwer, 2010). These are to keep customer data, especially sensitive data (such as
typically statistical methods that indicate potential biases social security and credit card numbers) secure can be il-
among customer attributes. legal and open to relevant government oversight agencies.
As the use of big data streams grows increasingly Finally, firms that make use of big data also need a clear
widespread in the forecasting research community, we big data strategy in order to keep their decisions free of
expect this to become a important research area. The fore- bias. For example, a multidisciplinary group of researchers
casting research community, with its expertise in statis- known as the Fairness, Accountability and Transparency in
tical methods, can contribute significantly to (a) fraud Machine Learning (FAT/ML) community lay out a set of five
detection (i.e., detecting outliers in data streams) in big principles that can potentially alleviate issues of bias: re-
data; (b) determining how ‘‘clean’’19 internet and user- sponsibility, explainability, accuracy, auditability, and fair-
generated content can be used in forecasting models; and ness. The aim of these principles is to ‘‘help developers
(c) establishing how big data or machine algorithms can and product managers design and implement algorithmic
be bias-free (for example, generating ‘‘bias" scores for systems in publicly accountable ways. Accountability in
forecasting models that predict customer behavior). this context includes an obligation to report, explain, or
justify algorithmic decision-making as well as mitigate any
5. Summary and conclusions negative social impacts or potential harms’’.20
A clearly articulated governance structure will lay out
Our intent in this paper has been to provide an overview what data are collected where and from whom; the way in
of the way in which the forecasting of sales in the supply which these data and the accompanying decisions’ quality
chain could be enhanced by customer analytics based on are validated; and how this data can be leveraged as an
big data and associated technologies. We have focused organizational asset (see for example Otto, 2011, Hashem
rather broadly on three aspects of the purchase decision: et al., 2015).
the research phase, the store experience, and finally the Despite the challenges, we see the next decade as being
sales transaction. From this perspective, we have reviewed the golden age of forecasting. We see a future where de-
the many ways in which big data can enhance aggregate mand planners can make use of large and varied datasets
forecasts. First, we show how the granular data that are in real-time; can provide a clearer picture of customer
now available through POS systems can help to overcome behavior; and can generate accurate forecasts that will help
demand censuring and provide a holistic view encompass- not only to improve the efficiency of their supply chains but
ing a multi-product perspective. Second, in-store technolo- also to enhance their revenues. We hope that our optimism
gies can be used to assess traffic and customer purchasing is shared by the forecasting community — we are looking
behaviors, which can inform aggregate sales forecasts. Fi- forward to the increased interest in big data technologies
nally, we show how researchers have used user-generated and the development of new methods and tools for turning
data from Internet searches and social media to enhance such data into more accurate forecasts for use in making
forecasting models. actionable operational decisions.
While the use of big data shows promise, it also intro-
duces significant challenges. First, the size and unstruc- References
tured content of these enormous data sets can be daunting.
Bangwayo-Skeete, P. F., & Skeete, R. W. (2015). Can Google data improve
Second, the data are sparse and non-repetitive and requires the forecasting performance of tourist arrivals? Mixed-data sampling
multiple methodologies for forecasting, which slows their approach. Tourism Management, 46, 454–464.
widespread acceptance. Finally, customer experiences and Bank, M., Larch, M., & Peter, G. (2011). Google search volume and its
supply chains are changing, making big data technologies influence on liquidity and returns of German stocks. Financial Markets
and Portfolio Management, 25(3), 239–264.
difficult to integrate into conventional SOP processes. We
Bertolini, M., Bottani, E., Rizzi, A., & Volpi, A. (2010). The benefits of RFID
concur with the assessment of the Associate Editor who and EPC in the supply chain: Lessons from an Italian pilot study. In The
reviewed this paper: ‘‘As yet, in demand forecasting we internet of things (pp. 293–302). Springer.
would see ‘big data’ mostly as potential with gains yet to Bobadilla, J., Ortega, F., Hernando, A., & Gutiérrez, A. (2013). Recommender
systems survey. Knowledge-based Systems, 46, 109–132.
be delivered’’.
Bollen, J., Mao, H., & Zeng, X. (2011). Twitter mood predicts the stock
Firms that embrace big data technologies must also market. Journal of Computer Science, 2(1), 1–8.
address three key related management issues: privacy, Boone, T., & Ganeshan, R. (2008). The value of information sharing in
security, and governance. First, since a multitude of sensors the retail supply chain: two case studies. Foresight: The International
are used to gather information about customer browsing Journal of Applied Forecasting, 9, 12–17.
Boone, T., Ganeshan, R., & Hicks, R. L. (2015). Incorporating Google trends
and purchase behaviors, it is imperative that firms ad- data into sales forecasting. Foresight: The International Journal of Ap-
dress and articulate a clear privacy policy. The growth of plied Forecasting, (38), 9–14.
technologies is outpacing privacy laws in many jurisdic- Boone, T., Ganeshan, R., Hicks, R. L., & Sanders, N. R. (2017). Can Google
tions (Machanavajjhala & Reiter, 2012), which adds the trends improve your sales forecast?. Production and Operations Man-
agement, 27(10), 1770–1774.
challenge that firms potentially may have to deal with
Broniatowski, D. A., Paul, M. J., & Dredze, M. (2013). National and local
privacy laws in different geographical locations. Second, influenza surveillance through Twitter: an analysis of the 2012–2013
these large datasets (especially if they are distributed) pose influenza epidemic. PLoS One, 8(12), e83672.

19 By this we mean that the content being captured correlates with the 20 https://www.fatml.org/resources/principles-for-accountable-
signal being measured. algorithms, accessed January 15, 2018.
 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180 179

Bruno, H., & Vilcassim, N. (2008). Structural demand esimation with Hui, S., Fader, P., & Bradlow, E. (2009). Research note — The traveling
varying product availability. Marketing Science, 27(6), 1126–1131. salesman goes shopping: The systematic deviations of grocery paths
Cachon, G. P., Daniels, K. M., & Lobel, R. (2017). The role of surge pricing on from TSP optimality. Marketing Science, 28(3), 566–572.
a service platform with self-scheduling capacity. Manufacturing and Jain, A., Misra, S., & Rudi, N. (2016). Sales Assistance, Search and Purchase
Service Operations Management, 19(3), 368–384. Decisions: an Analysis Using Retail Video Data, Working Paper. CUNY,
Calders, T., & Verwer, S. (2010). Three naive Bayes approaches for NY: Ziklin School of Business.
discrimination-free classification. Data Mining and Knowledge Discov- Jain, A., Rudi, N., & Wang, T. (2015). Demand estimation and ordering
ery, 21(2), 277–292. under censoring: stock-out timing is (almost) all you need. Operations
Castle, J. L., Fawcett, N. W., & Hendry, D. F. (2009). Nowcasting is not Research, 63(1), 134–150.
just contemporaneous forecasting. National Institute Economic Review, Karabati, S., Tan, B., & Öztürk, Ö. C. (2009). A method for estimating
210(1), 71–89. stock-out-based substitution rates by using point-of-sale data. IIE
Chen, X., Ma, W., Simchi-Levi, D., & Xin, L. (2016). Dynamic recommenda- Transactions, 41(5), 408–420.
tion at checkout under inventory constraint, MIT working paper. Kumar, V., Bhaskaran, V., Mirchandani, R., & Shah, M. (2013). Creating a
Chen, L., Mersereau, A. J., & Wang, Z. (2017). Optimal merchandise testing measurable social media marketing strategy: Increasing the value and
with limited inventory. Operations Research, 65(4), 968–991. ROI of intangibles and tangibles for hokey pokey. Marketing Science,
Choi, H., & Varian, H. (2009). Predicting initial claims for unemployment 32(2), 194–212.
benefits. Google Inc. Lam, S. Y., Vandenbosch, M., Hulland, J., & Pearce, M. (2001). Evaluating
Choi, H., & Varian, H. (2012). Predicting the present with Google Trends. promotions in shopping environments: Decomposing sales response
Economic Record, 88(s1), 2–9. into attraction, conversion, and spending effects. Marketing Science,
Chung, J., & Rao, V. R. (2003). A general choice model for bundles 20(2), 194–215.
with multiple-category products: application to market segmentation Lam, S., Vandenbosch, M., & Pearce, M. (1998). Retail sales force scheduling
and optimal pricing for bundles. Journal of Marketing Research, 40 based on store traffic forecasting. Journal of Retailing, 74(1), 61–88.
(2), 115–130. Lamb, A., Paul, M. J., & Dredze, M. (2013). Separating fact from fear: Track-
Cohen, M. C. (2018). Big data and service operations. Production and ing flu infections on Twitter. In Proceedings of the 2013 conference of the
Operations Management, 27(9), 1709–1723. North American chapter of the association for computational linguistics:
Cui, R., Gallino, S., Moreno, A., & Zhang, D. J. (2017). The operational value Human language technologies (pp. 789–795).
of social media information. Production and Operations Management, Lampos, V., De Bie, T., & Cristianini, N. (2010). Flu detector—tracking
27(10), 1749–1769. epidemics on Twitter. In Joint European conference on machine learning
Da, Z., Engelberg, J., & Gao, P. (2011). In search of attention. The Journal of and knowledge discovery in databases (pp. 599–602). Berlin, Heidel-
Finance, 66(5), 1461–1499. berg: Springer.
Da, Z., Engelberg, J., & Gao, P. (2014). The sum of all FEARS investor Lau, R. Y. K., Zhang, W., & Xu, W. (2017). Parallel aspect-oriented sentiment
sentiment and asset prices. The Review of Financial Studies, 28(1), 1–32. analysis for sales forecasting with big data. Production and Operations
Dimpfl, T., & Jank, S. (2016). Can internet search queries help to predict Management, 27(10), 1775–1794.
stock market volatility? European Financial Management, 22(2), 171– Lazer, D., Kennedy, R., King, G., & Vespignani, A. (2014). The parable of
192. Google Flu: traps in big data analysis. Science, 343(6176), 1203–1205.
Drake, M. S., Roulstone, D. T., & Thornock, J. R. (2012). Investor information Liu, T., Ding, X., Chen, Y., Chen, H., & Guo, M. (2016). Predicting movie
demand: Evidence from Google searches around earnings announce- box-office revenues by exploiting large-scale social media content.
ments. Journal of Accounting Research, 50(4), 1001–1040. Multimedia Tools and Applications, 75(3), 1509–1528.
Fantazzini, D., & Toktamysova, Z. (2015). Forecasting German car sales Lu, Y., Musalem, A., Olivares, M., & Schilkrut, A. (2013). Measuring the
using Google data and multivariate models. International Journal of effect of queues on customer purchases. Management Science, 59(8),
Production Economics, 170, 97–135. 1743–1763.
Feng, Q., & Shanthikumar, J. G. (2018). How research in production and Lu, R., Zhu, H., Liu, X., Liu, J. K., & Shao, J. (2014). Toward efficient
operations management may evolve in the era of big data. Production and privacy-preserving computing in big data era. IEEE Network, 28
and Operations Management, 27(9), 1670–1684. (4), 46–50.
Fildes, R., & Goodwin, P. (2007). Against your better judgment? How Machanavajjhala, A., & Reiter, J. P. (2012). Big privacy: protecting confi-
organizations can improve their use of management judgment in dentiality in big data. XRDS: Crossroads, The ACM Magazine for Students,
forecasting. Interfaces, 37(6), 570–576. 19(1), 20–23.
Fildes, R., Goodwin, P., Lawrence, M., & Nikolopoulos, K. (2009). Effective Mani, V., Kesavan, S., & Swaminathan, J. M. (2015). Estimating the impact
forecasting and judgmental adjustments: an empirical evaluation and of understaffing on sales and profitability in retail stores. Production
strategies for improvement in supply-chain planning. International and Operations Management, 24(2), 201–218.
Journal of Forecasting, 25(1), 3–23. Mishne, G., Glance, N. S., et al. (2006). Predicting movie sales from blogger
Franses, P. H., & Legerstee, R. (2010). Do experts’ adjustments on model- sentiment. In AAAI Spring symposium: Computational approaches to
based SKU-level forecasts improve forecast quality?. Journal of Fore- analyzing weblogs (pp. 155–158).
casting, 29(3), 331–340. Musalem, A., Olivares, M., Bradlow, E., Terwiesch, C., & Corsten, D. (2010).
Ganeshan, R. (2014). Clickstream analysis for forecasting online behavior. Structural estimation of the effect of out-of-stocks. Management Sci-
Foresight: The International Journal of Applied Forecasting, (33), 15–19. ence, 56(7), 1180–1197.
Ginsberg, J., Mohebbi, M. H., Patel, R. S., Brammer, L., Smolinski, M. S., & Nunnari, G., & Nunnari, V. (2017). Forecasting monthly sales retail time se-
Brilliant, L. (2009). Detecting influenza epidemics using search engine ries: A case study. In 2017 IEEE 19th Conference on Business Informatics,
query data. Nature, 457(7232), 1012–1014. vol. 1 (pp. 1–6). IEEE.
Goh, K. -Y., Heng, C. -S., & Lin, Z. (2013). Social media brand community O’Neil, C. (2017). Weapons of math destruction: how big data increases
and consumer behavior: Quantifying the relative impact of user- inequality and threatens democracy. Broadway Books.
and marketer-generated content. Information Systems Research, 24 Otto, B. (2011). Organizing data governance: Findings from the telecom-
(1), 88–107. munications industry and consequences for large service providers.
Hand, C., & Judge, G. (2012). Searching for the picture: forecasting UK cin- Communications of the Association for Information Systems, 29, 3.
ema admissions using Google Trends data. Applied Economics Letters, Pan, B., Chenguang Wu, D., & Song, H. (2012). Forecasting hotel room
19(11), 1051–1055. demand using search engine data. Journal of Hospitality and Tourism
Hashem, I. A. T., Yaqoob, I., Anuar, N. B., Mokhtar, S., Gani, A., & Khan, S. U. Technology, 3(3), 196–210.
(2015). The rise of ‘‘big data’’ on cloud computing: Review and open Panahifar, F., Heavey, C., Byrne, P. J., & Fazlollahtabar, H. (2015). A
research issues. Information Systems, 47, 98–115. framework for collaborative planning, forecasting and replenishment
Hruschka, H. (2017). Multicategory purchase incidence models for parti- (CPFR) state of the art. Journal of Enterprise Information Management,
tions of product categories. Journal of Forecasting, 36(3), 230–240. 28(6), 838–871.
Hui, S., Bradlow, E., & Fader, P. (2009). Testing behavioral hypotheses using Perdikaki, O., Kesavan, S., & Swaminathan, J. (2012). Effect of traffic on
an integrated model of grocery store shopping path and purchase sales and conversion rates of retail stores. Manufacturing and Service
behavior. Journal of Consumer Research, 36(3), 478–493. Operations Management, 14(1), 145–162.
180 T. Boone, R. Ganeshan, A. Jain et al. / International Journal of Forecasting 35 (2019) 170–180

Queenan, C., Ferguson, M., Higbie, J., & Kapoor, R. (2007). A comparison of She is a sought-after expert in the local community in the areas of
unconstraining methods to improve revenue management systems. sustainability, public service, and supply chain management.
Production and Operations Management, 16(6), 729–746. Tonya Boone has a Ph.D. in Operations and Technology Management
Rao, T., & Srivastava, S. (2012). Analyzing stock market movements us- from the University of North Carolina at Chapel Hill’s Kenan-Flagler School
ing Twitter sentiment analysis. In Proceedings of the 2012 interna- of Business; a MBA from the College of William and Mary; and a B.S. in
tional conference on advances in social networks analysis and mining Electrical & Electronics Engineering from the University of Kansas.
(pp. 119–123). IEEE Computer Society.
Rivera, R. (2016). A dynamic linear model to forecast hotel registrations
Ram Ganeshan is the D. Hillsdon Ryan Professor of Business at the
in Puerto Rico using Google Trends data. Tourism Management, 57,
Raymond A. Mason School of Business, The College of William and Mary,
12–20.
Williamsburg, VA.
Russell, G. J., & Petersen, A. (2000). Analysis of cross category dependence
Ram’s teaching, research and consulting interests are in the areas of
in market basket selection. Journal of Retailing, 76(3), 367–392.
supply chain management, data analytics, and logistics strategy, primarily
Sagaert, Y. R., Aghezzaf, E. -H., Kourentzes, N., & Desmet, B. (2018). Tempo-
in the chemical, hi-tech, and retail industries. He is a regular contributor to
ral big data for tactical sales forecasting in the tire industry. Interfaces,
academic and trade journals and is the co-editor of three books including
48(2), 121–129.
Quantitative Models for Supply Chain Management which is one of the most
Sanders, N. R. (2016). How to use big data to drive your supply chain.
highly cited books in supply chain management. In 2001, the Production
California Management Review, 58(3), 26–48.
& Operations Management Society (POMS) awarded him the prestigious
Sanguinet, M. E. (2016). Hashtags, tweets and movie receipts: Social media
Wickham Skinner Award for his research on how supply chains can be
analytics in predicting box office hits (Ph.D. thesis), San Diego State
efficiently managed. His current research projects extensively rely on
University.
big-data techniques to provide insights into making sound decisions for
Schaer, O., Kourentzes, N., & Fildes, R. (2019). Demand forecasting with
managing both manufacturing and service supply chains. They include:
user-generated online information. International Journal of Forecast-
(1) work with online retailers to harness and interpret clickstream data
ing, 35(1), 197–212.
from customer browsing behavior; (2) working with carriers in the trans-
Seifter, A., Schwarzwalder, A., Geis, K., & Aucott, J. (2010). The utility of
portation sector to reduce cost and carbon footprint; and (3) analyzing
‘‘Google Trends’’ for epidemiological research: Lyme disease as an
large scale project data in professional services to provide strategies for
example. Geospatial health, 4(2), 135–137.
productivity improvement.
Smith, P. (2016). Google’s MIDAS touch: Predicting UK unemployment
He received a Doctorate in Operations and Logistics Management
with internet search data. Journal of Forecasting, 35(3), 263–284.
from Penn State; a MSOR degree in Operations Research from the Uni-
Tan, B., & Karabati, S. (2004). Can the desired service level be achieved
versity of North Carolina at Chapel Hill; and an undergraduate degree in
when the demand and lost sales are unobserved?. IIE Transactions,
Industrial Management from the Birla Institute of Technology & Science
36(4), 345–358.
in India.
van Heerde, H. J., & Neslin, S. A. (2017). Sales promotion models. In
Handbook of marketing decision models (pp. 13–77). Springer.
van Ryzin, G., & Vulcano, G. (2014). A market discovery algorithm to Aditya Jain is assistant professor of Operations Management at Zicklin
estimate a general class of nonparametric choice models. Management Business School at Baruch College in New York city. His primary research
Science, 61(2), 281–300. interests include retail operations, supply chain management and health-
Wakamiya, S., Lee, R., & Sumiya, K. (2011). Towards better TV viewing care operations. Aditya earned his Ph.D. in Operations Management from
rates: exploiting crowd’s media life logs over Twitter for TV rating. the Simon School of Business at University of Rochester, and served as
In Proceedings of the 5th international conference on ubiquitous infor- a faculty member at Indian School of Business and the Kellogg School
mation management and communication (p. 39). ACM. of Northwestern University prior to his current appointment. His indus-
Wecker, W. (1978). Predicting demand from sales data in the presence of try experience includes extensive consulting work with fashion retailer
stockouts. Management Science, 24(10), 1043–1054. through the retail/supply chain analytics company he founded and ran for
Weller, M., & Crone, S. (2012). Supply chain forecasting: Best practices and several years. He holds bachelor’s degree in Chemical Engineering from
benchmarking study: Technical paper, Lancaster Centre for Forecasting. Indian Institute of Technology, Bombay.
Wu, L., & Brynjolfsson, E. (2015). The future of prediction: How Google
searches foreshadow housing prices and sales. In Economic Analysis of Nada R. Sanders is the Distinguished Professor of Supply Chain Manage-
the Digital Economy (pp. 89–118). University of Chicago Press. ment at the D’Amore-McKim School of Business at Northeastern Univer-
Zhang, X., Fuehres, H., & Gloor, P. A. (2011). Predicting stock market sity in Boston MA. Prior to that, she held the Iacocca Chair at the College
indicators through Twitter ‘‘I hope it is not as bad as I fear’’. Procedia- of Business and Economics at Lehigh University and as the West Chair at
Social and Behavioral Sciences, 26, 55–62. the M.J. Neeley School of Business.
Her research and teaching interests have been in forecasting and the
use of data analytics in decision-making within the supply chain context.
Tonya Boone is an Associate Professor at the Raymond A. Mason School She has authored over one hundred scholarly works and has served on the
of Business, The College of William & Mary. Prior to joining the faculty of editorial boards of prominent journals in her field, including the Journal of
the College of William and Mary, she was a faculty member at the Fischer Operations Management, Production and Operations Management, Decision
College of Business at Ohio State University. Sciences Journal, Journal of Business Logistics, and International Institute
Her research and teaching interests include sustainable operations, of Forecasting. She is a Fellow of the Decision Sciences Institute and was
knowledge management in professional service organizations, and the co-founder and Associate Editor of Foresight: The International Journal of
management of supply chains in data rich environments. Tonya’s research Applied Forecasting. She has authored multiple books with the most recent
has been published in the top academic management journals, including being Big Data Driven Supply Chain Management and has given numer-
Management Science, Journal of Operations Management, POMS Journal, and ous talks on the subject including a recent HBR webinar. She currently
Decision Sciences. She is also a co-editor of the recent book Sustainable serves on the Board of POMS, having served as both Program and General
Supply Chain Management: Methods, Models, and Policy Implications. Chair.
Tonya is active in her community, serving on the Board of the She holds a Ph.D. in Operations Management and Logistics, and an
Williamsburg Economic Development Authority, the Board of the Hamp- MBA, from the Fisher College of Business at The Ohio State University, as
ton Roads Incubator, and the Board of the Williamsburg Regional Library. well as a B.S. in Mechanical Engineering.