Business Horizons (2014) 57, 595605

Available online at www.sciencedirect.com

ScienceDirect
www.elsevier.com/locate/bushor

Supply chain analytics

Gilvan C. Souza
Kelley School of Business, Indiana University, Bloomington, IN 47405, U.S.A.

KEYWORDS
Supply chain
management;
Analytics;
Optimization;
Forecasting

Abstract In this article, I describe the application of advanced analytics techniques

to supply chain management. The applications are categorized in terms of descriptive, predictive, and prescriptive analytics and along the supply chain operations
reference (SCOR) model domains plan, source, make, deliver, and return. Descriptive
analytics applications center on the use of data from global positioning systems
(GPSs), radio frequency identification (RFID) chips, and data-visualization tools to
provide managers with real-time information regarding location and quantities of
goods in the supply chain. Predictive analytics centers on demand forecasting at
strategic, tactical, and operational levels, all of which drive the planning process in
supply chains in terms of network design, capacity planning, production planning, and
inventory management. Finally, prescriptive analytics focuses on the use of mathematical optimization and simulation techniques to provide decision-support tools
built upon descriptive and predictive analytics models.
# 2014 Kelley School of Business, Indiana University. Published by Elsevier Inc. All
rights reserved.

1. Why analytics in supply chain

management?
The supply chain for a product is the network of
firms and facilities involved in the transformation
process from raw materials to a product and in
the distribution of that product to customers. In
a supply chain, there are physical, financial, and
informational flows among different firms. Supply
chain analytics focuses on the use of information
and analytical tools to make better decisions regarding material flows in the supply chain. Put
differently, supply chain analytics focuses on

E-mail address: gsouza@indiana.edu

analytical approaches to make decisions that better

match supply and demand.
Well-planned and implemented decisions contribute directly to the bottom line by lowering
sourcing, transportation, storage, stockout, and
disposal costs. As a result, analytics has historically
played a significant role in supply chain management, starting with military operations during and
after World War IIparticularly with the development of the simplex method for solving linear programming by George Dantzig in the 1940s. Supply
chain analytics became more ingrained in decision
making with the advent of enterprise resource planning (ERP) systems in the 1990s and more recently
with big data applications, particularly in descriptive and predictive analytics, as I describe with
some examples in this article.

0007-6813/$ see front matter # 2014 Kelley School of Business, Indiana University. Published by Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.bushor.2014.06.004

596

G.C. Souza

The Supply Chain Operations Reference (SCOR)

model developed by the Supply Chain Council
(www.supply-chain.org) provides a good framework
for classifying the analytics applications in supply
chain management. The SCOR model outlines four
domains of supply chain activities: source, make,
deliver, and return. A fifth domain of the SCOR
modelplanis behind all four activity domains.
Furthermore, a key input of the supply chain planning process is demand forecasting at all time
frames: long, mid, and short term with planning
horizons of years, months, and days, respectively.
Table 1 illustrates different decisions in each of the
four SCOR domains that can be aided by analytics.
These decisions are further classified into strategic,
tactical, and operational according to their time
frame.
Analytics techniques can be categorized into
three types: descriptive, predictive, and prescriptive. Descriptive analytics derives information from
significant amounts of data and answers the question of what is happening. Real-time information
about the location and quantities of goods in the
supply chain provides managers with tools to make
adjustments to delivery schedules, place replenishment orders, place emergency orders, change transportation modes, and so forth. Traditional data
sources include global positioning system (GPS) data
on the location of trucks and ships that contain
inventories, radio frequency identification (RFID)
data originating from passive tags embedded in

Table 1.

pallets (even at the product level), and transactions

involving barcodes. Information is derived from the
vast amounts of data collected from these sources
through data visualization, often with the help of
geospatial mapping systems. RFID is a significant
improvement over barcodes because it does not
require direct line of sight. Accurate inventory records are critical in supply chains as they trigger
regular replenishment orders and emergency orders
when inventory levels are too low. Although RFID
technology helps in significantly reducing the frequency of manual inventory reviews, such reviews
are still needed because of data inaccuracy due to,
for example, inventory deterioration or damage or
even tag-reading errors.
Predictive analytics in supply chains derives demand forecasts from past data and answers the
question of what will be happening.
Prescriptive analytics derives decision recommendations based on descriptive and predictive
analytics models and mathematical optimization
models. It answers the question of what should be
happening. Arguably, the bulk of academic research, software, and practitioner activity in supply
chain analytics focuses on prescriptive analytics.
In Table 2, I provide a summary of analytics
techniquesdescriptive, predictive, and prescriptiveused in supply chains in terms of the four SCOR
domains of source, make, deliver, and return. I
elaborate on Table 1 and Table 2 in the next
sections.

SCOR model and examples of decisions at the three levels

SCOR Domain Source

Make

Deliver

Return

Activities

Order and receive

materials and
products

Schedule and
manufacture, repair,
remanufacture, or
recycle materials
and products

Receive, schedule,
pick, pack, and
ship orders

Request, approve, and

determine disposal of
products and assets

Strategic
(time frame:
years)

 Strategic

 Location of plants
 Product line mix

 Location of

 Location of return

Tactical
(time frame:
months)

 Tactical sourcing  Product line

 Supply chain
rationalization
 Sales and
contracts

Operational
(time frame:
days)

 Materials

Plan

Demand forecasting (long term, mid term, and short term)

sourcing
Supply chain
mapping

at plants

operations planning
requirement
planning
and inventory
replenishment
orders

 Workforce scheduling
 Manufacturing, order

distribution centers
Fleet planning

 Transportation and


distribution planning
Inventory policies
at locations

 Vehicle routing
(for deliveries)

tracking, and scheduling

centers

 Reverse distribution
plan

 Vehicle routing (for

returns collection)

Supply chain analytics

Table 2.

597

Analytic techniques used in supply chain management

Make

Deliver

Return

Analytics
Techniques

Source

Descriptive

 Supply chain mapping

 Supply chain visualization
 Time series methods (e.g., moving average, exponential smoothing, autoregressive models)
 Linear, non-linear, and logistic regression
 Data-mining techniques (e.g., cluster analysis, market basket analysis)
 Analytic hierarchy process
 Mixed-integer linear
 Network flow
 Game theory (e.g., auction design,
programming (MILP)
algorithms
 Non-linear programming
 MILP
contract design)
 Stochastic

Predictive

Prescriptive

dynamic
programming

2. Plan: Demand forecasting using

predictive analytics
Demand forecasting is a critical input to supply
chain planning. Different time frames for demand
forecasting require different analytics techniques.
Long-term demand forecasting is used at the strategic level and may use macro-economic data, demographic trends, technological trends, and
competitive intelligence. For example, demand factors for commercial aircraft at Boeing include energy prices, discretionary spending, population
growth, and inflation, whereas demand factors for
military aircraft include geo-political changes, congressional spending, budgetary constraints, and
government regulations (Safavi, 2005). Causal forecasting methodscalled such because they analyze
the underlying factors that drive demand for a
productare used at this level. Analytics causal
forecasting methods include linear, non-linear,
and logistic regression.
To illustrate demand forecasting for tactical and
operational supply chain decisions, consider the
production planning process for an original equipment manufacturer (OEM) such as Whirlpool. At the
product family level (e.g., refrigerators), the sales
and operations planning (S&OP) process uses aggregate demand forecasts in monthly time buckets to
establish aggregate production rates, aggregate
levels of inventories, and workforce levels. The
aggregate plan is revised on a rolling basis as new
data is available. The S&OP plan, as well as more
refined demand forecasts at the stock-keeping unit
(SKU) level, is used to derive the master production
schedule (MPS), which details weekly production
quantities at the SKU level for a typical planning
horizon of 812 weeks. The MPS and the bill of
materials are then used to plan production and
sourcing at the part level through a materials requirement planning (MRP) system that is embedded
in most ERP software. Time-based demands for parts

are derived from time-based demands for the

SKUs that use those parts, so parts have dependent
demand. In contrast, SKUs have independent
demand. Demand forecasts for items subject to independent demand require predictive analytics techniques, whereas forecasts for dependent demand items
are obtained directly from the MRP system. Demand
forecasts for independent demand items are also used
to plan for inventory safety stocks at other locations,
such as distribution centers and retailers.
Demand forecasting for independent demand
items is usually performed using time-series methods,
for which the only predictor of demand is time. Timeseries methods include moving average, exponential
smoothing, and autoregressive models. For example,
Winters exponential smoothing method incorporates
both trend and seasonality and can be used for both
short-term and mid-term forecasting. In an autoregressive model, demand forecast in one period is a
weighted sum of realized demands in the previous
periods.
Mid-term forecasting can also benefit from
causal forecasting methods, especially in nonmanufacturing industries or the manufacturing of
non-discrete items. For example, in order to forecast monthly demand for truckload (TL) freight
services, Fite, Taylor, Usher, English, and Roberts
(2002) considered 107 economic indexes as potential predictors, including the purchasing managers
index, the Dow Jones stock index, the consumer
goods production index, automotive dealer sales,
U.S. exports, the producer commodities price index
for construction materials and equipment, interest
rates, and gasoline production. They used stepwise
regression to identify the most relevant indexes and
found parsimonious models for predicting TL demand for specific industries and regions. Their model only predicts industry-wide demand for TL
services (nationally or by region); the connection
to demand forecasts at the firm level was made
using historical market shares.

598
Data mining has also been used for demand forecasting in conjunction with traditional forecasting
techniques (Rey, Kordon, & Wells, 2012). Usually,
the data-mining step precedes the use of causal
forecasting techniques by finding appropriate demand drivers (i.e., independent variables) for a
product that can be used in regression analysis.
For example, Dow Chemical uses a combination of
data mining and regression techniques to forecast
demand at the strategic and tactical levels (e.g.,
identifying demand trends), which is useful for its
pricing strategy and for configuring and designing its
supply chain to respond to these trends (Rey & Wells,
2013). Data-mining methods usually involve clustering techniques. So, if a retailer finds out, for example, that demand for cereal is strongly related to
milk sales, then the retailer may build a causal
forecasting model that predicts cereal sales with
milk sales as one of the predicting variables. Market
basket analysis is a specific data-mining technique
that provides an analysis of purchasing patterns at
the individual transaction level, so a retailer can
analyze the frequency with which two product categories (e.g., DVDs and baby products) are purchased together. Lift for a combination of items is
equal to the actual number of times the combination occurs in a given number of transactions divided
by the predicted number of times the combination
occurs if items in the combination were independent. Lift values above 1 indicate that items tend to
be purchased together. This kind of analysis can be
useful when building causal regression models for
demand forecasting. It can also aid in promotion
activities because the retailer can predict how much
sales of Product 1 would increase if there is a
promotion for Product 2 if the two products are
often purchased together.

3. Source
3.1. Source: Strategic decisions
Strategic sourcing is the process of evaluating and
selecting key suppliers. There is limited use of
analytics for strategic sourcing in practice even
though academics prescribe the use of sophisticated
multi-criteria decision-making techniques such as
analytic hierarchic process (AHP). AHP decomposes
a complex problem (e.g., selecting a supplier among
a diverse set) into more easily comprehended subproblems that can be analyzed separately. In the
supplier-selection problem, these sub-problems
might include distinct evaluations of factors like
cost, quality, delivery speed, delivery reliability,
volume flexibility, product mix flexibility, and sustainability. These evaluations are then weighed.

G.C. Souza
Firms are very familiar with their first-tier suppliers (i.e., those that directly supply them) and
perhaps their second-tier suppliers (i.e., those that
supply first-tier suppliers), but some of their lowertier suppliers may be unknown. A recent example is
the November 2012 fire at the Bangladesh factory
that killed more than 100 workers. An audit of the
factory by Walmart in 2011 ruled it out as a supplier.
However, one of Walmarts suppliers continued to
subcontract work to that factory (Tsikoudakis,
2013). The threat of disruptions like natural disasters, social and political unrest, and major strikes
makes it imperative for firms to map their supply
chains. For example, Cisco (2013) uses supply chain
mapping and enterprise social networking to identify its vulnerabilities to supply chain disruptions as
well as to collaborate with its suppliers and partners. The open source tool sourcemap.com, developed at the Massachusetts Institute of Technology,
allows one to visualize and map a supply chain; the
tool can also be used for purposes such as carbon
footprint estimation. An example is shown in
Figure 1.

3.2. Source: Tactical decisions

In contrast to strategic sourcing, tactical sourcing
refers to the process of achieving specific
objectivessuch as determining costs for parts,
materials, or servicesthrough structured procurement mechanisms like auctions. The central problem in procurement auctions centers on mechanism
design: How should one structure the rules of an
auction so that bidders (i.e., suppliers) behave in a
manner that results in minimal procurement cost
(and desired performance) for the buyer? Auctions
can be open (i.e., bidders can view and respond to
bids) or sealed and one shot or dynamic (which occur
over several rounds of bidding). Government auctions tend to be one-shot, sealed auctions, whereas
open, dynamic auctions are common in industrial
procurement (Beil, 2010). Buyers must consider the
total procurement cost as bidders usually bid on
contract payment terms only (e.g., unit cost). Additional logistics costs, if paid by the buyer, must be
taken into account in the bid price. The prescriptive
analytics used here is centered on game theory,
which is used to determine auction rules. Procurement auctions are widely used in practice.
A commonly used payment contract in sourcing is
wholesale price, via which the buyer (i.e., retailer)
pays the seller (i.e., manufacturer) a fixed price per
unit. Under this contract, retailers are exposed to
demand risk: they bear the entire costs of overstocking and therefore have an incentive to stock
less than what is optimal for the supply chain as a

Supply chain analytics

Figure 1.

599

Example of supply chain mapping using sourcemap.com

Source: free.sourcemap.com/view/6585/

whole. Recognizing this, academics have used a

combination of game theory and statistics to prescribe more sophisticated contracts that will improve product availability in retailers. For example,
in a buy-back contract, retailers can return unsold
units to the manufacturer and receive a partial
refund. Although such contracts can improve supply
chain performance, the wholesale price contract is
still widely used, perhaps due to its simplicity.

4. Make
4.1. Make: Strategic decisions
Network design determines the optimal location and
capacity of plants, distribution centers (DCs), and
retailers. The simplest form of the network design

problem can be illustrated when deciding where to

build DCs that serve as intermediary stocking
and shipping points between existing plants and
retailers. This problem is formulated as a mixedinteger linear program (MILP). Data requirements
include yearly aggregate demands for the product
family at each retailer, plant capacities, unit shipping costs between each pair of locations, and the
annual fixed cost of operating a DC at each potential
location. Decision variables include the quantity to
ship between locations and binary variables that
indicate if each DC should be open or closed. The
objective function minimizes total shipping and
fixed DC costs. Constraints ensure that demand is
met at all locations, that companies only ship products from a DC if it is open, and that all plant
capacities are respected. The solution provides
the location (i.e., where to open the DCs) as well

600
as the allocation of plants to the DCs, the allocation
of DCs to retailers, and the capacity of each DC.
Variations of this simple MILP formulation include
multiple products, transportation capacities between
locations, multiple transportation modes between
locations, a multi-year planning horizon, multiple
echelons (i.e., tiers in the supply chain), demand
uncertainty, supply uncertainty, and reverse flows
(e.g., the collection of used products for recycling
and remanufacturing). When the problem incorporates multiple products, the analysis also provides the
product mix at each plant. When many of the variations above are incorporated and the problem is large
(e.g., thousands of retailers and potential DC and
plant locations), the problem may become too difficult to solve to optimality using off-the-shelf optimization software. Therefore, many researchers have
proposed well-performing heuristics, such as genetic
algorithms, that ensure goodand sometimes
optimalsolutions. Genetic algorithms use a divide
and conquer (the feasible region) approach to finding
a good solution to the MILP as opposed to optimal
branch and bound algorithms, which are combinatorial in nature.
Some of the data necessary to perform such
analysis requires a preliminary level of analysis so
it can be extracted, cleaned, and aggregated from
ERP systems. Network design, however, is only performed infrequently for each firm, including during
mergers and acquisitions. As a result, it is not part of
standard ERP software. Specialized software makes
it easy to input this data, specify the constraints,
perform the optimization, and visualize the results,
especially for large problems.

4.2. Make: Tactical decisions

We have previously described the S&OP process,
which is used for planning aggregate workforce
and inventory levels on a medium planning horizon
based on demand forecasts, underlying costs, and
actual sales. Academics have proposed MILP models
for this process. For each month in the planning
horizon, decision variables include the amount to
produce for each product family using regular time,
overtime, and subcontracting and the number of
workers to be hired and laid off. The objective
function minimizes total cost, which comprises total
production cost (i.e., regular time, overtime, and
subcontracting), total inventory cost, total wages,
total hiring cost, and total layoff cost. Constraints
may come from, among others, inventory and workforce balancing, regular production capacity,
and overtime production. Many practitioners use
rules-based heuristics. For example, one heuristic
is a level production strategy, via which the firm

G.C. Souza
meets fluctuating demand by producing at a constant
rate and holding inventory to meet the peak demand.
Alternatively, the firm can use a chase strategy,
adjusting workforce levels monthly to meet fluctuating demand. Firms frequently use a hybrid strategy
between chase and level.
Product proliferation and mass customization have
been widely documented (e.g., Rungtusanatham &
Salvador, 2008). For product proliferation and mass
customization, the plant must adapt from a mass
production environmentdesigned for economies
of scale, with fewer products produced in dedicated
lines and setup costs spread over long production
runsto a flexible production environment. This
adaptation is made possible with the aid of flexible
manufacturing technology or changes in the product
and process design that support a postponement
strategy (Lee, 1996). In a postponement strategy,
the step in the manufacturing process in which product differentiation occursfrom gray boxes to
SKUsis located closer to the customer, which allows
the firm to carry inventory of gray boxes instead of
SKUs, and thus lessens differentiation time. Postponement mitigates the negative impacts of increased product proliferation, such as increased
forecasting uncertainty at the SKU level; increased
inventory costs; and complexity costs, such as research and development, testing, tooling, returns,
and obsolescence. Postponement requires changes in
product and process design, and it may not be feasible for products like automobiles, for which strict
quality guidelines in final assembly preclude significant customization at dealers. As an alternative,
firms may increase supply chain performance through
product rationalization using analytics, as shown in
Table 3.

4.3. Make: Operational decisions

Manufacturing scheduling is the last step in the
planning process after MRP plans are released. An
MRP plan specifies quantities and due dates for all
parts. Scheduling then sequences the jobs (i.e.,
parts) by the different resources necessary for
manufacturing the part in order to meet the due
dates. In general, there are n jobs to be scheduled in
m different resources, and the processing time, due
date, and weight (i.e., priority) of each job in each
resource are known. This problem takes different
forms depending on the decision makers objective,
the number of resources, and how the jobs are
processed with the resources. An objective function
minimizes the maximum completion time, or the
maximum lateness, across all jobs. There can be
precedence relationships, setup times, or even
sequence-dependent setup times (i.e., when the

Supply chain analytics

Table 3.

601

Product rationalization at Hewlett-Packard

Hewlett-Packard (HP) has developed optimization tools for product rationalization (Ward et al., 2010). One tool
requires proposed new product line extensions to meet minimum complexity return-on-investment (ROI)
thresholds. Complexity ROI is defined as the incremental margin minus variable complexity costs, divided by fixed
complexity costs. Variable complexity costs are largely driven by forecasting uncertainty and resulting increased
inventory costs, whereas fixed complexity costs are driven by criteria such as research and development, tooling,
and manufacturing setup costs. With another tool, HP uses a maximum flow algorithm on an existing product line to
perform product rationalization. The tool acknowledges that in firms with configurable product lines, some
products, such as power supplies, may generate little revenue on their own but are critical components for highrevenue orders and for overall order fulfillment. Order coverage is defined as the percentage of a given set of past
orders that can be met from the rationalized product portfolio. Similarly, revenue coverage is the smallest portfolio
of products that covers a given percentage of historical order revenue. This optimization tool revealed how HP can
offer only 20% of previously offered features in laptops and reach 80% revenue coverage. After implementing the
recommendations, HP realized significantly reduced inventory costs and increased gross margins.

setup time for a job at a resource depends on the

previous job there, such as in processing industries
like the chemical industry). Scheduling problems
can be formulated as MILPs, and the combinatorial
nature of these problems makes them very hard to
solve to optimality for large problems. As a result,
significant effort has been devoted to finding good
solutions through heuristics because other complications arise in practice, such as adding new jobs to
the existing pool of processing jobs as well as changing priorities and preferences. In terms of software,
some ERP systems have scheduling modules (e.g.,
the Applied Planning and Optimization module in
SAP) that use genetic algorithms to provide good
solutions to MILPs found in determinist scheduling.
These algorithms can provide good solutions to fairly
large problems, such as 1 million jobs over 1,000
resources (Pinedo, 2008). There are a few companies such as Taylor (www.taylor.com) that specialize
in providing scheduling software with many functionalities not present in ERP systems. Although the
discussion above has centered on manufacturing
scheduling, some of the same algorithms can be used
in other scheduling problems like assigning gates at
an airport or trucks at a cross-docking location.
Workforce scheduling can be challenging for service industries, such as call centers, hospitals, and
airlines, in which there is seasonal demand, not only
for time of the year (common in manufacturing), but
also for day of the week and hour of the day. A
common way of modeling these problems is
by defining tours. A tour is a combination of time
blocks within a day and within days of the week that
add up to the necessary work hours per employee.
An example of a tour would be Monday, 8 a.m.1
p.m.; Tuesday, 1 p.m.6 p.m.; Thursday, 8 a.m.6
p.m.; and Friday, 8 a.m.6 p.m. Tours should be
feasible; for example, it is not very convenient for
most people to work from 8 a.m.10 a.m. and
then from 3 p.m.5:00 p.m. on the same day.

The decision maker needs demand forecasts for

each time block (e.g., 12 p.m.1 p.m. on Monday),
which can be obtained through predictive forecasting models. This problem can be formulated as an
MILP in which decision variables include the number
of employees assigned to each tour and the number
of employees necessary to meet demands within
each time block. The objective function minimizes
total labor costs. Complications, such as workers
preferences, multiple locations, task assignments,
and so forth, increase the size of the MILP model to
such an extent that heuristics are almost certainly
needed. Some ERP vendors have workforce scheduling modules for specific applications like retail and
hospitality. There are also vendors for industry-specific software, such as call centers and health care
providers. Many airlines, which are heavy analytics
users, have developed their own scheduling algorithms.

5. Deliver and return

5.1. Deliver and return: Strategic
decisions
In Section 4, I presented the network design problem
of planning the location of DCs and return centers.
Another strategic decision here is fleet planning,
which can be described as the dynamic acquisition
and divestiture of delivery vehicles to meet the
demand for deliveries or returns collection. This
problem is formulated as an MILP, or dynamic programming, as in Table 4.

5.2. Deliver and return: Tactical decisions

In transportation and distribution planning, the firm
distributes a set of products from source nodes (i.e.,
supply points such as factories) to sink nodes

602
Table 4.

G.C. Souza
Fleet planning for Coca-Cola Enterprises

Coca-Cola Enterprises (CCE) has started replacing some of its fleet of diesel delivery trucks with diesel-electric
hybrid vehicle (HEV) trucks. How the company chooses to invest those dollars depends on volatile fuel costs, usagebased deterioration, and seasonal demand. Wang, Ferguson, Hu, and Souza (2013) have provided a prescriptive
analytics model that takes into consideration CCEs historical maintenance costs, purchasing costs for both diesel
and HEV trucks, CCE demand data, and historical diesel price data to calibrate a stochastic model that simulates
diesel prices dynamically. Using dynamic programming, the optimal policy is obtained, at each period of a planning
horizon and for each realization of diesel prices, that determines how many trucks of each type (diesel and HEV)
CCE should acquire and/or divest. Wang et al. found that at the current outlook of diesel prices, CCE should include
both HEV (54%) and diesel trucks (46%) in its capacity portfolio. In this regard, CCE could use HEV trucks to meet its
average baseline demand and then deploy diesel trucks to supplement the delivery fleet during peak demand
seasons.

(i.e., demand points such as retail locations)

through intermediary storage nodes (e.g., DCs). This
problem is solved using a multi-commodity network
flow model, which is a linear programming formulation with a special structure. In the network formulation, there can be multiple arcs between each pair
of nodes. Each arc represents a shipping mode with a
given capacity, such as rail, truckload (TL), less than
truckload (LTL), and air. The amount to ship in each
arc in the network for each commodity and time
period is considered. Constraints include capacity at
each arc, time period, and node, as well as flowbalancing at each node. Data requirements include
shipping costs in each arc, forecasts of supply available at each source node (provided by the S&OP
plan), point forecasts for demand at each sink node
(from predictive analytics models), and arc capacities. Economies of scale in shipping can also be
incorporated. Problems of realistic size have thousands of nodes, resulting in millions of decision
variables. However, such problems can be solved
efficiently with numerical algorithms based on the
network simplex method, which is embedded in
supply chain optimization software. Despite extensive planning, disruptions (e.g., traffic, weather)
and demand uncertainty often require plan modification, and descriptive analytics tools can be quite
valuable. For example, the Control Tower descriptive analytics system allows Procter & Gamble (P&G)
to see all the transportation occurring in its near
supply chain (i.e., inbound, outbound, raw materials, and finished product). With this technology,
P&G has reduced deadhead movement (i.e., when
trucks travel empty or not optimally loaded) by 15%
and thus has reduced costs (McDonald, 2011).
Another important decision is determining
supply levels at nodes in a distribution network
that is, setting inventory policies. The science for
setting inventory policies (i.e., reorder point and
order-up-to level or order quantity) for a product
at a single location, such as a DC, is mature, even
when demand is uncertain and non-stationary and

replenishment lead times are variable. Data requirements include historic demand and forecasting
data, replenishment lead times, the desired service
level (i.e., a desired fill rate or stock-out probability), holding cost, and the fixed cost of placing a
replenishment order. The inventory policy parametersreorder point and order quantitycan be
computed using exact algorithms or approximate
formulas, which are embedded in most supply chain
software, including in some ERP systems modules.
More often, the supply chain has multiple stocking
points for the same product. For example, a product
can be stocked at a DC and multiple different retailers
in different regions. Although one can set inventory
policies at each location that use only local demand
and replenishment lead-time information, this local
optimization approach is not optimal for the supply
chain. Due to risk pooling, it may be optimal to have
some level of inventory at the DC so that higher-thannormal demand in one retailer can be balanced against
lower-than-normal demand at another retailer. This
situation calls for an integrated inventory policy for
the entire supply chain; the theory that prescribes
these inventory policies is called multi-echelon inventory theory. The complication in multi-echelon inventory theory arises when the DC does not have sufficient
inventory to meet all incoming orders from retailers at
a given period. In that case, the optimal inventoryrationing policy is complex, and even more so if there
are more than two echelons. There are, however,
several well-performing heuristics that are computationally simple, such as the guaranteed service level
heuristic (Graves & Willems, 2000), which has been
implemented in software like Optiant. An example of
successful application is provided in Table 5.

5.3. Deliver and return: Operational

decisions
The vehicle routing problem (VRP) optimizes the
sequence of nodes to be visited in a route, for
example, for a parcel delivery truck, for a returns

Supply chain analytics

Table 5.

603

Multi-echelon inventory management at P&G

Before 2000, P&G used only single-location inventory models, which optimize inventory levels locally given that
locations own replenishment lead time. However, starting in 20052006, P&G started implementing multi-echelon
inventory models based on the guaranteed service level heuristic in its more complex supply networks. At a
particular stage in the supply chain, inventory is set to meet a desired service level based on a guaranteed delivery
time to the customer (S), its own replenishment lead time when ordering from a preceding stage (SI), and its
processing time (T). Essentially, the method sets safety stock levels as if it was a single location with a
replenishment lead time of SI + T - S. Note that SI for a stage is equal to S for a preceding stage. Through dynamic
programming, the method finds the optimal S for each stage to minimize holding costs across the supply chain. The
multi-echelon supply chain approach to inventory management was implemented at 30% of P&Gs locations using
Optiant software and consequently saved the company $1.5 billion in inventory costs in 2009 compared to the
single-location models previously in place (Farasyn et al., 2011).

Table 6.

Vehicle routing at Waste Management, Inc.

Waste Management, Inc. (WM) is a leading provider of solid waste collection and disposal services. It has a fleet of
more than 26,000 vehicles running nearly 20,000 routes. In 2003, the company implemented the WasteRoute
vehicle-routing software, which included GIS capabilities and navigational capabilities, and integrated it with a
relational database containing customer information. An origin-destination matrix was then developed that
considered constraints such as time and distance traveled between any two points, speed limits, and one-way
streets. By implementing the combined prescriptive and descriptive analytics software, the firm saved $44 million
in 2004.
Source: www.informs.org

collection truck, or for both. The optimal sequence

takes into account the distances between each pair
of nodes; expected traffic volume; left turns; and
other constraints placed on the routes, such as
delivery and pickup time windows. Known as the
travelling salesman problem (TSP), the classical VRP
problem only takes into account the distances between each pair of nodes: In what sequence should
nodes be visited, ending at the same starting point?
This problem can be formulated as an MILP. The TSP
problem is combinatorial in nature, and is hard
to solve beyond a few thousand nodes (Funke &
Gruenert, 2005). Among others, complications such
as multiple vehicles, vehicle capacities, tour-length
restrictions, and delivery and pickup time windows
result in an MILP that is very difficult to solve, thus
requiring heuristic approaches. In addition to heuristic approaches, vehicle-routing software incorporates descriptive analytics, as shown in Table 6.

6. Modulating demand to match

capacity: Revenue management
The SCOR model implicitly assumes that managers
plan their operationssource, make, deliver, and
returnbased on demand forecasts. Therefore, the
SCOR model plans capacity to match a given demand. Industries with perishable capacities, like
airlines, hospitality, and transportation, must take
a reverse approach, so firms modulate their demand

to match their fixed capacity through prices and

other mechanisms that will be described next. This
is known as revenue management.
Revenue management started in the airline industry after deregulation, with the problem of allocating seats in a flight to fare classes. Allocation
policies are nested. For instance, suppose there are
two fare classes: $150 (Fare Class 1) and $90 (Fare
Class 2). The decision maker sets a booking limit for
Fare Class 2 and then determines the booking limit
of Fare Class 1 based on the capacity of the flight.
Data requirements for the computation of booking
limits include demand forecasts for the different
classes (as a probability distribution) at different
times before departure, cancellation probabilities,
up-selling probabilities (i.e., the probability that a
customer will buy a higher fare if the lower fare is
unavailable), and fare values. The problem is significantly more complex in a network. For example,
one passenger goes from Indianapolis (IND) to
New York (JFK), whereas another passenger goes
from IND to Rochester (ROC) via JFK. In this case,
heuristic approaches, such as bid-price controls, are
used. The bid price for a resource (e.g., a seat in a
specific flight IND-JFK) is the marginal cost to the
network of consuming one unit of that resource.
When a customer demand arises (e.g., IND-ROC via
JFK), then the demands revenue is compared
against the sum of bid prices for all resources associated with the demand request (i.e., bid prices for
a seat IND-JFK and for a seat JFK-ROC). The demand

604

G.C. Souza

Table 7. Additional information on

techniques for supply chain management

analytics

 General overview: Snyder and Shen (2011)

 Network design: Funaki (2009)
 Auctions: Krishna (2002)
 Sales and operations planning: Jacobs, Berry,
Whybark, and Vollmann (2011)

 Transportation and distribution planning: Ahuja,

Magnanti, and Orlin (1993)

 Inventory management: Zipkin (2000)

 Dynamic pricing and revenue management: Talluri
and Van Ryzin (2004)

 Manufacturing scheduling: Pinedo (2008)

 Workforce scheduling: Campbell (2009)

is accepted if the revenue is higher than the sum of

bid prices. Bid prices can be approximated through
linear programming.
In capacity allocation, fare prices are given as
they are determined by market forces. Another way
to manage uncertain demand for fixed capacitybe
it flight seats, hotel rooms, rental cars, or inventory
in a retail environmentis through pricing. As argued by Talluri and Van Ryzin (2004, p. 175), the
distinction between quantity and price controls is
not always sharp (for instance, closing the availability of a discount class can be considered equivalent
to raising the products price to that of the next
highest class). However, using price as a direct
mechanism to match demand with capacity is an
important enough practical problem to merit special
treatment. Dynamic pricing has gained significant
traction lately, particularly in retailing (i.e., markdown pricing), e-commerce, and even manufacturing (e.g., Fords offering of incentives at its auto
dealers). The key is to find a good predictive demand
model: At price p, what is the expected demand
d( p) for the product? Demand models may be linear
(d( p) = a-bp), exponential (i.e., constant elasticity), logit (i.e., S-curve), or discrete-choice. There
are many vendors of dynamic pricing software, and
software calibrates the demand models using historical point-of-sale data. In addition, data on available
inventories is necessary for the price-optimization
algorithm. Different price-optimization algorithms
are embedded in these packages based on non-linear
and dynamic programming.

7. Conclusion
Supply chain management is a fertile area for the
application of analytics techniques, which has historically been the case through the use of operations

research, particularly linear programming and

optimization. For example, inventory theory is more
than 50 years old, and there were significant
contributions to production planning in the 1980s.
Therefore, analytics in supply chain management is
not new. More recent applications include the integration of price analytics and supply chain management in the field of revenue management, for which
the problem revolves around managing demand in
an environment with fixed and perishable capacity.
Revenue management research and practice (particularly in manufacturing) is relatively new because
many demand models can only be calibrated with
significant amounts of data, which just recently
became available from modern ERP systems and
web technologies.
With big data, new opportunities arise. I have
heard consultants praising the potential of harnessing social networks for supply chain management,
for example, by detecting local trends in demand to
adjust inventories and prices. There is indeed potential there, although many firms still struggle to
match basic supply and demand in a world with
increased product proliferation, competition, and
globalization (i.e., longer lead times). Among other
benefits, big data has the potential to improve
demand forecasting methods, detect supply chain
disruptions, and improve communications in supply
chains that are often global (see Table 7).

References
Ahuja, R., Magnanti, T., & Orlin, J. (1993). Network flows:
Theory, algorithms, and applications. Upper Saddle River,
NJ: Prentice Hall.
Beil, D. (2010). Procurement auctions. In J. J. Hasenbein (Ed.),
2010 tutorials in operations research: Risk and optimization
in an uncertain world. Linthicum, MD: INFORMS.
Campbell, G. (2009). Overview of workforce scheduling software.
Production and Inventory Management Journal, 45(2), 722.
Cisco. (2013). Building a collaboration architecture for a global
supply chain. Retrieved June 14, 2013, from http://www.
cisco.com/en/US/solutions/collateral/ns340/ns858/c22613680_sOview.html
Farasyn, I., Humair, S., Kahn, J., Neale, J., Rosen, O., Ruark, J.,
et al. (2011). Inventory optimization at Procter & Gamble:
Achieving real benefits through user adoption of inventory
tools. Interfaces, 41(1), 6678.
Fite, J., Taylor, G., Usher, J., English, J., & Roberts, J. (2002).
Forecasting freight demand using economic indices. International Journal of Physical Distribution and Logistics Management, 32(4), 299308.
Funaki, K. (2009). State of the art survey of commercial software
for supply chain design. Retrieved June 18, 2013, from http://
www.scl.gatech.edu/research/supply-chain/GTSCL_
scdesign_software_survey.pdf
Funke, B., & Gruenert, T. (2005). Local search for vehicle routing
and scheduling problems: Review and conceptual integration.
Journal of Heuristics, 11(4), 267306.

Supply chain analytics

Graves, S., & Willems, S. P. (2000). Optimizing strategic safety
stock placement in supply chains. Manufacturing and Service
Operations Management, 2(1), 6883.
Jacobs, R., Berry, W., Whybark, D. C., & Vollmann, T. (2011).
Manufacturing planning and control for supply chain management. New York: McGraw-Hill.
Krishna, V. (2002). Auction theory. New York: Academic Press.
Lee, H. (1996). Effective inventory and service management
through product and process redesign. Operations Research,
44(1), 151196.
McDonald, R. (2011, November). Inside P&Gs digital revolution.
McKinsey Quarterly. Available at http://www.mckinsey.com/
insights/consumer_and_retail/inside_p_and_ampgs_digital_
revolution
Pinedo, M. (2008). Scheduling theory, algorithms and systems
(3rd ed.). New York: Springer.
Rey, T., Kordon, A., & Wells, C. (2012). Applied data mining for
forecasting using SAS. Cary, NC: SAS Institute.
Rey, T., & Wells, C. (2013, March 7). Integrating data mining and
forecasting. Analytics. Retrieved June 17, 2013, from http://
www.analytics-magazine.org/march-april-2013/765integrating-data-mining-and-forecasting
Rungtusanatham, M. J., & Salvador, F. (2008). From mass production to mass customization: Hindrance factors, structural

605
inertia, and transition hazard. Production and Operations
Management, 17(3), 385396.
Safavi, A. (2005). Forecasting demand in the aerospace and
defense industry. Journal of Business Forecasting, 24(1),
27.
Snyder, L. V., & Shen, Z-J. M. (2011). Fundamentals of supply
chain theory. Hoboken, NJ: John Wiley & Sons Inc.
Talluri, K., & Van Ryzin, G. (2004). The theory and practice of
revenue management. New York: Springer.
Tsikoudakis, M. (2013, January 13). Supply chain disasters and
disruptions can cause lasting reputation damage. Business
Insurance. Retrieved June 14, 2013, from http://www.
businessinsurance.com/article/20130113/NEWS06/
301139968
Wang, W., Ferguson, M., Hu, S., & Souza, G. C. (2013). Dynamic
capacity investment with two competing technologies.
Manufacturing and service operations management, 15(4),
616629.
Ward, J., Zhang, B., Jain, S., Fry, C., Olavson, H., Amaral, J.,
et al. (2010). HP transforms product portfolio management
with operations research. Interfaces, 40(1), 1732.
Zipkin, P. (2000). Foundations of inventory management.
New York: McGraw-Hill.