1 views

Uploaded by Sunny Raj

gslmm

- India at a Tipping Point
- B-Plan -Organic Food Store.pptx
- YTC03447USEN
- Small Loans.docx
- RockNRoll
- MMM
- CME Economic Calendar
- GORAMI-BUSINESS-PLAN.docx
- Articles of Partnership for General Partnerships
- Retail Analytics [Anna-Lena Sachs]
- SMU Project Guidelines
- Chapter 1
- 53609621 Supply Chain and Order Management Analytics
- Kieso_15e_SGV1_Ch09
- Internship
- Manufacturing Value Proposition by Ginesys
- davis grant dye modern resume
- New-Product-Survey.docx
- Adrian Tan
- Samsung

You are on page 1of 169

by

Arzum Akkas

M.Eng., Massachusetts Institute of Technology (2004)

B.S., Istanbul Technical University (2000)

in partial fulfillment of the requirements for the degree of

at the

September 2015

○

c Massachusetts Institute of Technology 2015. All rights reserved.

Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Arzum Akkas

Institute for Data, Systems, and Society

August 31, 2015

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

David Simchi-Levi

Professor of Engineering Systems and Civil and Environmental Engineering

Thesis Supervisor

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Vishal Gaur

Professor of Samuel Curtis Johnson Graduate School of Management

Associate Dean for MBA Programs, Cornell University

Committee Member

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Roy Welsch

Professor of Engineering Systems and Sloan School of Management

Committee Member

Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Munther A. Dahleh

William A. Coolidge Professor of Electrical Engineering and Computer Science

Director of Institute for Data, Systems, and Society

2

Strategies to Reduce Product Waste in the Consumer

Packaged Goods Industry

by

Arzum Akkas

on August 31, 2015, in partial fulfillment of the

requirements for the degree of

Doctor of Philosophy in Engineering Systems

Abstract

The cost of waste for products such as soft drinks, shelf stable dry food, and dairy in

the consumer packaged goods industry is massive, about $15 billion annually in the

U.S.A.

This thesis focuses on waste associated with product expiration since this type of

waste involves both manufacturers and retailers as well as different functional areas

such as production, warehousing, sales, procurement, and store operations. As a

result, the industry has not made much progress in reducing this type of waste.

We study three problems related to product expiration. Chapter 2 presents a

descriptive study examining the root causes of product expiration and their impact on

expiration. Using econometrics and our collaborator’s data, we find that the amount

of expiration can be reduced considerably via a case size reduction. We identify the

next important opportunities in the areas of inventory aging in the manufacturer’s

supply chain and sales incentives, and thus the remainder of this thesis focuses on

these two areas.

Chapter 3 examines the manufacturer’s sell-or-dispose decision for aged inventory.

We develop an optimization model to find the minimum remaining shelf life below

which the manufacturer does not sell the product since the cost of expiration is more

than the sunk cost of production. We use machine learning to approximate optimum

values which can be used as a low cost alternative method. If supply chain managers

are held accountable for the cost of disposed items, they will have an incentive to

better manage inventory. As a result, expiration will be reduced.

Chapter 4 analyses sales-force compensation schemes from the perspective of prod-

uct expiration caused by overselling. We develop a game theoretic model of the deci-

sion process of the manufacturer and the sales representative. We find a compensation

scheme that aligns the interests of the manufacturer and the sales representative pre-

venting overselling while achieving full profit potential for the manufacturer.

Title: Professor of Engineering Systems and Civil and Environmental Engineering

3

4

Acknowledgments

First and most importantly, I would like to thank my advisor David Simchi-Levi.

David has been a valuable source of support. He provided me with opportunities to

develop and excel in research and teaching. Next, I would like to extend my thanks

to the other members of my committee. Throughout the short period in which I

had the great chance to work with him, I learned a lot from Vishal Gaur at Cornell

University. Aside from being a great teacher and mentor to me, he has always been

very kind, caring, and encouraging. I have endless respect for him as a person and as

a scholar. Roy Welsch at MIT provided valuable expertise in statistics and machine

learning. He was always available when I needed help and advice. I enjoyed his warm

and friendly personality and always looked forward to our meetings.

possible without their support. Through my discussions with Paul Hamilton, I found

my research topic. Paul introduced me to the right people in the company, so I could

get a better understanding of the problem. Brian Spearman accepted my research

proposal and provided me with data, financial support, and resources within his

organization. Thank you for your trust, generosity, and kindness. Lisa Sarneso and

Mark Howl, it was truly a great pleasure to work with you and thank you for all your

help and patience.

I am also very grateful to the members of the joint industry unsaleables committee,

in particular to Ted Lechner, Gary Regina, and George Thrower. Thank you for

supporting me, educating me, welcoming me in your committee meetings, and giving

me tours of your facilities. Your passion, dedication, and good heart inspired me.

Systems Division, Operations Research Center, and Sloan School of Management:

Alexandre Jacquillat, Vivek Sakhrani, Ross Collins, Erica Gralla, Jesse Sowell, Tom

Heaps-Nelson, Nataly Youssef, Chaithanya Bandi, Gerardo Pelayo, Cecilia Zenteno,

Ozge Karanfil, and Gokce Basbug. Thank you for providing feedback on my ideas

and my research presentations. Above all, thank you for being good friends.

5

I am also very grateful to the faculty, researchers, students, and staff at the MIT

Center for Transportation and Logistics (CTL). Even though it has been eleven years

since I graduated from the CTL master’s program in Supply Chain Management, you

still made me feel part of the CTL family throughout my doctoral years. Thank you

for letting me attend CTL’s research seminars and awarding me a fellowship.

My family and friends at MIT, in Boston, in Turkey, or elsewhere, thank you for

your love, support, and all the good times: Hasan Arslan, Yasemin Dalgali, Edan

Bashan, Adil Akbay, Michael Lester, Michel Alexandre-Cardin, Aziz Abdellahi, Firat

Guder, Claudia Octaviano, Goksin Kavlak, Burcu Balcik, Fusun Demirci, Ece Sancak,

Jale Boyacioglu, Zeynep Kirmizigul, Namik Kemal Caliskan, and Safiye Erdemir.

Finally, I would like to thank my father, Yuksel Akkas, for my inheritance which

made it possible for me to complete my education without financial stress. Thank

you and rest in peace.

6

Contents

1 Introduction 17

1.1 General background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1.2 Financial and environmental impact . . . . . . . . . . . . . . . . . . . 21

1.3 Industry efforts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1.4 Research questions and thesis outline . . . . . . . . . . . . . . . . . . 29

1.5 Industry collaborator . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.3 Research context and hypotheses . . . . . . . . . . . . . . . . . . . . 42

2.3.1 Unsaleables cost and reduction efforts at AlphaCo . . . . . . . 42

2.3.2 Simulation benchmark . . . . . . . . . . . . . . . . . . . . . . 43

2.3.3 Drivers of expiration . . . . . . . . . . . . . . . . . . . . . . . 44

2.4 Data description and estimation model . . . . . . . . . . . . . . . . . 49

2.4.1 Data description . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.4.2 Zero-inflated negative binomial regression model . . . . . . . 54

2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.5.1 Estimation results . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.5.2 Effect of sales incentives . . . . . . . . . . . . . . . . . . . . . 61

2.5.3 Endogeneity between supply chain aging and expiration . . . . 62

2.5.4 Counterfactual analysis . . . . . . . . . . . . . . . . . . . . . . 63

2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

7

Appendix A: Calculation of rotation measure . . . . . . . . . . . . . . . . 70

Appendix B: Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

ucts 81

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

3.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

3.3 The remaining shelf-life problem . . . . . . . . . . . . . . . . . . . . 85

3.3.1 Impact of manufacturer’s operations on perishability at the re-

tailer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

3.3.2 Supply chain cost dynamics and incentives . . . . . . . . . . . 87

3.3.3 Shipment policies in practice . . . . . . . . . . . . . . . . . . . 90

3.4 Problem definition and data description . . . . . . . . . . . . . . . . . 93

3.5 Determining targets for the remaining shelf life . . . . . . . . . . . . . 97

3.5.1 Solution approach and results . . . . . . . . . . . . . . . . . . 97

3.5.2 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . 102

3.5.3 Machine learning approximation . . . . . . . . . . . . . . . . . 110

3.6 Comparison with current practice . . . . . . . . . . . . . . . . . . . . 118

3.6.1 Existing practice at AlphaCo . . . . . . . . . . . . . . . . . . 118

3.6.2 Optimization versus existing practice . . . . . . . . . . . . . . 119

3.6.3 Machine learning approximation versus existing practice . . . 123

3.7 Conclusion and future research . . . . . . . . . . . . . . . . . . . . . 125

Expiration 131

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

4.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

4.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

4.4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

4.5 Conclusion and future research . . . . . . . . . . . . . . . . . . . . . 151

8

5 Concluding Remarks 155

9

10

List of Figures

1-2 Unsaleales types by channel . . . . . . . . . . . . . . . . . . . . . . . 19

1-3 Disposition methods of unsaleables . . . . . . . . . . . . . . . . . . . 20

1-4 Impact of unsaleables on company profits . . . . . . . . . . . . . . . . 22

1-5 Unbalanced incentives with existing unsaleables policies. . . . . . . . 24

1-6 AlphaCo’s outbound supply chain. . . . . . . . . . . . . . . . . . . . 32

2-2 Simulated versus actual expiration for a sample of 40 SKUs. . . . . . 44

2-3 Occurrence of different case sizes at AlphaCo. . . . . . . . . . . . . . 50

2-4 Coefficient estimates of store type indicators. . . . . . . . . . . . . . . 60

2-5 Histogram of supply chain age measure. . . . . . . . . . . . . . . . . . 65

2-6 Impact of addressing order inflation due to minimum order rule. . . . 66

3-2 Impact of an unsynchronized product introduction on remaining shelf

life. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

3-3 Circumstances under which manufacturer is responsible for expiration

cost. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3-4 Profit implications of sales and disposal decisions when remaining shelf

life is very short. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

3-5 Types of shipment policies in practice. . . . . . . . . . . . . . . . . . 91

3-6 Impact of demand rate on expiration with regards to remaining shelf

life. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

11

3-7 Net profit from sales and disposal for one SKU as remaining shelf life

changes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

3-8 Histograms of algorithm’s output. . . . . . . . . . . . . . . . . . . . . 101

3-9 Sensitivity of no-shipment threshold to variation in demand. Square

(in red), diamond (in blue), and triangle (in green) points represent the

results with the decreased, baseline, and increased demand scenarios,

respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

3-10 Histogram of store counts by SKU. . . . . . . . . . . . . . . . . . . . 105

3-11 Variation in no-shipment thresholds across SKUs with different store

samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

3-12 Variation in no-shipment thresholds across SKUs based on different

rotation assumptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

3-13 Examples of SKUs where rotation has no impact (first graph) and

rotation has an impact (second graph) on the no-shipment threshold. 109

3-14 Approximation of optimum values with machine learning. . . . . . . . 112

3-15 Prediction process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

3-16 Performance of regression methods in prediction of no-shipment thresh-

old. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

3-17 Performance of regression methods in prediction of profitability threshold.117

3-18 Existing shipment policies at AlphaCo. . . . . . . . . . . . . . . . . . 119

3-19 Existing versus optimum minimum remaining shelf life. . . . . . . . . 120

3-20 Existing versus optimum target remaining shelf life. . . . . . . . . . . 121

3-21 Dependence of simulation approach on optimum and existing minimum

shelf life. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

3-22 Impact of optimization on profits. . . . . . . . . . . . . . . . . . . . . 123

3-23 Simulation interval in comparison of three policies: existing, optimum,

and approximation based. . . . . . . . . . . . . . . . . . . . . . . . . 124

3-24 Impact of machine learning approximation on profits. . . . . . . . . . 125

12

4-2 Assumed functions of order quantity, 𝑄(𝑒), and effort cost, 𝐶(𝑒). . . . 142

4-3 Global optimization versus two-stage problem. In the example, 𝑝 = 1,

𝑐 = 4, 𝑥 = 2, and 𝑟𝑎 = 0.45. . . . . . . . . . . . . . . . . . . . . . . . 146

4-4 Relationship between coordinating commission rate and penalty fee for

𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0. In the example, 𝑝 = 1 and 𝑐 = 2. . . . . . . . . . 148

4-5 Relationship between optimum commission rate and penalty fee when

𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 < 0. In the example, 𝑝 = 1 and 𝑐 = 10. . . . . . . . . . 149

13

14

List of Tables

2.2 Characteristics of the sales incentives included in the models. . . . . . 73

2.3 Summary statistics of the variables included in the model. . . . . . . 74

2.4 Estimation results for the ZINB model specified in Section 2.4.2. . . . 75

2.5 Parameter estimates of alternative models . . . . . . . . . . . . . . . 76

2.6 Results of the analysis of interaction between demand, sales incentives,

and expiration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

2.7 Results of the Hausman test for endogeneity. . . . . . . . . . . . . . . 79

3.2 Distribution of difference in no-shipment threshold with varying as-

sumption on rotation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

3.3 Features used in prediction of no-shipment and profitability thresholds 113

15

16

Chapter 1

Introduction

The cost of consumer packaged goods products, such as soft drinks, shelf-stable dry

food, and health and beauty aids, that turn into waste at retail stores is estimated to

be $15 billion in 2008 which represents 1 to 2% of retail gross sales in the USA [30].

The consumer packaged goods (CPG) industry usually uses the term unsaleables to

refer to certain categories of product waste. Traditionally, unsaleables are defined as

products that are removed from a primary distribution channel, e.g. supermarkets,

mass merchants, pharmacies, for reasons such as damage, discontinuation, or expira-

tion [30]. This definition excludes product recalls, regular returns, or products that

are damaged or expired at secondary markets, as those items are managed outside of

unsaleables policies. A typical unsaleables policy defines the reverse logistics processes

for returning items as well as the terms of the reimbursement that the manufacturer

provides the retailer. The term unsaleables usually does not apply to produce and

fresh meat categories, since manufacturers typically do not reimburse retailers for un-

sold products in these categories. Figure 1-1 provides the distribution of unsaleables

cost across different product categories based on a recent industry survey [30]. Both

for manufacturers and retailers, the majority of the unsaleables cost (51% and 37%,

respectively) is coming from the edible shelf-stable dry category which includes prod-

ucts such as pasta, cereal, and canned food. The second largest category is refrigerated

17

grocery (e.g., dairy) which generates 16% and 24% of the total cost for manufacturers

and retailers, respectively. The third largest category is health & beauty aids (e.g.,

shampoos, cosmetics, vitamins and supplements, etc.). Health & beauty aids make up

17% of the unsaleables cost for manufacturers and 16% for retailers. Frozen grocery

is the fourth largest category generating 4% and 11% of the cost for manufacturers

and retailers, respectively. Finally, non-edible shelf-stable (e.g., paper towels, laundry

detergents, dog food, etc.) and general merchandise (e.g., clothing, light bulbs, etc.)

are the smallest categories each making up approximately 6% of the total unsaleables

cost. Observe that a difference exists between manufacturers and retailers since the

numbers are based on survey data from a sample of firms.

60%

Percent of total unsaleables cost by category

50%

51%$

37%$ Retailer cost

30%

20% 24%$

16%$ 17%$16%$

10%

11%$

7%$ 6%$ 4%$ 5%$ 6%$

0%

Edible shelf Non-edible Refrigerated Frozen grocery General Health &

stable dry shelf stable grocery (dairy) merchandise beauty aids

18

Types of unsaleables

Generally, unsaleables span three different types: damage, expiration and product

discontinuation. Each type has different sources of root causes. For instance, damages

typically occur due to weak packaging or poor handling practices, whereas reasons of

expiration include poor inventory management or forecasting, aging of products at

the warehouses, or lack of shelf rotation1 . Product discontinuation decisions are made

either by the manufacturer or the retailer, typically to make space for new products, at

the times of planogram2 resets, or as part of an SKU rationalization process. Figure 1-

2 provides the breakdown of different types of unsaleables by channel based on a 2008

industry study [30]. Accordingly, damage type constitutes the majority of unsaleables,

consistent across different channels (grocery stores, drug stores, wholesalers, and mass

merchants). However, as we will discuss later, the composition has been changing over

the years, with the damage portion declining and expiration growing.

100% 4% 4% 8% 10%

19% 20%

22% 34% 17%

80%

21% 12% 5%

60% 25%

45% 58%

56% 21%

40% 48%

40%

31%

20%

0%

Grocery Drug Wholesaler Mass All retailers

Damaged Discontinued/Seasonal Expired Unknown/Other

1

Shelf rotation is the practice of first-in-first-out inventory management at retail shelves. In other

words, rotation activity involves stocking the newer products to the back of the shelf while pulling

the older ones to the front.

2

Planograms are layouts of shelf configuration.

19

Reverse logistics

Reverse logistics of unsaleables can be very expensive. In forward logistics, prod-

ucts are handled in bulk either in pallets or cases. In reverse logistics, products are

handled in units resulting in high labor cost. The CPG company we have been col-

laborating with found that the reverse logistics cost of a product is as large as its

cost of goods sold plus the cost of transporting it to the store. Unlike durable goods

products, such as printers or appliances, product recovery or recycling is typically

not cost effective for unsaleables. Salvageable unsaleables, usually all categories ex-

cept dairy and frozen, are picked from retail stores and are forwarded to dedicated

warehouses called reclamation centers or return centers which are run either by third

party logistics companies or by retailers. At these centers, they are sorted based on

each product’s method of disposal, according to its unsaleables policy. From here,

the majority of them are donated to food banks (35%), some are sold to secondary

markets (26%), and a sizable portion of them are disposed of in landfills (17%). Fig-

ure 1-3 provides distribution of the common disposition methods of unsaleables [30].

1% 2%

Salvage

19% 26%

Donation

Destroyed (landfill)

17% Return to vendor

Recycle

35%

Unknown

20

1.2. Financial and environmental impact

implications. As mentioned earlier, the total cost of unsaleables is estimated to be

1 to 2 percent of retail gross sales. These are significant numbers considering the

tight profit margins in the CPG industry. For example, for a retailer with 2% profit

margin -typical for successful retailers- 1 to 2% of gross sales would be equivalent to

50% to 100% of its profit. In practice, retailers are reimbursed by manufacturers for

this cost in the USA via unsaleables policies; therefore, in reality the impact on the

retailer is not as dramatic. However, it still can be significant because manufacturers

do not always fully reimburse retailers. For instance, if a retailer were to be paid

back 80% of the cost, the remaining cost would be equivalent to 10 to 20% of its

profit. To illustrate this, consider a mid-size regional retailer such as Giant Eagle

with $9B annual revenue. In this case, the retailer’s remaining cost corresponds to

$18M to $36M annually. On the other hand, for a national retailer such as Kroger with

$80B revenue, this number equals $160M to $320M. These are remarkable numbers.

Furthermore, our informal conversations with retailers reveal that for some of them

the cost of unsaleables exceeds their annual profit. Perhaps those are retailers with

profit margins less than 2%, which is not uncommon.

The impact of the cost of product waste is equally noteworthy for manufactur-

ers. Consider a manufacturer with about 10% profit margin, which is typical for

CPG manufacturers. For an 80% reimbursement level paid by the manufacturer to

the retailer, 1 to 2% of retail gross sales could be equivalent to 11 to 21.3% of the

manufacturer’s profit3 .

Figure 1-4 illustrates the impact of the cost of unsalables on the retailer and

manufacturer’s profits as a function of the reimbursement percentage provided by

the manufacturer. The x-axis represents the percentage of the unsaleables cost that

is reimbursed by the manufacturer; the y-axis represents the impact of observed

3

Cost of Goods Sold (COGS) at retailers’ income statement usually corresponds to how much

they pay manufacturers for the goods they purchase. COGS/Revenue ratio is typically 0.75 and

accordingly, we assume that manufacturer’s wholesale price is 75% of retail price.

21

unsaleables cost on the retailer or manufacturer’s profit due to reimbursement policy.

As evident in this graph, the impact of unsaleables on bottom-lines is not trivial,

both for manufacturers and retailers.

100%

90%

80%

70% Manufacturer

60% Retailer

on profit

50%

40%

30%

20%

10%

0%

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Percentage of unsaleables cost reimbursed by

manufacturer

impact, consider the large CPG manufacturing company that we collaborated with.

We estimate that this company dump in landfills approximately 500,000 tons4 every

year of waste associated with unsalebales. This is quite substantial since according

to the Environmental Protection Agency, despite all recycling efforts in the US since

1990, the total amount of municipal solid waste going to landfills was reduced by

only 400,000 tons per year [15]. Hence, unsaleables is not an issue that only impacts

corporations’ profits, but also society due to a massive burden on the environment.

Given the substantial cost of unsaleables, the industry spends considerable effort for

the problem through annual conferences and research studies. Two industry organiza-

4

The warehouse manager we interviewed revealed that one full truckload of waste is picked up at

his facility every week to be disposed at landfills. Assuming 350 warehouses of this company follow

the same disposal method, the total amount of annual waste would be 350x52=18,200 trailers which

approximately corresponds to 500,000 tons.

22

tions, the grocery manufactures association (GMA) and the food marketing institute

(FMI), formed a joint-industry unsaleables committee with representatives from both

manufacturers and retailers. The committee members include national and regional

retailers such as Walmart, Harris Teeter, H-E-B, Winn-Dixie, etc. and manufactur-

ers of varying sizes such as Procter & Gamble, Kellogg and Heinz. The committee

organizes the annual unsaleables conference and determines the research studies.

At the annual unsaleables conference, manufacturers and retailers share best prac-

tices in unsalebles reduction and reverse logistics; also, collaboration awards are given

for successful partnerships between manufacturers and retailers that improve un-

saleables by attacking specific root causes. Further, conferences have been market-

places for secondary market buyers to meet retailers and arrange purchase agree-

ments. Since retailers do not want to cannibalize their own sales, it is common that

a regional retailer sells unsaleables to a secondary market grocery store at a distant

location (e.g., a Texas based retailer sells to discount stores in North Carolina). The

conference also functions as a trade show for third party reverse logistics companies

(e.g., Inmar, Genco, etc.) and packaging suppliers.

Manufacturer-retailer conflict

Generally, a tension exists between manufacturers and retailers arising from bad

practices. This tension is apparent at the annual unsaleables conference, joint in-

dustry meetings, and in overall interactions between manufacturers and retailers in

reverse logistics management. For example, manufacturers object to the practice

where retailers buy products at discount (due to trade promotions) but reclaim the

full price from the manufacturer when the product becomes unsaleable. Retailers

respond by saying that their business systems lack tracing capabilities to associate

an expired item with a historical shipment. Moreover, retailers find it damaging for

the channel that manufacturers sometimes, as a punishment, deduct the cost of un-

saleables (which manufacturers are billed for) from the marketing fund manufacturers

provide to retailers. This fund is normally used for promoting retail sales. In addition

to the mutual discontent about bad business processes, it is typical for manufacturers

23

and retailers to blame each other for the root causes of unsaleables. For example,

according to some manufacturers, expiration occurs since retailers do not rotate the

shelves and damage occurs because retailers do not handle products carefully. Simi-

larly, retailers complain that products have very little shelf life remaining when they

arrive at retailers’ supply chains resulting in expiration; also, manufacturers’ poor

packaging makes products vulnerable to damage.

Another problem, from the perspective of retailers, is the increasing adoption of

an unsaleables policy, called the adjusted-rate-policy (ARP), that favors the manu-

facturers. Historically, manufacturers have reimbursed the retailers for the full cost of

observed unsaleables under the joint-industry-report (JIR) policy. The ARP policy

simply provides a wholesale price discount for the future cost of unsalebales, sup-

posedly based on supply chain audits aimed at determining how much each party

contributed to the problem. Retailers see this policy as manufacturers washing their

hands of the problem. These two policies represent two extreme situations in which

the risk associated with observed unsaleables is allocated among channel partners.

See Figure 1-5.

0% 100%

ARP policy: Manufacturer Retailer

Further, audit methodologies usually are not transparent to retailers which raises

suspicion about their efficacy among retailers. Also, some retailers find the newer

unsaleables policy unjust since the policy does not differentiate among channels. For

example, drug stores typically have higher rates of unsaleables compared to the gro-

cery channel but the manufacturer’s policy provides the same discount rate to all

retailers (e.g., 2%). As a result, some retailers shut down their facilities to auditors

refusing to collaborate. Over the years, the representation in the joint-industry un-

24

saleables committee from the retailer side has diminished, perhaps due to discontent,

perceived lack of progress and manufacturer dominance in the issue.

In addition to the unsaleables conference, industry studies are another source of

effort aimed at alleviating unsaleables. GMA and FMI collaborate with consulting

companies for research studies to understand unsaleables and provide suggestions for

firms to mitigate the problem. Most existing studies use survey methods to collect

data from retailers and manufacturers to show the overall state of unsaleables in the

industry. For example, studies report unsaleables distribution by types (e.g. 50%

damage vs 43% expiration), unsaleable rate by channel (e.g., 1.1% in supermarkets

vs 5.1% at drugstores), and unsaleable rate by product category (e.g., 0.8% non-

refrigerated food vs 1.5% dairy products) [30]. The surveys also collect information on

respondent belief on leading root causes of unsaleables. Of course, survey methods do

not provide objective insights on leading causes of unsaleables. Indeed, manufacturers

and retailers have different views on the leading causes of unsaleables, as we discussed

earlier, with each party associating the main cause of a certain type of unsaleable with

the other party.

Industry reports also include case studies to demonstrate practices that have been

successful at reducing unsaleables [49]. Recent case studies give us a good picture

of where the industry stands at tackling the unsaleables problem. Table 1.1 lists a

sample of case studies from a joint industry project called "Reverse Supply Chain

Improvement" conducted by Raftery Resource Network company in 2011.

25

Table 1.1: Examples of industry case studies.

Involving sales peo- manufacturer education, incen- Unsaleables reports are reviewed for improvement op-

ple to solve prob- tives portunities at salesforce performance reviews.

lems

Shelf life manage- retailer data reporting By analyzing the return data for one manufacturer,

ment the retailer finds that a high percentage were expired

products. Unsaleables policy is updated to shift more

of the unsaleables cost to the manufacturer.

Sales-potential retailer shelf space man- Shelf space is allocated according to sales potential.

planograms agement Slow-moving products are discontinued. Instead of

filling the shelf space with inventory, space fillers and

product fronting techniques are used to reduce on-

shelf inventory.

Engaging the store retailer incentives All stores receive a standard credit for unsaleables

leaders and are charged for what they return. Stores that

return excessive levels are penalized; below average

stores are rewarded.

Six sigma reclama- retailer business process Instead of shipping discontinued products to reclama-

tion optimization tion centers, the retailer uses markdown management

at stores to reduce reclamation center expenses.

Keeping good prod- manufacturer business process Salespeople call customers after promotions to man-

uct out of reclaim age excess inventory either through markdown funds

or returns.

Partnership studies manufacturer audit Opportunities to reduce damage were identified by

collaborating with retailers. E.g., Reduced case

headspace, which reduced top crushing; A retailer

changed unloading practices, which reduced dock

damage.

Demand sales fore- manufacturer analytics Daily store-level POS data is used to improve fore-

casting casting which resulted in lower inventory at the cus-

tomer warehouse and stores.

Increasing available manufacturer data reporting Reclamation center data is collected to understand

shelf life types of unsaleables. Over 80% of the product was

found to be expired. Research is being conducted to

identify acceptable methods for extending shelf life.

26

Based on the existing case studies, some of which we listed in Table 1.1, we make

the following observations regarding the current state of unsaleables.

∙ The industry is struggling with basic issues such as unsaleables data collection.

∙ The distribution of unsaleables type (e.g, 70% expiration, 20% damage) varies

by the supply chain (i.e., manufacturer-retailer combination). The composition

of different unsaleables types is not known by every firm.

∙ Firms pick one issue and attack the related problem (e.g., forecasting, planograms,

damage, discontinuation, incentives, etc.). The associated project is not nec-

essarily part of a roadmap prioritizing opportunities based on a cost-benefit

rank.

cesses, data collection and reporting. Existing analytics is limited with sum-

mary statistics (e.g., volume or percentage of expired/damaged unsaleables by

location/SKU). Use of advanced analytical techniques is not common.

∙ There are many more projects on product discontinuation and damage and

than on expiration. Perhaps, discontinuation and damage are relatively easier

matters to address using simple tools such as business processes, data reporting,

and audits.

∙ Audits are effective at understanding the root causes of damage. Also, root

causes of damage are relatively straightforward matters (e.g., specific packaging

issues, loading/unloading environment, etc.)

∙ There are more examples from retailers than manufacturers leveraging incen-

tives and accountability to reduce unsaleables.

27

∙ Manufacturers see the new unsaleables policy as a tool to reduce expiration

since the policy addresses the retailer’s incentive problem. However, this policy

does create a new incentive problem about the remaining shelf life.

Given the importance of unsaleables for firms’ profitability, ironically, the current

state of unsaleables is rudimentary. For example, firms still struggle with basic issues

such as collecting unsaleables data. There are several impediments to progress. First,

scanning and sorting unsaleables (into categories as damaged, expired, discontinued)

is costly. When the unsaleables policy does not reimburse the retailer for the cost

of observed unsaleables (product cost plus the reverse logistics cost), some retailers

choose to compile different unsaleables types (without sorting) during the disposal

process to save labor cost. As a result of this practice, unsaleables type (i.e., damage,

expiration, discontinuation) data is lost. Since each type of unsaleables occurs due to

completely different dynamics, unavailability of granular unsaleables data makes it a

challenge to take effective actions to reduce unsaleables.

Another obstacle to progress is incentive issues arising from the existing un-

saleables policies. The policy that reimburses the retailer for observed unsaleables

creates a motivation problem for the retailer. For example, the retailer may not care

to handle products cautiously (to prevent damage) or to better manage inventory (to

prevent expiration) when the manufacturer provides full reimbursement. Likewise,

the newer policy reimbursing the retailer for future unsaleables through a wholesale

price discount causes incentive issues for the manufacturer. For example, the man-

ufacturer may not have an interest to use durable packing (to prevent damage) or

to ship fresh products (to prevent expiration). Similar incentive issues exist within

a firm since different functions contribute to the problem while usually one function

absorbs the cost (e.g., logistics, sales, etc.). Perhaps, the reason for such unbalanced

accountability is the lack of understanding about the precise drivers of unsaleables

(i.e., what portion of unsaleables is caused by one particular function or firm versus

the other).

Finally, the unsaleables problem has not been addressed effectively in past because

traditionally reverse logistics has not been considered a C-level subject, and thus un-

28

saleables have not received the attention and support it deserves from top executives.

At some firms, in particular at manufacturers (as we are told by industry leaders dur-

ing informal conversations), unsaleables fall under the responsibility of unpromising

managers suggesting that reverse logistics is not a function that a successful man-

ager aspires to be in. We see a parallel between today’s reverse logistics and the 80s

procurement organization. Before the 80s, procurement was generally seen as an ad-

ministrative function. During the 80s and 90s, procurement gained importance (and

got rebranded as "strategic sourcing") as firms started to realize that improvements

in procurement enable firms to cut costs and increase profits substantially. We hope

to see a similar shift with unsaleables in the future.

expiration since it is a more pressing issue for the industry compared to the other

two types of unsaleables (i.e., product damage and discontinuation). As evident from

existing industry studies, damage can be reduced either through durable packaging

or careful handling; waste due to product discontinuation can be mitigated by im-

proving business processes and communication between channel partners. Drivers

of product expiration, however, are complex spanning across many functions (man-

ufacturing, warehousing, transportation, forecasting, sales, procurement, and store

operations) which makes it a challenge to determine an efficient strategy to reduce

expiration. Furthermore, perhaps due its complexity, the product expiration cate-

gory of unsaleables has been exhibiting growth over the years and quickly becoming

a larger portion of the problem, while unsaleables rates have stayed steady over the

years [30]. For example, in case of our industry partner, product expiration accounts

for 65% of total unsaleables.

In this thesis, we aim to provide suggestions for practitioners to reduce product

expiration occurring in retail supply chains. Specifically, we address the following

research questions:

29

1) What are the root causes of product expiration and how substantial are they

relative to each other?

2) How can we set up shipment rules for the manufacturer (to control the remain-

ing shelf life of products moving into the retailer’s supply chain) in order to mitigate

the cost of expiration?

salespeople from overselling (which results in product expiration at the retailer) while

achieving full profit potential for the manufacturer?

start Chapter 2 with a descriptive study analyzing the drivers of product expiration

to get a thorough understanding of the issue. Industry audits have been effective at

understanding damage but not expiration. We use a method that is novel (i.e., econo-

metrics based on historical transaction data) to practitioners to analyze the drivers

of expiration. Our main finding is that expiration can be reduced with reduction

in case sizes. In addition, our framework of analysis provides a roadmap to reduce

expiration based on the impact of each expiration driver on savings. The second and

third research questions arose from our first study. Some of the drivers of expiration

are straightforward to address (e.g., reduce case sizes, reduce store replenishment

frequencies to address minimum order sizes, improve rotation compliance through

regular audits at sample stores). However, other root causes, such as aging of inven-

tory in the manufacturer’s supply chain or sales incentives, are more complex and

cannot be addressed with a simple change in operations. Therefore, the remainder of

this thesis focuses on the latter types of expiration drivers that are complex. Chapter

3 addresses the second research question focusing on the remaining shelf life problem.

Chapter 4 addresses the third research question analyzing sales-force compensation

schemes from the perspective of product expiration. Finally, Chapter 5 completes the

thesis with our concluding remarks.

30

1.5. Industry collaborator

Overview

We collaborate with a large CPG manufacturer, which we will refer to as AlphaCo

in this thesis. AlphaCo is a multi-billion dollar food and beverage company operat-

ing over 50 manufacturing locations and 400 distribution centers in North America.

AlphaCo operates under the direct-store-delivery (DSD) sales & distribution model,

which involves delivering products directly to retail stores bypassing retailers’ distri-

bution centers, as well as managing store inventory. Through DSD, AlphaCo’s over

5,000 sales representatives and 5,000 truck drivers service 200,000 consumer points.

AlphaCo has a multi-tier supply network consisting of plant warehouses, satellite

warehouses, and retail outlets. High velocity items are typically produced in all plants,

whereas low velocity items are produced in a subset of plants. Each retail outlet is

supplied by one warehouse. Also, each satellite warehouse is supplied by one plant

warehouse in full truckloads. Plants that produce low velocity products distribute

these items in full truckloads to other plants. Therefore, each satellite warehouse

receives all items from a single source plant location. Plants also serve as warehouses

servicing local markets. Products that AlphaCo does not produce but buys from

contract manufacturers are either shipped to plant warehouses or directly to satellite

warehouses, only if the volume is high enough for a full truckload. Figure 1-6 depicts

AlphaCo’s outbound supply chain.

31

AlphaCo plant

AlphaCo satellite warehouse

Retail store

Contract manufacturer

Each AlphaCo sales representative is responsible for a fixed set of stores, called a

route, and makes regular visits to the stores according to a fixed schedule. At each

visit, the sales representative creates a return order for damaged or expired products

and moves them to the back room for pick up, then observes the on-hand inventory

and creates replenishment orders. The sales representative is also responsible for re-

stocking shelves from the back room, rotating the shelf, and setting up promotional

displays. Deliveries are made the following day by a different employee, a truck driver,

who also picks up the returns from the back room. A store always receives all of its

deliveries from one warehouse. Stores are categorized by annual sales volume into

two groups: small format and large format stores. Large format stores typically are

supermarkets and mass merchants. Small format stores are drug stores, convenience

stores, and gas stations.

Collaboration

Our first research study, on the drivers of product expiration (Chapter2), as well

as the second study, about the remaining shelf life problem (Chapter 3), are empirical

analysis primarily based on AlphaCo’s archival data (e.g., shipments, returns, orders,

32

inventory counts, point-of-sale purchases, product cost, and product characteristics).

Our third study, on sales-force compensation schemes (Chapter 4), is a mathematical

modeling analysis; however, the study also uses AlphaCo’s data to test some of the

assumptions we make for our model. Overall, through all three studies, we develop

strategies to reduce expiration for AlphaCo as well as for the CPG industry.

33

34

Chapter 2

supply chains

2.1. Introduction

industry costing retailers and manufacturers approximately $6 billion annually1 . This

cost impacts firms’ profits substantially since the industry operates with tight profit

margins (retailers typically have 1-2% gross margin and manufacturers 10%). Chapter

1 provides a detailed discussion on the financial and environmental impact of product

waste (i.e., unsaleables) including the category of expired products.

In this study, we investigate the root causes of product expiration. The CPG

industry typically uses audits and surveys to diagnose the occurrence of unsaleables

[49, 19]. In audit studies, unsaleables are visually inspected at sampled stores or

return centers and a reason code is recorded for each instance of an unsaleable. Such

a visual inspection usually reveals the cause of damage, such as a packing failure

(e.g., weak plastic, case handle, carton burst, nail damage on pallet, etc.). But a

visual inspection of an expired product is not informative. A product can expire

1

Annual cost of unsaleables is $15 billion according to a 2008 industry estimate. Based on industry

surveys, we estimate that about 40% of the total unsaleables volume comes from the expired products

category.

35

on the shelf due to causes that have occurred anywhere during the sojourn of the

product from the factory to the shelf. Batching in production or transportation,

inefficient inventory management at the warehouse, or suboptimal shelf allocation at

the retail store could all cause product expiration. These causes cannot be identified

by examining a product on the shelf after it expires. Thus, audits have been successful

in addressing the causes of damage and product discontinuation, but not expiration.

Surveys, on the other hand, collect information about respondent beliefs on the

causes of unsaleables. Not surprisingly, manufacturers and retailers have different

views on the leading causes of unsaleables [30]. Manufacturers rank rotation prac-

tices at retailers as the major cause of expiration2 , whereas retailers rank code dating3

standards and procedures. Similarly, manufacturers rank product handling as the

leading root cause of damage whereas retailers rank package design. Thus, each asso-

ciates the main cause of certain types of unsaleable with the other. Our interviews at

our collaborator AlphaCo suggest that such differences persist even within the same

company—the sales organization identifies operational practices as the driver of prod-

uct expiration whereas the operations organization argues that sales incentives are the

main reason. Thus, due to dependent events and lack of transparency in the supply

chain, expiration remains an unsolved problem with cost and waste implications for

manufacturers and retailers.

The objective of our study is to propose an econometric model to improve the un-

derstanding of the root causes of expiration in the CPG industry. Unlike the methods

employed by existing studies, our analysis is based on archival data collected from

the entire supply chain to examine the extent to which the occurrence of expiration is

associated with store operations, supply chain performance, and product characteris-

tics. Our primary data source is AlphaCo’s archival system which includes deliveries

2

Rotation refers to the practice of putting the fresher products to the back of the shelf and pulling

the older ones to the front.

3

Manufacturers use either open codes or closed codes on products to assist the store determine

how long to display the product for sale. Open codes are calendar dates that take forms such as

"best buy", "sell by", "use by", etc. whereas a closed code represents a series of numbers. Retailers

contend that it is harder to manage rotation with closed codes. Also, they claim that with the recent

trend to switch from closed code to open code practices, manufacturers have reduced the shelf life

to be more conservative.

36

to and returns from 66,867 retail stores, warehouse inventory counts, product de-

ployment at 449 AlphaCo locations, and shelf life and case size information for 768

products. The data set is from the year 2011.

expect due to the randomness of demand, or is there an opportunity to reduce ex-

piration? We first compare observed expiration with a theoretical benchmark con-

structed by simulating a base stock policy on the sample paths of demand observed

in detailed transaction level data for a small subsample of products. We find that

the average actual expiration is 95.83 times the simulated expiration, significantly

different at 𝑝 < 0.01. This result suggests that expiration occurs due to reasons other

than the randomness of demand, and thus, could be reduced by improving operations

at manufacturers and retailers.

We identify six potential drivers of product expiration: case size, supply chain ag-

ing, sales incentives, forecasting complexity, minimum order rules, and shelf rotation.

Here, we define supply chain aging as the elapsing of a product’s life in the supply

chain before the product reaches the retail shelf. These variables represent different

aspects of practical supply chains, including store execution, back-end supply chain

operation, and product characteristics. We select these variables based on interviews

with AlphaCo, industry reports, and models of perishable inventory management. In

addition, we control for mean demand rate, store type, and product shelf life in the

analysis. Our estimation method is based on count models because expiration is a

nonnegative integer and is bounded above by the total shipment quantity. Using data

from 870,493 store-product combinations, we evaluate several econometric specifica-

tions: binomial, Poisson, negative binomial, and zero-inflated models. We find that

the zero-inflated negative binomial (ZINB) model yields the best fit and the most

unbiased residuals by addressing two characteristics of our data: probability mass at

zero and overdispersion.

Our main result is that case size, supply chain aging, sales incentives, forecasting

complexity, and minimum order rule are all statistically significant determinants of

product expiration. The control variables, demand rate, shelf life, and store type

37

also affect expiration significantly. These results are robust to different clustered

standard errors. Thus, our study shows that expiration occurring at retail shelves

can be caused by both manufacturers and retailers. To refine our results, we analyze

the interaction of sales incentives with demand rate—sales incentives can increase

demand rate, which can lead to underestimating the effect of sales incentives on

expiration. We also examine a potential inverse relationship between supply chain

aging and expiration, i.e., that more expiration may result in more inventory upstream

in the supply chain, which may increase supply chain aging. The Hausman test does

not reveal any significant evidence of endogeneity.

can be achieved by addressing various types of causes of product expiration. We

find that reducing the case size to 12 units for products that are currently packed

in 24 units yields a 41% decrease in expiration volume and a $7.66M decrease in

expiration cost. This constitutes an opportunity for CPG companies to reduce waste

by reducing case sizes. Reducing the days of supply in the supply chain by one week

corresponds to a 0.98% decrease in expiration volume and a $912.4K decrease in

expiration cost. Relaxing the minimum order rule can reduce expiration volume by

0.87% and expiration cost by $610.7K. Accordingly, visit frequencies can be reduced

at stores that have a high frequency of orders equal to their minimum order sizes.

Last, we find that three sales incentives cost $645K, $2,634K, and $25.7K with more

expiration of 14%, 9%, 7% in volume, respectively.

Our study is the first descriptive study of product expiration in the inventory

management academic literature. While the existing literature on perishables inven-

tory management has focused on inventory policies, our research contributes to the

literature by developing an econometric model of expiration, and providing insights

into its sources beyond demand uncertainty. Our analysis is distinct in combining

sku-level data from stores and supply chain and marketing data from manufacturers.

The results of this analysis can facilitate solutions to the problem of expiration by

improving the understanding of its root causes. Our expiration model identifies the

contribution of different actors to the expiration problem, and presents evidence that

38

expiration can be alleviated with better management at retailers and manufacturers.

operations, and sustainability models.

The management of perishables is an important problem in many industries, such

as blood banks, food, and pharmaceuticals. Seminal research in this area was con-

ducted by Nahmias [47] and Fries [16], who analyze the optimal inventory policy

considering expiration and shortage costs under a cost-minimizing dynamic program

and show that the optimal policy is non-stationary and is dependent on the age dis-

tribution of inventory. Nahmias [48] presents an extensive review of the issuance

and replenishment decision models for perishable inventory. Most of this literature

has focused on single-location models and ignored aging in the supply chain, i.e., a

product is available for its full life upon receipt at the retailer.

Ketzenberg et al. [33] consider a two-stage system with supply chain aging and

order batching in order to evaluate the benefit to the retailer from the availability of

product life information at the supplier. This benefit manifests in the retailer ordering

more product when the supplier has fresher inventory available, and less otherwise.

Through a simulation study, the authors show that sharing product life information

increases the retailer’s profit by an average of 4.4%, increases the average remaining

shelf life of retail inventory at the time of replenishment by 8%, and decreases the

incidence of product expiration by 40%. The retailer benefits the most from infor-

mation sharing when the variability of the demand or the remaining shelf life of the

items is high, product lifetimes are short, and the cost of the product is high.

Ketzenberg and Ferguson [34] focus on slow-moving items that are ordered in

single case pack sizes. For a two-stage supply chain with one supplier and one retailer,

they evaluate the value of two supply chain improvements—sharing of inventory and

replenishment information by both partners in a decentralized supply chain, and

centralized control. Using a numerical study, they find that, compared to the base

39

scenario, the total supply chain expected profit increases by 4.2% with information

sharing and by 5.6% with centralization. Further, the benefits of information sharing

or centralized control are minimal when an optimal case size is chosen.

in a real-world context, integrating data from the manufacturer and many retailers

to identify the role of each player, and identifying the role of supply chain execu-

tion variables such as rotation compliance, supply chain aging, case size, and sales

incentives. Whereas the literature has developed optimal policies and useful heuris-

tics for inventory management for perishable products, we develop insights into the

relative effects of different supply chain variables that can be managed to reduce the

occurrence of expiration.

The recent literature on retail operations and supply chain execution is also rel-

evant to our research. Several papers in this literature have studied real-world phe-

nomena through empirical research or analytical models. For example, Dehoratius

and Raman [14] identify the sources of inventory record inaccuracy using hierarchical

linear modeling and emphasize the need to incorporate these sources into inventory

planning tools. Kok and Shang [38], Dehoratius [13] and several others have since

developed inventory planning algorithms under inventory data inaccuracy. Gaur et

al. [61] study the ordering behavior in a supermarket chain and show that store

managers deviate predictably from automated replenishment systems due to factors

such as in-store handling cost, case size, demand rate, produce variety, and demand

variability. Accordingly, they devise a method to improve automated replenishment

systems. Kesavan et al. [32] investigate the relationship between flexible labor re-

sources and financial performance. Corstjens and Doyle [11] study optimal shelf space

allocation among multiple products, considering main and cross space elasticities, in

order to minimize procurement, inventory carrying, and out-of-stock costs. Kok and

Fisher [39] study the optimal allocation of shelf space to an assortment of substi-

tutable products in a category.

unexplored issue in the CPG industry, and identifying its causes, which have only been

40

partially analyzed in theoretical research. We exploit the statistical characteristics of

expiration in practice, i.e., count data with frequent zero observations, to employ zero-

inflated count models for hypothesis testing. Some variables used in our research are

based on the literature. For instance, excess shelf space, whether due to suboptimal

shelf allocation or case size considerations, can lead to excess inventory, which then

can cause product expiration. Thus, our research builds on these topics in retail

operations by showing the waste implications of different aspects of supply chain

execution.

Other aspects of supply chain execution include sales incentives, order batching,

and order inflation. Examples of this work span the literature in operations man-

agement and economics. Chen [8] proposes a salesforce compensation package that

induces a smooth ordering behavior to match the production cycle. He compares this

to a widely used compensation plan based on annual quotas which causes salesper-

sons to concentrate their efforts in the last period. Oyer [50] empirically shows that

salespersons and executives influence the timing of customer purchases, driven by

nonlinear incentive contracts, resulting in business seasonality with high sales at the

end of the fiscal year and low sales at the beginning. Our work adds to this litera-

ture by identifying sales incentives and order batching as factors that cause increased

product expiration.

41

2.3. Research context and hypotheses

The cost of unsaleables consists of the procurement cost of the product, sales & de-

livery cost to place it in the store, and reverse logistics cost. An internal unsaleables

study conducted at AlphaCo in 2010 suggested that the reverse logistics cost is ap-

proximately equal to the sum of the other two cost components. Thus, even though

unsaleables make up only 0.87% of the total sales volume at AlphaCo, the total cost

of unsaleables is equivalent to 50% of the net profit (profit margin at AlphaCo is

approximately 3%). Further, unsaleables have indirect costs such as the opportunity

cost of occupying shelf space that would otherwise be used for saleable products and

the cost of lost goodwill due to consumers switching to competing products. As a

result, unsaleables are a matter of great importance. Expiration comprises about 65%

of the unsaleables volume while damage makes up the remaining 35% at AlphaCo.

Discontinued products go through a phase-out process and eventually enter reverse

logistics once they expire.

AlphaCo conducted a comprehensive study of product waste in 2010. It involved

audits at sample stores to document the root cause for each instance of unsaleables.

The study was able to identify the root causes for damaged products. However, the

causes for expired products were not definitively established. Figure 2-1 presents the

causes of expiration as identified in the study. Note that the second most frequently

cited root cause for expiration is unknown. This shows the inability of audits to

diagnose the causes of expiration.

42

40%

35%

30%

25%

20%

15%

10%

5%

0%

Rota-on
Unknown
Over
Other
Ship
to
Retail
price

issue
ordered
trade
short
point

coded
increase

Figure 2-1: Root causes of expiration based on AlphaCo’s internal audit study.

obtained from one retail store. These products include all items carried at this store

supplied by AlphaCo, i.e., whose UPC codes match AlphaCo inventory IDs. The

products vary in their shelf lives, case sizes, and demand rates. The median, max-

imum, and minimum values of these variables are respectively as follows: shelf life,

14, 104, and 12 weeks; case size, 12, 24, and 1 units; and daily demand rate, 0.12,

0.65, and 0.01 units.

We receive daily point of sale data covering 604 days. Since the store is serviced

once a week by AlphaCo, we aggregate the daily point of sale data by week. 26

out of 40 products had zero sales for some of the weeks, most likely due to product

introductions and discontinuations. To account for such scenarios, we discard obser-

vations prior to the first week of positive sales and after the last week of positive sales.

We construct an empirical demand distribution from the observed sales, and use it to

generate a sample path of 10,000 demand occurrences. Order quantities for each week

are computed based on a heuristic order-up-to inventory replenishment policy with

a 95% in-stock rate. We use a heuristic because the optimal policy for perishables

suffers from the curse of dimensionality and is hard to compute. Each week, an order

equal to the difference between the order-up-to level and the sum of the inventories

43

of different ages is created. We allow single-unit replenishment in the simulation in

order to focus on the amount of expiration due to the randomness of demand; case

size will be included as a variable in the regression model in Section 3.2. We also

assume that inventory is depleted under the FIFO rule.

The average weekly expiration quantity per product in the simulation is 1.9917 × 10−3

units. The actual expired quantity, according to AlphaCo’s product return records, is

1.908 × 10−1 units/week/product, which is 95.83 times the simulated expiration, sig-

nificant at 𝑝 < 0.01 in a one-tailed t-test. Further, only 10/40 products have non-zero

expiration in the simulation result, whereas 23 out of 40 products have expiration in

AlphaCo’s data. Figure 2-2 shows the difference between simulated expiration and

actual expiration for all 40 products. The graph shows that the observed expiration

is substantially larger than the simulated expiration for most of the products. Thus,

demand uncertainty cannot be the only source of expiration at this store.

1.4

Actual
Expira5on
(average/week)

1.2

1

0.8

0.6

0.4

0.2

0

0
0.005
0.01
0.015
0.02
0.025

Simulated
Expira5on
(average/week)

uct at a retail store. Let 𝑆 denote the shelf life of the product, 𝐷𝑡 denote the random

demand in period 𝑡, and 𝑄 denote the amount of inventory shipped to the store at

the beginning of period 1 with zero starting inventory. Thus, the amount that will

44

eventually expire from this batch of shipment under the FIFO issuing policy will be

𝐸𝑄 = [𝑄 − 𝑆𝑡=1 𝐷𝑡 ]+ , which depends on the demand realization, the shelf life of

∑︀

the product, and the shipment quantity. For the same demand realization, a larger

shipment quantity or a shorter shelf life correspond to a higher amount of expira-

tion. Therefore, we expect practices or circumstances that reduce shelf life or inflate

shipments to cause expiration. We discuss these practices identified through our in-

terviews and industry reports, and set up our hypotheses.

case size.

CPG manufacturers ship items in multiples of case size to stores.

In the simulation study in Section 2.3.2, we find that the target inventory level

calculated based on empirical demand distribution is less than case size for 33/40

products. Thus, stores must round up their shipment quantities to one case. Further,

case size cover (i.e., case size divided by mean demand) is greater than shelf life for

15/40 products. In other words, a case of inventory is expected to last longer than

the shelf life for many products, which would result in expiration. Low demand rates

are not unique to this store. According to an industry study [63], nearly half of the

SKUs at retail stores sell less than one unit a week. For such products, shelf life does

not need to be very short for expiration to occur. For instance, expiration will occur

if case size is 24 units and shelf life is less than 6 months or if case size is 12 units

and shelf life is less than 3 months.

A test of this hypothesis is valuable because higher case sizes need not necessarily

correspond to a higher occurrence of expiration if they are correlated with demand

rate and shelf life. Moreover, manufacturers resist decreasing case size because it

increases handling and packaging costs and may even require upgrading production

lines. Due to these costs, it is important for manufacturers to verify the value of

smaller case size in order to make an informed optimal case size decision.

45

Hypothesis 2: The probability of expiration is positively correlated with

supply chain aging.

The probability of expiration is positively correlated with supply chain aging.

Every product has a fixed shelf life at the time of production. Time spent in the

supply chain erodes this shelf life. We call this supply chain aging. It can occur due to

reasons such as production and transportation batching, high safety stocks, and poor

forecasts. By reducing effective shelf life, supply chain aging can increase expiration

for any given level of demand and supply. Further, the effect of supply chain aging

on expiration can be expected to depend on the demand rate, the case size, and the

shelf life.

We measure supply chain aging for each product-store combination as the cumu-

lative average days of supply of that product in its supply chain. AlphaCo has a

multi-tier supply network consisting of plant warehouses, satellite warehouses, and

retail outlets. High velocity items are typically produced in all plants, whereas low

velocity items are produced in a subset of plants. We map this supply chain for each

product-store combination and compute the total average days of supply of the prod-

uct across the stages of the supply chain.

with rotation.

The probability of expiration is negatively correlated with rotation.

Shelf rotation is the practice of placing fresher products in the back of the shelf

while pulling older ones to the front. Rotation facilitates first-in-first-out issuing of

inventory. Perishable inventory theory on the issuance of inventory in LIFO or FIFO

order suggests that rotation affects the occurrence of expiration [48]. According to

the 2008 Joint Industry Report, CPG manufacturers believe lack of shelf rotation

to be the most common root cause of expiration. An internal study by AlphaCo,

however, found that 63% of the time when an expired product is found on the shelf,

the shelf was in fact rotated. Thus suggests that the impact of rotation may not be

as dramatic as believed in the industry.

46

Hypothesis 4: The probability of expiration is positively correlated with

compliance to minimum order rules.

The probability of expiration is positively correlated with compliance to minimum

order rules.

Inflation of order quantities to reduce transportation cost is a common practice.

AlphaCo imposes minimum order sizes for store replenishment orders in order to

reduce delivery costs. Our field trips reveal that original order quantities can be in-

creased to make the total order size equal to the required minimum. This behavior

inflates store inventory, which can be expected to increase expiration. Sales represen-

tatives are measured on their compliance to this rule, which we use as the measure

to test this hypothesis.

manufacturer’s sales incentives.

The probability of expiration is positively correlated with manufacturer’s sales incen-

tives.

CPG manufactures offer a variety of performance incentives to their sales force

which can increase the chances of product expiration. AlphaCo, for instance, offers

its sales representatives not only a sales commission applicable on the overall sales

volume, but also a second layer of rewards for growing sales volume or building

store displays for specific products. AlphaCo calls these reward programs incentives.

Products for which AlphaCo intends to improve market penetration are typically

selected for incentives. Each incentive is valid during a particular month and focuses

on a group of products.

Two types of incentives are offered. One involves competition for the best looking

in-store displays among sales representatives. Products not returned to the warehouse

at the end of the display period generate over supply of inventory at the stores. The

other incentive type involves a sales growth target by a fixed volume or a percentage

compared to the prior year. Typically, these targets are achieved by gaining additional

47

shelf space or permanent displays, which increases store inventory.

Incentives can inflate store inventory as well as stimulate demand. We test whether

incentives have a positive net effect on expiration. Section 2.5.2 discusses how de-

mand, incentives, and expiration interact with each other in our estimation model.

forecasting complexity.

The probability of expiration is positively correlated with forecasting complexity.

Each AlphaCo route is managed by one sales representative. Routes serve varying

number of stores. For instance, a sales representative assigned to large format stores

such as supermarkets may visit as few as 4 stores a day, whereas one assigned to

small format stores may visit up to 15 stores a day. Sales representatives initiate

replenishment orders at store visits by forecasting the demand up to the next replen-

ishment epoch. This forecasting task inevitably gets more complex with increasing

store-product combinations in the route and so does the likelihood for mismanaging

inventory. Thus, we expect that as the difficulty in forecasting increases, sales rep-

resentative will err on overstocking than understocking, which should increase the

chances of expiration. We measure forecasting complexity as the number of store-

product combinations that a sales representative is in charge of managing.

The above drivers of expiration can be classified through different perspectives.

For example, sources of expiration can be classified as (i) drivers reducing shelf life,

and (ii) drivers increasing shipment quantity. Among the root causes listed, case

size, minimum order rule, sales incentives, and forecasting complexity can raise the

shipment quantity beyond the amount required to match uncertain demand, and

therefore, can be considered as drivers increasing the shipment quantity. Supply chain

aging and non-compliance with shelf rotation on the other hand reduce effective shelf

life. Further, from a channel viewpoint, drivers can be categorized as (i) manufacturer-

related, and (ii) retailer-related drivers. Case size, supply chain aging, and sales

incentives can be categorized as manufacturer-related, whereas minimum order size,

rotation, and forecasting complexity can be considered as retailer-related drivers.

48

This classification is helpful in improving supply chain coordination issues in product

expiration.

We obtain delivery, return, supply chain, and marketing data for 768 SKUs and 66,867

stores in the United States for the year 2011. There are 8 store types in our dataset:

supermarkets, convenience stores & gas stations, other grocery (stores bigger than

convenience stores and smaller than supermarkets are categorized as other grocery

at AlphaCo’s business systems), dollar discount stores, drug stores, mass merchants,

club stores, and supercenters. For computational efficiency, we do not estimate our

models on the entire data set. Instead, we draw a random sample of 10,000 stores.

Thus, our final data set consists of 870,493 store-SKU level observations across 768

SKUs and 10,000 stores. We expect results to be generalizable to the full data set

because the sample is fairly large and representative of all store types.

Our data are obtained from three sources: data warehouse, spreadsheets, and

picture files. Different parts of the data have varying levels of aggregation with

respect to time (i.e., day, month, year) and supply chain structure (i.e., warehouse,

route, store). We construct the following variables from these data:

∙ 𝑟𝑒𝑡𝑢𝑟𝑛𝑝𝑠 is an integer-valued variable representing the total number of expired

units for store 𝑠 and product 𝑝 during 2011. 𝑟𝑒𝑡𝑢𝑟𝑛𝑝𝑠 is our dependent variable.

∙ 𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 is a discrete variable representing the net amount of product 𝑝 deliv-

ered to store 𝑠. A store receives all its inventory from a single warehouse and a single

route. Our data set includes total deliveries, total saleable returns, and total returns

due to damage aggregated for 2011 by shipping warehouse, route, store, and product.

Saleable returns are unsold display products that are returned to the warehouse at

the end of the display period. We deduct saleable returns and returns due to damage

from the delivery amount to obtain net delivery amount, 𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 . We use it to

49

represent the number of Bernoulli trials in our rate (i.e., binomial) model, and as the

exposure variable in our count models (i.e., Poisson and negative binomial).

∙ 𝑐𝑎𝑠𝑒 𝑠𝑖𝑧𝑒𝑝 , the explanatory variable associated with Hypothesis 1, represents

the number of consumer units contained in one case of product 𝑝. It is obtained from

the products table in the data warehouse. Case size varies by product. The case sizes

at AlphaCo are 1, 2, 3, 4, 6, 8, 12, 15, or 24 units of products. Figure 2-3 shows the

frequency distribution of case size across all 768 products in our data set.

160

140

120

Number
of
SKUs

100

80

60

40

20

0

1
2
3
4
6
8
12
15
24

Case
Size

∙ 𝑠𝑢𝑝𝑝𝑙𝑦 𝑐ℎ𝑎𝑖𝑛 𝑎𝑔𝑒𝑝𝑤 , the explanatory variable associated with Hypothesis 2, de-

notes the average number of days spent by product 𝑝 in the supply chain before

shipment to all stores served from warehouse 𝑤. We utilize several elements of infor-

mation obtained from the data warehouse to construct this measure. They include

shipments among AlphaCo warehouses for each product aggregated for 2011, individ-

ual physical inventory count records per warehouse-product for 2011, deliveries made

to retail stores by warehouse-product aggregated for 2011, and annual production

quantity by product aggregated for 2011. Using inventory count records, outgoing

shipments, and deliveries, we calculate days-of-supply for each product-warehouse

combination. We also derive the supply chain network using production and shipment

data. As discussed in Section 2.3.3, supply chain age is computed as the cumulative

days-of-supply across the supply chain.

50

∙ 𝑟𝑜𝑡𝑎𝑡𝑖𝑜𝑛𝑟 , the explanatory variable associated with Hypothesis 3, represents the

probability of at least one unrotated product on route 𝑟. We measure rotation at

the route level because it depends on process compliance by sales representatives who

manage different routes. Information about rotation is not stored in any business

system and is not collected for all routes. Therefore, we compute an instrument for

rotation utilizing audit data provided to us in spreadsheets. Using a logistic regression

model, we estimate the probability of rotation for 128 routes included in the audit

study. Coefficients from this model are then used to estimate the probability of

rotation for the remaining 2851 routes. Appendix A describes our method to compute

a measure for rotation.

∙ 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒𝑠 , the explanatory variable associated with Hypothesis 4, is de-

fined as the percentage of orders at store 𝑠 whose size is equal to the minimum order

quantity. Our data contain order ID, store ID, and order quantity aggregated by

order ID. We measure 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒𝑠 as the count of orders with quantity equal

to either 15 or 75, the two minimum quantities imposed by AlphaCo on small and

large format stores, respectively, divided by the total count of orders by store. Since

we have a large amount of transaction data, we calculate this measure using detailed

order data for the first quarter of 2011 only, assuming that the distribution of order

sizes remains the same throughout the year.

product 𝑝. AlphaCo stores sales incentive data in spreadsheets and picture files.

Spreadsheets contain names of the incentives, their dates of effectiveness, a list of

products or product groups covered, and rewards offered to sales representatives. As-

sociated with each incentive, a picture file illustrates the corresponding information in

a poster. We include nine incentives in our data set. We allow a time lag between the

dates of an incentive and the occurrence of expiration because expiration associated

with an incentive is likely to occur with a delay depending on the shelf life of the

product and on the time required to remove expired product from shelves and return

them to the manufacturer. Thus, we include incentives offered in the last six months

of 2010 and the first six months of 2011 in our data.

51

Table 2.2 provides information about the characteristics of incentives. The incen-

tives differ from each other in the times of the year when they are applied, the type

of growth targets, and the set of products to which they are applied. For instance,

incentive 1 is active in the eighth and ninth months of 2010, focuses on a specific

brand consisting of 15 products with relatively short shelf lives, and has an aggressive

growth target of 20%. The median, maximum, and minimum shelf lives are 13, 14,

and 12 weeks. Since this incentive takes place close to the end of 2010, we expect

most of the returns due to expiration to take place in 2011. Incentive 2 is valid dur-

ing the seventh and eight months of 2010. It has a narrower scope compared to the

other incentives and includes only nine products of a specific brand. Their median,

maximum, and minimum shelf lives are 30, 52, and 26. Due to their long shelf lives,

we expect their associated returns from expiration to occur in 2011. Incentive 2 has a

mild growth target with 5%. Similar details for all nine incentive types are presented

in Table 2.2.

∙ 𝑓 𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦𝑟 is the explanatory variable associated with Hypoth-

esis 6. It represents the complexity of forecasting demand as described in Section

3.2. Using deliveries by route, store and product aggregated for 2011, we measure

𝑓 𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦𝑟 as the count of store-product combinations in route 𝑟.

∙ 𝑑𝑒𝑚𝑎𝑛𝑑𝑝𝑠 is a control variable representing the demand for product 𝑝 at store

𝑠. It is a discrete variable and is calculated as the net deliveries (i.e. deliveries after

saleable and unsaleable returns are deducted) by store-product aggregated for 2011.

∙ 𝑠ℎ𝑒𝑙𝑓 𝑙𝑖𝑓 𝑒𝑝 is a control variable indicating the shelf life of product 𝑝 in weeks.

The data are obtained from the products table in the data warehouse.

∙ 𝑠𝑡(𝑘)𝑠 is a binary control variable indicating whether store 𝑠 is of type 𝑘. There

are eight different store types: supermarkets, convenience stores & gas stations, other

grocery, dollar discount stores, drug stores, mass merchants, club stores, and super-

centers.

∙ 𝑐𝑎𝑠𝑒 𝑠𝑖𝑧𝑒 𝑐𝑜𝑣𝑒𝑟𝑝𝑠 is defined as the ratio of 𝑐𝑎𝑠𝑒 𝑠𝑖𝑧𝑒𝑝 to 𝑑𝑒𝑚𝑎𝑛𝑑𝑝𝑠 . It represents

the expected amount of time to sell one case at the store.

Table 2.3 presents summary statistics of all variables. Gas stations and conve-

52

nience stores make up 46% of the data set, supermarkets 22%, other grocery channel

10%, dollar discount stores 4%, drug stores 11%, mass merchants 4%, super centers

3%, and club stores 0.024%. The annual number of returns due to expiration for a

store-product has an average of 2.86 units and a median of zero. About 80% of the

return data consist of zeroes indicating a large mass at zero. The annual number of

deliveries made per store-product has an average of 384.81 units and a median of 120.

Supply chain age has an average of 28.19 days and a median of 22.79 days. The av-

erage is higher than the median showing positive skewness. Case size ranges between

1 and 24, with an average of 12.6 and a median of 12. Shelf life ranges between 10

and 104 weeks, with an average of 21.6 and median of 27.71 weeks. The measure

for forecasting complexity has an average of 4046.64 and a median of 4463, meaning

that the median sales representative manages 4463 store-product combinations. The

average and median values of the rotation metric are 81% and 91%, respectively. The

average and median number of orders equal to the minimum order size are 4.29% and

0, respectively, with 58% of the stores having no orders equal to minimum order size.

Out of nine sales incentives, incentives 1, 2, 3, 4, 6, and 9 cover 3%, 3%, 4%, 6%,6%,

and 9% of the data points, respectively. Incentives 5, 7, and 8 are more prevalent and

include 22%, 13%, and 23% of the data points. The median case size cover is 0.09,

which implies that a one-case shipment covers 0.09 × 365 = 32.85 days of demand at

the median. On average, this number is 0.37, corresponding to 135 days. This large

gap is due to a skewed distribution of demand rate.

We exclude some of the routes due to data accuracy concerns. Prior to 2011, the

versions of the AlphaCo’s hand held system allows product returns to be recorded

only in full cases and not in units. For example, if there were only two units of

expiration found of a given product that has 12 units/case configuration, the sales

representative waited until there are 12 units (i.e. 1 case) of the same product in the

back room before they are returned to the warehouse. Alternatively, it was common

to combine different flavors of the same item together and record it as one case

return, since store managers usually do not want waste waiting in the back room.

This did not impact the credit given to the customer; since same price items were

53

being combined together, however, return records were not always accurate. In 2011,

AlphaCo upgraded the handheld software to allow single unit returns. We notice that

some sales representatives do not use this new feature. For this reason, we exclude

the routes that do not contain any single unit returns in 2011, which is about 10% of

all routes, from the analysis based on the belief that their return records may not be

reliable.

We examine different count models to find the most suitable specification for pre-

dicting expiration: binomial, Poisson, negative binomial and their zero-inflated gen-

eralizations, namely, zero-inflated binomial (ZIB), zero-inflated Poisson (ZIP), and

zero-inflated negative binomial (ZINB) models. A linear predictor that is common

across all six models forms the basis for these models. The predictor contains the

variables associated with six hypotheses about case size, supply chain aging, rotation,

minimum order rule, sales incentives, and forecasting complexity, as well as control

variables for demand, shelf life, and store types. Let 𝐵 (𝑖) denote the 𝑖-th row of the

data matrix 𝐵 and 𝛽 denote the vector of coefficients for the explanatory variables.

Then we set up the predictor as follows:

𝑠 + 𝛽1 · log 𝑑𝑒𝑚𝑎𝑛𝑑𝑝𝑠 + 𝛽2 · 𝑠𝑢𝑝𝑝𝑙𝑦 𝑐ℎ𝑎𝑖𝑛 𝑎𝑔𝑒𝑝𝑤

𝑝 + 𝛽5 · 𝑓 𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦𝑟

𝑟 + 𝛽7 · 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒𝑠 + 𝛽7+𝑗 · 𝑠𝑖(𝑗)𝑝 . (2.1)

Here, 𝑖 indexes observations in our data set, 𝑘 denotes store types, 𝑝 denotes prod-

ucts, 𝑟 indexes routes, 𝑠 denotes stores, and 𝛽𝑘 are the fixed effects for each store type

𝑘. We use 𝑑𝑒𝑚𝑎𝑛𝑑𝑝𝑠 with a log transformation because it reduces heteroscedasticity

and provides higher log-likelihood values and distribution of residuals for all six mod-

els. The other variables do not have this problem, and thus, we find that different

transformations (e.g., log, exponent, square root, square) do not improve the fit in

the same way as for demand.

54

Ideally, the functional specification of the predictor should be derived from a the-

oretical model of expiration. Unfortunately, this is difficult because of the complexity

of the theoretical model. Thus, we analyze different count models and transforma-

tions to improve the model specification. We start our analysis with the binomial

model. The structure of the binomial model inherently suits the nature of our data

because the estimate of expired volume should be limited by the amount of products

delivered to the store over an extended period. Binomial model fits this requirement

with the number of Bernoulli trials providing an upper bound to the estimate of the

response variable. The delivered volume serves as the number of Bernoulli trials and

the expired volume as the number of successes in the binomial model. Using the logis-

tic function to link the linear predictor to the response variable, we use the following

regression form:

𝑝𝑠 /𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 ] = 𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 · exp(𝐵

(𝑖)

𝛽)/[1 + exp(𝐵 (𝑖) 𝛽)] (2.2)

To find the maximum likelihood estimates of this model, we use the iteratively

reweighted least squares (IRLS) method under the generalized linear model (GLM)

framework [43]. An analysis of the residuals indicates a negatively skewed distribu-

tion. This leads us to explore alternative models. We test the Poisson model, the

other count model within the GLM family, also using the IRLS algorithm for param-

eter estimation. We utilize an exposure variable to establish an upper bound on the

estimate of the response variable. The exposure variable enters the data matrix as

an offset with a log transformation and its parameter is constrained to one. Using

the exponential link function and 𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝,𝑠 as the exposure variable, the regression

takes the form:

𝐸[𝑟𝑒𝑡𝑢𝑟𝑛(𝑖) (𝑖)

𝑝𝑠 ] = 𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 · exp(𝐵

(𝑖)

𝛽) (2.3)

olated with real data [26, 43]. Overdispersion in Poisson models refers to the cases

when the response variance is greater than the mean. A negative binomial model,

which is a gamma-Poisson mixture model, is often used as an alternative to model

55

overdispersed Poisson count data. An ancillary parameter in the model addresses

overdispersion. We test the negative binomial model utilizing the same regression

form as for the Poisson model. To estimate model parameters, we use an iterative

process between the IRLS method and the scoring algorithm.

We find that the negative binomial model performs better than Poisson or bino-

mial, as shown in Section 5. This indicates the presence of overdispersion in our data.

Overdispersion generally occurs due to violations in the distributional assumptions of

the data, excessive zero counts, positive correlations between responses, omitted vari-

ables that are important for the model, or a predictor that needs to be transformed

to another scale. This emphasizes two main issues. First, standard errors need to

consider correlations between responses. Thus, we compute clustered standard errors

using the HuberÐWhite method [42, 45] as presented in Section 5. The second is the

importance of a correct model specification. For this, we experiment with scaling ex-

planatory variables as described earlier and test models with different distributional

assumptions. Since zero counts account for 80% of the observations in our data set,

we expand our analysis to evaluate zero-inflated versions of the binomial, Poisson,

and negative binomial models.

mass at zero with a count distribution. A zero-inflation component of the model

captures the probability of zero observations from the point mass distribution. Let 𝛾

denote the coefficients’ vector and 𝐺(𝑖) denote the 𝑖-th row of the data matrix 𝐺 for

the ‘inflation model’. We define 𝐺(𝑖) 𝛾 as follows:

𝑠 +𝛾1 ·𝑐𝑎𝑠𝑒 𝑠𝑖𝑧𝑒 𝑐𝑜𝑣𝑒𝑟𝑝𝑠 +𝛾2 ·𝑠𝑢𝑝𝑝𝑙𝑦 𝑐ℎ𝑎𝑖𝑛 𝑎𝑔𝑒𝑝𝑤 +𝛾3 ·𝑠ℎ𝑒𝑙𝑓 𝑙𝑖𝑓 𝑒𝑝 .

(2.4)

We include these variables in the inflation model as they are expected to have a direct

effect on the occurrence of zero observations. If case size cover is small, supply chain

aging is small, or shelf life is long, we expect to have a higher probability of zero

expiration. Further, we expect large format store types to have more frequent zero

observations, and thus include the store type variables 𝑠𝑡(𝑘) in the inflation model.

56

We perform robustness tests by adding and subtracting different variables and find

that the specification in (2.4) produces the best fit.

Thus, our model specifications are as follows: we specify the count distribution

(𝑖) (𝑖)

component for the Poisson or negative binomial models as𝐸[𝑟𝑒𝑡𝑢𝑟𝑛𝑝𝑠 ] = 𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 ·

(𝑖) (𝑖)

exp (𝐵 (𝑖) 𝛽), the count distribution component for the binomial model as 𝐸(𝑟𝑒𝑡𝑢𝑟𝑛𝑝𝑠 /𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 ) =

(𝑖) (𝑖)

𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑝𝑠 ·exp(𝐵 (𝑖) 𝛽)/[1+exp(𝐵 (𝑖) 𝛽)], and the inflation component as 𝐸(𝑟𝑒𝑡𝑢𝑟𝑛𝑝𝑠 ) =

0, Probability[point mass distribution]=exp(𝐺(𝑖) 𝛾)/[1 + exp(𝐺(𝑖) 𝛾)].

We use the count model for hypothesis testing and the inflation model to ad-

dress the frequent zero counts in the data. We estimate our models in the statistical

language R version 2.14.2 [54] utilizing packages stats, MASS [62], and pscl [65].

We use the glm function for the binomial and Poisson regressions, glm.nb function

for the negative binomial regression, and zeroinfl function for the ZIP and ZINB

regressions. The zeroinfl function finds maximum likelihood estimates using R’s

optimizer function, optim. Since zeroinfl does not support the zero-inflated bino-

mial model, we develop our own function in R to estimate model parameters for the

ZIB model. Our approach is based on the Expectation-Maximization (EM) algorithm

[23, 40]. We treat the source of each zero observation as missing data. If this infor-

mation were present, the log-likelihood function could be decomposed into a binomial

and a logistic regression model as two distinct parts. Based on the expectation of

missing data, in the maximization step of the EM algorithm, we perform maximum

likelihood estimation using the IRLS method for the binomial and logistic regression

models. Using the parameter estimates from the maximization step, in the expecta-

tion step of the algorithm, we calculate the probability of a zero occurring due to the

zero-mass distribution. The iterations are repeated until convergence. This approach

is practical because it allows us to utilize the existing glm function in R without a

necessity to develop score and information functions for the full model. To calcu-

late the standard errors from this model, we use the hessian and numericGradient

functions in R that are part of the numDeriv and maxLik packages to compute the

variance-covariance matrix and the estimating functions.

57

2.5. Results

The estimates of the ZINB model shows that Hypotheses 1, 2, 4, 5, and 6 are sup-

ported by our data. We first present the results for these hypotheses, then discuss

Hypothesis 3 which is not supported by our data.

Table 2.4 shows the estimation results of the count and inflation parts of the ZINB

model. It includes the estimates of the coefficients and standard errors, as well as

standard errors clustered at the product, warehouse-product, store, and route levels.

For tests of statistical significance, we use the standard errors of the cluster that

is at the same level as the explanatory variable. For example, since supply chain

age is a warehouse-product level variable, we use standard errors clustered at the

warehouse-product for testing significance. Following this approach, standard errors

that correspond to the explanatory variable are highlighted in Table 2.4.

Case size in the count model has a positive coefficient and 𝑐𝑎𝑠𝑒 𝑠𝑖𝑧𝑒 𝑐𝑜𝑣𝑒𝑟 has a

negative coefficient in the inflation model, both significant at 𝑝 < 0.01, which supports

Hypothesis 1, showing that expiration increases with case size. The coefficient for

𝑠𝑢𝑝𝑝𝑙𝑦 𝑐ℎ𝑎𝑖𝑛 𝑎𝑔𝑒 is positive in the count model and negative in the inflation model,

also both significant at 𝑝 < 0.01. This supports Hypothesis 2 and shows that aging

in the supply chain increases expiration.

Hypothesis 4 is supported because the coefficient of 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒 is positive and

significant at 𝑝 < 0.01 showing that compliance to the minimum order rule increases

the likelihood of expiration. 𝑓 𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑖𝑛𝑔 𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦 has a positive coefficient and

is significant at 𝑝 < 0.01 which supports Hypothesis 6. It shows that the chance of

product expiration increases in forecasting complexity.

Among the sales incentive variables, 𝑠𝑖(1), 𝑠𝑖(8), and 𝑠𝑖(9) have positive coeffi-

cients and are statistically significant at 𝑝 < 0.01. Thus, Hypothesis 5 is supported

by three of the nine sales incentives. We observe that these incentives differ from

others in two aspects: they have aggressive growth targets and they promote growth

as opposed to display competition, as illustrated in Table 2.2. Hence, we conclude

58

that whether sales incentives lead to an increase in expiration depends on the type

and the intensity of the incentive.

The coefficient of 𝑟𝑜𝑡𝑎𝑡𝑖𝑜𝑛 has a negative sign, rejecting Hypotheses 3. This could

be because the rotation variable is correlated with store type 𝑠𝑡(𝑘). As discussed in the

Online Appendix, a binary variable representing route type is the biggest predictor

of rotation. Small format routes have higher chances of rotation non-compliance

and mostly include gas stations & convenience stores, dollar discount, other grocery,

and drug stores. Also, according to the ZINB model, these store types exhibit a

higher probability of expiration. There is more likelihood of finding rotation non-

compliance at small format stores, such as gas stations & convenience stores, dollar

discount, other grocery type, and drug stores. Also, the coefficient estimates for

these store types in the ZINB model are higher than other store types, indicating

higher probabilities of expiration. Figure 2-4 depicts these differences. Supercenters,

in particular, have much lower chances of expiration. This can also be explained

by rotation. Supercenters belong to large national retail chains which are important

customers for AlphaCo. It is possible that sales representatives service these stores

with extra care and comply more with rotation guidelines.

59

1.00

count

0.80

inﬂaCon

0.60

Coeﬃcient
esCmates

0.40

0.20

-‐

supercenters
convenience
other
dollar
supermarkets
mass
drug
stores
club
stores

(0.20)
&
gas
grocery
discount
merchants

(0.40)

(0.60)

(0.80)

Store
types

Notes. Coefficient estimates of the inflation model is multiplied with -1 for easier

graphical interpretation.

The fixed effects for store types may be correlated with other factors that are

missing variables in the model. For instance, drug stores are known to exhibit higher

expiration rates [30] for CPG products most likely due to excess shelf space allocated

to food & beverage items. In that case, better allocation of shelf space can help

reduce expiration.

Table 2.5 shows estimation results for the alternative models (i.e., binomial, Pois-

son, negative binomial, ZIB, ZIP, and ZINB models). The ZINB model fits our data

best with the highest log likelihood value of -981,344. We find that zero inflated ver-

sions of all count models perform better: ZIP regression performs better than Poisson

with log likelihood values of -1,746,544 vs. -4,134,935, ZINB regression performs bet-

ter than the negative binomial regression with log likelihood values of -981,344 vs.

-1,035,893, and ZIB performs better than the binomial regression with log likelihood

values of -1,621,489 vs. -4,125,456. This shows that zero-inflated models are effective

in addressing overdispersion caused by excess zero points in our data set. We pick

the ZINB model because it gives us the best log likelihood value and and the most

unbiased distribution of residuals.

60

2.5.2 Effect of sales incentives

In this section, we analyze the interaction of sales incentives, demand, and probability

of expiration. We redefine the notation for clarity of exposition. For product 𝑝 in

store 𝑠, let 𝐵𝑝𝑠 denote the baseline demand in the absence of a sales incentive, 𝑆𝑝

denote the sales incentive indicator, and 𝐷𝑝𝑠 denote the demand in the presence of a

sales incentive.

Sales incentives would stimulate demand but also increase expiration. We repre-

sent the first relationship as

𝐷𝑝𝑠 = 𝐵𝑝𝑠 + 𝑆𝑝 𝜃, (2.5)

where 𝜃 is expected to be positive. Also let 𝐸𝑝𝑠 denote the expiration amount observed

for product 𝑝 at store 𝑠, 𝑋𝑝𝑠 denote the matrix representing the explanatory variables

for expiration excluding the variables for sales incentives and demand, and 𝑓 () denote

the link function used in the regression model. In the absence of sales incentives, the

model of expiration is 𝐸𝑝𝑠 = 𝑓 (𝑋𝑝𝑠 𝛽 + 𝐵𝑝𝑠 𝛼).

Hypothesis 5 proposes that sales incentives impact expiration. Thus, including 𝑆𝑠

in the model for 𝐸𝑝𝑠 yields

with the observed demand 𝐷𝑝𝑠 gives us:

Note that 𝛾 and 𝜃 are expected to be positive. Therefore, our estimate for the

coefficient of 𝑆𝑠 presented in Section 2.5.1 reflects the net effect of sales incentives on

expiration as shown in (2.7). In fact, it is likely to understate the effect of incentives

on expiration. To compute the true effect of incentives, we estimate (2.5) and use the

estimates of 𝜃 and 𝛼 to obtain 𝛾. Table 2.6 presents the results of this estimation. We

find that only incentives 4 and 7 increase demand. Further, the true effect of five out

61

of nine incentives, namely 1, 5, 6, 8, and 9, on expiration is positive and significant.

This estimation refines our single equation results on sales incentives shown in Section

5.1, where we found incentives 1, 8, and 9 to have positive and statistically significant

effects.

The relationship between supply chain aging and expiration can be prone to reverse

causality. Expiration could lead to inflated estimates of mean demand at warehouses,

which could lead to higher warehouse inventory levels, and thus higher supply chain

aging. In this section, we examine the relationship between supply chain aging and

expiration, and test whether expiration causes supply chain aging.

There are several variables at the warehouse level that can be expected to be

correlated with supply chain aging but are not functions of product expiration in

retail stores. These variables are candidates to instrument for supply chain aging.

They include:

plan supply chain inventory. Safety stock levels at the warehouses are calculated

based on forecast error. Therefore, poor forecasting, through higher safety stock,

can cause excess warehouse inventory; as a result, products, on average, spend

longer time at the warehouse.

∙ Production batch size: Producing in bigger batch sizes lowers unit production

cost, but increases inventory. Such inventory would spend a longer time at

warehouses. We expect this relationship to be stronger for low velocity items.

∙ Full pallet shipments: Transporting products in full pallets from plants to dis-

tribution centers reduces handling costs, but can result in excess inventory at

the distribution centers, particularly for low velocity items.

before placing new products in their stores. The timing of approvals by different

62

retailers do not always coincide. Therefore, once a batch is produced, products

from that batch may enter different markets at different times. As a result,

some products wait and age at the warehouses.

level. We find that the correlation coefficient between warehouse aging and forecast

errors for each warehouse-product is 0.2. Table 2.7 reports the results obtained from

the omitted variable version of the Hausman test [31]. Model 1 in the table is the

baseline ZINB model. Model 2 augments Model 1 by including forecast error as an

additional explanatory variable. We find that including forecast error in the model

does not improve explanatory power because AIC values and coefficient estimates of

both models are near identical. Further, an F-test for 𝛽𝑓 𝑜𝑟𝑒𝑐𝑎𝑠𝑡 𝑒𝑟𝑟𝑜𝑟 = 0 fails to reject

the null hypothesis. As a result, we do not find any evidence of endogeneity.

A potential shortcoming of this analysis is that the theory behind this test is

based on the ordinary least squares model. Extending this theory to generalized

linear models would be a valuable future research area.

In this section, we utilize the estimation results to examine remedies that AlphaCo

can pursue to reduce expiration. We look at changes in four areas: (1) case size, (2)

supply chain aging, (3) frequency of replenishment, and (4) sales incentives.

To assess the reduction in the amount of expiration associated with a change, we

compare the expected expired volume for two scenarios, before and after the change.

For example, if current expected expired volume is 100 units and expected expired

volume with the change is 70 units, we conclude that the change reduces expiration

by 30%. Let 𝐵 𝑐 and 𝐺𝑐 denote data matrices for the count and inflation components

of the ZINB model associated with scenario 𝑐 and let 𝑖 index the data points that are

impacted by the change. We calculate the expected expired volume as:

∑︁

𝐸[𝑟𝑒𝑡𝑢𝑟𝑛𝑐𝑖 ] = 𝑑𝑒𝑙𝑖𝑣𝑒𝑟𝑦𝑖 · exp (𝐵 𝑐𝑖 𝛽)[1/(1 + exp(𝐺𝑐𝑖 𝛾))]. (2.8)

𝑖

63

To project the reduction in expiration amount onto monetary benefits for Al-

phaCo, we use data from AlphaCo’s 2010 waste study. The study gives us an estimate

of the total cost of unsaleables, including the cost of goods sold, sales & delivery cost,

as well as the reverse logistics cost associated with unsaleables products. In addition,

our data indicate that 65% of unsaleables occur due to expiration and the remaining

35% occur due to product damage. To extrapolate the monetary value of benefits, we

express the reduction in expiration as a proportion of the overall expired volume in

our dataset convert it into a monetary estimate. For instance, if a change corresponds

to a reduction in expired volume of 3,000 cases, and the total number of actual ex-

pired volume is 299,786 cases in our data set, then we multiply 3,000/299,786 with

65% of the total cost of unsaleables to find the monetary benefit of this change.

Case size: We find that reducing case sizes from 24 units to 12 reduces expiration

volume by 41% for products that are currently packed in 24 unit cases. Such prod-

ucts make up 26% of the data points in our data set. This corresponds to a $7.66M

reduction in expiration cost. AlphaCo concludes that this benefit is substantially

higher than the expected increase in product handling cost. Hence, the management

of case sizes is a significant opportunity for AlphaCo to improve its bottom-line. By

quantifying the benefits of smaller case sizes, our analysis provides a basis for a busi-

ness case to pursue changes in manufacturing and business processes. AlphaCo can

consider implementing this change either by reducing case sizes directly or developing

a modular case that can be split in half and also be ordered in half cases at stores

with low demand.

Supply chain aging: Our analysis shows that reducing days of supply in the

supply chain by 1 week corresponds to 0.98% reduction in expiration volume and

$912.4K in expiration cost. In this calculation, we consider no change for the data

points that already have less than one week of aging; such data points constitute

2.7% of the observations in our data set. Figure 2-5 illustrates the distribution of

𝑠𝑢𝑝𝑝𝑙𝑦 𝑐ℎ𝑎𝑖𝑛 𝑎𝑔𝑒 in our data set. The majority of the data points have more than 7

days of aging.

64

150000

100000

Frequency

50000

0

Supply chain age (in days)

Manufacturers and retailers often try to reduce inventory, usually under pressure

from their financial controllers, in order to decrease working capital requirements.

Our study suggests that the analysis for these efforts need not be limited to the

opportunity cost of the working capital. It is also important to consider the benefit

of inventory reduction on the occurrence of expiration in order to get a more thorough

picture of the benefits.

Frequency of replenishment: To assess the cost of order inflation that occurs

due to the minimum order rule, we examine the stores that have positive values for

𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒, indicating order(s) at the minimum order level. We estimate the

change in expected expiration volume for such stores when 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒𝑠 is set to

zero. These data points make up 41.6% of our data set.

We find that the expected expiration reduces by 0.87% when orders are not inflated

to meet the minimum order rule requirement. Further, this difference corresponds

to a monetary value of $610,769. As a remedy, AlphaCo can identify such stores

using 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒 as a metric on a periodic basis and reduce their visit frequency

65

accordingly. Considering the significant savings opportunity and relatively small effort

this remedy requires, this is a worthwhile initiative to pursue for AlphaCo.

AlphaCo’s sales representative visit stores according to a pre-determined schedule,

such as everyday, twice a week, once a week, once every other week, once a month,

etc. Reducing the visit frequency in certain cases, for instance from once a week to

once every other week, may pose a stockout risk because the available shelf space and

backroom space may not be sufficient to cover extra days of demand, especially in

small stores. Therefore, in pursuing this initiative, it is beneficial to focus on stores

with higher scores of 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒. Figure 2-6 represents the percentage reduction

in expiration at stores with different values of 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒. For example, when

orders are not inflated at stores whose value for 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒 ranges between 80

and 100, expiration is lessened by 14%. As seen on the chart, the saving opportunity

increases with the value of 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒. The saving opportunity at stores whose

values of 𝑚𝑖𝑛 𝑜𝑟𝑑𝑒𝑟 𝑟𝑢𝑙𝑒 ranging between 20 and 100 is $201.1K.

16.0%

Reduc2on
in
expira2on
volume

14.0%

12.0%

10.0%

8.0%

6.0%

4.0%

2.0%

0.0%

(0,20)
(20,40)
(40,60)
(60,80)
(80,100)

Interval
of
the
minimum
order
rule
measure

Figure 2-6: Impact of addressing order inflation due to minimum order rule.

icantly associated with higher expiration in Section 2.5.1. These incentives affect,

respectively, 2.8%, 22.9%, and 0.43% of the data points. We find that eliminating

these incentives, i.e., setting 𝑠𝑖(𝑗) = 0, results in 14%, 9%, 7% less expiration, and

monetary savings of $645,234, $2,634,663, and $25,715, respectively.

As a remedy, AlphaCo’s marketing organization can consider waste implications

66

in the design of these incentives. For example, sales targets for products with shorter

shelf lives can be selectively set to more moderate levels. For instance, a goal of 10%

increase in sales as opposed to a goal of 20% increase should reduce expiration. In

addition, sales representatives’ can be assisted with the execution of sales incentives.

For example, a decision support tool suggesting the stores that have higher likelihood

of selling the type of products included in the sales incentive can be useful in ensuring

that additional inventory in the market translates to consumer purchases.

2.6. Conclusion

facturers, and supply chain managers. Using data for a CPG manufacturer and a

network of retail stores, we show that expiration can occur due to various causes

related to the manufacturer or the retail stores, such as case size, supply chain aging,

shelf life of products, sales incentives, rotation compliance, minimum order rules, and

forecasting complexity. The amount of expiration also varies across retail store for-

mats and slow- vs. fast-moving products. Counterfactual analysis presented in our

study shows that a reduction in case size, even at the margin from 24 units to 12 units,

can lead to a substantial reduction in expiration. Other variables examined in the

counterfactual analysis include warehouse inventory, sales incentives, and frequency

of replenishment.

Overall, our research exposes and explains the complexity of expiration observed

in retail supply chains. We show that expiration does not occur solely due to demand

uncertainty, but is also a supply chain coordination problem. It can be reduced by

better management at both manufacturers and retailers. Our study suggests a number

of topics for future research. One topic is to examine perishable inventory from a

multi-location and channel perspective. The existing perishable inventory literature

focuses on optimal inventory policies at a single location. Also, supply chain aging

targets and issuance rules for inventory at upstream levels of supply chains have not

been analyzed in current literature. These are the focus of Chapter 3.

67

A second topic is to incorporate the implications of expiration in retail operations

models, such as the optimization of case sizes and shelf space allocation. The existing

literature studies the shelf space allocation problem considering cross-correlation of

demand and substitution effects. If excess shelf space is allocated to a slow-moving

item or an item with a short shelf-life, then its effect on expiration should be included

in the cost of shelf space allocation.

The study of product expiration can also have impact on the literature in sus-

tainable operations, which has studied topics ranging from green product design to

closed-loop supply chains [37]. Thus far, there have been many advancements in the

management of durable goods, such as modular product design [9, 58], remanufac-

turing [22, 36], and lean manufacturing [55]. These practices are useful for improving

sustainability in discrete manufacturing environments more so than in process manu-

facturing, which is the main manufacturing method in CPG companies. Thus, further

research can examine supply chain challenges related to expiration, such as disposal,

short term production planning, inventory replenishment, and package design.

Firms in the CPG industry can use our analysis as a framework to identify the

drivers of expiration that matter in their supply chain and construct business cases

for initiatives to reduce expiration. Current reimbursement policies practiced in the

industry are two extreme schemes in terms of incentives they offer to reduce un-

saleables. For instance, either the manufacturer does not have an incentive to supply

fresher products to the retailer or the retailer does not have an incentive to manage

store inventory better. This is most likely because poor understanding of the sources

of expiration makes it hard to share the cost of unsaleables in an effective and fair

way. Similar incentive issues exist within the same firm. Either no function is held

accountable for the cost or one function (e.g., sales or logistics) absorbs it regardless

of the source. Then, we see behaviors such as the sales-force flooding the market

with excess inventory or plant managers not having any regard to waste implications

when determining production batch sizes. These behaviors can be altered by design-

ing coordination mechanisms to reduce the occurrence of expiration. Thus, Chapter

4 examines incentive mechanisms focusing on the sales-force aiming to discourage

68

overselling.

69

Appendix A: Calculation of rotation measure

We seek to measure compliance with shelf rotation for all store-SKU combinations at

AlphaCo. However, this compliance is not recorded in a database because it cannot

be continuously monitored without incurring prohibitive cost. Therefore, we leverage

data from an audit study conducted by AlphaCo in 2010 to construct an instrument

for rotation.

The audit involved recording all incidences of unsaleable products in sampled store

visits. When found, auditors recorded a reason code associated with each occurrence

of unsaleables, also indicating whether the shelf or the back room inventory was

rotated. The audit covered 128 out of 2,851 routes in our data set. Stores selected for

the audit are representative of diverse geographies, store types, and warehouses. We

develop a logistic regression model to estimate rotation for the 128 routes examined

in the store audits. We utilize the coefficients from this model to predict rotation

for all 2,851 routes included in the data set. These predicted values constitute our

measure for rotation, and represent the probability of an unrotated store-product(s)

occurring on a route.

We aggregate the audit data at the route level and categorize each route into

two groups considering all store-products audited in the route: (i) routes where no

non-compliance is found; (ii) routes where at least one rotation non-compliance is

found in the shelf or the backroom. This binary measure is the dependent variable

in our logistic regression model. We use two explanatory variables, as follows, based

on interviews and observations from our field trips:

1. Route type: Small format routes are more likely to have rotation issues

compared to large format routes. At small format routes, both sales representatives

and drivers are responsible for shelf rotation, whereas at large format routes rotation

falls under the responsibility of only the sales representatives. This rule reflects the

difference in delivery processes between the two route types. At large format stores,

drivers drop off the pallets in the backroom and shelves are stocked from this inventory

by the sales representative or a merchandiser at a later time. Small format stores do

70

not have sufficient backroom space to accommodate bulk deliveries. Therefore, drivers

of small format stores place the products on the shelf. While doing so, they are also

expected to rotate the shelf.

A delivery supervisor stated that making two people accountable for the same

task reduces compliance because both people expect each other to perform the task.

Audit data support this claim; rotation issues are seen at 92% of the small format

routes and 47% of the large format routes audited. We define a binary variable, large

format route, indicating large format routes in the data set. The average value for

large format route is 0.68.

sentatives to have fewer rotation issues. As a proxy for process compliance, we use a

metric called Order Quality. AlphaCo provides the sales representatives with a tool

to suggest replenishment orders when they provide a set of information into their

handheld devices such as the retail price, the next delivery date, and the amount

of inventory-on-hand. Order Quality measures the extent to which the required in-

formation is provided to this tool. It, therefore, reflects the degree to which sales

representatives are compliant with AlphaCo’s sales processes. We expect complaint

sales representatives to be more likely to perform rotation. Thus, we include order

quality as an explanatory variable in the logistic regression model. It ranges between

59.3 and 99.5 with mean and median values of 87.9 and 90.25, respectively.

Table 2.1 summarizes the estimation results from the logistic regression. Large

format routes and routes with higher order quality have less incidence of rotation

non-compliance. Store format is statistically significant at 𝑝 < 0.01. Although the

p-value of order quality is 23%, not small enough to be considered significant, we

choose to keep it in the model because it reduces the residual deviance by 1%. The

null deviance and residual deviance for the model are 160.54 and 126.93, respectively,

yielding a pseudo R-squared value of 20%. Considering routes with probabilities

higher than 0.5 as having rotation non-compliance, our model classifies 70% of the

routes correctly. Thus, we use this model to construct an instrument for rotation

non-compliance across all routes in the data set.

71

Table 2.1: Estimation results for the logistics regression model.

intercept 5.5 2.65 2.078 0.0377 *

large format route -2.55 0.53 -4.795 1.63e-06 ***

order quality -0.035 0.029 -1.204 0.23

Notes. The null deviance is 160.54 and the residual deviance is 126.93. *, **, *** Statistically

significant at p=0.10, 0.05, 0.01, respectively.

72

Appendix B: Tables

year 2010 2010 2010 2010 2011 2011 2011 2011 2011

month(s) 8,9 7,8 6 5,6 1 2 4 6 3,4,5

products included 15 9 35 24 108 61 82 165 4

median shelf life 13 30 13 78 26 13 13 17 26

maximum shelf life 14 52 39 104 104 36 26 35 26

minimum shelf life 12 26 12 72 12 12 13 12 26

type volume volume volume volume display display volume volume volume

focus brand brand flavor brand-flavor broad flavor brand broad brand

target +20% +5% +15% +5% +5 to 25% fixed target

per brand in cases

73

Table 2.3: Summary statistics of the variables included in the model.

delivery (exposure variable) 384.81 1106.96 120 1 102008 -

return (dependent variable) 2.86 10.82 0 0 1408 80%

Main variables of interest:

supply chain age 28.19 18.89 22.79 0.06 247.05 -

case size 12.60 7.80 12 1 24 -

forecasting complexity 4046.64 1932.27 4463 14 9571 -

rotation 0.81 0.20 0.91 0.37 0.98 -

min order rule 4.29 8 0 0 100 58%

si(1) 0.03 0.16 0 0 1 97%

si(2) 0.03 0.16 0 0 1 97%

si(3) 0.04 0.21 0 0 1 96%

si(4) 0.06 0.24 0 0 1 94%

si(5) 0.22 0.42 0 0 1 78%

si(6) 0.06 0.24 0 0 1 94%

si(7) 0.13 0.34 0 0 1 87%

si(8) 0.23 0.42 0 0 1 77%

si(9) 0.004 0.07 0 0 1 99.6%

case size cover 0.37 1.79 0.09 0 24 -

Control variables:

demand 381.96 1106.95 120 0 102008 1.23%

shelf life 27.71 21.60 21 10 104 -

st(supermarket) 0.22 0.42 0 0 1 78%

st(gas station or convenience store) 0.46 0.5 0 0 1 54%

st(other grocery) 0.1 0.3 0 0 1 90%

st(dollar discount) 0.04 0.2 0 0 1 96%

st(drug store) 0.11 0.31 0 0 1 89%

st(mass merchant) 0.04 0.18 0 0 1 96%

st(club store) 0.00024 0.02 0 0 1 99.976%

st(supercenter) 0.03 0.17 0 0 1 97%

74

Table 2.4: Estimation results for the ZINB model specified in Section 2.4.2.

Mean Std. Error Std. Error Std. Error Std. Error Std. Error

(Product) (Warehouse-Product) (Store) (Route)

count:

intercept 0.49 0.018 *** 0.061 *** 0.025 *** 0.036 *** 0.046 ***

st(gas station or convenience store) -0.11 0.009 *** 0.022 *** 0.013 *** 0.022 *** 0.026 ***

st(other grocery) -0.11 0.012 *** 0.023 *** 0.015 *** 0.026 *** 0.029 ***

st(dollar discount) -0.06 0.015 *** 0.033 . 0.019 *** 0.030 * 0.034 .

st(drug store) 0.11 0.010 *** 0.029 *** 0.015 *** 0.025 *** 0.029 ***

st(mass merchant) -0.09 0.013 *** 0.023 *** 0.016 *** 0.038 * 0.037 *

st(club store) 0.86 0.105 *** 0.144 *** 0.121 *** 0.155 *** 0.155 ***

st(supercenter) -0.60 0.018 *** 0.042 *** 0.029 *** 0.052 *** 0.050 ***

case size 0.04 0.0003 *** 0.002 *** 0.001 *** 0.001 *** 0.001 ***

supply chain age 0.05 0.008 *** 0.003 * 0.001 *** 0.001 *** 0.002 ***

shelf life -0.01 0.0003 *** 0.001 *** 0.001 *** 0.001 *** 0.001 ***

demand -0.69 0.002 *** 0.011 *** 0.003 *** 0.003 *** 0.004 ***

si(1) 0.13 0.014 *** 0.066 . 0.022 *** 0.014 *** 0.015 ***

si(2) -0.04 0.022 . 0.108 0.027 0.026 0.027

si(3) -0.12 0.012 *** 0.056 * 0.016 *** 0.012 *** 0.013 ***

si(4) 0.16 0.034 *** 0.131 0.054 ** 0.051 ** 0.052 **

si(5) 0.04 0.008 *** 0.048 0.013 ** 0.008 *** 0.009 ***

si(6) -0.00 0.010 0.060 0.017 0.013 0.014

si(7) -0.08 0.008 *** 0.045 . 0.014 *** 0.008 *** 0.009 ***

si(8) 0.09 0.007 *** 0.056 0.014 *** 0.008 *** 0.009 ***

si(9) 0.11 0.031 *** 0.059 . 0.033 ** 0.032 *** 0.033 **

forecasting complexity 0.00001 0.000002*** 0.00001 . 0.000002 *** 0.000003 *** 0.000004 **

rotation -0.10 0.018 *** 0.047 * 0.028 *** 0.052 . 0.062

min order rule 0.002 0.0003 *** 0.0005 *** 0.0003 *** 0.001 ** 0.001 *

inflation:

intercept 0.12 0.009 *** 0.088 0.017 *** 0.027 *** 0.031 ***

st(gas station or convenience store) 0.59 0.008 *** 0.047 *** 0.012 *** 0.029 *** 0.035 ***

st(other grocery) 0.60 0.012 *** 0.040 *** 0.016 *** 0.041 *** 0.053 ***

st(dollar discount) 0.52 0.016 *** 0.061 *** 0.021 *** 0.046 *** 0.057 ***

st(drug store) -0.09 0.010 *** 0.054 0.015 *** 0.036 * 0.041 *

st(mass merchant) 0.01 0.016 0.029 0.018 0.069 0.071

st(club store) -0.86 0.160 *** 0.152 *** 0.160 *** 0.193 *** 0.194 ***

st(supercenter) 0.54 0.020 *** 0.048 *** 0.021 *** 0.069 *** 0.068 ***

supply chain age -0.01 0.009 *** 0.009 0.003 ** 0.003 *** 0.004 *

shelf life 0.05 0.0003 *** 0.003 *** 0.001 *** 0.001 *** 0.001 ***

case size cover -1.29 0.010 *** 0.061 *** 0.014 *** 0.016 *** 0.020 ***

Notes. .,*, **, *** Statistically significant at p=0.10, 0.05, 0.01,0.001 respectively. Dispersion parameter is

estimated as 1.0426.

75

Table 2.5: Parameter estimates of alternative models.

log likelihood: (1,035,893) (981,344) (4,134,935) (1,746,455) (4,125,456) (1,621,489)

count:

intercept 1.308 0.488 0.937 0.076 3.312 2.798

(0.029)*** (0.018)*** (0.005)*** (0.005)*** (0.006)*** (0.004)***

st(gas station -0.672 -0.114 -0.391 -0.063 -0.535 -0.281

or convenience store) (0.042)*** (0.022)*** (0.038)*** (0.025)* (0.042)*** (0.031)***

st(other grocery) -0.608 -0.11 -0.329 -0.03 -0.538 -0.273

(0.050)*** (0.026)*** (0.045)*** (0.029) (0.052)*** (0.037)***

st(dollar discount) -0.459 -0.062 -0.312 -0.029 -0.491 -0.163

(0.054)*** (0.030)* (0.052)*** (0.035) (0.057)*** (0.041)***

st(drug store) 0.068 0.109 0.106 0.116 0.138 0.028

(0.046) (0.025)*** (0.043)* (0.029)*** (0.049)** (0.037)

st(mass merchant) -0.163 -0.094 -0.107 -0.118 -0.127 -0.124

(0.066)* (0.038)* (0.065). (0.044)** (0.067). (0.045)**

st(club store) 1.55 0.857 1.042 0.64 1.832 1.115

(0.269)*** (0.155)*** (0.220)*** (0.137)*** (0.269)*** (0.226)***

st(supercenter) -0.794 -0.598 -0.806 -0.725 -0.542 -0.299

(0.081)*** (0.052)*** (0.088)*** (0.065)*** (0.088)*** (0.059)***

case size 0.055 0.039 0.031 0.024 0.055 0.047

(0.008)*** (0.002)*** (0.005)*** (0.002)*** (0.008)*** (0.004)***

supply chain age -0.08 0.047 0.221 0.171 0.016 -0.074

(0.004)*** (0.001)*** (0.003)*** (0.002)*** (0.004)*** (0.018)***

shelf life -0.041 -0.015 -0.056 -0.014 -0.069 -0.022

(0.003)*** (0.001)*** (0.006)*** (0.002)*** (0.008)*** (0.004)***

demand -0.998 -0.695 -0.849 -0.608 -1.221 -1.034

(0.003)*** (0.002)*** (0.000)*** (0.000)*** (0.001)*** (0.0004)***

si(1) 0.758 0.127 0.357 0.045 0.327 -0.05

(0.189)*** (0.066). (0.106)*** (0.043) (0.158)* (0.090)

si(2) -0.197 -0.039 -0.215 -0.013 -0.128 -0.119

(0.270) (0.108) (0.214) (0.074) (0.217) (0.090)

si(3) -0.347 -0.123 -0.222 0 -0.36 -0.156

(0.113)** (0.056)* (0.084)** (0.043) (0.103)*** (0.052)**

si(4) -0.103 0.156 0.645 0.038 1.22 0.503

(0.296) (0.131) (0.399) (0.124) (0.496)* (0.229)*

si(5) -0.138 0.038 0.091 0.128 0.024 0.039

(0.118) (0.048) (0.072) (0.035)*** (0.103) (0.047)

si(6) -0.003 0 0.016 0.054 -0.061 -0.042

(0.140) (0.060) (0.111) (0.060) (0.143) (0.099)

si(7) 0.11 -0.079 -0.075 -0.147 -0.037 -0.137

Continued on next page

76

Table 2.5 – continued from previous page

Negative Binomial ZINB Poisson ZIP Binomial ZIB

(0.118) (0.045). (0.076) (0.054)** (0.110) (0.067)*

si(8) 0.528 0.085 0.284 -0.004 0.334 -0.002

(0.147)*** (0.056) (0.091)** (0.035) (0.124)** (0.070)

si(9) 0.227 0.106 0.2 0.093 0.135 0.062

(0.128). (0.059). (0.111). (0.059) (0.137) (0.076)

forecasting complexity 0.000032 0.000013 0.000036 0.000016 0.000028 0.000007

(0.000013)* (0.0000043)** (0.00001)** (0.0000001)*** (0.000014)* (0.000008)

rotation -0.541 -0.095 -0.362 0.137 -0.684 -0.113

(0.114)*** (0.062) (0.106)*** (0.062)* (0.122)*** (0.082)

min order rule 0.002 0.002 0.002 0.002 0.001 0.001

(0.001). (0.001)** (0.001). (0.001)** (0.001) (0.001)

inflation:

intercept NA 0.12 NA 0.226 NA 0.229

(0.009)*** (0.008)*** (0.008)***

st(gas station NA 0.587 NA 0.549 NA 0.54

or convenience store) (0.029)*** (0.028)*** (0.028)***

st(other grocery) NA 0.604 NA 0.58 NA 0.571

(0.041)*** (0.040)*** (0.040)***

st(dollar discount) NA 0.521 NA 0.518 NA 0.513

(0.046)*** (0.044)*** (0.044)***

st(drug store) NA -0.086 NA -0.086 NA -0.09

(0.036)* (0.034)* (0.034)**

st(mass merchant) NA 0.014 NA 0.027 NA 0.035

(0.069) (0.066) (0.066)

st(club store) NA -0.859 NA -0.929 NA -0.943

(0.193)*** (0.180)*** (0.180)***

st(supercenter) NA 0.54 NA 0.555 NA 0.571

(0.069)*** (0.068)*** (0.068)***

supply chain age NA -0.063 NA -0.051 NA -0.062

(0.003)*** (0.003)*** (0.019)***

shelf life NA 0.048 NA 0.05 NA 0.05

(0.003)*** (0.003)*** (0.003)***

case size cover NA -1.289 NA -1.146 NA -1.1

(0.010)*** (0.009)*** (0.009)***

77

Table 2.6: Results of the analysis of interaction between demand, sales incentives,

and expiration.

𝜃 𝛽 𝛼 𝛾

si(1) -0.1016 0.1274 -0.6947 0.1980

(0.0117)*** (0.0659). (0.0017)*** (0.0718)**

si(2) -0.2688 -0.0387 -0.6947 0.1480

(0.0104)*** (0.1083) (0.0017)*** (0.1135)

si(3) -0.0599 -0.1227 -0.6947 -0.0811

(0.0081)*** (0.0557)* (0.0017)*** (0.0605)

si(4) -0.2922 0.1559 -0.6947 -0.0471

(0.0092)*** (0.1309) (0.0017)*** (0.1341)

si(5) -0.4859 0.0381 -0.6947 0.3757

(0.0060)*** (0.0476) (0.0017)*** (0.0504)***

si(6) -0.4479 -0.0002 -0.6947 0.3110

(0.0070)*** (0.0600) (0.0017)*** (0626)***

si(7) 0.5087 -0.0790 -0.6947 -0.4324

(0.0051)*** (0.0451). (0.0017)*** (0.0478)***

si(8) -0.0330 0.0851 -0.6947 0.1080

(0.0050)*** (0.0564) (0.0017)*** (0.0580).

si(9) -0.8946 0.1061 -0.6947 0.7276

(0.0255)*** (0.0586). (0.0017)*** (0.0733)***

Notes. .,*, **, *** Statistically significant at p=0.10, 0.05, 0.01,0.001 respectively.

78

Table 2.7: Results of the Hausman test for endogeneity.

Model 1 Model 2

AIC: 1652049 1652051

count:

intercept 0.44321 0.44239

st(gas station or convenience store) -0.11542 -0.11573

st(other grocery) -0.11070 -0.11098

st(dollar discount) -0.05143 -0.05177

st(drug store) 0.12091 0.12065

st(mass merchant) -0.06833 -0.06840

st(club store) 0.94795 0.94765

st(supercenter) -0.64138 -0.64141

case size 0.03837 0.03837

supply chain age 0.00342 0.00339

shelf life -0.01439 -0.01439

demand -0.70125 -0.70123

si(1) 0.14096 0.14107

si(2) -0.03875 -0.03861

si(3) -0.15007 -0.15015

si(4) 0.13489 0.13494

si(5) 0.04673 0.04663

si(6) -0.07340 -0.07330

si(7) -0.05414 -0.05412

si(8) 0.10848 0.10851

si(9) -0.05169 -0.05228

forecasting complexity 0.00002 0.00002

rotation -0.05374 -0.05254

min order rule 0.00209 0.00209

forecast error 0.00002

(0.0000267)

inflation:

intercept 0.16745 0.16745

st(gas station or convenience store) 0.55990 0.55990

st(other grocery) 0.55934 0.55933

st(dollar discount) 0.42890 0.42889

st(drug store) -0.15336 -0.15336

st(mass merchant) -0.02335 -0.02337

st(club store) -0.89077 -0.89083

st(supercenter) 0.48764 0.48764

supply chain age 0.00121 0.00121

shelf life 0.04570 0.04570

case size cover -1.29043 -1.29044

Notes. The value in parenthesis represent the standard error for forecast error. .,*, **, *** Statisti-

cally significant at p=0.10, 0.05, 0.01,0.001 respectively.

79

80

Chapter 3

Chains For Perishable Products

3.1. Introduction

The study of the drivers of product expiration presented in Chapter 2 motivates this

research. The analysis shows evidence that inventory aging in the upstream supply

chain increases the probability of expiration at retail stores, since aging erodes shelf

life. We might conclude that we should reduce the supply chain inventory to reduce

the amount of expiration. However, inventory reduction can lead to stockouts and

lost sales for the manufacturer. Further, it is a complex task to determine how much

inventory to reduce and by what means, since excess inventory is driven by multiple

sources. Examples of these sources include production or transportation batching,

forecasting inaccuracy, unrotated products at distribution centers, and delayed prod-

uct launches.

One way a manufacturer can control the amount of product expiration is through

the choice of its shipment policies. A shipment policy prescribes the minimum re-

maining shelf life below which the product will not be released for shipment to the

81

retailer and will be discarded. Properly determined shipment policies help manufac-

turers reduce cost in two ways: 1) by intelligently making sell versus dispose decisions,

and 2) if coupled with organizational incentives, by providing motives to supply chain

managers to rotate the inventory, better manage batch sizes, or improve forecasting

accuracy. When the disposition cost impacts the profit and loss performance of the

supply chain manager, the shipment policy would provide incentives to better man-

age inventory, which increases the effective shelf life and thus reduces the amount of

expired items. This study aims to alleviate the inventory aging issue by establishing

shipment policies for the manufacturer.

Currently, the remaining shelf life is an area of contention in the consumer pack-

aged goods industry. Retailers express discontent that products coming into their

supply chain have very little shelf life remaining [64]. As a result, manufacturers are

increasingly looking at the area of shelf-life management. Typically, existing rules

either dictate the same number of days across a product category (e.g., 30 days of

minimum remaining shelf life) or express the minimum in terms of a percentage of the

manufactured shelf life (e.g., minimum required shelf life is 70% of the manufactured

shelf life). The problem with existing rules is that they do not consider important

factors such as the demand rate or product cost. Consequently, this research asks the

following question: how should we set up shipment policies?

To answer this question, we develop an optimization model which minimizes the

remaining shelf life for a given product with respect to the constraint that the net

profit from sales exceeds the loss from disposition (i.e., disposing of inventory through

donation to food banks, landfills, etc.). In the model, the net profit has two compo-

nents: gross profit from sales and cost of expiration, which depends on the remaining

shelf life. We solve this problem using our collaborator AlphaCo’s historical shipment

and product cost data for 2013.

We find that the optimum remaining shelf life ranges between 1 and 26 weeks

82

with a mean value of 5 weeks for the products included in our dataset. These findings

contrast with our collaborator’s existing shipment policy, which requires a uniform

minimum remaining shelf life of 5 or 6 weeks across products. Comparing our opti-

mum policy with the existing shipment policy, we find that optimization increases the

net profits by 41% in scenarios where products are shipped to retailers’ supply chains

with the required minimum remaining shelf life. This significant increase shows that

a homogenous shipment policy is ineffective; manufacturers can improve profits with

an optimization-based shipment policy that is customized to the product’s profit, ex-

piration cost, and sales rate. Furthermore, our analysis has implications for supply

chain management. Products that require a high degree of freshness (indicated by a

high optimum remaining shelf life) can be considered for discontinuation unless fre-

quent production or distribution in small batch sizes is feasible. Regarding warehouse

management, products requiring high freshness can be rotated more frequently or can

be positioned on the warehouse floor in a way to facilitate rotation.

Simulation based optimization can be costly in practice due to multiple dimensions

(time, retail stores, inventory age). For this reason, we evaluate several regression

techniques, including machine learning tools, to approximate the optimum remaining

shelf life. Approximation of optimal values provides a 14% improvement in net profits

compared to the existing shipment policy at our collaborator. This improvement ac-

counts for 28% of the savings that optimization provides. Consequently, we conclude

that machine learning tools, such as the random forest algorithm, can be used as a

low-cost alternative to optimization.

With this study, we provide a framework (an optimization method and machine

learning approximation of optimal values) to solve a complex problem; such a frame-

work is valuable in practice. Currently, the industry lacks standards on shipment

policies. Further, alignment problems exist due to discordant shipment rules across

manufacturers and mismatched expectations among retailers. Our framework can ad-

83

dress this alignment problem by providing a quantitative-based method to establish

shipment policies.

In this research, we examine the rules of the manufacturer with regard to the shipment

of perishable products to the retailers. Thus, this study is related to the operations

literature that focuses on two-tier supply chains for perishable products. Related

research studies a range of topics but overlooks the management of effective shelf

lives in the manufacturer’s supply chain.

Goh et al. [20] study two-stage perishable inventory models for blood banks. Two

stages in the supply chain represent fresh and older blood units with separate sources

of demand. When time passes, units are transferred from the first stage to the second.

The analysis is concerned with whether the second-stage demand should be served by

the first stage when the second stage has no inventory. Their study is different from

this thesis in two major aspects. First, in this thesis, the sources of demand for the

manufacturer and the retailer are not independent; the retailer is the sole source of

demand for the manufacturer. Second, the time after which items pass from the first

stage to the second is an input in their study, whereas it is a decision variable in this

thesis.

Fujiwara et al. [17] establish an optimal ordering policy for a two-stage inventory

system for perishable products. In their setting, both stages belong to the retailer.

The retailer is concerned with finding an order-up-to policy and emergency shipment

rules in case of stock-outs for multiple products. In this thesis, orders from the retailer

are an input, not a decision variable.

Ketzenberg et al. [35] are concerned with the value of information sharing in

a two-tier supply chain. In the setting they study, the ages of perishable products

84

that the manufacturer ships to the retailer varies, as in this thesis. However, while

this thesis establishes rules about minimum ages, they compare two scenarios: 1) age

information is shared with the retailer, and 2) age information is not shared with the

retailer. They focus on the impact of these scenarios on the retailer’s profit.

Blackburn and Scudder [6] develop supply chain strategies for fresh produce. With

fresh produce, shelf life varies depending on the picking rate at the harvest stage and

batch size at the cooling stage of the supply chain. The authors find optimum picking

rate of the grower and optimum batch size of the distributer to minimize the total

cost. Although this thesis is concerned with fixed product life times, as with consumer

packaged products, it is similar to their research, since both studies are concerned

with finding suitable product aging levels as products move in the supply chain.

As we can see, the focus of the existing research studying the two-tier supply chain

for perishable products differs from this thesis. This study builds on the operations

literature by introducing and analyzing the following shipment problem of the manu-

facturer: the minimum remaining shelf life below which it is not cost-effective for the

manufacturer to release the inventory for sale to the retailer.

the retailer

Certain supply chain activities of the manufacturer erode the product shelf life, in-

creasing the chances of product expiration at the retailer. Three main sources for

shelf-life erosion are the following: 1) high inventory, 2) neglect of product rotation,

and 3) delayed and unsynchronized product introductions.

1- High inventory: Consider two opposite scenarios. In the first scenario, the man-

ufacturer holds an inventory of 10 days-of-supply at its distribution center. Assume

85

that the product’s manufactured shelf life is 45 days. In this case, the product may

have 35 days of remaining shelf life when it reaches the retailer. In contrast, in the

second scenario the manufacturer has an inventory of 30 days-of-supply. Then, a

product can have as little as 15 days remaining shelf life when it arrives at the re-

tailer. For the same demand rate, we expect to see a higher probability of expiration

in the second scenario. See Figure 3-1.

by forecast inaccuracy. Big production batches reduce the unit production cost by

minimizing costly changeovers. Similarly, bulk shipments in pallets reduce the han-

dling cost. Clearly, batching is advantageous. However, it erodes the product shelf

life, leaving less time for the retailer to sell through the inventory.

2- Neglect of product rotation: Rotating the products at the manufacturer’s

warehouse to ensure first-in-first-out issuing of inventory reduces the variation in the

remaining shelf life. When products are not rotated, we see instances with very fresh

or very old inventory in shipments to the retailers. Obviously, older inventory has

higher chances of expiration. Further, when rotation is neglected, the product can

even expire on the manufacturer’s warehouse floor. Automated warehousing systems

can perform rotation; however, they require high capital investment. Also, manual

86

rotation takes effort increasing the labor cost, in particular, if the warehouse space is

tight.

3- Delayed and unsynchronized product introductions: For new products,

the manufacturer typically produces one big batch to satisfy the demand of the in-

troductory weeks to prevent stock-outs and/or costly emergency interventions to the

production plan. In the absence of a solid sales and operations process, the new

product can be introduced to different retailers at sparse time intervals. Imagine that

the first batch of production is made on February 1st. Further, the product is intro-

duced to Walmart stores on February 10th, to Walgreens stores on March 1st, and

to 7-Eleven stores on March 15th. If the product has a total of 3 months of shelf life,

there will be only one and a half months remaining time for the 7-Eleven stores to

sell through the first shipment. See Figure 3-2. Synchronized product introductions

avoid this problem.

The extent to which the manufacturer is impacted by the cost of expiration occurring

at the retailer depends on two factors. The first is the sales and distribution model

of the manufacturer. The second is the unsaleables policy.

Consumer packaged products are distributed to retail stores through two types

of sales and distribution models: 1) the direct-store-delivery (DSD) model, and 2)

87

the traditional model. With DSD, the manufacturer delivers the product directly to

the retail store, skipping the retailer’s distribution center, and also generates replen-

ishment orders for the store. Since the manufacturer manages the store inventory,

when an item expires, the retailer is given full credit for the wholesale price. The

manufacturer also bears the reverse logistics costs. With the traditional model, the

unsaleables policy determines the rules of the reimbursement for the cost of expired

items.

Three unsaleables policies are practiced in the industry: 1) the swell policy, 2) the

adjusted-rate policy (ARP), and 3) the joint-industry-report policy (JIR). The swell

policy provides a wholesale price discount (e.g., 1-2%) as a compensation for future

expiration. The ARP policy provides a similar discount; however, the rate changes

every year based on annual audits in the supply chain. These audits aim to find

how much the manufacturer has contributed to the problem. In contrast, the JIR

policy provides full reimbursement to the retailer for the ex-post cost of expiration,

which comprises the wholesale price as well as the cost of reverse logistics. Figure 3-3

summarizes the situations in which the manufacturer is directly responsible for the

cost of expiration.

88

Examples of manufacturers:

Direct-store-delivery channel

Manufacturer A

Examples of

manufacturers:

(with joint-industry-report unsaleables policy)

cost.

Retailers argue that manufacturers lack incentives to ship fresh products under

the swell or ARP policies. Possibly in an attempt to address this issue, some retailers

impose minimum shelf life requirements and have rejection policies refusing shipments

that do not meet the requirement. These rejection policies are not executed strictly

due to the cost or to avoid conflict in the relationship. However, retailers expect that

rejection policies encourage manufacturers to improve the remaining shelf life.

With the DSD model and the JIR policy, the manufacturer has an incentive to

provide fresh products. However, intracompany dynamics of the manufacturer can

counter this incentive. For example, if the plant manager’s cost center is not charged

for the cost of expiration, he/she will have no incentive to produce in smaller batches.

Therefore, in order for different functions to serve the interests of the firm, appropriate

incentive mechanisms need to be in place within the manufacturer’s organization.

Under the DSD model and the JIR policy, product disposition cost of the unsold

product is less than the expiration cost of the sold product for the manufacturer.

Let us consider a case in which it costs $2 to produce an item and transfer it to

a distribution center. When the manufacturer discards this item, the company will

89

lose $2 due to the sunk cost of production and, suppose, $0.10 due to the cost of

reverse logistics. Now, let us imagine the manufacturer sells this product to the

retailer for $2.20, making a gross profit of $0.20, and the product expires on the retail

shelf. Then, the manufacturer will pay the retailer $2.20 and incur a reverse logistics

cost of, perhaps, $1. Notice that the cost of reverse logistics is higher from retailer

stores than it is from distribution centers. In summary, the manufacturer loses $3 if

it sells a product that can expire. However, the loss from disposition will be $2.10

if the product is not sold. Figure 3-4 shows these two scenarios for the sell and

dispose decisions of the manufacturer. Thus, when the remaining shelf life is short,

discarding the product can be a better decision for the manufacturer than selling it

to the retailer.

Cost-of-goods-sold: $2

Wholesale price: $2.20

Gross profit:

sales

$0.20

$2.20

Manufacturer Retailer

disposal expiration

- $2.10 = - $2 - $0.1 - $3 = $2.2 - $2 - $2.2 - $1

gross profit reimbursement

from sales

Figure 3-4: Profit implications of sales and disposal decisions when remaining shelf

life is very short.

According to our informal discussions with retailers, it is not uncommon that a prod-

uct with nearly a year long manufactured shelf life arrives at the retailer with only a

few months remaining. Manufacturers acknowledge this issue and pledge to improve

the situation [64]. Industry reports also emphasize the importance of remaining shelf

90

life to alleviate product expiration [64, 49, 30]. These reports recommend that man-

ufacturers set up policies to prevent shipments containing products with little shelf

life left. Two types of practices prevail in the industry. These are based on 1) a fixed

number of days and 2) a percent of manufactured shelf life as shown in Figure 3-5.

With the first policy type, the manufacturer determines a fixed time as the min-

imum remaining shelf life across different products, usually per product category.

Suppose this time is 60 days; then, products with a remaining life of less than 60 days

are not sent to the retailer. With the second policy type, the minimum remaining

shelf life varies with the manufactured shelf life. During an informal conversation, a

retailer stated that their requirement is 80% for perimeter products and 70% for aisle

products. Consider two perimeter products with 30 and 60 days of manufactured shelf

life. Then, this retailer will refuse shipments that contain items with less than 24 and

48 days of remaining shelf life for the first and the second product, respectively.

The main challenge with current practices is that requirements of retailers and

policies of manufacturers are often not aligned. It is a problem for manufacturers that

retailers impose different requirements for the same product, since most warehouse

management systems are not capable of handling different rules for different customers

[25]. It can also be confusing for a retailer when, for the same type of product, policies

of different manufacturers vary. Then, understandably, the retailer may doubt the

91

efficacy of the policies of some manufacturers.

The other key issue with existing policies is that they fail to account for the de-

mand rate and other relevant factors impacting profits (e.g., expiration cost, profit

margin). Imagine two products in the same product category with the same man-

ufactured shelf life. However, their demand rates differ. For the first product, it

takes one month for the retailer to sell through a pallet of inventory. In contrast, the

second product is a slower moving item that takes 4 months for the retailer to sell

through the same amount of inventory. Suppose the manufactured shelf life for both

products is 5 months and we apply the same shipment policy requiring 3.5 months

(70% of manufactured shelf life) of minimum remaining shelf life. In this case, we

could see expiration with the second product, but not with the first product. This

comparison suggests that it is prudent to consider the demand rate in determining

shipment policies.

inventory

1 pallet

product 1

product 2

expiration

amount

time

1 month 4 months

3.5 months

remaining shelf life

Figure 3-6: Impact of demand rate on expiration with regards to remaining shelf life.

Next, we consider a different scenario in which most attributes of the two products

are the same, including the demand rate, shipment size, shipment policy (e.g., 30 days

of minimum remaining shelf life), probability of expiration at the retailer (e.g., 0.01%),

and gross margin (e.g., $1). However, the cost of expiration of these products differs.

Suppose the cost of expiration for the first product is $4, while it is $2 for the second

92

product. In this case, it may not be reasonable to have the same shipment policy for

both of them. To reduce the probability of expiration for the first product, we can

consider changing its shipment policy to a more conservative one such as a 60 day

minimum remaining shelf life.

In conclusion, existing practices overlook relevant factors, such as the demand

rate and expiration cost. The industry seems to lack appropriate methods to establish

shipment policies. Thus, it is not surprising that a disconnect occurs between retailers

and manufacturers resulting from mismatched expectations. Currently, no framework

exists to set up standards that would align retailers and manufacturers.

As discussed in Section 3.3.2, manufacturers reimburse retailers for the cost of ex-

piration in most situations. Accordingly, we propose that when the remaining shelf

life is short, it may be less costly for the manufacturer to dispose of the product

at its distribution center as opposed to selling it to the retailer. Consequently, the

manufacturer faces the following problem: for a given remaining shelf life, should the

inventory at the distribution center be disposed of or be sold to the retailer?

If the inventory is disposed of, the manufacturer incurs a disposal cost per unit

which we denote by 𝑑. The disposal cost consists of the cost-of-goods-sold of the

product and the reverse logistics cost. If the inventory is used to fulfill a retailer order,

the manufacturer earns a profit of 𝑝, which is the difference between the product’s

wholesale price and cost-of-goods-sold. When this inventory expires on the retailer’s

shelf, the manufacturer incurs an expiration cost of 𝑐. This cost includes the credit

given to the store for the wholesale price and the cost of reverse logistics. Therefore

𝑐 > 𝑝, since wholesale price + reverse logistics cost > wholesale price - cost-of-goods-

sold. This relationship suggests that sales would not be profitable if the product

93

expires. In this situation, the manufacturer is interested in finding the minimum

remaining shelf life when it is economical to sell the product as opposed to disposing

of it from its distribution center over an extended horizon time.

We formulate this problem as follows:

max 𝑎

𝑎

∑︁ ∑︁ ∑︁

s.t. 𝑝 𝑄𝑠 − 𝑐 𝐸𝑠 (𝑎, 𝑓 (𝐷𝑠 ), 𝑄𝑠 ) ≥ −𝑑 𝑄𝑠 (3.1)

𝑠 𝑠 𝑠

In this model, 𝑎 denotes the threshold of remaining shelf life at which the manufacturer

stops selling the inventory to the retailer. 𝑄𝑠 denotes the long-run average shipment

quantity, 𝑓 (𝐷𝑠 ) denotes the demand distribution, and 𝐸𝑠 denotes the long-run average

expiration quantity at store 𝑠. Further, 𝐸𝑠 depends on 𝑎, 𝑓 (𝐷𝑠 ), and 𝑄𝑠 and is

nondecreasing in 𝑎1 . In practice, replenishment orders and corresponding shipments

are determined independent of the shelf life of inventory at the distribution center

(𝑄𝑠 is the solution to the multi-period inventory problem ignoring the shelf life).

Therefore, in our model, shipments are treated as external, while expiration quantity

depends on the remaining shelf life of the shipment and the consumer demand at the

retail store. The left-hand side of the inequality in (3.1) represents the net profit from

selling, while the right-hand side represents the net profit from not selling. Essentially,

optimum 𝑎 is a threshold for the remaining shelf life below which the manufacturer

loses more money from selling than from not selling. We refer to this threshold as

the no-shipment threshold.

1

𝑄𝑡 and 𝐷𝑡 denote the order quantity and demand, respectively, at time period 𝑡 for a given store.

Let us assume zero store inventory at 𝑡 = 0. First, consider the case of first-in-first-out inventory

issuing policy at the store. For the shelf life of 𝑎 = 1, 𝑎 = 2, and 𝑎 = 3, the expiration quantity at

the end of the shelf life will be [𝑄1 − 𝐷1 ]+ , [[𝑄1 − 𝐷1 ]+ − 𝐷2 ]+ , and [[[𝑄1 − 𝐷1 ]+ − 𝐷2 ]+ − 𝐷3 ]+ ,

respectively. Notice that 𝐸𝑥𝑝𝑖𝑟𝑎𝑡𝑖𝑜𝑛(𝑎 = 1) ≥ 𝐸𝑥𝑝𝑖𝑟𝑎𝑡𝑖𝑜𝑛(𝑎 = 2) ≥ 𝐸𝑥𝑝𝑖𝑟𝑎𝑡𝑖𝑜𝑛(𝑎 = 3).

Next, consider the case of last-in-first-out inventory issuing policy. The corresponding expiration

quantities are [𝑄1 − 𝐷1 ]+ , [[𝑄1 − 𝐷1 ]+ − [𝐷2 − 𝑄2 ]+ ]+ , and [[[𝑄1 − 𝐷1 ]+ − [𝐷2 − 𝑄2 ]+ ]+ − [[𝐷3 −

𝑄3 ]+ − [𝑄2 − 𝐷2 ]+ ]+ ]+ . Similarly, 𝐸𝑥𝑝𝑖𝑟𝑎𝑡𝑖𝑜𝑛(𝑎 = 1) ≥ 𝐸𝑥𝑝𝑖𝑟𝑎𝑡𝑖𝑜𝑛(𝑎 = 2) ≥ 𝐸𝑥𝑝𝑖𝑟𝑎𝑡𝑖𝑜𝑛(𝑎 = 3).

94

The following modified version of the problem gives us the minimum remaining

shelf life at which the manufacturer achieves a positive profit from sales.

max 𝑎

𝑎

∑︁ ∑︁

s.t. 𝑝 𝑄𝑠 − 𝑐 𝐸𝑠 (𝑎, 𝑓 (𝐷𝑠 ), 𝑄𝑠 ) ≥ 0 (3.2)

𝑠 𝑠

products leave the manufacturer’s distribution center with a minimum shelf life that

is equal to the profitability threshold. In a shipment policy, the profitability threshold

can be thought of as the target remaining shelf life, while the no-shipment threshold

can establish the sell-versus-dispose rule.

The extent to which 𝑎 impacts 𝐸𝑠 depends on the rotation practices at the store.

If the inventory is rotated on the shelf during replenishment and the consumer chooses

the outermost product on the shelf, we can assume that the inventory is issued ac-

cording to the first-in-first-out (FIFO) policy. If the product is front-loaded to the

shelf without rotation and the consumer still selects the outermost product, the last-

in-first-out (LIFO) policy can be assumed. It is reasonable to expect that rotation

is not performed at all times. As a result, products can be sorted with varying ages

on the shelf. Therefore, we can consider three cases with regards to shelf rotation:

1) FIFO, 2) LIFO, and 3) mixed issuing policy. In our main analysis, we assume

the mixed policy in which the inventory on the shelf is issued based on the uniform

distribution, where the consumer selects a product among different ages of inventory

with equal probability.

We solve this problem using data obtained from our industry collaborator Al-

phaCo. AlphaCo provided us three datasets:

1) Daily shipment and return data for 603 SKUs and 3,025 stores from one distribu-

95

tion center over a one year period.

2) Wholesale price and cost-of-goods-sold data for 1,094 SKUs.

3) Shelf life and case size data for 8,278 SKUs.

We connected these datasets using the identification numbers of SKUs. Further, we

eliminated SKUs that contain a negative cost or a negative wholesale price. As a

result, our final dataset contains 219 SKUs and 2,670 stores, in which a store carries

an average of 27 SKUs and a median of 15 SKUs.

To calculate 𝑝, 𝑐, and 𝑑, we leverage estimates of the following two costs: the cost

of sales and distribution to retail stores and the cost of reverse logistics from retail

stores.

Calculation of 𝑝: AlphaCo approximates a single number as the sales and distribu-

tion cost per case, identical across SKUs. If we suppose this number is $1, we arrive

at 𝑝 by deducting $1 and SKU’s cost-of-goods-sold from its wholesale price.

Calculation of 𝑐: AlphaCo estimates that the cost of reverse logistics at retail stores

is as high as the product’s cost-of-goods-sold, driven by costly handling in units at

stores as opposed to cases or pallets. Accordingly, we compute a uniform number

for the reverse logistics cost as the average of the cost-of-goods-sold across all SKUs.

Imagine that this number is $8. Then, to calculate 𝑐, we add $8 and the wholesale

price of the SKU.

Calculation of 𝑑: We stated that 𝑑 consists of the cost-of-goods-sold and the re-

verse logistics cost at the distribution center. We let the cost-of-goods-sold represent

𝑑, since the reverse logistics cost at distribution centers is negligible at AlphaCo.

96

3.5. Determining targets for the remaining shelf life

To find the optimum 𝑎 specified in our optimization model, we use a line search

starting from week one. At each iteration, we calculate the net profit from sales. We

∑︀ ∑︀

retain the value of 𝑎 at the iteration where 𝑝 𝑠 𝑄𝑠 − 𝑐 𝑠 𝐸𝑠 (𝑎, 𝑓 (𝐷𝑠 ), 𝑄𝑠 ) exceeds

∑︀ ∑︀

−𝑑 𝑠 𝑄𝑠 . This value is the no-shipment threshold. The value of 𝑎 at which 𝑝 𝑠 𝑄𝑠 −

∑︀

𝑐 𝑠 𝐸𝑠 (𝑎, 𝑓 (𝐷𝑠 ), 𝑄𝑠 ) exceeds 0 is the profitability threshold. The following graph in

Figure 3-7 represents an example for one SKU.

$3.00

sales disposal

$2.00

$1.00

NET PROFIT (per day for 10 stores)

$-

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39

$(1.00)

profitability threshold

$(2.00)

$(3.00)

no-shipment threshold

$(4.00)

$(5.00)

$(6.00)

$(7.00)

$(8.00)

Remaining shelf life (in weeks)

Figure 3-7: Net profit from sales and disposal for one SKU as remaining shelf life

changes.

The manufactured shelf life of this product is 39 weeks. The x-axis in the graph

represents the remaining shelf life, 𝑎, while the y-axis represents the net profit. The

thick black line shows the net profit for the case where the product is not sold to the

retailer, which is represented by the right-hand-side of the inequality in optimization

∑︀

model 3.1, −𝑑 𝑠 𝑄𝑠 . The curved gray line shows the net profit from sales, which is

97

∑︀

represented by the left-hand-side of the inequality in model 3.1 and 3.2: 𝑝 𝑠 𝑄𝑠 −

∑︀

𝑐 𝑠 𝐸𝑠 (𝑎, 𝑓 (𝐷𝑠 ), 𝑄𝑠 ). Notice that the net profit from sales is increasing with the

remaining shelf life. This increase occurs because as the remaining shelf life increases,

the amount of expiration, 𝐸𝑠 , decreases, eventually reaching zero in this example. The

optimum 𝑎 in model 3.1 (no-shipment threshold) is the minimum remaining shelf life

where the net profit from selling exceeds the net profit from not selling, which is

5 weeks here (marked with red color on the graph). In model 3.2, the optimum

𝑎 (profitability threshold) is the minimum remaining shelf life where the new profit

from selling is positive, which is 13 weeks for this product (marked with blue color).

Generally, there is no guarantee that optimum 𝑎 exists either for model 3.1 or 3.2,

since in the extreme case, in which 𝑄𝑠 = 𝐸𝑠 , both 𝑝 − 𝑐 ≥ −𝑑 (model 3.1) and

𝑝 − 𝑐 ≥ 0 (model 3.2) are not true.

In the example in Figure 3-7, when the remaining shelf life is less than 5 weeks, it

is more economical for AlphaCo to dispose of this product at the distribution center as

opposed to selling it to retailers, since the amount of expiration will be very high if it is

sold. Further, when the remaining shelf life is between 5 and 13 weeks, AlphaCo does

not generate a positive profit from sales due to the amount of expiration; however,

it is viable to sell since the loss from selling is less than the loss from not selling.

Finally, a remaining shelf life that is higher than 13 weeks is ideal for AlphaCo, since

in this case sales generate positive profit for this product.

Notice that the shipment policy we establish here, through computation of the

no-shipment threshold, does not vary across stores. It is possible to find a separate

optimum 𝑎 for each SKU-store. This customization simply requires dropping the 𝑠

subscript in the model 3.1. An SKU-store level shipment policy can yield higher prof-

its but is difficult to implement. In this study, we focus on shipment rules established

at the SKU-distribution center level since 1) it is an improvement compared to the

existing rules that are at the SKU level and 2) in practice, most manufacturers (in-

98

cluding AlphaCo) are not capable of handling different rules for different customers

[25].

The following are important elements of the algorithm regarding the time horizon,

consumer demand, and set of retail stores:

Time horizon: First, we replicate the original shipment data to obtain a long

time series of shipments. We replicate as opposed to randomize the orders since in

practice the sequence of orders is not random; orders follow the pre-determined visit

schedule for the store. For each SKU, remaining shelf life, and store combination, the

algorithm stops at the time period at which the average expiration quantity converges.

Through experimentation, we developed the following criteria for convergence: 1) the

number of time periods needs to be 1,000 at the minimum and 4,000 at the maximum

and 2) the average expiration for the last 200 time periods and the previous 200 time

periods needs to be less than 0.001. Once convergence is achieved, long-run average

expiration quantity and shipment quantity are obtained.

Consumer demand: Point-of-sale (POS) data would provide the ideal informa-

tion to understand store demand. However, AlphaCo does not have the POS data

for most customers. For example, convenient stores and gas stations usually lack

the technology to collect or transfer POS. Therefore, in the absence of POS data,

we leverage shipment data. We assume Poisson arrival of consumer purchase based

on literature with similar settings [2, 27]. We calculate the daily demand rate, 𝜆𝑠 ,

as [total shipments-total returns]/365. At each time period during computation, de-

mand is drawn from the Poisson distribution using the demand rate of the associated

SKU-store.

Retail stores: In the dataset, the number of stores that carries a given SKU

ranges between 1 and 1,772 with mean and median values of 271.6 and 163, respec-

tively. Because including the full set of stores increases the run-time substantially,

we chose a random sample of 10 stores for each SKU to speed up the computation.

99

To conclude, the algorithm is the following. In this program, 𝑖, 𝑎, 𝑑, 𝑠, and 𝑡 index

the SKU, remaining shelf life (in weeks), age of inventory (in days), store, and time

(in days), respectively. In addition, 𝐼 denotes the inventory.

for i in 1:219 do

a=1

while 𝑝𝑟𝑜𝑓 𝑖𝑡𝑎,𝑖 < 0 do

daily=7a

for 𝑠 in 1:10 do

t=1

difference=a large number

while (difference>0.01 and 1000<t) or t<4000 do

for d in 1:daily do

capture expired quantity: 𝐸𝑎,𝑖,𝑠,𝑡 = 𝐼𝑖,𝑠,1,𝑡

update inventory from yesterday: 𝐼𝑖,𝑠,𝑑−1,𝑡+1 = 𝐼𝑖,𝑠,𝑑,𝑡

update inventory based on shipment arrival: 𝐼𝑖,𝑠,𝑑,𝑡+1 = 𝑄𝑖,𝑠,𝑡

update inventory based on customer demand, 𝐷𝑖,𝑠,𝑡

end for

1 ∑︁

calculate average expiration until time period t: 𝐸𝑎𝑣𝑔𝑎,𝑖,𝑠,𝑡 = 𝐸𝑎,𝑖,𝑠,𝑡

𝑡 𝑡

𝑡−200 𝑡−400

1 ∑︁ 1 ∑︁

if t>1000 then difference= 𝐸𝑎𝑣𝑔𝑎,𝑖,𝑠,𝑡𝑖𝑚𝑒 − 𝐸𝑎𝑣𝑔𝑎,𝑖,𝑠,𝑡𝑖𝑚𝑒

200 𝑡𝑖𝑚𝑒=𝑡 200 𝑡𝑖𝑚𝑒=𝑡−200

t=t+1

end while

retain expired quantity per store at convergence: 𝐸𝑐𝑜𝑛𝑣𝑎,𝑖,𝑠 = 𝐸𝑎𝑣𝑔𝑎,𝑖,𝑠,𝑡

end for

∑︁ ∑︁ 1 ∑︁

𝑝𝑟𝑜𝑓 𝑖𝑡𝑎,𝑖 = 𝑝𝑖 𝑄𝑖,𝑠,𝑡 − 𝑐𝑖 𝐸𝑐𝑜𝑛𝑣𝑎,𝑖,𝑠

𝑠 𝑡

𝑡 𝑠

a=a+1

end while

100

∑︁ ∑︁ 1

no-shipment threshold(i)= index of a where 𝑝𝑟𝑜𝑓 𝑖𝑡𝑎,𝑖 > −𝑑 𝑄𝑖,𝑠,𝑡

𝑠 𝑡

𝑡

profitability threshold(i)= index of a where 𝑝𝑟𝑜𝑓 𝑖𝑡𝑎,𝑖 > 0

end for

Running this program for 219 SKUs in our dataset, we obtained the following

results. The no-shipment threshold ranges between 1 and 26 weeks with mean and

median values of 7.2 and 5 weeks, respectively. Further, the profitability threshold

ranges between 2 and 52 weeks with mean and median values of 17.6 and 17 weeks,

respectively. The histograms in Figure 3-8 depict the frequency of the results.

60

50

50

40

40

Frequency

Frequency

30

30

20

20

10

10

0

0 5 10 15 20 25 0 10 20 30 40 50

No−shipment threshold Profitability threshold

Here, we see SKUs with very high thresholds on the right tail of both histograms.

AlphaCo can consider discontinuing these SKUs, since it may not be profitable to

maintain product freshness at these levels (e.g., 30 weeks of remaining shelf life).

In making these decisions, one can take into account the manufacturing cycles. For

example, if a slow moving product is usually produced every 4 weeks and this product

already has a manufactured shelf life of 12 weeks, it may be hard to maintain a

101

freshness level of 10 weeks remaining shelf life. Our results can also be used as a

feedback for the production plan. Products that require high levels of freshness, as

indicated by high thresholds of no-shipment or profitability, can be produced more

often with smaller batch sizes. Furthermore, SKUs that require high freshness levels

could be given more emphasis for warehouse rotation or could be placed at convenient

locations on the warehouse floor to make rotation easier.

areas: 1) the demand rate, 2) sampling of stores, and 3) the assumption regarding

shelf rotation.

Since we use an estimate of demand, it is useful to look at the sensitivity of results with

varying demand rates. We run our algorithm for two additional demand scenarios:

demand rate that is half a standard deviation less than the baseline rate and demand

rate that is half a standard deviation more than the baseline rate. We obtain the

results in Figure 3-9 for the no-shipment threshold.

102

25

baseline

minus 0.5 standard deviation

20 Example of an SKU sensitive

no-shipment threshold

plus 0.5 standard deviation

to demand variation

15

(in weeks) 10

0

SKUs (50)

red), diamond (in blue), and triangle (in green) points represent the results with the

decreased, baseline, and increased demand scenarios, respectively.

As seen in Figure 3-9, some SKUs are sensitive to demand variation (those with

a wider gap between the red, blue, or green points), while others are not. To make

visualization easier, we include only the first 50 SKUs from our list on this graph.

Results for the full dataset (219 SKUs) are summarized in Table 3.1.

+ 1 week 67 31% - 1 week 87 40%

+ 2 weeks 28 13% - 2 weeks 24 11%

+ 3 weeks 24 11% - 3 weeks 15 7%

+ 4 weeks 9 4% - 4 weeks 17 8%

+ 5 weeks 14 6% - 5 weeks 3 1%

+ > 5 weeks 15 7% - > 5 weeks 3 1%

As Table 3.1 shows, when the demand rate is reduced by half a standard deviation,

the threshold remains the same for 28% of the SKUs. Similarly, the threshold does

not vary for 32% of the SKUs when the demand rate is increased by half a standard

103

deviation. The change is one week for 31% and 40% of the SKUs for the lower and

higher demand scenarios, respectively. Then, for majority of the SKUs, the results

are robust (with either no change or one week of change) to moderate changes in

demand rate.

For 41% of the SKUs with the reduced demand rate scenario and 28% of the

SKUs with the increased demand rate scenario, the threshold changes by more than

2 weeks. The magnitude of the change differs across SKUs. For example, for 7% of

SKUs, the threshold increases by more than 5 weeks with the lower demand scenario.

To understand what drives the sensitivity of results, we evaluate the relationship

between the range of the threshold across two demand scenarios and several prod-

uct attributes. We find that the range is mildly and negatively correlated with the

product’s gross margin, with a correlation coefficient of -0.15. This means that the

no-shipment threshold for products with higher gross profit can be more robust to

variation in demand. This result can be explained such that demand is relatively

less impactful in the optimization model for products with higher gross margin. Fur-

ther, the correlation coefficient between the range and the (baseline) median demand

rate across sampled stores is -0.21. This relationship suggests that the no-shipment

threshold for faster selling products can be less sensitive to demand variation. This

can occur because for SKUs with high demand, moderate perturbations in demand

are less likely to change the optimization result. Similar results hold when we use

coefficient of variation instead of range to measure variation.

We expect demand variations to impact the optimization results. Our analy-

sis shows that we find that some SKUs are more sensitive than others to moderate

changes in demand, while some are robust. To be conservative, we can use the results

of a reduced demand scenario in practice.

104

Variation in store samples

In our dataset, the number of stores that carry a given SKU varies greatly. The store

count per SKU is between 1 and 1772 stores, with median and mean values of 163

and 271.6, respectively. Figure 3-10 shows the frequencies of the number of stores per

SKU.

80

60

Frequency

40

20

0

Store count

For our analysis, we select a random sample of 10 stores to speed up the compu-

tation. In this section, we test the sensitivity of results to varying samples. For this

purpose, we run our algorithm using 5 distinct samples, each containing 10 stores.

Then, we exclude the SKUs that are carried at less than 50 stores from our dataset,

which reduces our number of data points from 219 to 181 SKUs.

The following graph shows the results for a selection of SKUs. Each sample is

represented by a different shape and color on the figure. For each SKU, results from

5 samples are sorted to make the interpretation easier.

105

25

2nd lowest value across 5 samples

3rd lowest value across 5 samples

no-shipment threshold 20

4th lowest value across 5 samples SKUs sensitive to

highest value across 5 samples store sample variation

(in weeks)

15

lowest value across 5 samples

10

0

SKUs (50)

Figure 3-11: Variation in no-shipment thresholds across SKUs with different store

samples.

As seen in Figure 3-11, the variation for some SKUs is more than others. To assess

the sensitivity of the optimization outcome to varying store samples, we calculate the

gap between the highest and the lowest no-shipment threshold across 5 samples for

each SKU. We find that the gap is smaller than 2 weeks for 28% of the SKUs, which we

can consider as robust to sample variation. Exactly 34% of the SKUs are moderately

sensitive to variation in store samples, with a gap that is between 3 and 5 weeks. The

remaining of the SKUs (38%) are sensitive to sample variation, with a gap that is

higher than 5 weeks.

The sensitivity may be explained by the small size of the overall pool from which

we select the sample. For example, an SKU that is carried at 1,000 stores could be

less sensitive to sample variation than an SKU that is carried at 60 stores. To test

this hypothesis, we calculate the correlation between the store count per SKU and the

gap between the highest and lowest optimum values across 5 samples. We find that

the correlation coefficient is -0.33, which supports our prior belief that the results are

less sensitive with large representations.

Based on this analysis, during implementation, one can consider two possible

106

strategies to reduce variation due to sampling of stores: 1) increase the sample size

which could depend on the store count, or 2) run the algorithm with more samples,

selecting the median (or mean) value across different runs to establish the shipment

policy.

Our initial analysis is based on a mixed inventory issuing policy on shelf rotation,

which assumes that the consumer selects a product among different ages of inventory

with equal probability. This assumption is conservative considering that store em-

ployees are expected to perform rotation each time they place new items on the shelf.

However, from industry reports [30] and based on our conversations with AlphaCo,

we know that neglect of rotation is not uncommon.

In this section, we analyze two extreme scenarios regarding rotation: full rotation

(FIFO) and no rotation (LIFO). After running our algorithm for these two scenarios,

we find that the impact of rotation on the results varies across SKUs. Figure 3-12

displays the results for a sample of SKUs. In the figure, square (in red), diamond (in

blue), and triangle (in green) points represent the results for the LIFO, mixed issuing,

and FIFO scenarios, respectively.

107

25

Mixed issuing

LIFO

20

no-shipment threshold

FIFO

SKUs not impacted

by shelf rotation

15

(in weeks)

10

0

SKUs (50)

rotation assumptions.

The following table summarizes the results for the full dataset of 219 SKUs.

tion on rotation.

# of SKUs % of SKUs

r=m 55 25%

m=nr 19 9%

r<>m<>nr 34 16%

In Table 3.2, r, m , and nr denote the full rotation, mixed issuing, and no rotation

scenarios, respectively. The equality r=m=nr represent the case where results are

identical for all 3 scenarios, which we find in 51% of the SKUs. For 16% of the SKUs,

results vary for all 3 scenarios, denoted by r<>m<>nr in the table. For 25% of

the SKUs, full rotation and mixed issuing yield the same results, shown by r=m.

Also, for 9% of the SKUs, results of the mixed issuing and no rotation scenarios are

identical (m=nr).

It is surprising that for a large portion of the data points (111 SKUs), rotation

does not have an impact. To get a better understanding of how rotation affects opti-

mization results, we select two SKUs. For the first SKU, the no-shipment threshold

108

for all three rotation scenarios is identical, 5 weeks. For the second SKU, there is a

wide gap in the no-shipment threshold (10 weeks) between the no rotation and full

rotation scenarios. Figure 3-13 shows the outcome of the optimization algorithm.

SKU$1$ SKU$2$

$4.00 $10.00

$5.00

$2.00

NET PROFIT (per day for 10 stores)

$-

1 11 21 31 41 51 61 71 81 91 101

$-

$(5.00)

1 4 7 10 13 16 19 22 25 28 31 34 37

$(2.00) $(10.00)

week 20 Sales (LIFO)

$(15.00) Sales (Mixed issuing)

$(4.00) week 14 Sales (FIFO)

$(20.00) Disposal

$(6.00) Sales (LIFO) week 10

$(25.00)

week 5 Sales (Mixed issuing)

$(8.00)

Sales (FIFO) $(30.00)

Disposal

$(35.00)

$(10.00)

$(40.00)

$(12.00) $(45.00)

Remaining shelf life (in weeks) Remaining shelf life (in weeks)

Figure 3-13: Examples of SKUs where rotation has no impact (first graph) and rota-

tion has an impact (second graph) on the no-shipment threshold.

In Figure 3-13, the three curved lines with green, red, and blue colors represent

the net profit from sales for the FIFO (full rotation), mixed issuing (some rotation),

and LIFO (no rotation) scenarios, respectively. The straight line in purple represents

the loss from disposal. As the graph shows, rotation does impact net profits for both

SKUs since three curved lines representing the three rotation scenarios are separate.

Notice that the net profit with full rotation is higher than (or equal to) the net profit

with some rotation; also, net profit with some rotation is higher than (or equal to)

the net profit with no rotation. Thus, rotation impacts profits favorably.

However, the impact of rotation on the no-shipment threshold is inconsistent. For

SKU 1, there is no impact; this means that the net profit associated with each scenario

exceeds the loss from disposal at the same remaining shelf life, which is 5 weeks. This

case occurs because the difference between net profits of the three rotation scenarios

is very narrow at lower values of the remaining shelf life. On the contrary, for SKU 2,

109

net profits for the three rotation scenarios intersect the loss from disposal at different

values of the remaining shelf life (weeks 10, 14, and 20 for the FIFO, mixed issuing,

and LIFO scenarios, respectively), producing different no-shipment thresholds.

We observe that product attributes of these two SKUs differ. For SKU 2, the

gross profit, cost of expiration, and cost of disposal are higher than they are for SKU

1 by a large margin, which explains the narrow difference in net profits across three

rotation scenarios for SKU 1. To see whether this observation is general, we calculate

the correlation between product attributes and the gap in the no-shipment threshold

between the no rotation and full rotation scenarios. We find the correlation coeffi-

cients to be 0.16, 0.48, and 0.26 for the gross margin, cost of expiration, and cost

of disposal, respectively. This result suggest that for SKUs with higher cost of ex-

piration, the impact of rotation on the no-shipment threshold will be more substantial.

costly; thus, we investigate approximation methods to reduce run time and imple-

mentation effort.

Our analysis includes only 219 SKUs at a single distribution center. Still, the run

time of the algorithm is considerable (approximately 4 hours on a personal computer)

due to multiple dimensions of time periods, stores, and age of the inventory. The

number of time periods varies between 1,000 and 4,000 for each combination of the

remaining shelf life, store, and SKU. Further, even after sampling stores, the number

of store-SKU combinations (2,190) is sizable. The age of inventory also increases the

problem substantially. For example, for a product with 10 weeks of remaining shelf

life, the array containing the inventory information has 10 x 7=70 elements.

The full set of data points requiring shipment rules at AlphaCo consists of 289,727

110

SKU-distribution center combinations across 390 distribution centers. Implementing

our solution for the full dataset is possible; however, it requires considerable effort

and resources. Thus, in search for a less expensive alternative to optimization, we

assess regression analysis as a heuristic to substitute for optimization.

We evaluate four regression techniques: 1) random forest, 2) regression tree, 3)

support vector machine, and 4) Poisson regression. Among these methods, random

forest, regression tree, and support vector machine are machine learning algorithms,

whereas Poisson regression is a parametric model. We consider both nonparametric

and parametric methods in our assessment to understand their relative performance.

Among the parametric methods that restrict predicted values to nonnegative inte-

gers (which is appropriate in our application), we choose Poisson regression since it

is commonly used in situations that do not require special considerations such as

overdispersion, zero-inflation, or zero-truncation.

Recently, use of machine learning in solving operations management problems has

gained popularity. In the limited existing literature, machine learning typically is used

to estimate unknown parameters of a model (e.g., parameters of a demand model).

Next, this model is used in an optimization problem (e.g., revenue maximization) [57].

In this research, we use machine learning in the reverse order. We first obtain results

from a complex optimization problem; later, the output of optimization is used as an

input to the machine learning algorithm to predict optimum values, as portrayed in

Figure 3-14.

111

data

(features)

approximated

Optimization optimum values (targets) Machine learning optimum values

algorithm algorithm (predicted values)

Our goal is to predict the optimum values that were obtained through our opti-

mization algorithm using regression techniques. As features (called independent vari-

ables in classical statistics), we leverage the parameters of the optimization model

(e.g., cost of expiration, gross margin). Our unit of analysis is an SKU-distribution

center combination. Among the parameters used in the optimization model, ship-

ment quantity and demand are SKU-store specific attributes. Therefore, they need

to be aggregated at the SKU-distribution center level. On the other hand, cost of

expiration, gross margin, cost of disposal, and shelf life are SKU specific attributes.

Then, they can be used as they are. Regarding shipment quantity and demand, we

develop the following features: total shipments, average demand rate, and standard

deviation of demand rate across stores. In addition, we include case size as a feature,

since many of the shipments are made in single cases. Table 3.3 lists the features used

in predicting the no-shipment threshold and the profitability threshold. Notice that

the cost of disposal is not used in the prediction of the profitability threshold since

this feature is not a part of the optimization model.

112

Table 3.3: Features used in prediction of no-shipment and profitability thresholds

Total shipments X X

Average demand rate across stores X X

Standard deviation of demand rate across stores X X

Case size X X

Cost of expiration X X

Gross profit X X

Cost of disposal X X

Shelf life X X

Case size X X

training and a test dataset. The training dataset includes 150 of the SKUs, while the

test dataset includes the remaining 69 SKUs. Typically, an algorithm is trained on

the training dataset, which creates a function (i.e., predictor) that maps features to a

target (called the dependent variable in classical statistics). This function is applied

to the test dataset to make predictions. The error obtained at this step is used to

evaluate different algorithms. Our training step involves two procedures: cross vali-

dation and bagging. Figure 3-15 depicts our overall prediction approach.

113

Training Testing

(150 SKUs) (69 SKUs)

Prediction

Cross validation Bagging

Prediction

(5-fold) Best algorithm (100 samples)

100 Prediction

parameters

predictors . x100

Validation Training .

.

1 .

2 .

3

4

5 Average

prediction

Cross validation: Our goal with cross validation is to tune the algorithm param-

eters (e.g., complexity parameter in regression tree) to improve accuracy. We apply

a 5-fold cross validation process which is performed as the following. The training

dataset is split into 5 partitions. At each of 5 iterations, a different partition is picked

as the validation dataset and the remaining 4 partitions are combined as the training

dataset. Also, at each iteration, the algorithm is trained using the training dataset

and is tested using the validation dataset over an array of varying parameter values.

The average error (across 5 iterations) is used to evaluate a given parameter value.

The goal is to pick the parameter value that minimizes the error. Regression tree

and support vector machine algorithms can benefit from parameter tuning more than

the other models since they include parameters that adjust for overfitting. Poisson

regression uses the iteratively reweighted least squares algorithm which typically does

not require parameter adjustments. The random forest is a sophisticated algorithm

consisting of bagging and feature selection procedures to control variance and over-

fitting.

For all computations in this research, we use the statistical language R version

3.1.2 [54] utilizing packages stats, randomForest [41], rpart [60], and e1071 [44].

For cross validation of the regression tree algorithm, we tune the complexity param-

114

eter using R’s rpart function. The complexity parameter prevents overfitting by

influencing the tree pruning process. We find that the best complexity parameter

is 0.037. For the support vector machine algorithm, we tune the gamma and cost

parameters using R’s tune and svm functions. Combinations of 0.5 and 8 for the

gamma and the cost parameters work best for the no-shipment threshold; 0.5 and 16

are the best parameter combinations for the profitability threshold.

Bagging: Bagging generates many samples from the training dataset with re-

placement. The size of the sample is the same as the size of the training dataset.

The goal is to avoid dataset overfitting and thus improve the accuracy of prediction.

Since the random forest method already includes its own bagging procedures that are

part of the algorithm, we perform bagging only for regression tree, support vector

machine, and Poission regressions. We produce 100 samples. Then, we train each

of the three algorithms on 100 different training datasets. Bagging is an important

step in our training process to mitigate the limitations of a small dataset (219 data

points).

Prediction: Because we do not perform bagging for random forest, the prediction

step for this method involves testing one predictor on the test dataset. For the other

three methods, we perform testing using the 100 predictors we produced for each

algorithm. Then, we aggregate the results by averaging the predictions over 100

cases.

We find that the random forecast algorithm performs significantly better compared

to the other regression methods, producing the smallest mean squared error (MSE)

value in predicting both the no-shipment threshold and the profitability threshold.

For the no-shipment threshold, the MSE values are 8.4, 10.1, 19.4, and 53.1 for

the random forest, regression tree, support vector machine, and Poisson regression

methods, respectively. Figure 3-16 shows the prediction results for the 69 SKUs used

in the test dataset. The x-axis represents the optimum value, while the y-axis shows

115

the predicted value. The MSE results are noted on the upper right corner of each

graph. As Figure 3-16 shows, optimum and predicted values are more aligned with the

random forest method than in other methods. The regression tree and support vector

machine methods are the second and third best performing methods, respectively. As

we can see, Poisson regression performs very poorly.

25 25

20 20

Predicted

Predicted

15 15

10 10

5 5

0 0

0 5 10 15 20 25 30 0 5 10 15 20 25 30

Optimum Optimum

25 25

20 20

Predicted

Predicted

15 15

10 10

5 5

0 0

0 5 10 15 20 25 30 0 5 10 15 20 25 30

-5 -5

Optimum Optimum

old.

3-17. The MSE values for the random forest, regression tree, support vector machine,

and Poisson regression methods are 16.8, 31.3, 25.9, and 272.2, respectively. Accord-

ingly, the random forest algorithm performs best. In contrast with the prediction of

the no-shipment threshold, the support vector machine method performs better than

the regression tree method here. Similar to the previous case, the performance of

Poisson regression is worse than the other three methods by a large margin.

116

Random(Forest( MSE=16.8 Regression(Tree( MSE=31.3

40" 40"

35" 35"

30" 30"

Predicted(

Predicted(

25" 25"

20" 20"

15" 15"

10" 10"

5" 5"

0" 0"

0" 10" 20" 30" 40" 50" 0" 10" 20" 30" 40" 50"

Op+mum(( Op+mum((

40" 40"

35" 35"

30" 30"

Predicted(

Predicted(

25" 25"

20" 20"

15" 15"

10" 10"

5" 5"

0" 0"

0" 10" 20" 30" 40" 50" 0" 10" 20" 30" 40" 50"

Op+mum(( Op+mum((

old.

features. On the contrary, the random forest, support vector machine, and regression

tree are non-parametric machine learning tools that do not restrict the features in

any way. The support vector machine uses hyperplanes to separate the data points in

multi-dimensional space. Regression trees (random forest is a type of regression tree)

partition the space into rectangles grouping observations within those rectangles. The

nonrestrictive nature of machine learning tools may explain why they perform well in

our application in which we predict the outcome of a complex optimization algorithm.

In addition, unlike parametric methods, machine learning tools do not need trans-

formation or scaling of features. Feature transformation is already embedded in the

support vector machine algorithm through kernel functions; also, tree methods inher-

ently do not require data transformation since the algorithm works based on the rank,

not on actual data values. Our prediction process does not include a data preparation

117

step to scale or transform the features, which could partially explain the large margin

in performance between Poisson regression and machine learning methods. Based on

our application, we conclude that machine learning tools are advantageous methods

to approximate complex optimization models with high predictive power and easy

use without a need for prior data manipulation.

1) The minimum remaining shelf life required for shipments that are from plants

to distribution centers. AlphaCo refers to this threshold as ship to satellite.

2) The minimum remaining shelf life required for shipments that are from distribu-

tion centers to retail stores, which is called ship to trade. This rule is comparable

to the no-shipment threshold we defined in this thesis.

3) The ideal remaining shelf life when the product arrives at the store, referred to

as target to trade. This threshold is similar to the profitability threshold described in

this research.

AlphaCo’s shipment policy is based on the product’s manufactured shelf life. For

example, for all SKUs with a manufactured shelf life of 10 weeks, ship to satellite,

ship to trade, and target to trade values are 7, 5, and 8 weeks, respectively. Graph

3-18 shows how the shipment policy changes with the manufactured shelf life. The

x-axis represents SKUs with distinct manufactured shelf lives, while the y-axis shows

the shipment policy along with the manufactured shelf life.

118

120

Manufactured shelf life

100 Target to trade

Ship to satellite

80 Ship to trade

(in$weeks)$ 60

40

20

0

SKUs with distinct manufactured shelf lives

As seen on the graph, target to trade values follow the manufactured shelf life.

However, ship to satellite and ship to trade initially increase with the manufactured

shelf life but stay constant after a certain point. Basically, AlphaCo’s policy can be

seen as a mix of two industry practices discussed in Section 3.3.3 (minimum remaining

days and percent of manufactured shelf life). Unlike the rules of some manufacturers

practicing the minimum remaining days policy, shipment rules do not change across

different product categories at AlphaCo. The policy is determined based on the SKU’s

manufactured shelf life and is uniform across different distribution centers carrying

the SKU. Also, AlphaCo’s shipment policies are built upon managers’ experience and

are not based on a rigorous empirical analysis.

This section evaluates the value of optimization. Figure 3-19 compares AlphaCo’s

ship to trade and our no-shipment threshold values for 219 SKUs included in our

analysis.

119

30

as-is (ship to trade)

25 optimization (no-shipment trehshold)

20

(in weeks)

15

10

0

SKU (219)

As the graph shows, the minimum remaining shelf life is uniform across SKUs

(either 5 or 6 weeks) with the current practice at AlphaCo. However, the optimum

remaining shelf life we found in our analysis varies across SKUs from 1 to 26 weeks.

Figure 3-20 compares AlphaCo’s target to trade (y-axis) and our profitability

threshold (x-axis) values for 219 SKUs. The data points are concentrated in the

upper diagonal of the graph, which shows that the target to trade values are gener-

ally higher than the profitability threshold. Although they both represent the ideal

amount of remaining shelf life, the difference exists since AlphaCo’s criteria for es-

tablishing target to trade is not based on profitability; target to trade values simply

follow the manufacturered shelf life as we see in Figure 3-18.

120

Target to trade (as-is) in weeks

80

60

40

20

0

0 20 40 60 80

Profitability threshold (optimum) in weeks

considering AlphaCo’s ship to trade policy and our no-shipment policy. We calculate

the net profit for sales and disposal decisions, considering only the interval between

the existing policy and the optimum policy values. We take into account only this

interval, because existing and optimum strategies are identical for values outside of

this range. Let us consider two scenarios. In the first one, AlphaCo’s existing policy

requires a minimum shelf life of 5 weeks while optimization suggests 9 weeks. For the

remaining shelf life that is less than 5 weeks, both the current and optimum strategies

suggest discarding the product; further, for the remaining shelf life that is higher than

9 weeks, both strategies suggest selling the product. However, strategies vary for

the remaining shelf life that is between 5 weeks and 9 weeks: optimization suggests

discarding the product while the existing strategy suggests selling the product. Next,

consider the opposite scenario in which the current value is higher than the optimum

value for the minimum required shelf life, 5 weeks and 3 weeks, respectively. In this

case, strategies outside of the range of 3 and 5 weeks are identical. However, for the

values within the range, optimization suggests selling the product while the existing

121

rule suggests discarding the product. The following figure explains these two scenarios

and how their difference relates to the simulation approach.

Scenario 1: Scenario 2:

interval considered in interval considered in

simulation simulation

optimization: discard discard sell discard sell sell

shelf life.

We adopt the following approach for simulation. Considering the first scenario

depicted in Figure 3-21, we simulate the remaining shelf life for 5, 6, 7, 8, and 9 weeks.

For each value, we compute the net profit from sales and from disposal decisions. We

record the average net profit from sales under the as-is policy, while the average net

profit from disposal is recorded under the optimum policy. For example, for the SKU

that is considered in this scenario, average net profits from sales and disposal are

$-1.29 and $-0.97, respectively. Next, consider the second scenario. Similarly, we

simulate the remaining shelf life for 3, 4, and 5 weeks. In this case, the average net

profit from sales is recorded under the optimum policy, while the average net profit

from disposal is recorded under the as-is policy, $-3.15 and $-4.57, respectively for

the SKU selected in this scenario. Graph 3-22 shows the cumulative net profit for the

as-is and optimization based policies across 219 SKUs in our dataset.

122

Comparison across 219 SKUs

0.00

as-is optimization

-100.00

-200.00

-300.00

-400.00

41%

-500.00 improvement

-600.00

-700.00

-800.00

this number into savings in monetary value since we do not have data on the actual

remaining shelf life in historical shipments. As a result, we do not know how often

current rules apply in actual shipments. Neverthless, the simulation approach we use

here gives us a good idea about the relative value of optimization.

the same prediction process followed in Section 3.5.3, we train the random forest

algorithm using 150 SKUs and make predictions on the remaining 69 SKUs. Ap-

plying a similar simulation approach used in comparison of the current practice and

optimization, we compare the existing policy, optimum policy, and the policy based

on approximation. We evaluated two scenarios when we compared the existing and

optimum policies: 1) optimum value is higher than the existing value and 2) opti-

mum value is smaller than the existing value, as shown in Figure 3-21. In this case,

we have 12 different scenarios. Let e, o, and a denote the existing, optimum, and

approximate values, respectively, for the minimum required remaining shelf life. The

12 scenarios are 1)e>o>a, 2)o>e>a, 3)a>o>e, 4)a>e>o, 5)o>a>e, 6)e>a>o,

123

7)e>o=a, 8)o>e=a, 9)a>o=e, 10)o=a>e, 11)e=a>o, and 12)o=e>a. Figure

3-23 explains the simulation approach for one example.

Scenario: o>a>e

interval considered in

simulation

optimization: discard discard discard sell

approximation: discard discard sell sell

5as-is 7approximation 9optimization

and approximation based.

For all three policies, we calculate the average net profit between the minimum

and maximum of e, o, and a. In the example depicted in Figure 3-23, the optimum,

approximate, and as-is values are 9, 7, and 5 weeks, respectively. The net profit for

the as-is policy is calculated as the average net profit from sales between weeks 5 and

7. For the optimum policy, we compute the loss from disposal for the same interval.

The net profit for approximation is calculated as the weighted average of the loss

from disposal between weeks 5 and 7 and the profit from sales between weeks 7 and

9. Each SKU falls into one of the 12 scenarios listed earlier. Lastly, we add the net

profit across 69 SKUs to compute the overall net profit for the as-is, optimization,

and approximation based policies, which is shown in Figure 3-24.

124

Comparison across 69 SKUs

$0.00

as-is optimization approximation

-$200.00

-$400.00

-$600.00 48%

improvement

-$800.00

14%

-$1,000.00

improvement

-$1,200.00

As seen in Figure 3-24, compared to the current policy, optimization provides 48%

savings, while machine learning approximation provides 14% savings for a sample of

69 SKUs considered in the analysis. Accordingly, machine learning approximation

gives us 28% of the benefit we can achieve with optimization. This is a considerable

improvement considering the minimal time (i.e., a few seconds) required to run the

machine learning algorithm.

Currently, the industry lacks standards to manage the remaining shelf life of prod-

ucts entering the retailers’ supply chain. As a result, we see a mismatch between

retailers’ expectations and manufacturers’ practices. Further, requirements of retail-

ers and rules of manufacturers can vary across different firms for the same type of

product. Existing shipment policies are simply rules determined arbitrarily or based

on managers’ experience. Here, it seems that the industry is in need of quantitative

methods to establish shipments policies based on rigorous analysis. Rules that are

built on a solid foundation can be adopted more easily in the industry, for they can

help achieve alignment across manufacturers and retailers.

125

This research develops a framework (an optimization method and machine learning

approximation of optimal values) to establish targets for the minimum remaining shelf

life in manufacturers’ supply chain. Our methodology takes into account relevant

factors such as the demand rate or cost of expiration, which are disregarded in current

industry practices. We built an optimization model that produces the minimum

required remaining shelf life for a given SKU subject to the constraint that the net

profit of the sold product from sales exceeds the net profit of the unsold and disposed

product. Using shipment and product cost data from a manufacturer, we found the

optimum value to range between 1 and 26 weeks with mean and median values of 7.2

and 5 weeks, respectively, across 219 SKUs. A modified version of the model finds

the minimum remaining shelf life that ensures a positive net profit from sales. The

optimum value in this case ranged between 2 and 52 weeks with mean and median

values of 17.6 and 17 weeks. Comparing our results with the firm’s existing policy,

which requires a minimum remaining shelf life of 5 or 6 weeks, we show that the

shipment policy based on our methodology is superior by potentially improving the

savings by 41% in situations where products are shipped to retailers’ supply chain

with the designated minimum remaining shelf life.

Moreover, we perform sensitivity analysis in three areas: 1) the demand rate, 2)

sampling of stores, and 3) the assumption regarding the rotation practices on the retail

shelf. In all three areas, we found that while the results did change for some SKUs

with varying magnitudes, they remained the same for others. For the demand rate, we

assessed two scenarios: plus and minus half a standard deviation of the initial demand

rate. For nearly 65% of the SKUs, the results are robust to the varying demand rate.

Further, we evaluated 5 different store samples. We found that for 28% of the SKUs,

the results are robust, while the results for 38% of the SKUs are sensitive to varying

samples. The magnitude of the sensitivity is moderately and negatively correlated

with the total number of outlets that carry the SKU. Accordingly, we can consider

126

adjusting the sample size or the number of samples depending on the size of the pool

from which we select the sample. Regarding the assumption on shelf rotation, we

analyzed two extreme scenarios: full rotation and no rotation. We found that the

results for 51% of the SKUs are robust to varying assumptions about rotation. This

result is unexpected, as we know from literature that rotation matters in inventory

management [48]. After further analysis, we find that rotation does impact the net

profit. However, when the cost of expiration is low, the difference in net profits across

different rotation scenarios is so narrow that the minimum required remaining shelf

life for the sell versus dispose decision remains the same. Considering our findings with

the sensitivity analysis, we expect a certain level of error with our savings calculation.

However, 41% improvement is substantial; thus, we expect the value of optimization

to remain valid.

To alleviate the implementation effort, we use a novel approach in approximating

optimization, which uses machine learning tools as a heuristic. We found that the

random forest algorithm performs well predicting the optimum remaining shelf life for

the two models we study: 1) the model that finds the optimum threshold for sell or

dispose decisions and 2) the model that finds the optimum threshold for profitability.

For both models, we use the same test dataset consisting of 69 data points. For

the first model, we found that the mean squared error (MSE) values are 8.4, 10.1,

19.4, and 53.1 for the random forest, regression tree, support vector machine, and

Poisson regression methods, respectively. For the second model, MSE values are

16.8, 31.3, 25.9, and 272.2 for the same sequence of methods. Based on the first

model, we found that savings with the approximation method are 14%, accounting

for 28% of the benefit we can achieve with optimization. These results suggest that

machine learning performs well as a substitute for optimization. Additionally, the

run time of machine learning algorithms is a few seconds compared to four hours of

the optimization algorithm based on our analysis. This difference shows that machine

127

learning can be a low-cost alternative to optimization. In practice, several projects

compete for the same budget, which aim to improve various aspects of operations.

Managers need to justify why their project should be selected. This justification is

provided through a cost-benefit analysis. The higher the benefit and cost ratio, the

more chances a project has for a funding. Remarkably low cost of machine learning

stands out in this competition providing moderate benefits with minimal cost.

Overall, this thesis contributes to the operations literature that studies perishable

products in two-tier supply chains in two ways. First, we introduce the industry’s

remaining shelf life problem. Second, we develop a methodology to establish shipment

policies to set up minimum required shelf life for products entering the retailer’s

supply chain. Also, we contribute to the overall operations management literature by

applying an innovative approach using machine learning algorithms to approximate

optimum solutions.

This research provides guidelines for manufacturers in determining shipment poli-

cies of perishable products and also helps increase profitability. Our framework can

aid the industry in developing standards in managing the remaining shelf life as

products move in the supply chain. It is important to note that our model solves the

problem from the perspective of the manufacturer. The model can also be used from

the perspective of the retailer. However, since retailers’ gross margin, a key param-

eter in the model, differs from the manufactures’ gross margin, the shipment policy

the model produces can differ. Thus, it still might be a challenge to establish stan-

dards that please both retailers and manufacturers. Nevertheless, our framework can

help eliminate the discrepancy in shipment policies within manufacturers or within

retailers.

With the empirical analysis in Chapter 2, we showed that inventory aging is

positively correlated with product expiration. Our motivation for this research is to

address the inventory aging problem in the supply chain. If shipment policies are cou-

128

pled with appropriate incentive mechanisms within the manufacturer’s organization,

they can address the aging problem to a certain degree. For example, if the supply

chain manager’s performance metric is impacted by the cost of unsold disposed prod-

ucts that are a result of shipment policies, he/she will be incentivized to manage batch

sizes better, improve forecast accuracy, or rotate the supply chain inventory. Man-

agers also need instructions and methods to accomplish these improvements. Future

complementary research can investigate the drivers of supply chain inventory and,

accordingly, address inventory issues. For example, if production batching is found to

be the main driver of inventory, we can investigate optimum batch sizes considering

the relevant trade-offs involved in the supply chain, including the impact on product

expiration. Analysis of warehouse rotation is also an important future research area.

Considering the labor cost of rotation, it may not be cost-effective to perform full-

rotation at all times for all products. Future research can examine the circumstances

at which it is worthwhile to perform warehouse rotation.

129

130

Chapter 4

Schemes Considering Product

Expiration

4.1. Introduction

thesis, this research investigates sales-force compensation schemes that can alleviate

product expiration. During our interviews at our research collaborator, we found

that operations managers attributed product expiration to the overselling behavior

of the sales-force. Then, in our empirical analysis, we find that certain sales incentive

programs are positively correlated with the amount of expiration.

The overselling behavior of the sales-force and its impact on product expiration

is not unique to our research collaborator. The industry recognizes that overselling

is a problem generating product expiration. Furthermore, according to surveys and

qualitative studies, overselling is driven by sales-force incentives. As we can see

from the following quotes, the industry studies imply that product expiration can be

131

mitigated with proper sales-force incentive mechanisms.

"Many of the decisions that have been made in the past by the manufacturer’s

sales staff and the retailer’s procurement staff never considered the unsaleables aspect

of the procuring process. Many manufacturer sales teams have incentives for sales

without penalty for unsaleables." [1]

"There is a general consensus that corporate growth strategies and personal perfor-

mance and compensation packages often drive over-buying and over-selling decisions,

when metrics exclude data about unsaleables." [49]

"Corporate sales goals or quotas (i.e., making a number) are mentioned as the

top reason that sales people over-sell products. Personal compensation and bonus

incentives which are based on gross sales rather than net sales (i.e., minus unsaleables)

are number two." [49]

Despite the understanding that unsaleables need to be incorporated into sales-

force compensation plans, the industry does not prescribe any specific tool or precise

strategy as to how. Seemingly the only industry study dedicated to the issue, titled

"The Impact of Sales and Procurement on Reverse Logistics Management", does

not go beyond emphasizing the importance of accountability and suggesting that

manufacturers’ sales organizations need be educated about the financial implications

of unsaleables and how their role impacts results [64].

Our goal with this research is to find a sales-force compensation mechanism that

can alleviate product expiration while helping the manufacturer achieve full profit

potential. Also, we aim to understand whether currently practiced compensation

plans can fulfill this goal.

The most commonly practiced sales-force compensation scheme in the industry

is the sales-volume-based commission. Accordingly, the sales person is paid a fixed

monetary award for each unit sold (e.g., $0.30/case). Meanwhile, expired volume may

or may not be deducted from the compensation. The sales-volume-based commission

132

scheme does not seem to align incentives since the earnings of the manufacturer

and the sales representative are not based on the same performance measure. As

discussed in Section 3.3.2, expiration erodes the manufacturer’s profit significantly.

In contrast, the sales representative’s commission is not affected similarly even after

expired volume is deducted from the compensation. Let us consider a case in which the

manufacturer’s wholesale price is $4. With a gross margin of 10%, the manufacturer

earns $0.40 for each unit sold. Let us assume that the sales representative earns

$0.10 for each unit sold. If the product expires, we deduct $0.10 from the sales

representative’s compensation, while expiration costs the manufacturer at least $4 if

the retailer gets full credit for expiration. According to this example, there is a one-to-

one relationship between what the sales representative earns and loses if expiration

occurs. On the other hand, this relationship for the manufacturer is one-to-ten.

As we can see in this example, expiration hurts the earnings of the manufacturer

and the sales representative with different magnitudes under the sales-volume-based

compensation scheme.

This asymmetrical effect of expiration on the manufacturer and the sales repre-

sentative leads us to consider a profit-sharing compensation scheme as a potential

mechanism aligning the incentives. The profit-sharing scheme can be aligning since

the cost of expiration impacts the sales representative’s commission the same way it

impacts the manufacturer’s profit under the profit-sharing scheme. Consequently, we

ask the following research questions.

profit and waste outcome for the manufacturer?

manufacturer and the sales representative?

133

problem?

principal-agent framework from microeconomic theory. Our model represents the

direct-store-delivery (DSD) setting in which retailers’ inventory levels are determined

by sales effort. DSD is important for the consumer packaged goods industry since

products distributed through the DSD system represent 24% of unit sales and 52%

of retail profits in the grocery channel [52]. In our model, the effort has a cost for the

sales representative; in addition, compensation depends on the shipment quantity.

We assume a two-stage decision-making process. In the first stage, the manufacturer

selects a compensation scheme; in the second stage, the sales person selects an effort

level. Based on our assumption of newsvendor-based ordering, the inventory that is

in excess of the consumer demand expires after one period.

For this study, we design a new compensation scheme with two parameters: a

penalty fee for a unit of expired item and a commission rate. The sales-volume-based

commission (both the gross-sales-volume and the net-sales-volume) and profit-sharing

schemes are special cases of this new scheme.

In our analysis, we use global optimization as a benchmark, in which inefficiency

due to moral hazard does not exist (i.e., effort can be observed) resulting in the

manufacturer achieving full profit potential. Jointly solving the global optimization

problem and the sales representative’s problem, we find a closed-form relationship

between the penalty fee and the commission rate. Based on this relationship, we

derive the following main results.

∙ Contrary to our prior belief, the profit-sharing compensation scheme does not

align the incentives of the manufacturer and the sales representative.

cost of effort is high or the cost of expiration is low relative to the gross margin

134

of the product.

∙ The compensation scheme designed for this research can align the incentives of

the manufacturer and the sales representative. Based on this scheme, we derive

the following strategy. When the cost of effort is high (e.g., at corporate chain

stores or competitor-dominated markets) or the cost of expiration is low relative

to the gross margin of the product, charge the sales representative a penalty

fee that is lower than the unit cost of expiration. In the reverse case in which

the cost of effort is low (e.g., at individually owned convenience stores) or the

cost of expiration is high relative to the gross margin of the product, charge the

sales representative a higher penalty fee than the unit cost of expiration.

This research clarifies the decision process of the sales representative, specifically

the impact of sales-effort cost on ordering decisions. Earlier, disregarding the role of

effort cost, we mistakenly expected the profit-sharing scheme to align the incentives

since the earnings of the manufacturer and the sales representative are based on

the same performance measure (i.e., profit). The sales representative might indeed

take into account the effort cost. At AlphaCo, the effort for creating the order is

less costly at small stores since the sales representative’s job at small stores (e.g.,

convenience stores, gas stations) is limited to generating the replenishment order

while the driver completes the remaining work (stocking the shelf). At large stores,

the sales representative is also responsible for stocking the shelf. Then, based on

our model, we expect over-ordering at small stores where effort cost is low. This

expectation is supported by our empirical analysis presented in Chapter 2, since we

find that expiration is more likely to occur at small stores even after controlling for

other factors such as case size, which can partially be explained by over-ordering.

Therefore, effort cost is a factor that needs to be considered in designing sales-force

compensation schemes.

This research also shows us that a compensation scheme with homogenous parame-

135

ters (commission rate, penalty fee for expiration) across products, channels, or regions

may not be effective. Manufacturers can benefit from customizing the compensation

parameters based on the cost of expiration and varying effort cost at different chan-

nels (convenience stores versus supermarkets) and markets (where the competitor is

dominant versus weak).

To the best of our knowledge, our research is first in academia analyzing sales-

force compensation mechanisms for settings in which sales effort determines inventory

levels and thus directly impacts product expiration. Further, our study is valuable for

practitioners since the industry lacks the knowledge about how to best incorporate

expiration into sales-force compensation schemes. Although our model represents

the DSD system in which the retailer’s inventory levels are solely determined by

manufacturers’ sales-force, our findings can also apply to the traditional channel.

Our results are relevant for the traditional channel because manufacturers’ sales teams

influence retailers’ purchasing decisions and accordingly inventory levels.

Our research is related to the following literature: agency theory in economics, sales-

force compensation plans in marketing, marketing and operations interface, supply

chain coordination, and vendor-managed inventory.

Agency theory in Economics The analytical foundation of this work is pro-

vided by the principal-agent model. In the setting we analyze, the principal is the

manufacturing firm that operates through the direct-store-delivery channel; the agent

is the sales representative. Moral hazard occurs because the sales representative may

place too much inventory in the retail store knowing that he/she will not be fully

accountable for the cost of product expiration. This is due to the information asym-

metry such that the actions of the sales representative at retail stores are not observ-

136

able by the manufacturer. There is a vast amount of literature in economics on the

principal-agent problem. The seminal papers in agency theory include [28], [24], [56],

and [21].

Sales-force compensation plans in Marketing Basu et al. [4] are the first to

apply the agency theory to the sales-force problem. They formulate the problem and

derive results that are in alignment with Holmstrom’s [28] analysis. Assuming dimin-

ishing agent utility with pay and gamma or binomial distributions for the demand,

they derive a compensation plan as a function of sales. This represents the salary plus

commission scheme. Special cases of this scheme are: 1)straight commission, 2)slid-

ing commission (varying commission rates based on different sales brackets), 3)salary

only, 4) salary plus commission, and 5)salary plus commission beyond a target. Using

Pratt’s [53] risk tolerance definition, they show how increase, decrease, or constancy

of commission rates depend on the risk aversion characteristics of the salesperson.

The same as in Basu et al.’s [4] study, sales is a function of effort in this research.

Also, in both studies, the agent’s utility function for income and effort is additively

separable. However, our study is distinct in a number of ways. In their setting,

sales represents demand and its probability distribution is conditioned on sales effort.

In our setting, sales represents deterministic shipment quantity (i.e., store inventory

level), which is a function of sales effort; further, demand is stochastic and inde-

pendent. Moreover, we focus on product expiration and strategies to appropriately

penalize the sales representative for product expiration. This analysis is done under

a straight commission scheme.

Marketing and Operations interface A number of studies in the Operations

literature analyze the sales-force compensation problem considering inventory or pro-

duction issues. The following are notable examples.

Chen [8] proposes a sales-force compensation package to induce sales that smoothes

production considering multiple periods. This scheme evaluates and compensates the

137

sales force on a moving-time-window basis, where the length of the time window is

determined by the production lead time.

Dai and Jerath [12] develop sales-force compensation and inventory strategies

under the coordinating contract. The analysis is based on a newsvendor type model.

In the setup, sales effort impacts uncertain demand, but sales is limited by inventory.

They find that under the optimal contract, the agent is paid a bonus for meeting a

sales quota. Further, disregarding excess inventory scenarios, they focus on stockout

possibilities and accordingly find several counterintuitive results with regard to the

amount of compensation bonus and optimum inventory levels. These results are not

expected if we were to ignore the limited inventory aspect of operations.

Similar to these studies, we also analyze the sales-force compensation problem

considering inventory issues. In addition, the focus of our analysis is product expira-

tion and the setting is based on direct-store-delivery operations. Accordingly, effort

determines the deterministic order quantity and demand is stochastic. This is in

contrast with Chen’s [8] and Dai et al.’s [12] analysis in which sales effort impacts

uncertain customer demand.

Supply chain coordination The goal of channel coordination is to improve the

performance of the supply chain by aligning the objectives of channel partners (e.g., a

manufacturer and a retailer). The literature usually focuses on ordering decisions and

inventory management. Likewise, the objective of this research is to align the incen-

tives of the manufacturing firm and manufacturer’s sales representative. The setting

involves an ordering decision to manage a retailer’s inventory based on a newsvendor

type model, which is the underlying model in most supply chain coordination anal-

yses. The following are examples of the most relevant studies from the supply chain

coordination literature.

Taylor [59] evaluates channel rebates considering retailers’ sales effort. In a linear

rebate contract, the manufacturer pays the retailer for each unit sold to the consumer.

138

In a target rebate contract, the payment occurs only if the retail sales exceed a

certain threshold. When demand is not influenced by sales effort, only a target

rebate can achieve coordination that is implementable. However, in the more realistic

scenario where demand is an outcome of sales effort, a target rebate contract needs

to be coupled with a return contract for coordination. Further, the authors show

that provision of returns strengthens incentives for sales effort, contrary to the view

expressed in the literature.

Bandyopadhyay and Paul [3] explain full return policies practiced in industry,

citing examples from the direct-store-delivery channel. They model the environment

where two capacity-constrained manufacturers compete for the shelf space at the same

retailer and show that a complete-credit returns policy is the only equilibrium of the

game. This finding is in contrast to Pasternack’s [51] analysis, a seminal study in

the field, suggesting that only less than full-credit returns can coordinate the supply

chain.

Vendor managed inventory The context of this research is direct-store-delivery

(DSD) operations. Essentially, DSD is a type of vendor managed inventory (VMI)

business model. In industry, the term VMI is typically used for the traditional channel

where the manufacturer manages the inventory of retailers’ distribution centers. With

VMI, inventory levels at retailers’ distribution centers are usually communicated to

the manufacturer via an electronic data interchange; accordingly, the manufacturer

generates replenishment orders for the distribution center.

The existing literature focuses on different aspects of VMI. Cetinkaya and Lee

[7] present an analytical model for jointly coordinating inventory and transportation

decisions in VMI systems. Mishra and Raghunathan [46] focus on the brand com-

petition aspect and suggests that VMI is beneficial for the retailer since it promotes

competition and, as a result, better in-stock performance. Bernstein et al.[5] focus

on the role of VMI partnerships on supply chain coordination under simple wholesale

139

pricing schemes. Research that specifically focuses on DSD operations is scarce. One

example is the study of Chen et al. [10] analyzing the switching of Gatorade from

the traditional distribution channel to the DSD channel belonging to Pepsi. The

study determines the circumstances in which it is cost effective for Pepsi to switch

the distribution of Gatorade from the traditional channel to the DSD channel.

In summary, the sales-force compensation problem has been studied in the lit-

erature, but not from the perspective of product waste and in the context of DSD

operations. Sales operations in the DSD environment are unique, because the man-

ufacturer’s sales representative manages the store inventory, and thus sales effort

impacts inventory levels. This study builds on marketing and operations literature

analyzing the salesforce compensation problem for perishable products in DSD oper-

ations.

4.3. Model

one retail store. The manufacturer employs a sales representative who creates replen-

ishment orders for the store. Typically, the store manager, employed by the retailer,

is not involved with the sales representative’s basic ordering decisions; however, it is

common to get the store manager’s consent for additional inventory placement such as

in-store displays or expansion of shelf space. If the store is large (i.e., supermarket),

the delivery is made in bulk at the store’s backroom. In this case, the sales rep-

resentative is responsible for unwrapping the pallet(s) in the backroom and moving

the inventory from the backroom to the shelf. If the store is small (i.e., convenience

store), the sales representative’s job requires less effort and is limited to visiting the

store to generate the order (the driver moves the inventory from the truck directly to

the shelf at the time of delivery).

140

We assume a newsvendor type single-period inventory model. Although the prod-

ucts we study in this thesis have multiple periods of shelf life, it is appropriate to

assume a single-period model in our analysis because the problem of interest in this

research is to analyze compensation schemes with respect to over-ordering behavior.

For this purpose, the single-period model simplifies our analysis while capturing the

effect of excess inventory on expiration.

The events occur according to the sequence shown in Figure 4-1.

Stage 1 Stage 2

compensation scheme. effort level determining the order Excess inventory expires and is

quantity. returned to the manufacturer.

At the first stage, the manufacturer selects a compensation scheme, which we de-

note by 𝑠. At the second stage, with the knowledge of the compensation scheme, the

sales representative selects an effort level, denoted by 𝑒. The effort represents the

activities to place inventory at the store (e.g., create order, replenish the shelf from

backroom inventory, build displays, acquire additional shelf space). The sales repre-

sentative’s effort determines the order quantity, 𝑄(𝑒), which is concave and increasing

in 𝑒, as shown in Figure 4-2. We assume a concave relationship since each incremental

activity to increase inventory at the store requires more effort. For example, the sales

representative first replenishes the depleted inventory on the shelf. Additional inven-

tory placement, through building displays or acquiring more shelf space, increasingly

gets harder since typically the store manager is involved in these decisions. Also, the

shelf space is limited at the store, and thus we expect the order quantity eventually

to reach a limiting value.

141

Q(e) C(e)

e e

Figure 4-2: Assumed functions of order quantity, 𝑄(𝑒), and effort cost, 𝐶(𝑒).

thus, the shipment quantity is equal to the order quantity. After the delivery is made,

consumer demand, 𝐷, is observed. At the end of one period, excess inventory expires,

[𝑄(𝑒) − 𝐷]+ , and is returned to the manufacturer. Effort has a cost for the sales rep-

resentative, which is denoted by 𝐶(𝑒). The cost of effort represents the value of time

spent for the effort. 𝐶(𝑒) is linear and increasing in 𝑒, as shown in Figure 4-2.

Let 𝑝 denote the unit gross profit from sales (= wholesale price - cost of production

and distribution) and 𝑐 denote the unit cost of expiration (= wholesale price + cost

of reverse logistics) for the manufacturer. Thus, 𝑐 > 𝑝. The profit function of the

sales representative and the manufacturer are denoted by Π𝑆𝑅 and Π𝑀 , respectively.

For a given compensation scheme, the sales representative’s second stage problem is

as follows.

max Π𝑆𝑅 (𝑒) = [𝑐𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛(𝑠) − 𝐶(𝑒)] (4.1)

𝑒

𝑠

(4.2)

*

subject to 𝑒 (𝑠) ∈ arg max𝑒 Π𝑆𝑅 (𝑒) = [𝑐𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛(𝑠) − 𝐶(𝑒)]

the net-sales-based commission, 3) the profit-sharing compensation, and 4) an al-

ternative scheme designed by us. These schemes are denoted by 𝑔𝑠, 𝑛𝑠, 𝑝, and 𝑎,

respectively. In addition, the commission percentages for these schemes are repre-

142

sented by 𝑟𝑔𝑠 , 𝑟𝑛𝑠 , 𝑟𝑝 , and 𝑟𝑎 . For the manufacturer and the sales representative to

participate, we assume that 0 < 𝑟𝑠 < 1 for all 𝑠. The alternative scheme involves two

parameters: the commission rate, 𝑟𝑎 , and the penalty fee that the sales representative

is charged for a unit of expiration, denoted by 𝑥. The following equations show the

compensation for the four schemes we analyze.

𝑐𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛(𝑔𝑠) = [𝑄(𝑒)𝑝]𝑟𝑔𝑠

(4.3)

+

𝑐𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛(𝑝) = [𝑄(𝑒)𝑝 − 𝐸𝐷 [𝑄(𝑒) − 𝐷] 𝑐]𝑟𝑝

sharing compensation schemes are special cases of the alternative scheme. This is

because 𝑥 takes the values of 0, 𝑝, and 𝑐 for the schemes 𝑔𝑠, 𝑛𝑠, and 𝑝, respectively.

Earlier we describe 𝑔𝑠 and 𝑛𝑠 as schemes rewarding a fixed monetary reward for a

unit of sales (= 𝑝𝑟𝑛𝑠 ). According to our model in (4.3), 𝑔𝑠 and 𝑛𝑠 can also be thought

of as gross-profit sharing schemes. Lastly, it is important to note that with scheme

𝑎, only a portion (= 𝑟𝑎 ) of 𝑥 penalizes the sales representative for expiration.

4.4. Analysis

the sales representative’s effort and pays the sales representative a fixed monetary

award to compensate for his/her effort. We refer to the manufacturer’s problem in

this scenario as the global optimization problem, denoted by 𝐺, which is shown by

the following formulation.

𝑒

143

In this problem, the manufacturer selects an effort level that maximizes its profits.

Alternatively, we can think of the global optimization scenario as the manufacturer

performing inventory placement at the retailer, without hiring a sales representative

for the job (i.e. the manufacturer and the sales representative are the same entity).

Hiring can produce inefficient outcomes since the manufacturer cannot observe the

sales representative’s behavior and, as a result, the sales representative can over-order

knowing that he/she is not fully accountable for the cost of expiration, deviating from

the manufacturer’s interest. This situation is referred to as moral hazard.

A compensation scheme that produces the same effort level as the global optimiza-

tion aligns the incentives of the manufacturer and the sales representative, eliminating

moral hazard, and helps the manufacturer realize the full profit potential. Therefore,

in our analysis, we look at the first-order conditions of the global optimization prob-

lem and the sales representative’s profit maximization problem.

Theorem Let 𝑟𝑎* and 𝑥* denote the pair of the commission rate and penalty fee

that achieves the optimum effort level of the global optimization problem and 𝐹 (.)

denote the cumulative distribution function of demand. The relationship between 𝑟𝑎*

and 𝑥* is determined by

𝑝 − 𝐹 (𝑄(𝑒* ))𝑐

𝑟𝑎* =

𝑝 − 𝐹 (𝑄(𝑒* ))𝑥* (4.5)

* +

𝑒 ∈ arg max𝑒 𝑄(𝑒)𝑝 − 𝐸𝐷 [𝑄(𝑒) − 𝐷] 𝑐 − 𝐶(𝑒)

Proof: To find the optimum effort level in (4.4), we differentiate Π𝐺 (𝑒) with re-

spect to 𝑒 and set this amount equal to 0. This gives

𝑝− 𝑐− =0 (4.6)

𝑑𝑒 𝑑𝑒 𝑑𝑒

144

Based on the Leibniz rule

∫︁ )︁

= [𝑄(𝑒) − 𝐷)]𝑓 (𝐷)𝑑𝐷

𝑑𝑒 𝑑𝑒 0

∫︁ 𝑄(𝑒)

𝑑𝑄(𝑒) 𝑑𝑄(𝑒) (︀ )︀

= 𝑓 (𝐷)𝑑𝐷 + (𝑄(𝑒) − 𝑄(𝑒) 𝑓 (𝑄(𝑒)) − 0

0 𝑑𝑒 𝑑𝑒

𝑑𝑄(𝑒)

= 𝐹 (𝑄(𝑒)) (4.7)

𝑑𝑒

𝑑𝑄(𝑒) 𝑑𝐶(𝑒)

[𝑝 − 𝐹 (𝑄(𝑒))𝑐] − =0 (4.8)

𝑑𝑒 𝑑𝑒

Next, we consider the two-stage problem of the manufacturer and the sales repre-

sentative under the alternative compensation scheme. To obtain the optimum effort

𝑑Π𝑆𝑅 (𝑒)

level for the sales representative in the two-stage problem, we solve for =0

𝑑𝑒

and get

𝑑𝑄(𝑒) 𝑑𝐶(𝑒)

[𝑝𝑟𝑎 − 𝐹 (𝑄(𝑒))𝑥𝑟𝑎 ] − =0 (4.9)

𝑑𝑒 𝑑𝑒

𝑝 − 𝐹 (𝑄(𝑒* ))𝑐

𝑟𝑎* =

𝑝 − 𝐹 (𝑄(𝑒* ))𝑥*

Different values of 𝑟𝑎* and 𝑥* result in the manufacturer achieving the full profit

potential. Then, the alternative compensation scheme is said to be a coordinating

mechanism; also 𝑟𝑎* and 𝑥* are called the coordinating commission rate and penalty

fee.

The following graph shows a numerical application of the coordinating scheme.

145

!$60.00!!

Global optimization

Sales representative

!$40.00!!

Manufacturer

!$20.00!!

Profit

!$*!!!!

0.05! 0.15! 0.25! 0.35! 0.45! 0.55! 0.65! 0.75! 0.85! 0.95!

Effort level

!$(20.00)!

!$(40.00)!

!$(60.00)!

𝑐 = 4, 𝑥 = 2, and 𝑟𝑎 = 0.45.

The x-axis in Figure 4-3 represents the effort level while the y-axis represents

the profit. The blue line on the top represents the profit for global optimization,

formulated in (4.4). The green and red lines show the profits for the manufacturer and

the sales representative, represented with models (4.2) and (4.1), with the alternative

scheme (𝑠 = 𝑎). In the graph, the maximum profit for all three lines intersect at the

same effort level, which indicates that coordination is achieved and the incentives of

the manufacturer and the sales representative are aligned. A coordinating mechanism

eliminates the inefficiency occurring due to moral hazard1 . Therefore, here, the sum

of the profits for the manufacturer and the sales representative is equal to the profit

achieved with global optimization (Π𝐺 = Π𝑀 + Π𝑆𝑅 ).

1

Moral hazard can occur when sales representatives place too much inventory at retail stores

knowing that they will not be fully accountable for the cost of product expiration.

146

The implications of this theorem are as follows.

Proposition 1 When the cost of expiration is low relative to the gross margin or the

cost of effort is high (𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0), in order to realize the full profit potential,

the manufacturer should charge the sales representative a lower penalty fee than the

cost of expiration (𝑥* < 𝑐).

Proof: Based on (4.5), 𝑥* < 𝑐 must be the case for 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0 since

0 < 𝑟𝑎* < 1, which completes the proof.

We observe 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0 in two situations: 1) the expiration cost is low

relative to the gross margin, or 2) the cost of effort is high, which leads to lower

optimum effort and thus lower 𝑄(𝑒* ) and 𝐹 (𝑄(𝑒* )).

The cost of effort varies across different stores, channels, or regions. It is higher at

supermarkets than at gas stations since the sales representative’s job involves stocking

the shelves and building displays at large stores (e.g., supermarkets, supercenters).

Also, it is higher at chain stores since corporate standards may restrict a sales rep-

resentative’s inventory decisions. For example, convincing the store owner to build a

permanent display at an individually owned convenience store is usually easier since,

knowing that the store takes no risk associated with excess inventory (due to full-credit

returns), store owners typically welcome display requests. In contrast, at a corporate

chain store, the store manager can resist a display request due to centrally-planned

planograms. The cost of effort is also higher at competitor-dominated markets since

retailers usually favor popular brands in shelf space allocation or when granting end-

caps and displays.

Moreover, when 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0, there is a codirectional relationship between

𝑟𝑎* and 𝑥* because the derivative of 𝑟𝑎* with respect to 𝑥 is positive.

= >0 (4.10)

𝑑𝑥 [𝑝 − 𝐹 (𝑄(𝑒* ))𝑥]2

147

The following figure shows a numerical example of the relationship between the

coordinating commission rate and penalty fee when 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0.

1"

0.8"

0.6"

r_a*%

0.4"

0.2"

0"

0" 0.5" 1" 1.5" 2"

x*%

Figure 4-4: Relationship between coordinating commission rate and penalty fee for

𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0. In the example, 𝑝 = 1 and 𝑐 = 2.

penalty fee by charging a higher fee, the commission rate needs to be increased as

well to prevent the sales representative from reducing his/her effort level.

Proposition 2 When the cost of expiration is high relative to the gross margin or the

cost of effort is low (𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 < 0), in order to realize the full profit potential,

the manufacturer should charge the sales representative a higher penalty fee than the

cost of expiration (𝑥* > 𝑐).

Proof: 𝑥* > 𝑐 needs to be true for 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 < 0, since 0 < 𝑟𝑎* < 1.

Proposition 2 states the exact opposite of Proposition 1. We see 𝑝−𝐹 (𝑄(𝑒* ))𝑐 < 0

occurring when 1) the expiration cost is high relative to the gross margin, or 2) the

cost of effort is low, which leads to higher optimum effort and thus higher 𝑄(𝑒) and

𝐹 (𝑄(𝑒)).

Here, the derivative of 𝑟𝑎* with respect to 𝑥 is negative thus 𝑟𝑎* decreases with 𝑥* .

= <0 (4.11)

𝑑𝑥 [𝑝 − 𝐹 (𝑄(𝑒* ))𝑥]2

148

Figure 4-5 shows a numerical example of the relationship between 𝑟𝑎* and 𝑥* for

𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 < 0.

1$

0.8$

r_a*%

0.6$

0.4$

10$ 10.5$ 11$ 11.5$ 12$

x*%

Figure 4-5: Relationship between optimum commission rate and penalty fee when

𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 < 0. In the example, 𝑝 = 1 and 𝑐 = 10.

this scenario, not only are we penalizing the sales representative for expiration more

than what expiration costs for the firm (𝑥* > 𝑐), but also we further increase the

penalty fee if we decrease the coordinating commission rate. Intuitively, the sales

representative compensates for the high penalty fee by selling more through in-

creasing the effort. Then, the manufacturer can afford to offer a lower commission

rate. The sales representative does not respond the same way in the reverse scenario

(𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0) in which expiration is inexpensive for the sales representative.

Proposition 3 Effective penalty charge is less than the unit cost of expiration, 𝑐 >

𝑥* 𝑟𝑎* . Further, the weighted average of the penalty fee and gross profit is less than or

equal to the cost of expiration, 𝑝(1 − 𝑟𝑎* ) + 𝑥* 𝑟𝑎* ≤ 𝑐.

𝑝 − 𝑝𝑟𝑎*

Proof: Re-arranging the terms in (4.5), we derive 𝐹 (𝑄(𝑒*)) = . Since 0 ≤

𝑐 − 𝑥* 𝑟𝑎*

𝑝 − 𝑝𝑟𝑎*

𝐹 (𝑄(𝑒*)) ≤ 1, then 0 ≤ * *

≤ 1. Accordingly, 𝑐 > 𝑥* 𝑟𝑎* and 𝑝(1−𝑟𝑎* )+𝑥* 𝑟𝑎* ≤ 𝑐,

𝑐 − 𝑥 𝑟𝑎

which completes the proof.

𝑐 − 𝑥* 𝑟𝑎* > 0 also suggests that the sales representative is fully charged for the cost

of expiration for a portion of the expired items. Imagine 100 units expire, costing

149

the manufacturer 100𝑐. Sales representative is charged for the cost of expiration for

less than 100 units, suppose 80 units. Then, the manufacturer deducts 80𝑐 from

compensation, while sharing the gross profit with the sales representative (= 𝑄𝑝𝑟𝑎 ).

Proof: Based on (4.3), profit sharing scheme is a special case of the alternative

compensation scheme in which 𝑥 = 𝑐. Then, according to (4.5), the commission rate

needs to be 1, an infeasible value for the commission rate.

Intuitively, the sales representative exerts suboptimal effort under the profit shar-

ing compensation scheme since he/she is paid only a portion of the profits while

incurring an effort cost.

Proposition 5 The gross sales based commission and net sales based commission

schemes can be coordinating only when 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 > 0.

Proof: According to (4.3), both the gross-sales-based commission and the net-

sales-based commission schemes are special cases of the alternative compensation

scheme where 𝑥 = 0 and 𝑥 = 𝑝, respectively.

* 𝑝 − 𝐹 (𝑄(𝑒*))𝑐

Based on (4.5), coordinating commission rates are determined by 𝑟𝑔𝑠 =

𝑝

* 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐

and 𝑟𝑛𝑠 = . Then, under coordinating schemes, the firm needs to pay

𝑝 − 𝐹 (𝑄(𝑒* ))𝑝

a higher commission rate to the sales representative with the net-sales-based commis-

* *

sion (𝑟𝑔𝑠 < 𝑟𝑛𝑠 ).

Sales-based commission schemes can only be coordinating when 𝑝 − 𝐹 (𝑄(𝑒* ))𝑤 >

0, since:

1) If 𝑝 = 𝐹 (𝑄(𝑒* ))𝑐, then 𝑟𝑔𝑟 = 0 and 𝑟𝑛𝑟 = 0 (not feasible).

2) If 𝑝 − 𝐹 (𝑄(𝑒* ))𝑐 < 0, then 𝑟𝑔𝑠 < 0 and 𝑟𝑛𝑠 < 0 (not feasible).

In summary, there exists a compensation scheme that aligns the incentives of the

manufacturer and the sales person preventing overselling while achieving full profit

150

potential for the manufacturer. Under the coordinating scheme, the sales person is

charged with a penalty fee which varies depending on product cost and market char-

acteristics. Moreover, sales-volume based compensation scheme and profit-sharing

scheme are special cases of the coordinating scheme. Profit-sharing scheme is never

coordinating; sales-volume based compensation scheme can be coordinating under

certain circumstances.

The industry acknowledges the existence of over-selling and its impact on product

expiration yet no specific strategy has been prescribed to alleviate the problem. In

this research, we study sales-force compensation schemes from the perspective of

product waste. We build a theoretical model representing the decision process of the

manufacturer and the salesperson based on the principal-agent framework. Through

the model, we evaluate several compensation schemes. These schemes are the sales-

volume-based commission scheme, profit-sharing scheme and an alternative scheme

we design for this study. The alternative scheme consists of two parameters: the

commission rate and the penalty fee that is imposed on the sales representative for

each expired item. We find that the alternative scheme can align the incentives of

the manufacturer and the sales representative when the commission rate depends on

the penalty fee through a closed-form relationship. We find that the profit-sharing

scheme is never coordinating and sales-volume-based compensation can be coordi-

nating in some situations, which are unexpected findings. Earlier we expected the

reverse coordination outcome since the sales representative’s and the manufacturer’s

earnings are based on the same performance measure under the profit-sharing scheme,

unlike sales-volume-based compensation. Finally, we derive the following strategy for

practitioners. When the cost of expiration is low (high) relative to the gross margin

151

of the product or when the cost of effort is high (low) as in competitor dominated

markets or corporate chain stores, the penalty fee should be lower (higher) than the

product’s expiration cost.

Our model assumes that the demand and order quantity are independent, which

is a limitation in our analysis since the general belief in the industry is that inventory

drives sales. Academic research supports this belief [18]. In the more realistic situation

where demand is an increasing function of inventory, we expect the optimum effort

level to be higher in the manufacturer’s problem than in the scenario with the current

model assumptions. Optimum effort level would be higher since expiration would be

lower if demand increases with order quantity. Nevertheless, our finding regarding

the two strategies for the penalty fee (that the penalty fee should be high when the

cost of expiration is high or effort cost is low; it should be low in the reverse scenario)

might still hold true. In a future study, we intend to extend our analysis considering

the dependency of demand on order quantity.

In addition, two future empirical studies can complement our research. The first

study can examine how sales representatives make inventory decisions in practice.

Our analysis (as well as many academic studies in similar settings) is based on the

assumption that the sales representative chooses an optimum effort level considering

the trade-off between earnings and effort cost. We do not know whether this as-

sumption is true. According to a manager at our collaborator, some salespeople even

consider the product’s weight (thus the required physical effort to process inventory)

relative to the commission per case of shipment. Accordingly, these salespeople will

not promote heavy products which do not generate sufficient income while taking

higher risk with products with high commission amount. Our preliminary analysis

supports this belief. During our interviews at AlphaCo, we found out that the state of

California is an exception in the country practicing a commission-only compensation

plan (required by labor unions) while the rest of the U.S. practices a base salary plus

152

commission scheme. Accordingly, a sales representative in California earns more per

unit of shipment than in other states. Using a binary indicator for California and con-

trolling for shelf life, case size, and store types (e.g., supermarket, convenience store,

drug store, etc.), we performed a regression analysis predicting the amount of expira-

tion, similar to the analysis presented in Chapter 2. We find that the probability of

expiration is higher in California, which could be explained by over-ordering due to

the higher commission rate. This analysis is high level (based on a binary indicator

for one geography) and thus is limited. A comprehensive study, using historical order,

commission, and product characteristic data (e.g., weight per case, packaging type),

is needed to understand whether salespeople favor items that require little effort and

generate high earnings. Such a study will be valuable for practitioners as well as

academics doing theoretical research.

The second study can test the theory this thesis generates. Our theory suggests

that we can reduce expiration, while maintaining profit, through a compensation

scheme that penalizes the sales representative for expiration with a penalty fee varying

depending on product cost and market characteristics. An experimental study can

examine whether the compensation scheme we offer with this research is effective at

reducing expiration.

153

154

Chapter 5

Concluding Remarks

This thesis studies the unsaleables (i.e., damaged, expired, and discontinued products)

problem of consumer packaged goods (CPG) in retail supply chains. Unsaleables is

an important problem for the CPG industry due to their substantial financial and

environmental implications. At our collaborator, a food and beverage manufacturer,

the cost of product waste is equivalent to 50% of the firm’s annual profits. Financial

significance of waste is not unique at our collaborator. The annual cost of unsaleables

is $15 billion for the CPG industry, which constitutes 1-2% of gross retail sales [30].

Industry leaders tells us that annual waste cost can be even higher than the net profit.

The reimbursement provided by manufacturers alleviates this impact; as a result, the

cost of waste affects both manufactures’ and retailers’ profits alike.

Financial impact of product waste affects firms’ bottom-lines, and thus their share-

holders’ income. However, the environmental impact of waste affects all of us as a

society. Indeed, 17% of the industry’s waste is disposed at landfills creating a sig-

nificant burden for the environment [30]. Let us consider our collaborator. The

warehouse manager we interviewed revealed that one full truckload of waste is picked

up at his facility every week to be disposed at landfills. This firm has approximately

350 warehouses in North America. If other warehouses were to follow the same dis-

155

posal method (landfills as opposed to, for example, donation at food banks), roughly

350x52=18,200 trailers of waste may be dumped at landfills every year. This estimate

is for only one company. The overall volume of waste disposed at landfills by CPG

companies must be massive.

Given how much unsaleables impact firms’ profits, it is surprising that industry

has not progressed much on this problem. For example, firms still struggle with col-

lecting unsaleables data since product handling in reverse logistics is very expensive.

There has been more progress on damage and discontinuation types of unsaleables

since audits and improvements in business processes have been effective at identifying

the root causes with these types and accordingly addressing the problems. However,

product expiration category remains relatively untouched since expiration is complex

spanning multiple functions (e.g., manufacturing, warehousing, sales and procure-

ment, store operations). Our collaborator conducted an audit study to identify the

root causes of product waste. Product damage was successfully sorted (e.g., nails

on the pallet, insufficient glue, etc.) but audits revealed only one insight regarding

product expiration which is whether the product on the shelf was rotated. Contrary

to the general belief, they found that in majority of cases where an expired product

is found, the shelf was actually rotated. Regardless, because the other root causes

were not understood, the action taken as a result of this study focused on rotation

(packages are color coded in production; the color changes every month to make it

easier to spot the older products). Overall, the expiration category of unsaleables has

been growing over the years while the damage and discontinuation categories decline.

Consequently, we focus on product expiration in this thesis.

We start our research with a descriptive study on the root causes of product

expiration; this study is presented in Chapter 2. We compile a list of hypothesized root

causes of product expiration based on interviews and site visits at our collaborator,

industry reports, unsaleables conference, and our supply chain knowledge. These

156

root causes are product case sizes, aging of inventory in the manufacturer’s supply

chain, negligence of inventory rotation on the retail shelf, rules on minimum order

sizes, sales incentive programs, and forecasting complexity. Using our collaborator’s

archival data (e.g., shipments, returns, inventory counts, shelf life, case size, etc.)

and an econometric model, we estimate the relative impact of different root causes on

product expiration. Our main finding is that the best opportunity to reduce expired

product volume lies in case sizes reduction. This is a surprising finding given that

no interviewee at our collaborator mentioned the product case size as one of the root

causes of product expiration. At the same time, it is not surprising if we consider

that (according to an industry study) most items at grocery stores are slow moving

products (more than half of products sell less than one unit per week) [63]. At our

collaborator, approximately, 80% of the orders are single case orders. Then, we expect

that reduction in case size (e.g., from 24 units to 12 units) reduces the expired volume.

According to our analysis, the next best areas of opportunity to reduce expiration

are sales incentives and inventory aging in the supply chain. Unfortunately, there is

no straightforward remedy for these challenges as there is with case sizes. Therefore,

we devote the remainder of the thesis for the analysis of these two areas. Chapter 3

is concerned with inventory aging in the manufacturer’s supply chain. Chapter 4 is

on the subject of sales incentives.

In Chapter 3, we evaluate the manufacturer’s sell-or-dispose decision for aged in-

ventory. Some manufacturers practice shipment rules for old inventory such that the

product that has a remaining shelf life below a certain level is not released for sale

to the retailer and is disposed directly at the manufacturer’s distribution center. The

problem with existing rules is that they are arbitrarily set up and do not consider

relevant factors such as the demand rate and product costs. We develop an analytical

model to set up optimum inventory aging targets based on the demand rate, cost of

expiration, shelf life, and gross margin of products. Using our collaborator’s data,

157

we find that the optimum target largely varies across products in accordance with

cost parameters and demand. This variation is in contrast with existing rules which

are uniform across products. Further, we find that optimization generates substan-

tial savings since sell-or-dispose decisions are based on cost implications under the

optimization model. Ultimately, if the manager is held accountable for the cost of

unsold and disposed products, shipment rules can help reduce product expiration by

providing an incentive for supply chain managers to improve inventory freshness (e.g.,

through better management of batch sizes, more accurate forecasts, etc.).

In Chapter 4, we examine sales-force compensation schemes from the perspec-

tive of product expiration caused by over-ordering. Based on the principal-agent

framework, we develop a mathematical model of the decision making process of the

manufacturer and the sales person. Our model is directly applicable to the direct-

store-delivery sales and distribution model, in which manufacturers’ sales-force de-

termine store inventory levels. We design a sales-force compensation plan that aligns

the incentives of the manufacturer and the sales person. This plan holds sales people

accountable for the cost of expiration through a penalty fee to encourage them to

order just enough, achieving full profit potential for the manufacturer. This penalty

fee depends on the product cost and market characteristics. Our analysis suggests,

contrary to prior belief, that a profit-sharing payment scheme does not align the in-

centives of the manufacturer and the sales representative. We also find that, again in

contrast with prior belief, sales-volume based compensation plan (a commonly prac-

ticed scheme in industry) can align the incentives of the manufacturer and the sales

person.

The tools that inform these studies (econometrics, optimization, machine learning)

make this a methodologically diverse thesis that contributes to supply chain theory

and practice. Further, with this thesis, we introduce new problems to the operations

literature. To the best of our knowledge, the unsaleables problem was not stud-

158

ied before by the operations management research community; our descriptive study

on the relative impact of expiration drivers is first in literature studying perishable

inventory from a holistic perspective (analyze the impact of different organizational

functions such as sales, product design, production, warehousing, and store operations

on product expiration). Our descriptive study also exposes a number of prescriptive

research areas from case size optimization to production/transportation batching op-

timization. Similarly, the second research in this thesis is first in literature studying

shipment rules in a two-tier supply chain for perishable products. Lastly, our third

study on sales-force compensation plans is novel, applying the principal-agent frame-

work to a direct-store-delivery setting in which manufacturers’ sales people determine

store inventory levels via sales effort.

The three studies in this thesis are motivated by industry challenges. Our findings

can be directly applied as strategies to reduce the amount of product expiration.

Based on our overall understanding of unsaleables and our findings in this thesis,

we recommend the following processes and actions to practitioners:

1. Sort unsaleables into four categories as damaged, expired, retailer discon-

tinued and manufacturer discontinued products.

Sorting is expensive, but if done efficiently, is very beneficial. To make the process

efficient, for example, unsaleables could be collected in four different containers (each

labeled with the type) in the back room. Scanning and sorting processes and systems

can be co-designed to collect data efficiently. Since each unsaleables type occurs due

to different dynamics, sorting is a first step in identifying the root cause.

Existing conflict in the industry regarding unsaleables policies seems to occur since

the root cause by an instance of unsaleables is unknown. If it was known, the respon-

sible party could compensate for the cost. Therefore, sorting (as the first step to iden-

159

tify root causes) can form the basis for unsaleables policies. Unsaleables policies

can be established separately for four types of unsaleables. It is reasonable

to expect that the retailer absorbs the cost for retailer-discontinued products, and

likewise, the manufacturer absorbs the cost for manufacturer-discontinued products.

Regarding the damage type, if damage is observed for the same product consistently,

it most likely occurs due to a packaging failure as opposed to mishandling. Thus,

damage can be further categorized as "packaging issue" versus "mishandling" during

the scanning process; these subcategories can be compensated by manufacturers and

retailers, respectively. Subcategorization of damage could eliminate or lessen the need

for expensive audits.

2. GMA and FMI allocate funds for a system that helps facilitate unsalebales

data sharing. We recommend that retailers continue to inform manufacturers on

the root causes of unsaleables regardless of the reimbursement policy.

3. Focus on case size management. Our research finds that large case sizes are

the largest contributor to the amount of expiration. To obtain a rough understanding

about this affect in a given supply chain, one can perform a simple data summary

such as a report showing the expiration percentage by case size (e.g., 2% for 24 unit

cases, 1.7% for 12 unit cases, 1.5% for 6 unit cases).

One remedy to the case size problem is to reduce case sizes (e.g., from 24 units to

12 units). Another solution is to design modular cases that can be split, for example

in half, allowing for fractional ordering (e.g., 1/2 case order). We recommend that

retailers urge manufacturers to address the case size problem.

4. Establish rules for the remaining shelf life considering the product’s de-

mand rate. In this thesis, we developed an optimization model to determine the

minimum remaining shelf life of a product that is leaving the manufacturer’s supply

chain. Practitioners can directly replicate our approach. For practitioners who do

not have resources to undertake an optimization effort, based on the insight we gain

160

from our analysis, we recommend that the minimum required shelf life is established

in a fixed number of days or weeks, not in a percentage of the manufactured shelf life

(a common practice by retailers). This fixed number should depend on the product’s

demand rate with an inverse relationship (e.g., 3 weeks for a fast-moving item versus

7 weeks for a slow-moving item).

Some manufacturers have the capability to capture the remaining shelf life of

items as they move in their supply chain. Retailers can request this information

electronically, for example attached to the advanced shipping notice, for incoming

shipments. This information helps retailers to increase compliance to the minimum

required shelf life rules. In addition, retailers can use this information when negoti-

ating the unsaleables policies for expired products (e.g., the retailer can enforce the

full-reimbursement policy for a manufacturer that frequently sends old products).

5. FMI advocates for retailers to mandate manufacturers print barcodes

on the case of the product informing the product’s manufactured date. Current

manual process discourages retailers from capturing this important information. Some

manufacturers print this information on the single product (e.g., a bottle) as opposed

to the case; this makes it even more inefficient for retailers to capture the data.

6. Incorporate the unsaleables performance (on expired and discontinued prod-

ucts) in sales-force compensation schemes. Based on our analysis, manufacturers

at a minimum should deduct the expired and discontinued product volume from com-

pensation, in particular for products with a low to medium gross margin (the risk

from unsold inventory can be justified for high gross margin products). Our analysis

differentiates compensation strategies depending on the product’s gross margin and

market characteristics. Practitioners can follow our strategy, penalizing the sales-

force for expiration more for low margin products in competitor dominated markets

than for high margin products in markets with high brand penetration. For example,

under the sales-volume-based commission scheme, twice or three times of the expired

161

volume can be deducted from gross sales volume for low margin items and in markets

with high brand penetration.

7. Establish incentive systems to prevent expiration. Some retailers already

have effective incentive structures in place; see industry case studies in Table 1.1. For

example, a retailer established a system in which stores receive a standard credit for

unsaleables and are charged for what they return; also, stores that return excessive

levels are penalized while below average stores are rewarded. Manufacturers can sim-

ilarly incentivize supply chain managers (for example, using the remaining shelf life

as a metric) and sales-force, as we discussed earlier in recommendation 6.

Future work

Unsaleables is a complex and broad problem involving multiple organizational

functions and channel partners with conflicting objectives. In this thesis, we narrow

our focus to product expiration type of unsaleables and study problems in three areas:

root causes of product expiration, the remaining shelf life problem, and sales-force

incentives. There are still several more important issues that can benefit from future

academic research.

In our opinion, the most important issue is the incentive problems in the channel.

Unsaleables policies determine the terms of reimbursement provided by manufacturers

for the cost of unsaleables. Two main policies in practice favor either the manufacturer

or the retailer, depending on the balance of power. Typically, the benefited party

has little incentive to improve the practices that cause unsaleables. For instance,

either the manufacturer does not have an incentive to supply fresher products to the

retailer (e.g., when the policy provides a wholesale discount for future unsaleables) or

the retailer does not have an incentive to manage store inventory better (e.g., when

the policy provides full reimbursement). Future research can develop an unsaleables

policy that aligns the interests of manufacturers and retailers.

162

Similar incentive problems exist within a firm, manufacturer or retailer. Practices

leading to unsaleables span multiple functions (e.g., manufacturing, sales, warehous-

ing, logistics, procurement, store operations). In practice, either one function absorbs

the cost regardless of cause or no particular function is accountable for the cost of

unsaleables [64]. As a result of these misaligned incentives, we see behaviors such as

the sales force flooding the market with excess inventory or plant managers not hav-

ing any regard to waste implications when determining production batch sizes. These

behaviors can be altered by designing coordination mechanisms to reduce the occur-

rence of expiration. In this thesis, we address the incentive problem of the sales-force.

Similarly, unsaleables performance needs to be incorporated into the performance

metrics of other functions contributing to the problem.

Batching in manufacturing and transportation is another outstanding issue im-

pacting product expiration. Both in practice and in academia, batching problems

are solved independent of the impact of batching on product expiration in the down-

stream supply chain. Practitioners typically consider the handling cost and demand

in determining shipment sizes and consider the production-line productivity and de-

mand in production size decisions. To improve profitability for perishable products,

it is important to consider product expiration effects in batching decisions.

Another potential opportunity to reduce product expiration is shelf space op-

timization. The existing literature studies the shelf space allocation problem con-

sidering cross-correlation of demand and substitution effects disregarding expiration

outcomes. If excess shelf space is allocated to a slow-moving item or an item with

a short shelf life, then its effect on expiration should be included in the cost of shelf

space allocation. Shelf allocation can be optimized based on a multi-product model

considering profit, expiration cost, shelf life, and cross-correlation of demand. This

work is important for perishable products in retail operations since it incorporates an

important factor, shelf lives, to shelf space allocation decisions.

163

Furthermore, sustainable packaging practices are an important topic in unsaleables.

Sustainable packaging has implications both for product damage and expiration.

Manufacturers reduce plastic to decrease material cost and waste going to landfills,

but ironically reduced plastic makes products vulnerable to damage in the supply

chain increasing unsaleables. Also, biodegradable packages are believed to be pre-

ferred by consumers but reduce the shelf life of products increasing the probability of

expiration. Future research can investigate optimal package designs considering all

trade-offs involved in the supply chain.

Lately, retailers have been occupied with the problem of handling damaged prod-

ucts that are categorized as hazardous materials by law. Examples to these products

include batteries, bleach, paints, and pesticides. The environmental protection agency

requires that these items are collected at proper disposal facilities as opposed to be-

ing mixed with the rest of unsaleables which are collected at reclamation centers.

Recently, several retailers were fined significant amounts for mishandling hazardous

waste (e.g., Walmart paid $81.6 million in 2013, Target $22.5 million in 2011, Wal-

greens $16.6 million in 2012 [29]). This problem has been a substantial concern for

retailers. Thus, the industry is looking for efficient means to process these items in

reverse logistics.

164

Bibliography

[1] Adjustable rate policies: An unsaleables white paper, 2002. Food Distributors In-

ternational, Grocery Manufacturers Association, Food Marketing Institute, De-

loitte.

stock-out based substitution: An application to vending machine products. Mar-

keting Science, 17(4):406–423, 1998.

the presence of supplier competition. Marketing Science, 29(5):846–857, 2010.

[4] Amiya K Basu, Rajiv Lal, Venkataraman Srinivasan, and Richard Staelin. Sales-

force compensation plans: An agency theoretic perspective. Marketing science,

4(4):267–291, 1985.

[5] Fernando Bernstein, Fangruo Chen, and Awi Federgruen. Coordinating supply

chains with simple pricing schemes: the role of vendor-managed inventories.

Management Science, 52(10):1483–1492, 2006.

[6] Joseph Blackburn and Gary Scudder. Supply chain strategies for perishable

products: the case of fresh produce. Production and Operations Management,

18(2):129–137, 2009.

[7] Sila Çetinkaya and Chung-Yee Lee. Stock replenishment and shipment scheduling

for vendor-managed inventory systems. Management Science, 46(2):217–232,

2000.

& Service Operations Management, 2(2):186–202, 2000.

[9] Rosy Wei Chen, Dundee Navin-Chandra, et al. A cost-benefit analysis model

of product design for recyclability and its application. Components, Packaging,

and Manufacturing Technology, Part A, IEEE Transactions on, 17(4):502–507,

1994.

[10] Xinlei Chen, George John, and Om Narasimhan. Assessing the consequences of

a channel switch. Marketing Science, 27(3):398–416, 2008.

165

[11] Marcel Corstjens and Peter Doyle. A model for optimizing retail space alloca-

tions. Management Science, 27(7):822–833, 1981.

[12] Tinglong Dai and Kinshuk Jerath. Salesforce compensation with inventory con-

siderations. Management Science, 59(11):2490–2501, 2013.

[13] Nicole DeHoratius, Adam J Mersereau, and Linus Schrage. Retail inventory

management when records are inaccurate. Manufacturing & Service Operations

Management, 10(2):257–277, 2008.

[14] Nicole DeHoratius and Ananth Raman. Inventory record inaccuracy: an empir-

ical analysis. Management Science, 54(4):627–641, 2008.

[15] Municipal solid waste generation, recycling, and disposal in the United States,

2012. Environmental Protection Agency.

[16] B.E. Fries. Optimal ordering policy for a perishable commodity with fixed life-

time. Operations Research, 23(1):46–61, 1975.

[17] Okitsugu Fujiwara, Hanijanto Soewandi, and Dayani Sedarage. An optimal or-

dering and issuing policy for a two-stage inventory system for perishable prod-

ucts. European Journal of Operational Research, 99(2):412–424, 1997.

[18] Gallino. Does inventory influence demand? Exploring billboard and scarcity

effects. In Consortium For Operational Excellence in Retailing Conference, Har-

vard Business School, 2011.

[20] Chon-Huat Goh, Betsy S Greenberg, and Hirofumi Matsuo. Two-stage perishable

inventory models. Management Science, 39(5):633–649, 1993.

problem. Econometrica: Journal of the Econometric Society, pages 7–45, 1983.

[22] V Daniel R Guide Jr and Luk N Wassenhove. Managing product returns for

remanufacturing. Production and Operations Management, 10(2):142–155, 2001.

[23] Daniel B Hall. Zero-inflated poisson and binomial regression with random effects:

a case study. Biometrics, 56(4):1030–1039, 2000.

[24] Milton Harris and Artur Raviv. Some results on incentive contracts with appli-

cations to education and employment, health insurance, and law enforcement.

The American Economic Review, pages 20–30, 1978.

[26] Joseph M Hilbe. Negative binomial regression. Cambridge University Press, 2011.

166

[27] Teck-Hua Ho, Young-Hoon Park, and Yong-Pin Zhou. Incorporating satisfaction

into customer value analysis: Optimal investment in lifetime value. Marketing

Science, 25(3):260–277, 2006.

[28] Bengt Hölmstrom. Moral hazard and observability. The Bell journal of eco-

nomics, pages 74–91, 1979.

[29] EPA is Hitting Retail Stores for Hazardous Waste Violations, 2013.

http://www.mondaq.com/.

[30] Joint indsutry unsaleables report: The real causes and actionable solutions, 2008.

Grocery Manufacturers Association, Food Marketing Institute, Deloitte.

[32] Saravanan Kesavan, Bradley R Staats, and Wendell Gilland. Volume flexibility in

services: The costs and benefits of flexible labor resources. Management Science,

2014.

[33] Michael Ketzenberg and Mark E Ferguson. Information sharing to improve re-

tail product freshness of perishables. Production and Operations Management,

15(1):57–73, 2006.

in the grocery industry. Production and Operations Management, 17(5):513–521,

2008.

[35] Michael E Ketzenberg, Erwin Laan, and Ruud H Teunter. Value of information

in closed loop supply chains. Production and Operations Management, 15(3):393–

406, 2006.

[36] Michael E Ketzenberg, Gilvan C Souza, and V Daniel R Guide. Mixed assembly

and disassembly operations for remanufacturing. Production and Operations

Management, 12(3):320–335, 2009.

[37] Paul R Kleindorfer, Kalyan Singhal, and Luk N Wassenhove. Sustainable op-

erations management. Production and operations management, 14(4):482–492,

2005.

[38] A Gürhan Kök and Kevin H Shang. Inspection and replenishment policies for

systems with inventory record inaccuracy. Manufacturing & Service Operations

Management, 9(2):185–205, 2007.

[39] A.G. Kök and M.L. Fisher. Demand estimation and assortment optimiza-

tion under substitution: Methodology and application. Operations Research,

55(6):1001–1021, 2007.

in manufacturing. Technometrics, 34(1):1–14, 1992.

167

[41] Andy Liaw and Matthew Wiener. Classification and regression by randomforest.

R News, 2(3):18–22, 2002.

[42] Lloyd A Mancl and Timothy A DeRouen. A covariance estimator for gee with

improved small-sample properties. Biometrics, 57(1):126–134, 2001.

[43] P McCullagh and JA Nelder. Generalized Linear Models. Chapman and Hall,

1989.

[44] David Meyer, Evgenia Dimitriadou, Kurt Hornik, Andreas Weingessel, and

Friedrich Leisch. e1071: Misc Functions of the Department of Statistics (e1071),

TU Wien, 2014. R package version 1.6-4.

multilevel data using standard software. American journal of epidemiology,

165(4):453–463, 2007.

inventory and brand competition. Management Science, 50(4):445–457, 2004.

[47] S. Nahmias. Optimal ordering policies for perishable inventory ii. Operations

Research, 23(4):735–749, 1975.

30(4):680–708, 1982.

[49] Reverse supply chain improvement: A joint industry project, 2011. National

Association of Chain Drug Stores, Grocery Manufacturers Association, Raftery

Resource Network, Inc.

[50] Paul Oyer. Fiscal year ends and nonlinear incentive contracts: The effect on

business seasonality. The Quarterly Journal of Economics, 113(1):149–185, 1998.

[51] Barry Alan Pasternack. Optimal pricing and return policies for perishable com-

modities. Marketing science, 27(1):133–140, 2008.

[52] The Impact of Direct Store Delivery, 2007. The Grocery Manufacturers Associ-

ation, AMR Research, Clarkston Consulting and The Nielsen Company.

[53] John W Pratt. Risk aversion in the small and in the large. Econometrica: Journal

of the Econometric Society, pages 122–136, 1964.

[54] R Core Team. R: A Language and Environment for Statistical Computing. Vi-

enna, Austria, 2012. ISBN 3-900051-07-0.

[55] Sandra Rothenberg, Frits K Pil, and James Maxwell. Lean, green, and the quest

for superior environmental performance. Production and Operations Manage-

ment, 10(3):228–243, 2001.

168

[56] Steven Shavell. Risk sharing and incentives in the principal and agent relation-

ship. The Bell Journal of Economics, pages 55–73, 1979.

research. Manufacturing & Service Operations Management, 16(1):2–10, 2013.

[58] Jaya Singhal and Kalyan Singhal. Supply chains and compatibility among com-

ponents in product design. Journal of Operations Management, 20(3):289–302,

2002.

[59] Terry A Taylor. Supply chain coordination under channel rebates with sales

effort effects. Management science, 48(8):992–1007, 2002.

[60] Terry Therneau, Beth Atkinson, and Brian Ripley. rpart: Recursive Partitioning

and Regression Trees, 2014. R package version 4.1-8.

[61] Karel H Van Donselaar, Vishal Gaur, Tom Van Woensel, Rob ACM Broek-

meulen, and Jan C Fransoo. Ordering behavior in retail stores and implications

for automated replenishment. Management Science, 56(5):766–784, 2010.

New York, fourth edition, 2002. ISBN 0-387-95457-0.

Joint Industry Unsaleables Management Conference.

[64] The impact of sales and procurement on reverse logistics management, 2010.

Wipro Technologies.

[65] Achim Zeileis, Christian Kleiber, and Simon Jackman. Regression models for

count data in R. 2007.

169

- India at a Tipping PointUploaded byRahul Hanumante
- B-Plan -Organic Food Store.pptxUploaded byAnushka Sharma
- YTC03447USENUploaded bytradag
- Small Loans.docxUploaded byThomas Wee
- RockNRollUploaded bybinzidd007
- MMMUploaded bysetji
- CME Economic CalendarUploaded bynsbeaney
- GORAMI-BUSINESS-PLAN.docxUploaded byJonald A. Carbonell
- Articles of Partnership for General PartnershipsUploaded byLou Brad Nazareno Ignacio
- Retail Analytics [Anna-Lena Sachs]Uploaded byudayraj_v
- SMU Project GuidelinesUploaded byprinceramji90
- Chapter 1Uploaded byMira Nayrouz
- 53609621 Supply Chain and Order Management AnalyticsUploaded bykrishnendu sengupta
- Kieso_15e_SGV1_Ch09Uploaded byanilegna99
- InternshipUploaded bySunil K Sharma
- Manufacturing Value Proposition by GinesysUploaded bySachin Pilaniwala
- davis grant dye modern resumeUploaded byapi-215861771
- New-Product-Survey.docxUploaded byMariela Rossy Pérez Ortega
- Adrian TanUploaded bykkapur
- SamsungUploaded byLaxmi Kanth Reddy
- VP or RegionalUploaded byapi-79082696
- Respect is GoodUploaded byGsmd Gouvêa De Souza
- MBA Ankit ParikhUploaded byKetan Vador
- IT Trends in RetailUploaded byHitesh Paliwal
- IT in retail sectorUploaded byMegha Dawda
- CH03.pdfUploaded byMuhammad Jaenuddin
- An Insight Into the Needs of Patanjali and Pitching in for InvestmentUploaded byAnwesha Biswas
- niquisha-keysUploaded byapi-458735110
- PptUploaded byAvinash Kumar
- PRONTO-Xi R Retail Overview r2.0LoResUploaded byHe Ro

- VocabularyUploaded bySunny Raj
- wir2016_enUploaded byTheMyawadyDaily
- wir2016_enUploaded byTheMyawadyDaily
- Are Two CEOs Better Than One- WIPRO CaseUploaded byVaibhav Jain
- Who Killed Print Britannica-VSM Related Case Alongwith Lecture 5thUploaded bySunny Raj
- New Microsoft Excel Worksheet.xlsxUploaded bySunny Raj
- ChangedUploaded bySunny Raj
- Submission.docUploaded bySunny Raj
- The Parable of Sadhu -Bowen h. MccoyUploaded bySunny Raj
- CRM 1Uploaded bySunny Raj
- 11299 Extended EnterpriseUploaded bySunny Raj
- ZONE a SubmissionUploaded bysumit
- 11418_Balance scorecard.pdfUploaded bySubrat Rath
- SupplyChainFinanceforIBECconferencefinalversion18July2016Uploaded bySunny Raj
- 1_Example.pdfUploaded bySunny Raj
- Charts.xlsxUploaded bySunny Raj
- New Microsoft Excel WorksheetUploaded bySunny Raj
- Tea.pdfUploaded bySunny Raj
- Tea.pdfUploaded bySunny Raj
- Tea.pdfUploaded bySunny Raj
- Purpose of the Research Study and Limitation - CopyUploaded bySunny Raj
- SunnyUploaded bySunny Raj
- BE Section DUploaded bySunny Raj
- Gainsharing Consulting_ Gainsharing Versus Profit Sharing - Masternak & Associates.pdfUploaded bySunny Raj
- BS - Hamil.xlsxUploaded bySunny Raj
- Ankit Spss (1)Uploaded bySunny Raj
- Book1Uploaded bySunny Raj
- Gainsharing Consulting_ Gainsharing Versus Profit Sharing - Masternak & AssociatesUploaded bySunny Raj
- ExportUploaded bySunny Raj

- finalllllllllllUploaded byapi-433271934
- ARC-1212_1222__manualUploaded byderbot
- FEMA_FS_earthquake_508_081513.pdfUploaded byRizali AB Rahim
- chapter 9Uploaded bycarlo knows
- 09GUploaded byAaron Brandt
- Delta Differences DisadvantgesUploaded byIndia Kosam Puttanu
- 8_TheImpactOfMagnesium.pdfUploaded byRogerio Cannoni
- Cyberpunk 2020 - Reference Book 5Uploaded byCryptoKid
- WAM2Uploaded byTaufik Sign
- VasudharaUploaded bydrwilsonyong
- Study Skills HandbookUploaded byAdham Zuhairi
- A Celebration of the Private Press -- Artists' Books Catalogue -- Winter 2007/8Uploaded byscormeny
- 290984468-GB941-RA-Guide-V1-0.pdfUploaded bymajdi27
- reliabilityweb_uptime_200812Uploaded bycatraio
- A Study on Feasibility of Rapid Wall Panel for Building ConstructionUploaded byIRJET Journal
- 16_Energy_39Uploaded byener
- Circulatory System Notes v1 Student Version(1)Uploaded bybsmith_983738
- Quando Um Pastor Não é QualificadoUploaded byAnjoVision I
- 967Uploaded byB. Merkur
- The Great Pyramid's Blessed Curse _ Conscious Life NewsUploaded byMayLiu
- Electro Chemistry AssaignmentUploaded byGadde Gopala Krishna
- Future ConstruFuture Constructor & Architect - July 2013ctor & Architect - July 2013.pdfUploaded byMostapha Halim
- SystemQ DigitalIO ManualUploaded byeka prasetia
- National Parks of IndiaUploaded byPrakram Prasad
- Sample - Resume FormatUploaded bysandywaits4u31
- Dye Tracing DATA ENVIRORedi Blue NSF Tablet Powder LiquidUploaded byPromagEnviro.com
- Kolonko RefUploaded bySundas Fatima
- Role of ManagersUploaded byprashant_ingale
- Overview of VideoconferencingUploaded byMadhusuthan CS
- French physicist accused of plagiarism seems set to lose prestigious job | Science | AAASUploaded byJorge Lima