BIG DATA AND FOOD LOSS MITIGATION IN THE SUPPLY CHAIN

• Sponsoring Institution: National Institute of Food and Agriculture
• Project Status: COMPLETE
• Funding Source: AFRI COMPETITIVE GRANT
• Reporting Frequency: Annual
• Accession No.: 1018660
• Grant No.: 2019-67023-29417
• Cumulative Award Amt.: $499,999.00
• Proposal No.: 2018-08554
• Project Start Date: Jun 1, 2019
• Project End Date: May 31, 2022
• Grant Year: 2019
• Program Code: [A1643]- AERC: Economic Implications and Applications of Big Data in Food and Agriculture

Recipient Organization
ARIZONA STATE UNIVERSITY
660 S MILL AVE STE 312
TEMPE,AZ 85281-3670

Performing Department
Agribusiness, Morrison School

Non Technical Summary
The amount of fresh food lost or wasted between the farm and retail levels results in a substantial loss in economic value. Retailers reject, discard, or donate some 19.5 million metric tons of edible, perishable food products every year, representing a considerable loss of economic, social, and ecological value (Buzby and Hyman 2012). Food waste at the retail level is created by, among other things, retailers' minimum quality standards, over-purchasing by retailers to avoid costly stockouts, and by retailing strategies that induce unplanned purchases. Our proposed research aims to develop a set of empirical methods, applied to data on retail food purchase-and-sale transactions, that can help reduce the amount of food loss and food waste in the retail-consumer market. Our supporting objectives are to design a theoretical model of quality-based price discrimination, to empirically examine quality-assortment strategies for fresh produce retailers and the implications for retail food-loss, to develop a set of machine-learning algorithms to assist retailers in better matching demand-flow to wholesale purchases, and to analyze online and offline purchase patterns in order to better understand the implications of online purchasing for household waste. By focusing new analytical techniques specifically on the problem of food waste, our research seeks to advance the state of knowledge on how managing fresh-food supply chains can be both more sustainable, and profitable, for all stakeholders.

Animal Health Component

50%

Research Effort Categories

Basic

50%

Applied

50%

Developmental

(N/A)

Classification

Knowledge Area (KA)	Subject of Investigation (SOI)	Field of Science (FOS)	Percent
604	5010	3010	100%

Knowledge Area
604 - Marketing and Distribution Practices;

Subject Of Investigation
5010 - Food;

Field Of Science
3010 - Economics;

Keywords

big data

food retailing

food waste

machine learning

online retailing

price discrimination

Goals / Objectives
The overall objective of the proposed research is to develop a set of empirical methods, applied to large, high-dimensional data sets, that have the potential to reduce the amount of food waste at the retail and household levels. In order to achieve this larger objective, we aim to achieve a number of supporting objectives, each targeting a specific area of the overall problem, and using a different set of methodological tools. Our supporting objectives are to: 1. (Stage 1) -- Develop a theoretical model of perishable-food intermediation that demonstrates how retailers have rational incentives to generate food waste, through both quality-based price discrimination, and over-ordering in order to avoid stock-outs; 2. (Stage 2) -- Design and estimate an empirical model of a price-discriminating retailer that tests the hypothesis that imperfect matches between quality--preference and the quality produced on the farm are responsible for substantial food losses in the farm-to-retail segment of the fresh food supply chain. We also aim to use this model, estimated with big data generated by one of the largest US food retailers, to develop normative systems of ordering and preference-matching to help eliminate this source of food loss; 3. (Stage 3) -- Specify and estimate a machine-learning model to estimate the cost of a stockout, and test the bullwhip-effect in fresh produce ordering and sales, generated by the observation that the cost of a stockout is sufficiently large to induce retailers to typically order far more than they can sell on a weekly basis. We aim to use our empirical estimates in this stage to demonstrate how superior empirical estimates, using big data methods, can reduce the bullwhip effect by better matching consumer purchases downstream with retail purchases upstream; 4. (Stage 4) -- Use a large data set of household-panel grocery purchases to empirically examine the difference between food waste from households that purchase online relative to those who purchase food offline stores, comparing estimates using traditional econometric methods, and methods of pure pattern recognition developed in the machine learning literature. With this data, we will develop an empirical algorithm to develop better food-purchase planning tools, from either physical or online stores; 5. (Stage 5) -- In the final stage, we will synthesize our findings from the first four stages, and recommend a set of empirical tools, and data-types, that are able to address the primary cause of most food waste -- uncertainty in demand flow, whether it arises from variation in quality preferences, purchase-timing, or from impulse purchases that arise from unplanned purchases.

Project Methods
We demonstrate how a range of retailing strategies and practices result in substantial food waste throughout all stages of the supply chain by failing to completely mediate farm supply and consumer demand for perishable food. Doing so involves examining three, closely-related food retailing practices that cause cascading incentives for food waste to ripple from downstream retail markets to upstream stages of the food distribution channel. First, we examine the implications of retailers' price-discrimination strategies based on consumers' willingness-to-pay for quality. By setting quality standards that essentially truncate the distribution of willingness-to-pay above the minimum level supplied by farmers, retailers create unintentional food waste as a consequence of rational, profit-maximizing decisions (Verboven 2002; Leslie 2004; McManus 2007; Cohen 2008). Put simply, biological production processes in agriculture result in a quality distribution of fresh produce that is unlikely to match the distribution of quality preferred by consumers. Retailers accordingly have an incentive to price-discriminate in the consumer market by truncating the quality distribution at a level that results in excess food supply at the retailer's optimal price point.In the proposed research, we formally develop the hypothesis that retail quality standards contribute to food waste, and then empirically test the theory using highly granular panel data on fresh produce sales at a major supermarket chain, and a new non-parametric econometric approach. We extend our empirical method to develop recommendations for how data-driven assortment-management practices by retailers can better match consumer preferences to the offered food quality distribution to substantially reduce the amount of perishable produce that goes unsold on the shelf, ultimately to be donated or discarded as food waste. Our empirical test of this hypothesis uses a novel identification strategy to infer the amount of a perishable product that is left unpurchased by retail buyers. That is, we define the amount of food lost due to quality standards as the difference between the distribution of quality produced on the farm, and the distribution of quality-preference from consumers. While there is an extensive agronomic literature that documents the range of quality produced for many fresh produce items, we infer consumers' preference for quality using a non-parametric estimation method to recover the kernel distribution of quality preference. Our estimator accounts for price, and other marketing-mix elements, as well horizontal and vertical differentiation of different product-variants. Second, purchasing sufficient quantities of food to avoid costly stock-outs is a paramount consideration for retailers in developing inventory-replenishment decisions (Fitzsimons 2000; Campo et al. 2003; Anderson et al. 2006; Kök and Fisher 2007; Matsa 2011). Maintaining buffer stocks of fresh produce reduces the probability of a retailer stock-out, but can send ripple effects throughout the food distribution channel that generate food waste. The reason is that customers tend to choose stores on the basis of their product assortment (Bell and Lattin 1998; Richards and Hamilton 2015, 2018), including private-label products that are unique to their own store (Richards et al. 2014). If consumers cannot find an item on their shopping list at a store due to a stock out, retailers face the risk that consumers will switch an entire shopping basket of purchases to a rival retailer's store, or even permanently switch. Given the importance of store-loyalty in retailing (Briesch, Chintagunta, and Fox 2009), and the relatively small cost of maintaining buffer stocks of individual items, retailers typically over-order products in nearly every category they offer. For non-perishable grocery categories, this dynamic poses few issues beyond excessive storage and ordering costs; however, for perishable categories such as fresh fruit and vegetable products, this practice results in surplus food that is ultimately wasted by being redeployed to lower-valued uses or discarded. In this stage of the proposed research, we will estimate the cost of a stockout, test for the severity of the bullwhip effect, and seek to identify how retail inventory-management practices can be modified to mitigate the bullwhip effect's impact on food waste. This research relies on a large-scale data set secured through a data-sharing partnership between the Investigators and a large, regional US supermarket chain. These data will provide highly granular wholesale purchase data for perishable items procured by individual stores, in addition to traditional point-of-sale (POS) retail data. By combining purchase and sale data, we will be able to directly measure the amount of retail food loss due to a fundamental mismatch between demand flow from consumers, and purchasing by retailers. We will develop machine learning algorithms using this data to develop purchasing programs that better match consumer demand, with daily, seasonal, and local demand patterns. By better matching uncertain consumer-demand patterns with ordering and replenishment schedules, our approach will help retailers develop pricing strategies, promotional behavior, and inventory-management practices that have the potential to significantly reduce the amount of perishable-waste at the retail level. Third, the rise of online food retailing has dramatic implications for household food planning that has yet to be fully explored, and points to another way in which traditional forms of retailing contribute to the level of food loss in the farm-to-consumer channel. Namely, food merchandising and promotion through traditional retailers is commonly predicated on the observation that consumers tend to make unplanned food purchases, and over-purchase quantities on the assumption that large units are a better "value" than smaller packages (Inman et al. 2009). In this way, retailers facilitate and contribute to household food waste. When consumers purchase online, however, they are better able to follow pre-filled shopping lists that minimize the possibility of impulse shopping (Belavina 2014). Again, we will use machine learning methods to develop household-planning algorithms that can ultimately be adapted to user-friendly APIs, and integrated into retailers' digital marketing products. Our data for this stage of the proposed research is from InfoScout, Inc. InfoScout uses purchase receipts, both physical and digital, to create large-scale household-panel data sets that include a far larger range of vendors than traditionally covered by either IRI or Nielsen, or even the USDA's own FoodApps data set. InfoScout data includes purchases from all potential bricks-and-mortar outlets (club stores, dollar stores, mass merchandisers, etc) in addition to all online sources.

Progress 06/01/19 to 05/31/22

Outputs
Target Audience:Over the duration of the project, we sought to reach academic researchers, industry members, community members, and government officials charged with evaluating and addressing the problem of food waste. As we progressed through our research, we found that the problem of food waste has many more stakeholders than originally believed as it is now a problem that is being addressed not only at the level of the firm that we originally thought, but entire communities and regions are develping food-waste mitigation strategies. Changes/Problems:As described above, the COVID-19 pandemic, and the change of ownership of Bashas' stores represented real obstacles to achieving our proposed objectives in a timely way. We continue to work with the new ownership of Bashas' stores to acquire the data they originally promised, but none of the original management team who agreed to work with us remain employed by the new Bashas' Stores. That said, we believe that our alternative data strategies -- acquiring inventory ordering, sales, and loss data from Crescent Crown Distributors, and retailer inventory and sales data from DecaData -- allowed us to achieve our original objectives, and more. Specifically, our original proposal using Bashas' data addressed the drivers of food waste from only a single-store perspective. However, our subsequent research lead us to understand that over-ordering is more likely to be a competitive phenomenon as retailers seek to avoid stockouts in order to prevent losing customers to other retailers. We were able to directly test for this theory in our DecaData (confirmed), while we could not have tested this insight with our Bashas' data. Therefore, the delay we encountered likely ended up as a benefit to our ultimate research aims. What opportunities for training and professional development has the project provided?Over the period of the grant, we employed 2PhD students for different periods of time. In each case, the students went on to conduct some of their own resarch on food waste, although none of their own publications were directly targeted to the objectives of this project (they were employed as research assistants, and contributed work to the publications we describe here). In this regard, we believe that we have helped to create a new generation of researchers who intend to focus their efforts, at least in part, on the causes of food waste, and ways to mitigate loss in the food supply chain. We have 2 current PhD students, who were not employed on this grant, who want to make food waste the focus of their research as a result of our discussing this project with them. How have the results been disseminated to communities of interest?Over the period of the grant, we were able to present our research at three meetings of the Agricultural and Applied Economics Association, two in person and one online. At these meetings, we presented our research on retail price discrimination and food waste, on retailer inventory management and loss, and on household online / offline food purchases. We were able to present our findings as well to industry meetings of the Arizona Food Marketing Association (AFMA), which is a trade association for independent retailers in the state of Arizona, and discussed the effect of inventory management and food loss by food retailers. In our field, communicating research results through journal and book publications is considered the most impactful way to disseminate findings. In this regard, we have produced 5 journal publications, and 1 book chapter. Although these outlets are largely targeted to academic researchers, and less often practitioners and government officials, we hope that our findings will be cited by others, and our research will enter the body of knowledge regarding food waste that ultimately informs business and policy decisions. What do you plan to do during the next reporting period to accomplish the goals? Nothing Reported

Impacts
What was accomplished under these goals? At the end of the project, we have substantially achieved all of the proposed objectives listed above. We have had to modify our outputs for Stages 3 and 4 based on data availability, but we feel that our insights achieve what we had intended in the original proposal. Specifically, the data we had anticipated for Stage 3 -- retail ordering, inventory, and waste from Bashas' Family Stores -- did not materialized as the COVID-19 pandemic and subsequent takeover of Bashas' (by Raley's of California) meant that all of our data contacts and NDA were void. However, we managed to find similar data from Crescent Crown Distributors, which is the largest beer distributor in the state of Arizona. They provided store-level ordering, inventory, and loss data that we used to publish our 2022 Agribusiness paper that addresses the empirical objectives described in Stage 3. Also relevant to this objective, we obtained store-level data from DecaData, which is a startup data-syndication firm in San Francisco, that describes sales and inventory data for 6 competing retailers in the US Southeast. Although this data does not have waste information directly, we are able to infer how much perishable food they likely lost by comparing their sales and inventory data. With this data, we were able to fully achieve the objectives specified in Stage 3, and produce another paper that is currently under review. With respect to the Stage 4 objective, we were able to obtain a large data set of household online and offline purchases from Nielsen, Inc. and estimated a model of inventory accumulation and loss at the household level. This manuscript is also currently under review, but shows that online purchasing, of the sort that rose to prominence during the COVID-19 pandemic, can reduce household food loss as consumers are able to more closely monitor and adhere to "shopping lists" and minimize the extent of impulse purchases, and over-ordering that typically occurs in physical stores. Again, we feel that this manuscript addresses the objectives of Stage 4. Finally, we weave our synthesis of our findings into each of the publications that we have generated as a result of this grant. Perhaps most importantly, our Handbook of Agricultural Economics chapter (with Brian Roe), summarizes all of our findings, and presents a synthesis of the literature that we contribute to with each of our journal publications. Writing this chapter was a great opportunity to communicate the entirety of our findings that emerged from this research.

Publications

• Type: Journal Articles Status: Under Review Year Published: 2022 Citation: Richards, T. J., and S. F. Hamilton. "The Cost of a Stockout and Retail Food Waste." Working paper, W. P. Carey School of Business, Arizona State University, Tempe, AZ.
• Type: Journal Articles Status: Submitted Year Published: 2022 Citation: Yonezawa, K., M. Gomez, and T. J. Richards. "Household Food Waste and Online / Offline Food Purchases." Working paper, Dyson School of Applied Economics and Management, Cornell University, Ithaca, NY.
• Type: Journal Articles Status: Published Year Published: 2022 Citation: Richards, T. J., & Hamilton, S. F. (2022). Inventory management and loss in beer retailing. Agribusiness. Forthcoming.

Progress 06/01/20 to 05/31/21

Outputs
Target Audience:During this reporting period, we have presented the results of our research to professional research audiences at the Agricultural and Applied Economics Association (AAEA), INFORMS, Production and Operations Management Society (POMS), practitioners at Crescent Crown Distributing, and seminars at the University of Arizona, Univeristy of Massachusetts, Pennsylvania State University, Montana State University and the University of Idaho. In the past, we have also presented our findings at the National Academies of Sciences, Engineering, and Medicine, and to USDA-ERS officials in Washington, DC. Changes/Problems:Our progress in achieving the Stage 3 goals was initially delayed due to COVID-19-related problems with our food-retailing data partner (Bashas' Foods), but these issues have been resolved with our new data partnership with Crescent Crown Distributing. What opportunities for training and professional development has the project provided?Our initial work on the Stage 3 goals above is designed entirely to educate Crescent Crown staff on the importance of supply-chain factors as drivers of supply-chain waste. Our intent is to publish the initial manuscript out of this stage in a practitioner-oriented journal, which will allow for the rapid and broad dissemination of our findings to a large audience of key decision makers. How have the results been disseminated to communities of interest?To this point, our research results have been published through relevant academic journals to professional research audiences. What do you plan to do during the next reporting period to accomplish the goals?See the progress description above. Over the next 6 months, we intend to complete the three manuscripts targeting the Stage 3 goals, and the manuscript that focuses on the Stage 4 goals. With the new data partnership with Crescent Crown, we see no constraints to completing these goals.

Impacts
What was accomplished under these goals? Relative to these goals, we have substantially achieved the Stage 1 goal (published in AJAE in 2019), and the Stage 2 goal (published in ERAE in 2020). We are currently working on one project pursuant to the Stage 4 goal (target AJAE, late 2021) and threeprojects pursuant to the objectives described in Stage 3. Our progress on the Stage 3 goals was delayed because our original data partner, Bashas' Foods, had to step away from their partnership due to COVID-19-related time-commitment issues in the spring of 2020. However, we have since signed a new data agreement with Crescent Crown Distributing (a large beverage distributor in Arizona) which will allow us to complete all of our Stage 3 goals. We are currently working on three manuscripts in this stage, one of which is nearly complete. In this manuscript, we examine Crescent Crown sales-and-delivery data for the drivers of loss, from overstocking to competitive pricing, and over-commitment of shelf space. We will have this manuscript completed in May 2021, and will then move to the other two. The second project using the Crescent Crown data will examine the "cost of a stockout" concept described in the objectives, and will recommend specific stocking strategies the distributor can take to minimize loss. The third project, currently underway at Cal Poly, uses the Crescent Crown data in a machine learning environment to determine which factors are most important to retail loss, and ways to modify the supply chain to reduce loss in any perishable-product setting. With respect to the Stage 4 goals, wepurchased Nielsen consumer panel data and built a data set for estimation. Specifically, we chose 451 households nationwide who frequently purchased white milk products and has some experience in online milk purchase during the period of April 2018 - April 2020. We find that they purchased 6% less in online stores compared with brick-and-mortar stores. This is evidence that consumer make near-optimaldecisions that can leadto less household food waste. However, it is possible that this observation is not a result of their decision-making,but just changes in their interpurchase time, inventory, and consumption behaviors. To control of these effect, we are developinga forward-looking dynamic structural model of consumer stockpiling behavior. We are now working on writing an estimation code. Our model is based on Hendel and Nevo (2006), but further incorporates consumer heterogeneity in consumption via a Bayesian MCMC algorithm (Imai, Jain, and Ching 2009) and measures inventory depreciation by using an efficiency unit (Ching, Erdem and Keane 2020). The model will allow us to quantity impacts of product characteristics and purchase occasions on a level of household food waste and test whether the decrease in purchase quantity in an online environment is a behavior reducing household food waste.

Publications

• Type: Journal Articles Status: Published Year Published: 2020 Citation: Timothy J. Richards and Stephen F. Hamilton. "Retail Price Discrimination and Food Waste." European Review of Agricultural Economics, 47(2020): 1861-1896.
• Type: Book Chapters Status: Submitted Year Published: 2021 Citation: Stephen F. Hamilton, Timothy J. Richards, and Brian Roe. "Food Waste:: Farms, Distributors, Retailers, and Households." in Handbook of Agricultural Economics, Chris Barrett and David Just, eds. Springer. 2021.

Progress 06/01/19 to 05/31/20

Outputs
Target Audience:Our target audience is defined asother academicresearchers, industry practitioners, and government officials concerned with data and data analysis, food security, environmental degradation, the economic efficiency of food supply-chains, and how economic analysis can be brought to the understanding of food loss and waste. Our intent is to generate four separate manuscripts, present these manuscripts at national and international agricultural economics, industrial organization, supply chain, and general economics meetings, and issue press releases through the Arizona State University, California Polytechnic University, and Cornell University public relations offices. In this reporting period, we presented the results from Stage 1 (theoretical model) and Stage 3 (empirical model of price discrimination and food waste) at professional conferences in Europe (EAAE workshop in Garmisch, Germany) and in the US (AAEA annual meeting in Atlanta, GA) as well as to industry professionals in our own department seminars. Changes/Problems:In the report above, I describe how the COVID-19 outbreak has stalled our ability to acquire data from our regional retailer. Further, both Cornell and ASU have instituted data-security regulations that are delaying the receipt of our Homescan data for Stage 4, but this is a mere delay that will not prevent us from ultimately obtaining the data. What opportunities for training and professional development has the project provided?Our findings in this period provide us the opportunity to talk with community food waste leaders, through the Maricopa County Food Coalition, and the Arizona Food Marketing Alliance regarding their efforts to help industry partners reduce retail food waste. We have meetings with both organizations planned for later this spring, pending completion of the COVID-19 cycle. How have the results been disseminated to communities of interest?As explained above, we have presented papers to both the EAAE and AAEA professional associations. In both cases, the presentations were part of the ERAE publication described above. We presented the theoretical model once, and the empirical model on two separate occasions. I have also presented the empirical model at 3 different university seminar visits. What do you plan to do during the next reporting period to accomplish the goals?In the first box above, I describe what we intend to accomplish under Stages 3 and 4 of the proposed research. In both cases, we are prepared to start empirical estimation, pending the receipt of the data. Given that the models are in place, and we are well-experienced with these types of models, we antcipate no further issues in completing our goals.

Impacts
What was accomplished under these goals? In the current reporting period, we completed the goals described in Stage 1 and Stage 2 above, which were combined to produce the ERAE publication described above. This manuscript describes a novel, structural test of our theory of price discrimination and food waste using a large-scale scanner data set obtained from Nielsen, Inc. (free of cost to USDA). We have completed the conceptual development of the model in Stage 3, but are awaiting approval to receive data from our regional supermarket partner. Although they have signed an NDA, as have we, and they support the research, they simply are trying to find time to commit someone to provide us access to the data. The COVID-19 outbreak in spring 2020 did not help in this regard, so may ultimately end up pushing us beyond our proposed timeframe. That said, part of the objective is to estimate the cost of a stockout. While we are confident that we can find examples of stockouts in historical data, the COVID-19 outbreak is going to make our identification strategy extremely powerful, and much more general. We have also completed the conceptual model for the Stage 4 empirical model, and are awaiting delivery of the data from Nielsen. Our original plan was to use data from InfoScout, but the quality of their omni-channel data did not prove up to their promises. Fortunately, Nielsen has included an omni-channel element in their Homescan data product, so we will be using data from them to estimate our model. Once we have completed the Stages 3 and 4 models, we will be able to synthesize our findings for Stage 5.

Publications

• Type: Journal Articles Status: Under Review Year Published: 2020 Citation: Timothy J. Richards and Stephen F. Hamilton. "Price Discrimination and Food Waste." Accepted and forthcoming in European Review of Agricultural Economics, 2020