Matching Supply and Demand: Delayed Two-Phase Distribution at Yedioth Group—Models, Algorithms, and Information Technology

This paper describes collaborative work between the distribution organization of the Yedioth Group (we use the terms Yedioth Group and Yedioth interchangeably), the largest media group in Israel, Technion—Israel Institute of Technology in Haifa, Israel, and the Massachusetts Institute of Technology (MIT) in Cambridge, Massachusetts. This work is part of the PhD dissertation of the first author, who did his doctoral work at Technion during the period in which he held a senior leadership position at Yedioth as the CEO of the group’s technology company. This effort was also supported by the owner of the group who provided both financial support and extended periods of dedicated time for the first author to work on his research, including a three-month visit to MIT to work with the third author. The first author’s dissertation focused on several major operational challenges that Yedioth Group is currently facing. As we discuss in detail next, like many other major players in the print and electronic media, Yedioth is facing major challenges to sustain its various business lines, particularly the traditional core activities in the area of print magazines and newspaper. Cost reduction is a major goal across the entire group, and is particularly important in this product line. Yedioth constantly seeks to identify opportunities for efficiencies through process innovation, technology innovation, and sophisticated business analytics. The focus of this paper is on the distribution system of newspaper and magazines, which significantly impacts the overall operational cost of the group; savings in this area will have major ramifications for the company. As we discuss next, changing the process caused some significant cultural challenges, because the print industry is very traditional and somewhat conservative.

The work led to fundamental changes in how Yedioth distributes print magazines and newspapers. These changes include the reorganization of the sales agents’ responsibilities and schedules such that they now visit the retailers on Thursdays instead of Wednesdays. The changes generated major cost savings—in particular, a nine percent reduction in magazine production levels and a 35 percent reduction in the return levels. It is important to note that these savings were achieved while maintaining the same sales levels.

The implementation of the work included intensive analysis of sales data, the development of the new concept of two-phase distribution, the development and implementation of a decision support tool based on new models and algorithms, installation of a new information technology (IT) infrastructure, and significant changes in related business practices. Moreover, the models and algorithms that were developed are directly applicable to fundamental and widespread problems in the global print industry and other industries. We believe that this work will have a longstanding impact on the use of operations research (OR) methodologies within Yedioth and potentially more broadly in the print industry.

The practical problem addressed in this paper, though general in nature, is faced by Yedioth on an ongoing basis. The Yedioth Group, founded in 1939 and currently the largest media group in Israel, owns and operates, among others, the country’s major daily newspaper, its most popular news website, many print magazines, several television channels, and other online content services. Yedioth is also a visible player in the global print and media industry and provides IT and technological solutions to content providers around the world.

In recent years, the print industry has been facing constant and increasing competition from the double onslaught of the Internet and the proliferation of television channels. Nevertheless, it has maintained its leading role in the advertising publicity market and retains “eyeballs” worldwide. This competition has forced printing companies to become more operationally efficient, driving them to relentlessly improve their systems to maintain company profits.

Print products are usually distributed by publishers through networks of retailers. Retailer stock is typically replenished once at the beginning of the sales cycle (i.e., period); returns are collected at the end of the period. The length of the period is product dependent (e.g., day, week, or month). A key cost element in the industry is the mismatch between supply and demand. Supply-demand matching is particularly challenging in the print industry because each retailer often faces low-level, but highly variable demand. Publishers are unwilling to lose potential readers and also eyeballs for advertising (the lost-sales penalty cost in the industry is perceived as very high), and hence tend to oversupply retailers to ensure that lost sales are unlikely. This situation leads to high levels of unused supply that are returned and scrapped. The INCA-FIEJ Research Association (IFRA), an international publishing organization with more than 3,000 members, reports that return levels in the print industry are about 25 percent of total production.

The first phase of our work focused on the distribution of La’Isha, the most popular weekly women’s magazine in Israel. Similar to most other companies in the print industry, Yedioth distributes and sells magazines through a network of about 8,000 retailers. Prior to our work, Yedioth made production and distribution decisions in two stages. First, it decided how many magazines to print. Second, it determined the number of copies to supply to each retailer at the beginning of the week. Yedioth collected unsold magazines the following week and gave retailers a full refund. Retailers did not share information among each other or conduct inventory pooling or sharing. Moreover, given that most retail outlets are small mom-and-pop operations, they are not connected to an IT system that would enable easy data collection. This situation meant that determining the quantity to be delivered to each retailer involved the trade-off between overordering and underordering, which is captured by the well-known newsvendor model. Although the magazines are “sold” to retailers, because the retailers are given full refunds for returns, a special research team at Yedioth predetermines the quantities to be delivered. The quantities distributed to the various retailers were based on forecasted demand that was based on historical weekly sales reports. The actual delivery to each retailer is handled by a team of Yedioth sales agents, each covering a number of retailers in a geographical region.

In our work, we developed and implemented decision support tools that are based on new models and algorithms, which are enabled today through an electronic data interchange (EDI) system, and in the future will be supported by a specialized radio-frequency identification (RFID) technological solution. The underlying concept was to use real-time information about the sales at the retail outlets to enable pooling of inventory in the network through delayed distribution. In particular, we wanted to leverage this information to implement an additional resupply decision point during the week, delaying the distribution of some magazines until after partial sales data are obtained. We call this method of supply and resupply two-phase delayed distribution.

The sales agents usually handle the resupply; each holds pooled inventory and is responsible for appropriately resupplying his or her retailers. This situation imposes several operational constraints; for example, the number of retailers that can be grouped together for two-phase distribution has a limit; this limit is dictated by the effective resupply capacity of a single sales agent and ranges from around 80 to 100. In addition to the operational and business challenges involved with changing the distribution practices, the new concept required significant investments in new IT infrastructure and made the decisions about how to replenish each retailer dependent on each other. As a result, we needed to build a decision support tool to allow strategic cost-benefit analyses of the new concept, and to support for production and distribution decisions.

Taking these goals and constraints into account, we developed a new model that supports the following decisions.

1. What should the initial aggregated quantity that is produced for each group of retailers be? Production decisions must be made before any demand information becomes available, because producing additional quantities during the week is prohibitively expensive and effectively infeasible.

2. What should be the quantity supplied to each retailer in the group assigned to each agent at the beginning of the week and the quantity to be held by the sales agent?

3. Based on the demand information from the first part of the week, and given the quantity available, what should be the quantity resupplied to each retailer in the group during the week?

Our model can be viewed as a two-stage stochastic optimization problem with recourse. We showed that the resulting expected cost is jointly convex in the total quantity produced and the quantities distributed to each retailer. Arriving at this result, however, requires a multidimensional dynamic program. By formulating the second-stage subproblem as a linear program (LP), we developed a stochastic subgradient-based optimization algorithm that finds the optimal solution (specifically, how much to produce and how much to distribute to each retailer at the beginning of the week) in a matter of seconds. In particular, the subgradients are generated from the dual program of the second-stage LP.

After an initial and successful five-week implementation covering 50 retailers, Yedioth implemented the decision support tool we previously describe at more retailers for more magazines. Today, the Yedioth team uses it to supply 400 large retailers (mostly chains) that are connected through an EDI system that Yedioth has installed. In addition, the tool is being used to formulate distribution plans for all the group’s magazines, not only the one initially tested; for example, it now generates plans for magazines that focus on men, television, health, fashion, children, and teenagers. The results have been dramatic—a nine percent reduction in production levels while maintaining the same level of sales, and a corresponding 35 percent reduction in returns levels. The savings from printing costs alone amount to more than $250,000 a year, and this does not include the savings from reduced returns (the costs of excess inventory are strictly positive and are approximately $100,000 per year) and the savings from increased production capacity.

Extrapolating to the full 8,000 retailers (some of whom differ qualitatively from those in the initial implementation), Yedioth anticipates annual savings of $1,000,000 in printing costs. To realize these savings, it is currently piloting an RFID-based solution that will enable it to apply the new distribution concept to the rest of its retail network, including the mom-and-pop operations. As a result of its unbridled success with the distribution of weekly and monthly magazines, Yedioth is also piloting the new distribution concept on the shorter sales cycle of its weekend newspaper; if successful (as we predict it will be), the printing cost savings will be an additional $1,000,000 per year.

Yedioth has a distribution network that is typical of publishers in the print industry; therefore, we believe that the two-phase distribution concept could have a major impact beyond Yedioth. Other industries have similar distribution challenges and would likely benefit from implementing this concept. For example, an Israeli bread manufacturer and distributer is currently implementing the models and algorithms described in this paper to manage its specialty-bread distribution system. Finally, many interesting research and practical questions emerge from implementing this concept, as we describe throughout this paper.

We organized the remainder of this paper as follows: In the Literature Review section, we review the relevant literature. In the Business Problem Modeling section, we present our model of the business problem. In the Modeling and Solution Approach section, we describe, in general terms, both our model and the approach we take to solve it. In the Implementation at Yedioth section, we discuss both the initial and expanded implementation at Yedioth. Finally, we present our summary. In Appendix A: Mathematical Model, we provide mathematical details of both our model and the solution algorithm.

Literature Review

The fact that information has value in the management of supply chains is well accepted. It is a fundamental truth with many facets, some of which both academics and practitioners have studied. In particular, some aspects of information in supply chain management have been quantified. Lee et al. (2004) explored the impact of RFID technology on a generic supply chain. They developed a simulation model that showed the impact of RFID technology on the manufacturer–retailer supply chain. Lee and Özer (2007) studied whether RFID technology could affect generic supply chains. They surveyed previous research related to RFID technology and investigated the effect that RFID technology has had on each part of the supply chain. Liu and Miao (2006) proposed employing RFID technology as a production control tool in a discrete manufacturing system. Pramatari and Doukidis (2005) presented a case study involving customers, retailers, and suppliers. They compared the total costs of these systems before and after applying RFID technology, and reported a 10–20 percent reduction in total inventory. Atali et al. (2006) presented models that estimate the value of additional information as it relates to the discrepancy between the inventory record and physical stock. Lee et al. (2000) presented analytical models that explore the value of sharing information between retailers and their upstream suppliers. Gaukler et al. (2007), considering a supply chain with one manufacturer and one retailer, presented analytic models of the benefits of item-level RFID to both supply chain partners.

The literature notwithstanding, several aspects of information within supply chains have not yet been adequately explored and quantified. As we explain previously, our work focuses on distribution systems that are based on networks of retailers. Our goal is to explore and quantify the value of additional information in these systems. In particular, we explore the value of being able to more frequently review the state of the system, and make an additional supplementary distribution during the week (period). The concept of splitting the period into two subperiods is not entirely new. It goes back as far as Allen (1958), who presented an algorithm to allow redistribution when a review is made.

The literature is also replete with studies concerning risk-pooling methodologies in one-warehouse, multiple-retailer systems that resemble the Yedioth distribution setting. This literature is relevant because the system we offer allows a certain application of risk-pooling strategy. Eppen and Schrage (1981) and Federgruen and Zipkin (1984) investigated the one-warehouse, multiple-retailer (risk-pooling) system, and solved the problem using two complementary methods: relaxation (Federgruen and Zipkin) and restriction (Eppen and Schrage). Others, including Jackson (1988) and Jackson and Muckstadt (1989), made contributions to solving this problem.

McGavin et al. (1993) explored a system of multiple identical retailers. They constructed a model for determining warehouse inventory-allocation policies that minimize system-wide lost-sales per retailer between system allocations. They showed that in the case of two allocations, available stock in each interval should be divided to balance retailer inventories. McGavin et al. (1997) showed that for multiple identical retailers, balancing allocations minimizes expected lost sales and backorders over multiple independent intervals (in our case, subperiods). They also demonstrated that balancing is not necessarily optimal for nonidentical retailers. In both of these works, the authors presented heuristics for the location of the timing of the allocation.

Our work is also related to the concept of delayed differentiation (see Swaminathan and Lee 2003), in which one attempts to postpone the specialization step of a product. In our model, we postpone the decision of where inventory should be distributed.

Business Problem Modeling

Business Problem

Before describing the formal mathematical formulation of the model, we discuss some business aspects of the problem, particularly why Yedioth was interested in a solution to this problem. The description reflects the new distribution approach that we modeled and implemented and also provides a general background to the business challenge. Although various retailers buy the magazine from Yedioth each week, Yedioth buys any leftover stock back from the retailers at end of each period (week) and refunds them the price it charged them at the beginning of the week. Clearly, in this situation, if the retailers were left to determine their own order quantities, all retailers would buy many more copies than they would ever sell to ensure that they would not lose sales. As a result, the agreement between Yedioth and the retailers is that Yedioth controls and manages the order quantities.

Note that although Yedioth controls the replenishment policies, it does not own the retailers. The retailers are independent entities; thus, various strategies that have been investigated in the literature, such as physical pooling of inventories and transshipments, are not acceptable to them. Therefore, these strategies are not feasible in practice.

Weekly magazines cannot be held from week to week. Thus, each week is treated as independent of all other weeks. At the beginning of the week, Yedioth must determine how many magazines to produce. This quantity must be based on demand forecasts—not on actual demand. In particular, production is completed before deliveries are sent to the retailers. Moreover, even if demand in a particular week exceeds forecasts (by even a great amount), the fixed cost of a production batch is so high that Yedioth never makes a second production run. Therefore, the initial quantity produced at the beginning of the week—and only this quantity—can be used to satisfy the entire demand over the week.

Before Yedioth implemented the new distribution approach, it made only one delivery to each retailer at the beginning of the distribution cycle (i.e., on Sunday). In the two-phase distribution model, it distributes the magazines in two phases: one at the beginning of the week (Sunday) and another during the week (Wednesday). It splits the period (week) into two subperiods, Sunday–Tuesday and Wednesday–Friday. Note that the work week in Israel begins on Sunday and ends on Thursday or Friday; businesses generally are closed on Saturday. The Sunday distribution coincides with the traditional beginning of the distribution cycle that was in force before the implementation of the new model. In addition, we chose Wednesday for the second distribution for several reasons. The first is that Wednesday is in the middle of the week; this was the timing suggested by the heuristic of McGavin et al. (1993). Moreover, and most importantly for ease of adoption, Yedioth’s sales agents were already regularly visiting the retailers on Wednesday. Thus, adding the second distribution would not require changes to the distribution costs or practices. We numerically examined other days, as we describe in the Implementation section. Moreover, after both the initial and expanded implementations, Yedioth decided to reorganize the responsibilities of the sales agents to make their supplementary distributions on Thursday.

Using our two-phase distribution model, Yedioth must determine the initial quantity to be delivered to each retailer each week and the quantity to hold back and not distribute until the supplementary distribution in the second subperiod. Note that these two decisions also dictate the initial quantity produced. The sales agent, who is responsible for several retailers, holds the quantity kept for the supplementary distribution. After the initial quantities are delivered to each retailer, demand in subperiod 1 occurs. Unmet demand is lost and incurs a lost-sales penalty cost. This penalty cost includes the lost revenue from the sale; more importantly, it also includes the lost-advertising revenue. In this industry, the perceived lost-sales penalty is generally very high. If any magazines are left, they are held until subperiod 2. Holding inventory throughout the week incurs no cost.

At the end of subperiod 1 (Tuesday evening), Yedioth obtains inventory information from its retailers. It may gather this information via physical counts (as in our initial implementation), an EDI system (as in the current implementation at 400 of 8,000 retailers), or an RFID system (as in the planned future implementation). Physical counts are appropriate only for the initial implementation, EDI systems for relatively large retailers, and RFID systems for the smallest retailers. The Implementation section gives details on the RFID implementation.

Using this information, Yedioth determines the distribution quantities for subperiod 2 to be added to each retailer’s inventory (if any) left from subperiod 1. After this second distribution, the demand over subperiod 2 occurs. Unmet demand is lost, as in subperiod 1, because Yedioth does not differentiate between customers in the subperiods. If any magazines are left, they must be collected, shredded, and recycled, which implies a significant disposal cost at the end of subperiod 2. In particular, the company believes that it cannot allow retailers to throw out (or recycle) magazines, because a significant number of copies will end up in waiting rooms (e.g., in medical offices), effectively cannibalizing demand. This cost is identical for both the copies left with the retailers and those left undistributed. To implement the new distribution concept, Yedioth needed a new model to support its production and distribution decisions. Moreover, it had to determine whether the IT investments and business practice changes were worthwhile.

Modeling and Solution Approach

In this section, we provide a high-level overview of the modeling and solution approaches we used to solve the business problem described in the Business Problem Modeling section. For the interested reader, we provide a technically rigorous and detailed discussion of our approach in Appendix A: Mathematical Model. This problem can be classified as a two-stage stochastic program with recourse (see Birge and Louveaux 1997), a class of problems in which we make decisions in two stages. The first stage involves making decisions in the face of a specified uncertainty. After the uncertainty is resolved, second-stage decisions are made in response (recourse). In our problem, the first-stage decisions are the total production level, the initial shipment to each retailer, and the quantity left with the sales agent to be used for the supplementary distribution during the week. We make these decisions in the face of uncertain demand at each retailer in the first subperiod. Clearly, the total production level must equal the total amount shipped to the retailers plus the quantity kept with the sales agent. We construct the demand distributions for each day (and thus for each subperiod) using historical data collected daily over several weeks. The second-stage decisions are made after the sales at each retailer during the first subperiod are known; these decisions specify how the sales agents are to distribute to the retailers the quantity kept back. The second-stage decisions are also made in the face of uncertain demand. Our model’s goal is to minimize the total cost of the system, including linear production costs, linear scrapping costs of the unused supply at the end of the second subperiod, and linear lost-sales costs in both subperiods.

Note that we make the first-stage decisions under the assumption that the second-stage decisions (i.e., the supplementary distribution decisions) will be made to minimize the expected cost with respect to the uncertain demands in the second stage. Thus, the first-stage decisions are made to minimize the immediate and explicit costs of production and expected lost sales in the first subperiod plus the implicit expected cost that results from the optimal second-stage decisions. (The second stage expected cost is implicit because it is the result of another optimization problem; however, it is affected by the first-stage decisions.) The first theoretical question regarding the model is whether the objective function considered in the first stage is convex. This is important because convexity ensures that no local optimal solutions exist and allows the design of algorithms that are based on first-order information (i.e., subgradient information), as Boyd and Mutapcic (2007) discuss. We answer this question positively in Theorem 1 in Appendix A: Mathematical Model. We iteratively prove this by first showing that the second-stage optimization subproblem is convex.

The fact that the model is convex does not guarantee that it can be solved for large instances. The challenge arises because of the high-dimensional state space. In particular, consider a solution approach based on a dynamic programming algorithm that first computes the second-stage optimal value for each possible second-subperiod beginning inventory and then uses these values to solve the first-stage problem. The challenge is that the number of second-stage subproblems that the algorithm must solve increases exponentially as the number of retailers increases; thus, it quickly becomes computationally intractable. Instead, we exploit the convexity of the problem to develop a stochastic subgradient descent algorithm.

Subgradient descent algorithms are designed to work iteratively. Specifically, in each iteration, the algorithm generates a subgradient of the objective function at the current point and then moves to a new point by following this subgradient and using an appropriately chosen step length. This algorithm (provided that the step lengths are chosen appropriately) converges to a local optimum, which in the case of a convex objective function is guaranteed to be a globally optimal point (see Boyd and Mutapcic 2007). In stochastic gradient descent algorithms, the same logic applies; however, instead of generating a subgradient, the algorithm generates a stochastic unbiased estimator of the subgradient at the current point. Similar convergence results have been obtained (see Boyd and Mutapcic 2007).

Considering the first-stage objective function, the main challenge is to generate a subgradient of the implicit expected cost of the second stage, which is dependent on the optimization of the supplementary distribution decisions. To address this challenge for Yedioth, we use two properties. First, the subgradient of the expected cost is equal to the expected subgradient. This allows us to focus on finding the subgradient for a given realization of the demands during the first subperiod and then take the expectation. Second, for a given realization of the demands during the first subperiod, the demand distributions are discrete; therefore, we can formulate the second-stage problem as a LP that is unimodular (i.e., the solution is guaranteed to be integral). We then use known results to show that the optimal dual solution of this LP is the subgradient we require.

As a result, for each point of the first-stage decision variables (i.e., total production level, initial shipments to retailer, and quantity left with the sales agent), we sample from the demand distribution for the first subperiod and then consider the resulting second-stage recourse problem. We solve the corresponding LP and its dual, and use the average of the dual solution over repeated samples of the first subperiod demand as the stochastic subgradient estimator. We then employ an appropriate step in this direction to the next point. Yedioth coded and implemented this algorithm that provided optimal first-stage decisions in a matter of seconds. The supplementary distribution decisions during the week were computed using the aforementioned LP. We note that the second-stage problem is closely related to the well-known multiple-item newsvendor problem.

Implementation at Yedioth

Initial Implementation at Yedioth

As previously mentioned, in conjunction with Yedioth, we investigated the possibility of using lateral transshipments, as Krishnan and Rao (1965) and Herer et al. (2006) discuss; however, we ultimately abandoned this idea because of retailer opposition. The retailers were unwilling to part with inventory that was intended to assist their competitors—the other retailers. We also examined the possibility of using full risk pooling (i.e., placing the inventory of several locations in one place), but also found this to be infeasible. Although retailers are in close proximity to each other, each retailer is an independent entity and maintains its own stock. Sharing is not practical from organizational and logistical points of view. After extensive consultations with Yedioth and the retailers, we settled on the model analyzed in this paper.

Implementation of the two-phase distribution concept was initially impossible at most retailers, because it requires information, which was unavailable, about the sales during the week. Only a few retailers had EDI systems that enable the collection of sales information through scanning of barcodes at the point of sale and reporting this information to the information system. When the initial implementation began, we knew that existing RFID technology could be used to implement the model. Before developing the RFID solution, however, Yedioth examined an initial implementation to determine if installing such technology would be worthwhile.

Yedioth’s initial implementation included only one weekly magazine. In the expanded implementation described next, more weekly magazines were included easily because they are delivered to the same retailers at the same time by the same sales agents.

We compared two systems. The first is the traditional system: each retailer is independent, deliveries are made once a week, and demand information is obtained on a weekly basis. The second is our proposed two-phased delayed distribution system—mid-week information on the inventory level at each retailer is available and used to make a supplementary distribution. In practice, the inventory-level information could be gathered via reports from the sales agents (as in the initial implementation), by an EDI system (as done for the larger retailers), by an RFID system (as is now in an initial implementation stage at the smaller retailers), or by any other means. We note that the actual method of gathering the information is not germane to this study.

Yedioth has good weekly demand information, which it needs for its current system; however, it had never gathered the additional information needed to run our two-phase distribution system. Thus, the company began its initial implementation by gathering this information. In particular, it gathered data about the inventory levels at the retailers on a daily basis—not merely in the middle of the week as required by the algorithm. Yedioth wanted this information for reasons not connected with this research. Note, however, we used this data, which Yedioth gathered over five consecutive weeks, to perform a sensitivity analysis; see Table 2.

Using these data, we evaluated the added value of this information in augmenting our ability to improve order quantities and implement risk-pooling strategies. After both the gathering and evaluation stages, we implemented the two-phase distribution policy discussed previously.

Using historical data, Yedioth identified 10 groups of five retailers (i.e., 50 retailers); each group was handled by the same sales agent and its members (retailers) were geographically close to each other (at most a mile or two apart). Yedioth managers explained to the sales agents, who explained to the retailers, what information they needed to gather.

We recorded and analyzed the inventory level at each of the 50 retailers four times a week for five weeks. In analyzing this information, we discovered numerous opportunities for risk pooling (e.g., points in time when one retailer in a group of five had a stockout and another retailer had a surplus). This situation existed although the system as a whole had too much inventory, including a large number of end-of-period returns and only a small number of stockouts.

Our initial implementation tested the two-phase distribution algorithm at these 50 retailers for five weeks. We maintained close contact with Yedioth’s distribution department, which we found made manual adjustments to our model’s recommendation (e.g., it added copies for safety). Results for the five weeks were impressive (see Table 1).

Table 1 After implementing our model, system performance improved greatly.

The second column in Table 1 reflects the total quantities for five weeks before implementing our model; the third column reflects the total quantities for five weeks after implementing our model. These two columns, which describe the situation for the same period across different years, are comparable. We verified that (1) the demand for this magazine was stable across years, and (2) no special events fell within either five-week period.

We see from Table 1 that our model (after additions from Yedioth’s distribution department) reduced the production quantity for the five weeks from 7,438 to 6,764 copies (a nine percent reduction). As a result, the total returns decreased from 1,908 to 1,174 copies (a 38 percent reduction). Sales increased from 5,530 to 5,590 copies. Moreover, in 35 of 250 cases (i.e., 50 retailers for five weeks), the retailers did not return any copies; that is, they sold all copies and might have sold additional magazines if they had them. This compares to 62 returned copies before the initial implementation. One factor that might contribute to this stockout reduction is that the distribution department sometimes added copies to the model’s recommendation. We note that all the stockouts occurred at the end of the week; there were no midweek stockouts. We performed a sign test to compare the performance of the system using our model to the system prior to implementing the model. The model performed better (i.e., resulted in lower costs) in 206 of 250 cases; the resulting p value of the sign test was less than 0.01.

The process of changing the structure of the distribution chain was a delicate one; therefore, we implemented it with the appropriate sensitivity necessitated by the conservative nature of the organization. One main obstacle that we had to overcome was convincing the Yedioth employees at various levels in the organization that they could print less, while maintaining service and sales levels. Because sales agent compensation is based on sales, these agents were especially hard to convince. This was one of the reasons that Yedioth made manual changes to our model’s solution in the pilot tests. The successful completion of the pilot gave employees confidence in the new distribution structure, our model, and its associated solution.

Expanded Implementation at Yedioth

After the success of the initial implementation (i.e., the pilot), Yedioth decided to expand the implementation of the two-phase distribution model. The first question addressed was whether to continue the supplemental distributions on Wednesdays, when they could be carried out without additional cost, or change them to another day and incur additional costs. To address this problem, we carried out a sensitivity analysis; that is, we examined how changing the distribution day would affect the cost savings. Using the daily demand information that Yedioth gathered, we calculated the cost savings predicted by our model for each possible supplementary distribution day; see Table 2. Recall that the initial distribution day is Sunday and the last day of sales is Friday. For technical reasons, a supplementary distribution cannot be performed on Friday and sales information is not collected.

Table 2 The timing of the supplementary distribution has a significant effect.

As Table 2 shows, doing the supplementary distribution on Thursday results in the greatest savings; however, the savings from a Wednesday distribution are almost as high. Because the supplementary distribution can be made on Wednesdays without additional costs, Yedioth decided to use this day in the expanded implementation. However, after Yedioth saw the results of the expanded implementation and was convinced of the model’s reliability, it reorganized the sales agent responsibilities to move the supplementary distribution to Thursdays.

For the large-scale implementation of the model, Yedioth decided to include its entire magazine-distribution division for retailers that have the necessary technology. This division includes approximately eight magazine titles. For each title, Yedioth realized the savings achieved in the initial implementation for a single title. The company rapidly implemented the model in more than 400 (of the larger) retailers by using an EDI system that was installed specifically to allow the implementation of our algorithm. This new EDI system enables visibility of inventory levels and sales once each day. The implementations of the EDI system at these 400 retailers were straightforward because each retailer was EDI ready (i.e., all required hardware was in place and we needed only to hook the retailer up to the appropriate information system). Results in this large-scale implementation were similar to the results of our initial implementation; the number of copies produced decreased by nine percent without any negative effect on sales.

In its expanded implementation, Yedioth did not include the approximately 7,000 retailers who did not have an EDI system. To realize the benefits of the two-phased delayed distribution system, Yedioth developed and is currently testing a RFID technology solution to obtain the necessary information. In this solution, it places RFID chips on the wrapper of each magazine (see Figure 1) and integrates RFID readers into the magazine stands of retailers without EDI capability (see Figure 2).

In addition to the continuing implementation for its magazines, Yedioth is performing a pilot of our algorithm on its weekend newspaper. In this implementation, it divides the period (i.e., a single day) into two subperiods, and does the initial distribution at 5:00 am and the supplementary distribution at 11:00 am on the same day.

Using our work, Yedioth implemented a limited risk-pooling method in both time (one supplementary distribution) and space (five retailers pooled using a single sales agent). In the future, the company will likely increase the number of retailers pooled by one sales agent, thereby increasing the risk-pooling effect. As discussed previously, the limitation on the number of retailers being pooled resulted from Yedioth operational and organizational issues; algorithmically, the model can include many more retailers.

Implementation in Another Industry

Following the success of our implementation, an Israeli bread manufacturer and distributer, together with 3ID (a provider of RFID solutions to the retail industry in Israel), implemented the RFID system described above and our algorithms to control its bread distribution chain. In particular, this company is evaluating the system for its specialty breads whose shelf life is several days. Currently the system is operational, and we anticipate that the results will be similar to those achieved by Yedioth.

Summary

In this paper, we describe a large-scale implementation of a new distribution concept of two-phase delayed distribution within the print industry—specifically, the distribution of weekly and monthly print magazines by Yedioth, the largest media group in Israel. The major underlying concept is the addition of a resupply point during the week and the use of real-time sales data to delay the distribution of some magazines and effectively pool inventory among several retailers. The new distribution concept required the development of new models to optimize the distribution and shipment decisions within Yedioth’s distribution network. We analyzed the model and showed analytically that it is convex. Moreover, we developed novel stochastic gradient-based algorithms to solve the resulting optimization model. We implemented the algorithms as decision support tools at Yedioth and they are currently using them to guide the distribution decisions of all its weekly and monthly magazines. The implementation of the new concept included the installation of new IT infrastructure and EDI systems, and in the future will include RFID systems. The models developed were used to assess the cost and benefits of these technologies. The results of the implementation are dramatic. Yedioth achieved major reductions in its production and return levels, whereas sales levels remained the same as they were prior to the implementation.

The models discussed in this paper are relevant to other companies in the print industry and other application domains. Therefore, we believe that our work can lead to additional practical applications. In addition, this work raises interesting theoretical questions with practical importance. For example, for various operational constraints, we added the resupply point in the middle of the distribution cycle. However, computational experiments suggest that this is not necessarily optimal. A study that focuses on the optimal location of an additional resupply point would be interesting. Moreover, the value of additional multiple resupply points is an intriguing issue for future exploration.