a strategic empty container logistics optimization in a major

Vol. 42, No. 1, January–February 2012, pp. 5–16ISSN 0092-2102 (print) � ISSN 1526-551X (online) http://dx.doi.org/10.1287/inte.1110.0611

© 2012 INFORMS

THE FRANZ EDELMAN AWARDAchievement in Operations Research

A Strategic Empty Container LogisticsOptimization in a Major Shipping Company

Rafael Epstein, Andres Neely, Andres WeintraubDepartment of Industrial Engineering, University of Chile, Santiago, Chile

{[email protected], [email protected], [email protected]}

Fernando Valenzuela, Sergio Hurtado, Guillermo Gonzalez,Alex Beiza, Mauricio Naveas, Florencio Infante

Compañia Sud Americana de Vapores (CSAV), Valparaíso, Chile{[email protected], [email protected], [email protected],

[email protected], [email protected], [email protected]}

Fernando Alarcon, Gustavo Angulo, Cristian Berner,Jaime Catalan, Cristian Gonzalez, Daniel Yung

Department of Industrial Engineering, University of Chile, Santiago, Chile{[email protected], [email protected], [email protected],

[email protected], [email protected], [email protected]}

In this paper, we present a system that Compañía Sud Americana de Vapores (CSAV), one of the world’slargest shipping companies, developed to support its decisions for repositioning and stocking empty containers.CSAV’s main business is shipping cargo in containers to clients worldwide. It uses a fleet of about 700,000TEU containers of different types, which are carried by both CSAV-owned and third-party ships. Managing thecontainer fleet is complex; CSAV must make thousands of decisions each day. In particular, imbalances existamong the regions. For example, China often has a deficit of empty containers and is a net importer; SaudiArabia often has a surplus and is a net exporter. CSAV and researchers from the University of Chile developedthe Empty Container Logistics Optimization System (ECO) to manage this imbalance. ECO’s multicommodity,multiperiod model manages the repositioning problem, whereas an inventory model determines the safety stockrequired at each location. CSAV uses safety stock to ensure high service levels despite uncertainties, particularlyin the demand for containers. A hybrid forecasting system supports both the inventory and the multicommoditynetwork flow model. Major improvements in data gathering, real-time communications, and automation ofdata handling were needed as input to the models. A collaborative Web-based optimization framework allowsagents from different zones to interact in decision making. The use of ECO led to direct savings of $81 millionfor CSAV, a reduction in inventory stock of 50 percent, and an increase in container turnover of 60 percent.Moreover, the system helped CSAV to become more efficient and to overcome the 2008 economic crisis.

Key words : shipping; container repositioning; logistics; forecasting: applications; inventory; networks: flow;uncertainty; collaborative planning.

Compañía Sud Americana de Vapores (CSAV) isthe largest logistics operator in Latin America and

one of the world’s largest shipping companies in termsof capacity. Founded in 1872, it is the most global

Chilean firm. In addition to its headquarters in Chile,CSAV operates seven regional offices: Valparaíso,Chile; Sao Paulo, Brazil; Hong Kong, China; Hamburg,Germany; New Jersey, United States; Barcelona, Spain;

5

Epstein et al.: A Strategic Empty Container Logistics Optimization6 Interfaces 42(1), pp. 5–16, © 2012 INFORMS

and Mumbai, India. In 2010, CSAV had operations inover 100 countries and more than 2,000 terminals anddepots worldwide.

CSAV’s main business is shipping cargo using con-tainers, which are transported on ships operatedby both CSAV and other companies. The containerfleet consists of about 700,000 20-foot equivalent-unit(TEU) containers of over 10 types in terms of sizeand type of cargo (e.g., 20- and 40-foot dry vans,reefers, and oil tanks). This fleet is valued at approxi-mately $2 billion and serves customer needs of morethan 2.9 million TEU (about 50 million tons) per year.CSAV owns only about 5 percent of its container fleet;it charters the rest from leasing companies. In general,CSAV rents containers on long-term contracts thatallow it to receive the containers at specific depots ina specific time frame and return them to previouslyspecified locations within a predetermined period.

To cover its worldwide client needs, CSAV coordi-nates a multimodal network and offers about 40 con-tainer transport services. These services have fixedroutes and are carried out by more than 180 containerships that have capacities ranging from 2,000 to 6,500TEU each. Each week, CSAV handles approximately400 vessel calls, requiring hundreds of thousandsof container logistics decisions. Intermodal services,which include thousands of feeder ships, barges,trucks, and trains both from CSAV and third parties,complement CSAV’s transport services.

Managing this global operation requires 24/7 atten-tion. CSAV employs more than 6,500 people world-wide, and its decisions impact workers at thousandsof suppliers. The CSAV professionals responsible foroperations are of different cultures, nationalities, andbackgrounds, and live in different time zones, thusmaking it difficult for them to coordinate and planinterconnected activities.

The company previously managed its containerfleet based on decentralized decisions from itsregional offices. The high level of interconnectionof these decisions, the empty container imbalanceamong regions, and the uncertainty in main param-eters, such as demand for containers, led to signifi-cant shortcomings in container management, reflectedmainly by CSAV’s high level of container safety stock.In 2006, CSAV and researchers from the Universityof Chile began a project to address this problem.

As part of the project, we developed an operationsresearch (OR)-based system, Empty Container Logis-tics Optimization System (ECO), which optimizesCSAV’s empty container logistics by integrating theoperational and business decisions of all regionaloffices. Our objective was to minimize global emptycontainer-related costs while guaranteeing a high ser-vice level. ECO is also a collaborative Web-based opti-mization framework in which multiple agents whohave different local objectives and work in dissimilarbusiness conditions make decisions.

In late 2007, we implemented the first version ofECO in Chile and Brazil. When the financial crisishit the shipping industry particularly hard in 2008,CSAV decided to address the crisis by gaining a com-petitive advantage through excellence in managing itscontainer fleet; ECO played a key role in achievingthis goal. CSAV implemented the system rapidly andimplemented globally. In January 2010, ECO becameoperational worldwide.

In this paper, we use the following structure. TheEmpty Container Problem section presents the problemthat we address. In the Background section, we discusshow CSAV handled empty container logistics deci-sions prior to implementing ECO. The Empty ContainerLogistics Optimization System section explains the ORfeatures of ECO and how it improved coordinationand decision making by changing CSAV’s operationalstrategy. Implementing the ECO System describes theproject’s history and how CSAV successfully imple-mented this tool to support its global operations.We discuss quantitative and qualitative results in theImpact section and present final comments in the Con-clusions section.

The Empty Container ProblemThe literature discusses container routing and reposi-tioning problems as they relate to shipping and othermeans of transportation. Movement of empty con-tainers is usually the result of imbalances in movingcargo. Models often take the form of multicommodityflow problems to reflect the conservation of flow andsatisfaction of demand constraints. In Crainic et al.(1993), the authors present one of the first charac-terizations of these problems, as they develop sin-gle and multicommodity network formulations. Erera

Epstein et al.: A Strategic Empty Container Logistics OptimizationInterfaces 42(1), pp. 5–16, © 2012 INFORMS 7

et al. (2005) present a large-scale multicommodityflow problem on a time-discretized network modeldeveloped for the chemical industry; the model inte-grates container routing and repositioning decisions.Choong et al. (2002) discuss a tactical planning prob-lem for managing empty containers on barges forintermodal transportation. They use an integer pro-gram to analyze the effects of the length of theplanning horizon. Jansen et al. (2004) report an opera-tional planning system used to reposition empty con-tainers for a German parcel post firm.

In dealing with empty containers, the handling ofuncertainty, particularly of demand, can be a majorproblem. Erera et al. (2006) propose a robust opti-mization approach based on the ideas of Ben-Tal andNemirovski (2000). Cheung and Chen (1998) handlethe uncertainty issue by using a two-stage stochasticnetwork flow model, and develop procedures to solvethis problem.

To define the empty container problem at CSAV,we first describe its dynamics. The typical cycle of acargo container (see Figure 1) starts when a shippingcompany takes an empty container from a containerdepot. The empty container is loaded on a truck anddelivered to a shipper, who fills it with the merchan-dise to be shipped. Once the container has been filled,the company transports it to its final destination, asindicated by the consignee to whom the cargo is to

Empty containerfrom depot to shipper

Full containerfrom shipper to port

Loading

UnloadingFull containerfrom port toconsignee

Empty containerfrom consignee

to depot

Port oforigin

Port ofdestination

Vessel(includes severaltransshipments)

Emptycontainerlogistics

Destinationcontainer

depot

Origincontainer

depotShipper

Consignee

Figure 1: The graphic depicts the typical cycle of a cargo container, which starts empty in the origin depot andfinishes empty at the destination depot.

be delivered. Typically the company uses multipletypes of transport. The filled container is transportedby truck to the main port, where it will be loadedonto a vessel. Some shippers fill the container at theport, saving the travel from the depot to the ship-per’s location and then to the port. Once all customsrequirements for the filled container at the port havebeen satisfied, the container is then loaded onto a ves-sel. Once loaded on the vessel, the filled container istransported to the port at which it is to be unloaded.The shipping process might also include transship-ments, which involve moving the container from onevessel to another before it reaches its final destina-tion. The destination port is generally close to theconsignee’s location. The filled container is unloadedfrom the vessel and is transported by truck, train,or feeder ship to the consignee’s location, at whichthe consignee receives the container, unloads the mer-chandise, and returns the empty container to the ship-ping company, which performs maintenance on thecontainer if it is damaged or dirty. Consignees some-times use filled containers as warehouses, extendingthe time they keep the containers before returningthem to the company, and paying fines for the latereturn.

We observed four main problems related to man-aging the fleet of empty containers. The first prob-lem is an imbalance of demand among regions, as


we describe above. Some regions are net exporters ofempty containers; other regions are net importers. Thisregional imbalance of supply and demand for emptycontainers forces shipping companies to solve thisproblem by efficiently repositioning the containers.

A company must consider several solutions to mit-igate the imbalance. The most efficient and cost-effective solution is to transport empty containersfrom surplus to shortage locations using containerships and other available transport modes. Emptycontainers represent a firm’s biggest operational costafter ship fuel; therefore, empty container reposition-ing is a key element in a company’s performance. Forexample, during 2010 in one region in Asia, CSAV hada container imbalance—a deficit of more than 900,000TEU—that had to be repositioned empty to cover thisuneven demand.

The second problem is the multiple sources ofuncertainty. The main source of uncertainty is thedemand for empty containers, which depends onexternal factors such as market conditions. Thus, asignificant part of the challenge is forecasting thedemand for empty containers at each location, foreach equipment type, and for a specific date. Thetime and place at which consignees will return emptycontainers is uncertain because customers sometimesdelay returns. Travel times are also an element ofuncertainty. For example, the travel time from Shang-hai, China to Port Elizabeth, New Jersey in the UnitedStates is 37.9 days on average; the standard deviationis 4.1 days, which represents a coefficient of variationof 11 percent. A final source of uncertainty is the avail-ability of vessel capacity allocated to move emptycontainers. Filled containers have a higher priorityover empty containers because paying customers areawaiting them. Empty containers do not necessarilyhave a booking waiting for them at their destina-tion; however, they must eventually be repositionedto reduce the container imbalance.

A third major problem relates to handling andsharing operations information. Tracking worldwidecontainer activities, compiling the information, andmaking it available in real time to all decision mak-ers are technological challenges. In one year, CSAV’stracking systems record over 18 million containeractivities; these systems process over 400,000 trans-actions daily to update information related to the

container activities. At the beginning of the project,the regional offices used different information, whichthey obtained from various sources. This often forcedplanners to make decisions using outdated or inaccu-rate reports. Moreover, data gathering and processingwere done manually, which forced logistics plannersto spend much of their time processing worksheetsand database extracts rather than making decisionsand coordinating empty container activities.

The fourth problem we faced was that the deci-sion makers were distributed throughout the regionaloffices. Each regional office can handle intraregionaldecisions, which relate to trips between locationswithin a geographic area. However, because interre-gional decisions relate to trips between geographicareas, they require coordination between the regionaloffices. Coordinating both intraregional and interre-gional decisions in real time for regions with differ-ent cultures and dissimilar operational practices wasa problem. Prior to the ECO implementation, eachCSAV regional office had a team responsible for coor-dinating and making empty container logistics deci-sions, which the headquarters office in Valparaísocoordinated. Regional teams were supplemented byover 30 logistics planners worldwide, who coordi-nated activities with people both within the companyand from third-party companies.

The imbalance, uncertainty, data, and coordinationproblems inherent to CSAV’s business shaped theproblem. In addition, if we consider CSAV’s globaloperations at thousands of points worldwide, hun-dreds of vessels with scheduled itineraries, multiplecontainer types, and hundreds of thousands of emptycontainer-repositioning routes, this problem is hard tomanage by even the most skilled team of profession-als. Moreover, the seven regional offices coordinatingthis operation are in different time zones, adding com-plexity to daily planning and decision making.

BackgroundCSAV previously managed its container fleet usinga decentralized decision-making process throughregional offices, which have poor visibility of presentand future container flows and no clear definitionof stock levels. This process was the result of the


firm’s rapid growth, which made centralized con-trol difficult given the complexity of global opera-tions. Therefore, CSAV gave significant independenceto its regional offices. Decentralization, among otherfactors related to the size and nature of the problem,led to significant shortcomings in container manage-ment, mostly expressed by a high level of empty con-tainer safety stock held to satisfy expected demands.This was aggravated by the lack of flexibility in com-bining containers in different regions, difficulties inhandling the natural fluctuations of market condi-tions, and the lack of tools to support global decisionmaking. Because of this inadequate knowledge aboutcontainer stocks and future container needs, emptycontainers were not always repositioned efficiently.Regional offices focused on finding the best local opti-mum instead of finding a company-wide optimum.

CSAV’s philosophy is to achieve a superior levelof service (i.e., close to 100 percent customer satisfac-tion). In the shipping industry, efficiency is importantbecause margins are small and the container fleet rep-resents a firm’s major cost. Therefore, an optimizedpolicy for empty container storage and a smart strat-egy to reallocate the containers are crucial for opera-tions. The uncertainty in the main parameters and thelack of tools to quantify this uncertainty led decisionmakers to hold high levels of stock of empty contain-ers to maintain their high service standards.

Initially, we did not have a clear picture of the com-plexity, uncertainty, size, and required coordination ofthe issues to be addressed. After some work, we real-ized that our main contribution would be to assistCSAV in centralizing its empty container inventoryand repositioning decisions.

The Empty Container LogisticsOptimization SystemWe considered developing a single, integrated, androbust optimization model that would address theentire problem, including uncertainties (Bertsimasand Sim 2004). When we did some tests with smallinstances of the problem, we found that the timerequired to find an optimal solution was too long forour needs. Therefore, we opted for a two-stage solu-tion approach, based on a network flow model andan inventory model, as we describe below.

ECO is based on two decision models supportedby a forecasting system. First, an inventory modeladdresses the uncertainty problem and determines thesafety stock for each point. Second, a multicommodity,multiperiod network flow model addresses the imbal-ance problem and supports daily empty containerrepositioning and inventory levels. The service qual-ity is managed by imposing the safety stock as con-straints in the network flow model. Finally, to addressthe coordination problem, ECO uses a collaborativeWeb-based optimization framework in which multipleagents make decisions taking local objectives and dis-similar business conditions into consideration.

The main decision variables considered in the net-work flow model are (1) how and when CSAV shouldreposition empty containers to fulfill its needs atspecific locations using the various transport modes,(2) what the level of empty containers should be ateach point and period, including safety stock to han-dle the uncertainties, and (3) when and where CSAVshould lease and return containers.

Appendix A shows the multicommodity networkflow model. Its main constraints are container flow-balance equations, demand satisfaction, capacity con-straints (defined by both weight and number of slotsavailable), safety stock limits, and initial conditions.Because one goal of the model is to achieve a prede-fined service level, the inventory variables have thesafety stock as lower bounds. If the safety stock can-not be fulfilled with currently available empty con-tainers, the variables representing shortages accountfor the additional requirements for empty containers.Finally, we consider other constraints such as unfeasi-ble or undesirable repositioning options. The model’sobjective is to minimize costs associated with emptycontainer loading, transport, unloading, leasing, andinventory.

We determined that ECO should be run three timesper day to support regional offices located in dif-ferent time zones. To achieve three runs per day,we limited its run time to no more than two hours,using warm starts and other speed improvements.For example, the model incorporates only the mainlocation equipment-type combinations. The 2,000 ter-minals and depots are grouped into 600 locations.From the 600 locations and 10 equipment types, 2,500location equipment-type pairs account for more than


99 percent of the company’s operations. We consid-ered only these combinations. However, this is one ofthe many parameters that ECO system administratorscan fine-tune.

After logical preprocessing, which eliminatesunneeded constraints and variables, a typical instanceof the network flow model has approximately 3.7 mil-lion parameters, 1.2 million variables, and 600,000equations. To reduce the solution time, we first runa relaxed continuous version of the problem andthen round off the container flows. Given that theseare integer-friendly MIP problems, the error gapsare small.

To address the problem of uncertainties, we devel-oped an inventory model. We imposed the servicequality by including adequate safety stock in the net-work flow model at each location based on containertype and period. Both the network flow model andthe inventory model required that we generate emptycontainer demand and return forecasts.

After a lengthy analysis, we found that the bestforecasts required that we integrate time-series mod-els with specific knowledge from the sales force andlocal logistics planners. We based the safety stock onthe accuracy of the forecasts (Thomopoulos 2006). TheECO system has a module that keeps track of his-torical demand and return-forecast accuracy, whichallows us to have updated estimators for the standarddeviation and the mean forecast error for each loca-tion, equipment type, and period.

An important part of the project was understand-ing the relationship between forecast error and safetystock. In Figure 2, we present an example of a loca-tion and a specific equipment type in which we com-puted safety stock based on different service levelsand standard deviations of the demand-forecast error.For example, for a service level of 99 percent, a fore-cast with a standard deviation of 12 units requires asafety stock of 28; this rises to 77 when the standarddeviation is 33.

The results in Figure 2 showed CSAV manage-ment the importance of having the best possibledemand forecasts. After a long testing period, man-agement agreed that the best possible forecasts couldbe determined by a combination of approaches, as wedescribe below.

29

55

77

116

1631

43

64

1120

28

42

0

20

40

60

80

100

120

140

80 95 99 99.99

Saf

ety

stoc

k (c

onta

iner

s)

Service level (%)

Variation of safety stock on service level and demandforecast error standard deviation

33 Units18 Units12 Units

Demand forecast errorstandard deviation

Figure 2: The safety stock depends on the demand forecast accuracy givenby the standard deviation of its error and the service level given bythe company. This example shows the values for one location and oneequipment type, assuming that the only source of uncertainty is given bydemand of empty containers.

1. Time-series forecasts: We tested several time-series forecasting techniques, including ARIMA mod-els (Box and Jenkins 1976). However, we obtained themost accurate forecasts using seasonal and trendedmoving averages. Therefore, ECO uses two types oftime-series forecasts: (1) a moving average forecast,which uses the average of the past n days; (2) atrended seasonal forecast, which uses past demandfrom the same season and adds a yearly trend that iscomputed from the previous year’s figures.

2. Sales force forecast: CSAV developed a quali-tative sales forecast, the consensus forecast model(CFM), in which the sales agents worldwide regis-ter their demand expectations; we elaborate on thisforecast, filter it to ensure its accuracy, and use it tocomplement the time-series forecasts.

3. Logistics planners’ forecast: We realized thatlogistics planners sometimes have information thatmust be considered in the forecasts; therefore, weincluded the option of introducing manual demandforecast values.

To forecast returned containers, we consider thefirst two methods. Note that in each instance, over1,500,000 forecasts are calculated based on updatedinformation and new user settings. This calculationconsiders both demand and return forecasts, thelength of the horizon (i.e., 128 days), and the mainlocation-equipment type combinations.


Logistics planners determine which forecast willbe used at each location based on the forecast accu-racy and their own experiences. In addition, theyenter their demand expectations for specific locations,equipment types, and dates, which can replace theinformation in the ECO system forecasts. Then, theforecast type that the planner in each region config-ured is used in the network flow model.

To calculate the safety stock, we included thevariance of the forecasts errors and the uncertain-ties because of repositioning travel times and capac-ity availability for shipping empty containers (seeAppendix B).

The data problems mentioned above required sig-nificant changes in information quality and avail-ability. CSAV undertook several projects to improvecontainer-activity gathering and data transmission inits regional offices. Moreover, it designed several ini-tiatives to improve information quality (e.g., regard-ing container activities, demand forecasts, and vesselschedules). As a result, the information quality hasimproved in both the ECO system and in many othersystems and business units in the company.

As a result of implementing ECO, the average com-puter time required to track container activities fellfrom over 48 hours to fewer than 24 hours. In someareas, this resulted in dramatic reductions in the timebetween an activity’s occurrence and its entry intoCSAV databases. These reductions were the result ofprojects that CSAV undertook to improve the processthat dictates how to gather and transmit informationfrom the agencies and depots worldwide.

Container activities also include the equipment ID,the activity’s date and type (e.g., loading, discharge,demand, drayage), and complementary informationsuch as booking numbers, ports, and schedules. Thequality of this information has also improved signif-icantly, allowing the system to provide better infor-mation to the logistics planners about the status ofthe containers and the time each is expected to beavailable for use.

Data are sent to ECO directly from many of thecompany’s core databases, which transfer updatedinformation every minute. However, we were awarethat information is not perfect; therefore, we inte-grated a data-cleansing module into ECO. Informa-tion quality has improved significantly because of

CSAV’s data-quality initiatives, some of which wereborn out of problems that were detected during thedevelopment and testing of this project.

To tackle the coordination problem, we designedand implemented a framework that optimizes CSAV’sglobal operations based on a collaborative and partici-pative optimization model on the Internet. The team ateach regional office sets the parameters for each geo-graphical area; the optimization model then consid-ers all the settings simultaneously to deliver a globalresult. For example, users can configure the type offorecast to be used in the network flow model, setthe desired service level at each location for eachequipment type, and refine costs and some constraints.The network flow model solutions are integrated withupdated business information, and several reports anddynamic indicators are made available to all plan-ners and logistics people worldwide. Users from allregional offices review and share optimized plans anduse them as a common basis for planning, coordi-nation, and decision making. Bulk upload function-alities and integration with common office softwareallow parameter tuning and solution analysis in auser-friendly and timely way, given the amount ofinformation available and considered in each instance.

Implementing the ECO SystemImplementing the ECO system was a long pro-cess. In 2006, we evaluated the potential savingsto ensure that this approach would address CSAV’smain problems, while improving efficiency and per-formance. Our first approach was to graduallyimplement the system in all the regional offices,starting with the headquarters office in Valparaíso.We used a prototype-based development process,which required a large amount of field work; how-ever, it ultimately helped us to understand the majordecisions and problems that logistics planners faced.It also allowed us to gradually improve the networkflow model, the inventory model, and the forecastingmodule.

At that time, CSAV empty container logistics oper-ations did not include analysis of global processesor global systems; therefore, each region used a dif-ferent operational method to solve its logistics prob-lems. This became one of the most difficult parts of


the implementation; we had to coordinate differentcultures, processes, concepts, and languages to imple-ment a global unique process and tool.

Another challenge was implementing this new wayof handling empty container logistics simultaneouswith the existing process. The people who were test-ing, tuning, and starting to use ECO had to continuedoing their jobs as usual. We did not force its usage.We trained the users in the methodology and basicOR concepts, believing that using ECO would maketheir jobs easier.

In early 2007, we tested the first prototype using his-torical data; our tests showed significant opportunitiesfor reductions in storing and shipping empty contain-ers. This allowed us to move to the next step—usingdynamic information in the system and supportingreal decisions. In late 2007, we implemented a suc-cessful prototype in Valparaíso, Chile and Sao Paulo,Brazil. This prototype allowed us to work with tworegional offices to understand how they used the sys-tem and determine any problems. For example, werealized that coordinating the regional offices wouldrequire a collaborative Web-based interface. Based onthe results of the prototypes developed in Chile andBrazil, CSAV management committed to sequentiallyinstalling the system in all regions.

In 2008, the financial crisis hit the shipping indus-try particularly hard; international trading fell signif-icantly, and freight and charter rates plunged. Theshipping industry has a relatively large time lagbetween economic cycles and shipping orders. There-fore, the crisis resulted in a large surplus of ships;many shipping companies went bankrupt or becameinsolvent; they were no longer able to make theirassets profitable, and therefore required complex debtrestructurings. Between July 2008 and March 2009,CSAV’s freight cargo fell 18 percent, its sales dropped38 percent, and its market value fell 66 percent.

To address the crisis, CSAV determined that itwould gain a competitive advantage through excel-lence in managing its container fleet. Radically chang-ing its management strategy was imperative, andECO played a key role in achieving this goal. Man-agement decided to improve CSAV’s operations andinstill a standard of excellence in staff at all regionaloffices; it would be based on integrating and optimiz-ing the decisions regarding managing its containerfleet.

CSAV implemented ECO in 2009 and launchedit globally in January 2010. The plan required allregional offices to implement ECO simultaneously.Therefore, to test and use the system, the offices hadto be connected in real time. Once the main fea-tures were ready, a team from CSAV went to eachregional office to introduce the system and train logis-tics planners in the methodology and in using thecollaborative Web interface. In January 2010, testingand revisions in all regions were completed and thesystem became operational worldwide. The ECO sys-tem replaced the existing decision-making process,thus globally optimizing CSAV’s large and complexcontainer shipping system.

To generate and implement a global process to sup-port the tool that we were implementing worldwide,we coordinated weekly logistics meetings with staff ineach region; in these meetings, we defined the emptycontainer flows for the following weeks and targetedoptimal levels of safety stock at each location. Theprocess also incorporated the best practices in usingthe system. We consolidated all this information in anoperations manual that all CSAV offices use today.

ECO runs with the participation of all regionaloffices. The results of the runs are incorporated withbusiness information and presented on a Web plat-form to which all logistics planners have access. Theyconsider the system’s suggestions and complementthem with their information on the status of ports,ships, possible new demands, and other critical vari-ables, and generate the final repositioning plan. Theplanners create the inventory and shipping plans forempty containers based on the system’s suggestions,share them with the rest of the organization, and exe-cute them on time.

ECO was implemented using four servers, eachwith a specific function:

1. A database server stores and computes updatedparameters for the network flow and inventory mod-els. More than 60 million container activities arestored and used for computing necessary information(e.g., historical demand, empty container returns, andtravel times). This server uses a marginal approach togenerate all the parameters based only on the changesfrom the previous run.

2. A second database server stores the net-work flow model solutions combined with businessinformation.


3. A Web server generates dynamic Web pages tomeet user information needs.

4. An application server runs the network flowmodel through GAMS/CPLEX software and includesthe ECO system application, which controls the flowof information entered into the network flow modeland the database storing the solutions.

Today, ECO serves CSAV’s regional offices and isused as a framework to support empty containerinventory and repositioning decisions. The head officein Valparaíso uses ECO output as a common platformfor planning and coordinating interregional decisionsamong the regional offices.

ImpactThe project had both a wider and deeper impactthan originally intended. It led to significant improve-ments in data gathering, real-time communications,automation of data handling, and quality of the deci-sion processes; it allowed managers to make decisionswith better, standardized information. ECO allowedfor global decisions that were clearly superior to theones obtained locally by each region.

In quantitative terms, ECO resulted in savings of$81 million in 2010, compared with the 2006–2009average costs. In 2010, the cost reductions weremainly the result of reducing empty container inven-tory stocks by 50 percent and increasing containerturnover by 60 percent.

Figure 3 shows the cost per full voyage, consider-ing year 2010 as baseline with cost zero. The aver-age cost in 2006–2009 was $35 per full voyage morethan in 2010; we used this value to estimate the ECO-related savings. Considering the 2.9 million voyagesin 2010, the total savings were $101 million. We esti-mate the ECO-related impact to be 80 percent ofthese savings, considering that the demand forecasts,stocks visibility, return forecasts, and logistics solu-tions proposed are vital to the decision making. How-ever, other minor projects were ongoing; we assignthem the remaining 20 percent. This results in the $81million savings mentioned above.

CSAV’s flexibility to modify the size of its containerfleet shows that the savings come from improved effi-ciency because of better management of the fleet. Thecompany has the flexibility to return or hire leased

$ 32 $ 32

$ 51

$ 0–10

0

10

20

30

40

50

60

2006 2007 2008 2009 2010

Cos

t per

full

voya

ge

Average per voyage excess cost 2006–2009:

$35

Voyages 2010: 2.894 millionTotal savings: $35 · 2.9 = $101 million

Estimated ECO effect: 80%Total ECO savings: $81 million

$ 27

Figure 3: The graph shows the cost per full voyage from 2006 to 2009,compared to 2010. The cost during 2010 was considered as level zero,and marginal costs are shown.

containers and determine the size of the fleet basedon market conditions. Thus, the improved efficiencyin container management generates the savings.

The ECO implementation had an impact both byreducing empty container logistics costs and by mak-ing CSAV more profitable than it had been in theprevious decade. Average net income for the years2000–2008 was $60 million; however, CSAV lost $669million in 2009 because of the financial crisis. The sav-ings from this project are an important part of thecompany’s 2010 net income of $170 million. If we con-sider that CSAV had a demand for 2.9 million TEUin 2010, the net income per TEU was $59, an increaseof 91 percent over what it would have been withoutthe ECO-related savings. In addition, the savings gen-erated by the ECO implementation surpassed by farthe average net income of CSAV in the previous fouryears, even if we disregard the 2009 net income (seeFigure 4).

This efficiency improvements gave CSAV these sav-ings and a competitive advantage, which were keyelements in helping CSAV to overcome the crisis andposition itself as one of the major carriers in theworld. According to CSAV CEO Arturo Ricke,

The ECO system allowed us to manage the highlycomplex problem of our empty container fleet logis-tics very efficiently, which is fundamental in the globalmanagement of the firm. This improvement was a


117

–39

–800

–700

–600

–500

–400

–300

–200

–100

0

100

200

300

2006 2007 2008 2009 2010

Net

inco

me

(US

D m

illio

n)

–669

171

–58

Figure 4: The chart shows CSAV’s net income from 2006 to 2010.

major contribution in turning our company aroundand is now one of the key elements in CSAV’soperational strategy.

We estimate that the ECO system will generate anadditional $200 million in savings during 2011–2012.

On the qualitative side, OR made CSAV’s trackingand information systems useful; the optimization sys-tem tells operators what to look for and focus on ina global, complex, and changing network. ECO pro-vided decision makers with a robust and trustworthymethodology that gives them target inventory levels.In addition, the optimization model provides themwith a reliable source of alternative repositioningoptions, which can sometimes suggest new solutionsthat they have not considered.

ECO made contributions in at least four areas:• Data quality, management, and availability: ECO

development required a dramatic improvement ofinformation quality, management, and availability.These improvements positively affected other areas ofthe firm.

• Personnel efficiency: Before its ECO implementa-tion, CSAV used approximately 30 logistics plannersto manage the container fleet. In August 2008, theseplanners managed 550,000 TEU in the container fleet.During 2010, the 30 logistics planners used ECO tomanage and coordinate empty container logistics of a700,000 TEU fleet. These planners were able to absorbthis higher workload because the system significantlyreduced the time required to gather and process theinformation they used.

• Unification of processes: Prior to creating andimplementing ECO, each regional office used differentinformation and procedures to plan and coordinateempty containers logistics, thus making the coordi-nation and tracking of logistics decisions difficult.Today, ECO serves as a common platform for analy-sis and coordination, providing a standardized emptycontainer planning process.

• Better reporting and control: Prior to the ECOimplementation, empty container logistics planningrequired a mostly manual gathering of informationfrom different sources. ECO allowed operators moretime to focus on core tactical and strategic decisions.Logistics planners now spend most of their time mak-ing decisions and coordinating activities rather thanprocessing data and reports.

In addition, in the past five years, CSAV’s containerbusiness has increased significantly. Its use of ECOhas allowed CSAV to maintain the same quality ofservice while incrementing the number of empty con-tainers provided to customers well below the increasein the number demanded. The overall environmentaland economic effects have been positive.

ConclusionsThis project and the resulting OR-based solution pro-vided a decision support framework that changedhow logistics planners make their decisions and coor-dinate empty container logistics. This change pro-duced important qualitative and quantitative resultsand also plays a key role in CSAV’s operational strat-egy today. OR allowed CSAV to coordinate emptycontainer logistics and support decision makingunder uncertainty, changing conditions, and imper-fect information.

The ECO implementation demonstrated importantfeatures of OR usage—how it must integrate withinformation technology, data handling, and commu-nications. Most importantly, it fostered a genuinepartnership between the researchers at the universityand the users at CSAV throughout the entire develop-ment process. As CSAV managers became convincedof the significant benefits of an OR approach, we wereable to align the organization behind a change inits operational strategy, which led to the significantpositive benefits described.


The use of prototypes was also an important part ofthe project; at an early stage of the project, it allowedus to understand the real needs and challenges facedby logistics planners.

ECO enabled CSAV to make the organizationalchanges needed to centralize the coordination ofempty container decisions. The collaborative Webinterface assisted the distributed planning processover distinct time zones. The optimization of the net-work flow and the inventory levels subject to safetystock constraints proved to be the right approach tosupport the decision-making process. The improve-ment in data gathering and management needed torun ECO provided high value for users in otherfunctional areas at CSAV.

We believe that the main lesson of this project washow OR at CSAV facilitated a structural change thathad very significant positive results.

Appendix A. The MulticommodityNetwork Flow ModelWe present a simplified version of the network flowmodel considered in ECO. Neely (2008) provides adetailed description.

Decision Variableswik

t ∈�+: Number of containers of type k in inventoryat location i on day t.xijktsv ∈ �+: Number of containers of type k that are

moved by vessel v when it departs from location i onday t and arrives at location j on day s.

yiktsv ∈ �+: Number of containers of type k that are

unloaded from vessel v at location i when it arriveson day t and departs on day s.

yiktsv ∈ �+: Number of containers of type k that are

loaded into vessel v at location i when it arrives onday t and departs on day s.

zikt ∈�+: Number of additional containers of type k

that are required at location i on day t.zikt ∈�+: Number of containers of type k that are no

longer required at location i on day t.

Parametersdikt ∈�+: Expected demand for containers of type k at

location i on day t.r ikt ∈ �+: Expected return of containers of type k at

location i on day t.

gikt ∈ �+: Number of containers of type k already

scheduled to arrive at location i on day t and intransit.

ãikt ∈ �+: Safety stock for containers of type k at

location i on day t.lD ∈ �+: Lead time for unloading containers, num-

ber of days required to make containers available.lU ∈ �+: Lead time for loading containers, number

of days required to prepare containers for loading.

Main ConstraintsThe difference in the number of containers that areloaded and unloaded from a vessel in a given call isequal to the difference of containers that arrive andleave the location on this vessel.

∀ i1 k1 t1 s1v yiktsv − yik

tsv =∑

j1 r

xjikrtv −

∑

j1 r

xijksrv0

Inventory dynamics account for flow conservation,including lead times when loading and unloadingcontainers.

∀ i1 k1 t wikt+1 =wik

t + gikt + r ikt − dik

t −∑

v1 s

yik4t+lU 5sv

+∑

v1 s

yiks4t−lD5v + zikt − zikt 0

The inventory must at least include the safety stock.

ãikt ≤wik

t 0

The model also considers capacity constraints forthe various means of transport and the initial inven-tory available at each location, among other opera-tional constraints. In particular, given the detailed for-mulation of the model, containers might be unloadedat a given port and then loaded again at this portin some instances. To avoid this nonlogical step, weintroduced binary variables. The objective function isto minimize operational costs.

Appendix B. The Inventory ModelWe consider safety stock requirements—at period t,with orders at each period, and with a single sourceof uncertainty—in demand for empty containers. Thesafety stock depends on the demand forecast errorand is given by

St = max4��e1Dt + z��

e1Dt �1051


with the following notation:�e1D

t : Estimator of the mean demand forecast errort days ahead.

� e1Dt : Estimator of the standard deviation of the

demand forecast error t days ahead.z�: Safety factor to achieve a service level of � per-

cent in demand forecast coverage.The previous model was enhanced by assuming that

orders are placed at periods in which main vessels calland that demand is not the only source of uncertainty.Then, the safety stock in period t is given by

St

= max

(⌈

q4t5∑

l=t

4�e1Dl −�e1R

l 5+z�

√

√

√

√

q4t5∑

l=t

4� e1D2

l +� e1R2

l 5

⌉

10

)

1

with the following notation:q4t5= p4t5+ d+

⌈

�N4t5 + z��N4t5

⌉

.�e1D

l : Estimator of the mean demand forecast errorfor day l.

�e1Rl : Estimator of the mean return forecast error for

day l.�e1D

l : Estimator of the standard deviation of thedemand forecast error for day l.

�e1Rl : Estimator of the standard deviation of the

return forecast error for day l.z�: Safety factor to achieve a service level of � per-

cent in demand and return forecast coverage.z�: Safety factor to achieve a service level of � per-

cent in the coverage from estimated times of arrivals.d: Number of days required to unload containers.p4t5: Day of arrival of next vessel, after day t.N4t5: Next vessel, after day t.�N4t5: Estimator of the mean error of the time of

arrival of vessel N4t5.�N4t5: Estimator of the standard deviation of the

time of arrival of vessel N4t5.In the previous expression, q4t5 represents the day

of the next call plus the time needed to cover foruncertainty in this next call. In addition, note that weassumed that the demand and return forecast errors

have normal distributions and are independent vari-ables. We also assumed that delay times have normaldistributions. To achieve a specified service level, weneeded to calculate both � and �.

To include the uncertainty in empty containercapacity on vessels, we added a procedure in theinventory model. Safety stocks are adjusted using afactor calculated using the standard deviation of theempty container capacity of each vessel that has callsat each location considered.

This inventory model was used to provide dailysafety stocks, which were included as the minimumempty container requirements at each location in thenetwork flow model. The basics of inventory manage-ment are in Silver and Peterson (1985). Neely (2008)gives a detailed description of the model used.

ReferencesBen-Tal, A., A. Nemirovski. 2000. Robust solutions of linear pro-

gramming problems contaminated with uncertain data. Math.Programming 88(3) 411–424.

Bertsimas, D., M. Sim. 2004. The price of robustness. Oper. Res. 52(1)35–53.

Box, G., G. Jenkins. 1976. Time Series Analysis: Forecasting and Con-trol. Holden-Day, San Francisco.

Cheung, R., C. Chen. 1998. A two-stage stochastic network modeland solution methods for the dynamic empty container alloca-tion problem. Transportation Sci. 32(2) 142–162.

Choong, S. T., M. H. Cole, E. Kutanoglu. 2002. Empty con-tainer management for intermodal transportation networks.Transportation Res. E 38(6) 423–438.

Crainic, T., M. Gendreau, P. Dejax. 1993. Dynamic and stochasticmodels for the allocation of empty containers. Oper. Res. 41(1)102–126.

Erera, A., J. C. Morales, M. Savelsbergh. 2005. Global inter-modal tank container management for the chemical industry.Transportation Res. E 41(6) 551–566.

Erera, A., J. C. Morales, M. Savelsbergh. 2006. Robust optimizationfor empty repositioning problems. Oper. Res. 57(2) 468–483.

Jansen, B., P. Swinkles, G. Teeuwen, B. van Antwerpen de Fluiter,H. Fleuren. 2004. Operational planning of a large-scale multi-modal transportation system. Eur. J. Oper. Res. 156(1) 41–53.

Neely, A. 2008. Políticas de inventario de contenedores vacíos en laindustria naviera. Unpublished master’s dissertation, Depart-ment of Industrial Engineering, University of Chile, Santiago,Chile. [English translation: Inventory policies for empty con-tainers in the shipping industry.]

Silver, E. A., R. Peterson. 1985. Decision Systems for Inventory Man-agement and Production Planning. John Wiley & Sons, New York.

Thomopoulos, N. T. 2006. Supplier lateness, service level and safetytime. J. Res. Engrg. Tech. 2(4) 291–301.

a strategic empty container logistics optimization in a major

Documents