informs is located in maryland, usa publisher: … et al.: marriott international increases revenue...

This article was downloaded by: [149.164.92.211] On: 25 March 2016, At: 12:29Publisher: Institute for Operations Research and the Management Sciences (INFORMS)INFORMS is located in Maryland, USA

Interfaces

Publication details, including instructions for authors and subscription information:http://pubsonline.informs.org

Marriott International Increases Revenue by Implementinga Group Pricing OptimizerSharon Hormby, Julia Morrison, Prashant Dave, Michele Meyers, Tim Tenca,

To cite this article:Sharon Hormby, Julia Morrison, Prashant Dave, Michele Meyers, Tim Tenca, (2010) Marriott International Increases Revenue byImplementing a Group Pricing Optimizer. Interfaces 40(1):47-57. http://dx.doi.org/10.1287/inte.1090.0482

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial useor systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisherapproval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitnessfor a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, orinclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, orsupport of claims made of that product, publication, or service.

Copyright © 2010, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, managementscience, and analytics.For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org

http://pubsonline.informs.org

http://dx.doi.org/10.1287/inte.1090.0482

http://pubsonline.informs.org/page/terms-and-conditions

http://www.informs.org

Vol. 40, No. 1, January–February 2010, pp. 47–57issn 0092-2102 �eissn 1526-551X �10 �4001 �0047

informs ®

doi 10.1287/inte.1090.0482©2010 INFORMS

THE FRANZ EDELMAN AWARDAchievement in Operations Research

Marriott International Increases Revenue byImplementing a Group Pricing Optimizer

Sharon Hormby, Julia Morrison, Prashant Dave, Michele Meyers, Tim TencaMarriott International, Inc., Bethesda, Maryland 20817 {[email protected], [email protected],

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

Marriott International’s Group Pricing Optimizer (GPO), a decision support system, provides guidance to Mar-riott personnel on pricing hotel rooms for group customers. GPO uses demand segmentation, price-elasticitymodeling, and optimization techniques to recommend an optimal rate. In operation since late 2006, the sys-tem has improved Marriott’s hotel profitability and enhanced the sales process for both sales managers andcustomers.

Key words : pricing; optimization; revenue management; lodging/hotel industry; segmentation; group business.

Marriott International implemented its GroupPricing Optimizer (GPO), a group pricing sys-

tem that helps its sales force price hotel rooms forgroup customers. The system uses price-elasticitymodels for each statistically derived market segmentto recommend an optimal rate and negotiating range.To assist the sales manager during the negotiations,GPO also displays additional data, including avail-ability of sleeping-room inventory, potential displace-ment of more valuable business, probability of thecustomer accepting the rate, and evaluation of alter-nate dates. In its first two years of operation, GPOhas met its objectives to drive profitable revenue andimprove the sales process for both the customer andthe sales manager.

Marriott InternationalMarriott International, Inc., a leading hospitality com-pany with more than 3,300 hotels in nearly 70 coun-tries and territories, operates and franchises hotelsunder the Marriott, JW Marriott, The Ritz-Carlton,Renaissance, Residence Inn, Courtyard, TownePlaceSuites, Fairfield Inn, SpringHill Suites, and Bulgari

brand names. It also develops and operates vacationownership resorts, operates Marriott Executive Apart-ments, provides furnished corporate housing throughits Marriott ExecuStay division, and operates confer-ence centers. Headquartered in Bethesda, Maryland,the company has more than 140,000 employeesworldwide.Marriott’s heritage can be traced to a root beer

stand that J. Willard and Alice S. Marriott opened inWashington, DC in 1927 (Marriott and Brown 1997).Today, Fortune magazine ranks Marriott as the lodg-ing industry’s most admired company and one of thebest companies for which to work.

Revenue ManagementRevenue management has long been recognized asa critical business practice that contributes increasedrevenues across various industries (Talluri and vanRyzin 2004, Ingold et al. 2001, Cross 1997). The birthof revenue management is largely attributed to theairline industry; its first application was optimizingrevenue associated with individual passengers on asingle flight leg (Belobaba 1987). The scope of the

47

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Hormby et al.: Marriott International Increases Revenue by Implementing a GPO48 Interfaces 40(1), pp. 47–57, © 2010 INFORMS

revenue-management discipline has expanded greatlyover the last two decades, as have the underlyingmodels and technology. The models, which initiallyprovided inventory-allocation recommendations, nowalso generate inventory and pricing controls. We rec-ognize that the models and technology have evolvedsignificantly over these last two decades; however,the user’s interaction with the system still remains animportant consideration.Marriott International was a pioneer in implement-

ing a revenue-management system for the hospital-ity industry. For more than 20 years, it has appliedautomation to performing revenue management forindividual bookings. Today, 97 percent of Marriott’shotels use its One Yield (Overby 2005), a revenue-management system developed for individual book-ings, to provide detailed demand forecasts, optimalinventory allocations, and a seamless interface to thereservation system that executes these control poli-cies. Marriott’s reservation system handles more than75 million transactions per year, making it a richsource of information about individual transactions.A typical revenue-management system uses a sta-

tistical model to forecast unconstrained demand, anoptimization model to generate optimal control poli-cies based upon optimal inventory allocation, anda reservation system to execute these policies (seeFigure 1). One Yield helps Marriott manage millionsof bookings a year. It is primarily a batch system thatproduces a demand forecast for each rate category

Reservationsystem

Revenue managementsystem

BookingsBookings

Inventory

Inventory Demand

data forecastForecastOptimization

Rates

AllocationsRestrictions

Figure 1: The diagram illustrates the relationship between a reservationsystem and a revenue-management system.

and length of stay for each arrival day up to 90days in advance and establishes inventory allocationsthat are published to the reservation system. Subse-quently, various channels sell the controlled inventoryof hotel rooms; these include Marriott.com, Marriott’s800-reservations telephone number, the specific hotel,and global distribution systems to which airlines, carrental companies, and travel agents have access.The current generation of revenue-management

systems in the lodging industry performs demandforecasting and optimization for the individual cus-tomer segment of the hotel business. The next gener-ation, referred to as total hotel revenue-managementsystems, will span the entire hotel, including bothmeeting space and guest rooms, and will opti-mize both individual and group demand segments(Cross et al. 2009). Marriott is currently developingTotal Yield, the next generation of its hotel revenue-management systems to improve profitability for allthe major revenue streams of a hotel: guest rooms soldto individuals, guest rooms sold to groups, meetingspace sold to groups, and meeting space sold for localuse without associated guest rooms (see Figure 2).

Performance MeasurementTo establish the size of the total hotel revenue-management opportunity and to define the specificfunctionality needed to address this opportunity,we enhanced an existing performance metric, therevenue-opportunity model. The concept behind thiseffectiveness metric is to compare the actual decisionsmade with those of an optimal model with perfectknowledge. The optimization model maximizes prof-itability given all available information about demandand supply for a period in the past. The demand

Customertype

Localcateringevent

GroupIndividual

Inventoryrequested Sleeping

roomsMeetingspace

Figure 2: Multiple customer segments compete for rooms and space.

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Hormby et al.: Marriott International Increases Revenue by Implementing a GPOInterfaces 40(1), pp. 47–57, © 2010 INFORMS 49

provided to the model includes booked individualand group business, the statistically unconstrainedindividual demand beyond that which was booked,and the lost or turndown information about groups(i.e., the customer or the hotel declined the book-ing). Because the optimal policy has the advantage ofperfect hindsight and the ability to remake all deci-sions about which business to take or turn down, itproduces a theoretically optimal solution. The ratioof actual revenue to optimal revenue measures theeffectiveness of the revenue-management controls, asimplemented by Marriott staff, processes, and sys-tems. An advantage of using this metric is that it iso-lates the effects of revenue-management controls fromthose of exogenous factors, such as local, regional,national, and global macroeconomic variables, whichcan influence lodging demand, supply, and costs.The gap between actual and optimal revenue shouldlessen as better revenue-management decisions aremade.A final enhancement to this revenue-opportunity

model includes the effect of using price controls forthe group segment. We used a Monte Carlo-based sim-ulation technique, which we refer to as the Pricingand Inventory Revenue Opportunity Model (PROM),to obtain a distribution of the enhanced metric. ThePROM findings established the magnitude of the busi-ness case and enabled us to prioritize the functionalityneeded for a total hotel revenue-management system.

Group BusinessGroup business contributes significant revenue toMarriott International. For full-service brands (i.e.,Marriott, JW Marriott, Renaissance, and The Ritz-Carlton), it can represent more than half the hotel’srevenue.Group business contains characteristics that make

modeling it difficult. The longer booking windowsassociated with groups (compared with individuals)lead to greater statistical uncertainty because deci-sions taken far in the future are more often subject tochange. Groups demand blocks of rooms, introducinga combinatorial aspect to the optimization problem.In addition, transaction data, especially for the largestgroups and smallest hotels, are sparse.To forecast accurately, we need information about

both actual demand and demand that might have

been booked under different control policies. For theindividual customer, we can use statistical methods toinfer unconstrained demand (Orkin 1998, Zeni 2001);however, we rely on sales managers to record lostand turndown data for the group segment. In somecases, certain group inquiries with a short bookingwindow and small number of rooms simply are notlogged, resulting in incomplete information. In casesfor which we do have turndown information, it is alsooften incomplete. For example, group attrition—thedifference between the number of actual (stayed andpaid) rooms and the number contracted—is some-times significant. This information is available onlyfor groups that have materialized and is lacking forlost or turndown groups.Group business is highly negotiated; therefore, con-

tracted group room rates differ from retail rates andfrequently differ from target, budget, or forecastedrates. The other source of variation in group ratesrelates to the group’s requirements. For example, therates of groups that need only guest rooms differ fromthose of groups that also need meeting space.

Group Sales ProcessesThe sales process for groups begins when a hotel’sleadership team, which comprises the general man-ager and personnel from revenue management, sales,and marketing, agrees upon a group sales strategy.The process includes establishing group target rates,goals, and ceilings (maximum rooms to be allocatedfor groups) by season. The revenue manager conductsa historical review of past group data to understandpatterns that can be expected to repeat.Although the team conducts considerable analysis

as part of the historical review, it sets rate and volumetargets manually. Because the manual process cannotmanage the complexity of daily updates and rapidchanges in demand over a very long booking horizon,the target rates are largely static and independent ofgroup size.This manual approach has two serious deficits:

(1) target rates do not reflect individual customers’willingness to pay, and (2) target rates do notreflect the potential for displacement or the trade-offbetween one type of booking and another that mighthave higher value.Although the existing individual revenue-manage-

ment system performed all of its calculations during

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


an overnight batch process, we determined that thedeficits of a batch approach were too limiting. Manag-ing group blocks and calculating displacement usingreal-time pricing would be a better approach.

Group Pricing AnalyticsThe group reservation request requires a response totwo questions: Does the hotel have sufficient roomsto house the group? What rate is recommended forthe rooms?One traditional method for calculating an optimal

price to offer to a customer is to combine the probabil-ity to win a particular piece of business as a functionof price and the corresponding profit (Phillips 2005).We describe one such model in the appendix.To implement this model in a system, we must have

a way of computing the profit function while simul-taneously evaluating a particular group request, amethod of computing displacement cost, a techniquefor estimating pricing curves, and a one-dimension(1-D) optimization algorithm that maximizes expectedprofit. Although a 1-D optimization approach mightappear trivial, it is necessary if the solution is to run inreal time and provide subsecond response to the user.In GPO, we implemented the standard pricing modeldescribed above; however, for successful applicationto the hotel group pricing problem, we needed to findcreative solutions to several parts of the model.When pricing a block of hotel rooms, we generally

know the direct costs of selling one room in a hotel(e.g., the cost of room cleaning, supplies, etc.). We alsoknow the sales cost, which usually incorporates com-missions or rebates, for each group request.Our challenge in the profit function is estimating

the displacement cost. A hotel has limited capacity;to calculate displacement we need to determine thevalue of the business that we will have to turn away ifwe give the inventory to a candidate group. We devel-oped a displacement model that uses the followinginput to calculate displacement cost:• Forecast of remaining unconstrained individual-

customer guest-room demand (the demand we wouldhave seen between today and the arrival date if wehad an unlimited supply of rooms) and associatedrates for individual customers. These forecasts must

be very detailed and incorporate customers’ antici-pated lengths of stay. Our revenue-management sys-tem already had a model that provided forecasts90 days in advance; as part of the GPO project, weextended that period to two years.• Forecast of cancellations for bookings already

received.• Forecast of guest-room demand and associated

rates for additional group customers. Ideally, wewant these forecasts to be unconstrained and verydetailed. Our revenue-management system alreadyincluded constrained group room projections that rev-enue managers in the field were maintaining in thesystem for other purposes. The role of a field revenuemanager includes oversight of forecasts for individualcustomers (these are automated), analysis and inputof forecasts for group business, pricing for all cus-tomer segments, and review of sleeping-room inven-tory levels in the reservation system as recommendedby the revenue-management system. For the initialimplementation, we used the revenue-manager inputsof constrained group projections as a proxy for the“ideal” unconstrained forecasts. In later phases, wereplaced them with unconstrained group room fore-casts and built the displacement model to accommo-date the detailed group forecasts when they becameavailable.• A count of guest rooms that have been sold to

other guests and of any rooms that are not availablefor each stay date (e.g., rooms under renovation).• A count of guest rooms the candidate group

needs for each stay date.Using the data above, we apply the following

methodology to calculate the displacement cost:• We compute the forecasted maximum profit with-

out the current group for that hotel by solving an opti-mization model for a window long enough to coverthe group’s stay plus surrounding days.• Next, we solve the same optimization model for

the same period—after we have removed the candi-date group’s rooms from inventory. In practice, weset the revenue contribution of the group to zero andforce it in.• We compute the difference between maximum

profit without the current group and maximum profitwith the current group; this gives us the estimate forthe displacement cost, which we include in the con-tribution function.

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


We augmented the model to account for cases inwhich we do not initially have enough availableinventory to accommodate the entire group. In sucha situation, we apply a much higher penalty to anyroom that will be overbooked; this feature is very use-ful because it allows GPO to return a price for groupswhen only a small portion of the rooms will need to beoverbooked. GPO quickly pulls correct values for thedirect costs and marginal cost of sales and then runsthe displacement model to obtain the displaced profit.With these values, GPO then computes the contribu-tion function and its derivative for any price.

Estimating Pricing CurvesThe estimation of the pricing curves was at the heartof the GPO effort. We hypothesized that customerprice sensitivity varies based on many variables,including occasion, group size, and season. Becausewe did not know a priori the factors that influencedcustomer price sensitivity, we needed to segment ourcustomers and transactions by price sensitivity beforeestimating the curves. The traditional approach to thisproblem involves the following steps:• Create a database that houses requests for group

business with an indicator of whether it is actual(won) business or lost business, the price quoted (orpaid), and any other variables that might impact pricesensitivity.• Clean the data.• Perform univariate and regression analysis to see

which variables correlate most highly with win rate.• Use CART (classification and regression trees),

CHAID (chi-squared automatic interaction detector),or similar tree algorithms to segment the data, usingwin rate as a target variable (Hill and Lewicki 2005).• Fit pricing curves by segment using some stan-

dard functional form of the curve (e.g., logistic curve).Marriott is a data-rich company, and we were able

to create a database containing 180 descriptors for800,000 group requests relatively quickly. However,when we attempted to implement other steps of thetraditional approach, we faced several challenges forwhich we needed to find innovative solutions.One challenge was that data collected at different

points in the group’s life cycle had characteristicswhose value was unavailable: if the group requestwas turned away at the initial call, we had lim-ited information available about it; however, we had

very detailed data for stayed-and-paid groups. Forexample, when we naively looked at our data, wefound that audiovisual revenue was one of the mostimportant drivers of group win rates. The irony withthis finding was that we were winning 99.9 percentof groups with audiovisual revenue greater than $1,revealing that this field was not filled in for anylost or turndown business. This is not surprising;a group generally determines its audiovisual needsonly when it knows and has agreed to most otherbooking details, including the room rate.After we began working on the segmentation, we

quickly realized that the most “statistically signifi-cant” segmentation did not make intuitive sense andwould be hard to communicate to field personnel.Because user acceptance hinged on the field’s abil-ity to understand the segmentation, this was a criticaljuncture; however, the new system disrupted long-held assumptions about the segmentation, which wefound that the data did not support. Reviewing andtest-pressuring all the tree splits by a joint statistical-business team helped us to develop segmentation thatnontechnical personnel would consider to be businessreasonable and sufficiently statistically significant.The most challenging activity was fitting the pric-

ing curves. We decided to use the logistic model forthe shape of the curve (Talluri and van Ryzin 2004,Phillips 2005). However, we needed a way to aggre-gate pricing data for different hotels by establishinga reference price. Additionally, we needed the pric-ing curves, in combination with the profit function, toyield business-reasonable rates.To address the challenge of fitting the pricing curves,

we defined a proprietary reference price. Unlike inother industries that use a combination of a retailor manufacturer suggested retail price for a productand its competition, in the hospitality industry thesame room has different values for different days ofweek, seasons, and specific arrival dates; in addition,a competitive price differential exists among hotels. Toensure that the forecasts would be accurate, we neededto find a reference price that we could compute histor-ically and forecast into the future, and that would befairly stable. Before settling on the most desirable met-ric, the team evaluated at least 11 different candidates.The second challenge of making the rates realis-

tic led to the development of the dependent pricing-curve model. When we attempted to fit curves to

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


1.91.71.51.31.1

Price ratio0.90.70.5

0

0.1

0.2

0.3

0.4

Pro

babi

lity

to w

in

0.5

0.6

0.7

0.8

0.9

1.0

Figure 3: This graph illustrates the challenge of making pricing curvessharp enough. The bubbles on the chart represent the win rate of observeddata (the size of the bubble corresponds to the number of observations).The solid line represents the logistic regression curve for this data—itresults in optimal rates two to three times greater than the average rateshistorically offered by the hotel; the broken line illustrates the pricingcurve that would produce optimal rates within range of the rates the hotelis selling today.

the existing data, using standard logistic regression,the resulting pricing curves were not sharp enough—they implied price-insensitive demand. Implement-ing these curves into the 1-D optimization modelreturned optimal prices that were too high to be busi-ness reasonable when compared with other prices onthe market. Figure 3 illustrates this problem.This problem of pricing curves not being suffi-

ciently sharp happens often when analyzing pric-ing data, especially in the presence of active sup-ply and demand management. This can be addressedby considering endogenous processes that have astrong effect on price elasticities and curve estimation(Bijmolt et al. 2005).Another way to address the problem is to assume

that an endogenous process exists that influencesquoted prices, so that the price that the salespersonquotes to the customer is correlated with some othervariable Y that we have in our database (the valuewe ultimately used for the variable Y is proprietary).For our group pricing, we developed a new “modelwith dependence” that addresses our curve-sharpnessissue.

Price-Optimization-Model SummaryIn GPO we implemented, in one system, the profitfunction (including the displacement model), pricing

curves by segment, and a 1-D optimization model thatcomputes the optimal price. The pricing segmentationand the curves are validated and refreshed regularlyas a separate process and then loaded into GPO.

Field Collaboration and AcceptanceAfter our pricing research was complete, the tech-nical team built a basic Java Web-based prototypefor 28 hotels. In developing the prototype, our objec-tive was to gather feedback from revenue and salesmanagers at these hotels on the quality of the ratesbeing returned and the features they would like tosee incorporated into a production system. The rev-enue managers would input sample group leads andthe prototype would return a price and a negotiatingrange. The result of this prototype phase was that theusers deemed the rates were reasonable; however, inthe system we implemented, the revenue managersneeded more controls around the rate and the salesmanagers wanted more supporting details displayedon the pricing response screen.Based on the positive feedback from our prototype

users, the team decided to deploy a system pilot;we built the pilot as we would have built an oper-ational production system (in a much more robustform than the prototype), put it in the hands of salesand revenue-management users, and used it to gatheradditional feedback. Assuming that the pilot did nothave any critical deficiencies, we planned on transfer-ring it to a production system and rolling it out tothe entire user base. At the end of the pilot phase, thepilot hotels had experienced successful use of GPOby their group sales managers located at the hotelsand in regional sales offices. Their sales and revenue-management users provided excellent feedback onenhancements that would need to be included in theoperational, fully deployed production version.During the pilot phase, several obstacles to a suc-

cessful rollout became evident. Some required tech-nical enhancements; some required business processchanges. To receive a pricing response from GPO, thebusiness process required that sales managers enterthe group request information into the sales systemin real time as they talked on the telephone with thecustomer.To ensure adequate data with integrity, the team

decided to build a real-time link to pull group

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


requests from the sales-CRM system. This approachrequired sales managers to capture complete infor-mation on group business, both won and lost, ensur-ing quality data for individual pricing requests andfor maintaining the pricing models. Although thisseemed to be a small change to the business process,the sales organization expressed great concern that itnot negatively impact the customer’s experience. Thiswas addressed by ensuring that the system linkagewas fast through performance tuning and by teachingthe sales managers to log the details while conversingwith their customers. The sales managers were ableto ultimately accommodate this change with no inter-ruption to the sales process.The Total Yield team learned a lot about the infor-

mation needed to speedily conclude a group sale andaugmented the basic information coming from thepricing engine to support that end. Because GPO onlysolves the room-rate problem, we included additionalinformation to communicate the strategy for meet-ing space when it is requested in conjunction withgroup sleeping rooms. To do this, we gave the rev-enue manager controls for two key statistics to useas thresholds—the rooms-to-space ratio and catering

Figure 4: GPO provides a recommended rate, negotiating range, and additional information to support the groupsales process.

revenue per group room night. For each of thesethresholds, GPO calculates whether the group underevaluation is within bounds. If not, the sales man-ager receives an alert message that the group doesnot meet the desired threshold. We also created spe-cial control levers to allow revenue managers to stopthe automated pricing for very large groups or spe-cial situations. These situations required a thoroughreview of proposed pricing and contractual terms bythe sales-strategy team. The Total Yield team also cre-ated controls for special events and group-patterncontrols. These allow a revenue manager to indicate,by day, if no additional group business is desirableor if it must meet minimum length-of-stay require-ments to be considered. Controls are then communi-cated to the sales managers to help them understandwhy rates are not provided or why the recommendedrates are high.By integrating GPO with Marriott’s reservation sys-

tem, GPO is able to provide real-time informationon guest-room inventory. The information is morerobust than a simple yes or no, as Figure 4 illus-trates. Using this real-time inventory information andadditional group information that is available in the

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


revenue-management system, GPO is also able to dis-play displacement details to the sales manager. Thisfrequently answers the question, “Why is the rate sohigh?” Coupled with special-event comments, salesmanagers can be knowledgeable about market con-ditions. This helps them work with their customers.For example, a sales manager might say “The weekyou requested is a high-rate week because there isa citywide medical convention. May I suggest somealternate dates that will fit your budget?” They reportthat this added knowledge enables them to keep cus-tomers from contacting competitors and helps themto close sales during initial calls.Early in the GPO design, the team designed critique

tools that help revenue managers assess how well thesystem is being used. These views measure and com-pare contracted group rates and recommended rates.The tools also show divergence, by sales manager,allowing the sales-strategy team to decide if retrainingis necessary and if the customer set is being servedappropriately. We also developed additional analyti-cal tools to track win rates against the theoretical winrate. This helped us to adjust to a rapid decline ingroup demand in 2008, as the recession began to havean impact. Usage tracking is an important indicator ofsales force acceptance. The team tracks this in parallelwith win rates to understand how well the sales teamis complying with GPO recommendations. Becausethe team has been able to closely watch win rates, wehave been able to keep sales usage and compliancehigh.

ImplementationMany distinct pieces of Marriott’s infrastructureneeded to be integrated to provide the real-timepricing response to the customer. These includesales, event-management, forecasting, and reservationsystems.Group request details for the set of GPO hotels are

stored in two systems: the sales-CRM system and anevent management system for managing the details ofexisting contracted customers. A real-time feed fromthe sales-CRM system, which was absolutely criti-cal to the mechanics of the GPO system, was devel-oped to price new sales leads immediately. A detailedovernight feed of the contracted group data already

existed in the One Yield revenue-management appli-cation. In its final state, GPO seamlessly pulls datafrom all the data sources. The displacement modelrelies on forecasts of individual customer and grouproom demand for the entire two-year GPO pricingwindow. This function also exists in One Yield.Another critical requirement of the GPO system

is checking to ensure that the rooms requested bythe group are indeed available to sell. This technicallink to query Marriott’s master sleeping-room reser-vation system was built prior to the system rollout.As a result, each time a sales manager prices a grouprequest, the system performs a real-time availabilitycheck that references the same inventory database asMarriott’s other sales channels.GPO was designed to accommodate thousands of

users selling rooms for hundreds of hotels. To max-imize the system’s accessibility, the GPO interfacewas built as a Java Web-based application, whichis securely hosted behind the Marriott firewall. Thisallows sales managers to use GPO as a pricing toolfrom any computer (with an Internet connection)worldwide.Rolling out GPO was challenging. As we men-

tioned in the previous section, we had many hur-dles to overcome in the business processes to makethe substantial changes to the revenue-managementand sales processes to sell group business. The teamfelt it was critical to provide high-level training tosenior managers in the sales, revenue-management,and event-management disciplines to enable them toassist with change management.In June 2006, we conducted initial facilitated train-

ing. Attendees included 600 revenue-management,event-management, and sales managers from150 hotels; an additional 1,000 sales managers usedWeb-based training, which is ongoing as new salesmanagers and hotels begin to use GPO. Revenuemanagers spent several months using the administra-tive setup and “test pricing” to become comfortablewith the recommended rates. By September 2006,sales managers began using GPO in the sales process;by the end of 2006, we considered the system to beramped up.We undertook one additional integration effort.

During the period in which we were implement-ing GPO, Marriott began its QuickGroupSM initiative,

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


with the goal of selling small-group sleeping-roomblocks online—“the way the customer wants to buy.”In late 2007, we rolled out the initiative with a real-time link to the GPO pricing engine; customers arenow able to book small groups (i.e., 10 to 25 rooms)using this self-service feature on Marriott.com and toreserve the rooms using a credit card.

Success to DateBecause of the challenges associated with measur-ing the impact of good revenue-management deci-sions, Marriott has invested in a set of tools we callrevenue-opportunity models. As we discussed ear-lier in the Performance Measurement section, Marriottbegan using the revenue-opportunity model for indi-vidual bookings shortly after it originally imple-mented automated revenue management. Once weimplemented GPO, we tracked the change in PROM’smeasures of revenue and profit. We were fortu-nate that the same demand levels existed in 2006(our before-GPO data set) and in 2007 (our after-GPO data set). This allowed us a good before-and-after view without concerns that changes in groupdemand caused any changes in demand levels. Theaggregate actual revenue/optimal revenue score for2007, for those hotels that had implemented GPO,was 1.1 percent higher than for the same hotels in2006. This equates to $46 million in revenue for2007. As is the case with most revenue-managementfunctions, these incremental revenues do not incuradditional cost; therefore, they directly add toprofit.Completing this same estimate for 2008 is more dif-

ficult. If we took the native improvement from PROM,we would claim an even higher profit improvementbecause of implementing GPO. However, that fig-ure is inflated by the lower demand we saw as therecession spread. We must restate demand relative toa baseline, so that we will always be able to con-sider one year’s results against the baseline. An initialrestatement of 2008 results from PROM indicates thatGPO produced 1.8 percent incremental revenue andprofit over the baseline 2007 adjusted for the differentdemand levels seen in these two years. Additionally,an increase in the number of hotels using GPO offsetthe decline in overall demand in 2008. Therefore, our

estimate of the profit improvement from GPO in itsfirst two years of use is over $120 million. By the endof 2008, Marriott had approximately 1,600 GPO users;they had used GPO to price more than 525,000 indi-vidual group opportunities and to book $1.3 billionof group business (already stayed and contracted forfuture arrival).

Supporting an Evolving Sales OrganizationIn 2005, Marriott’s sales organization began to rede-fine a successful sales process from the customer’spoint of view. Customer feedback revealed that cust-mers did not want individual sales managers repre-senting unique hotels to call upon them. Therefore,we created Marriott’s Sales Force One initiative totransform the sales process to focus around the cus-tomer rather than the hotel and to redeploy the com-pany’s sales managers to more effectively reach ourthousands of customers. In summary, the companyis evolving the sales force to sell the way our cus-tomers want to buy. To accomplish this goal, Marriottis investing in technology to make it possible forone sales manager to sell for multiple hotels to anycustomer.Automation is critical because the sales managers

are being removed from their hotels and placed inregional sales offices. They are no longer down thehall from their colleagues who establish the pric-ing and sales strategy. The information they needto close a sale must be at their fingertips and inmuch more detail than they required previouslywhen their offices were in the hotels. GPO sup-ports these regional sales offices by providing quick,rational pricing, thus allowing a sales manager toprice and check availability at multiple hotels for thesame group request. Because GPO provides muchadditional information beyond the recommendedprice, it enables these sales managers to feel intouch with each hotel they sell. With the assistanceof GPO, Sales Force One has enabled the samesize sales force to cover 10 times the number ofaccounts.

Other Indications of SuccessGPO use continues to grow. Despite a declining econ-omy and the need for exceptions to cover distressedinventory, the percentage of eligible opportunities

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


priced with GPO remains very high. From the startof 2007 to the end of 2009, the number of partici-pating hotels increased 38 percent at their request.Revenue managers continue to rely on GPO to com-municate the sales strategy despite, or rather becauseof, the challenging economy and sales managers con-tinue to feel confident in the rates they quote usingGPO. Finally, as our Sales Force One initiative is intro-duced to new markets, hotels in that market continueto request GPO. The sales team acknowledges thatthis technology is the key to enabling sales managersin the new organization to price and sell the way ourcustomers want to buy.

Conclusion and Future PlansMarriott International has successfully applied oper-ations research techniques to a wide variety ofproblems. The group pricing functionality is a lead-ing portion of a larger technology investment thatMarriott is making to further empower our sales,revenue-management, and service associates. Thissystems effort includes GPO and adds to it the fullmeeting-space inventory and all the details neededfor hotel staff to deliver group meeting space.To complete Total Yield, the GPO team is building

group forecasting functionality. This includes mod-ules that forecast group room demand with all of itscomplexity, group room rates, and the future valueof meeting space. The team will apply the total-hotel-optimization algorithm, which PROM uses, to solvefor maximum profit by considering all the revenuecategories for this problem: sleeping rooms, meetingspace, food and beverage, audio visual, and others.Despite the economic downturn, Marriott Interna-tional remains committed to these innovations, whichreflect the wise application of operations research tochallenging problems in the hospitality industry.

AppendixLet c denote the direct cost of providing this piece ofbusiness, which is usually not dependent on its saleprice, to the customer; let D be any associated oppor-tunity costs; and let � denote any sales or advertisingcosts as a fraction of the ultimate selling price. Then,

50

Exp

ecte

d pr

ofit

($)

Pro

fit fu

nctio

n ($

)P

roba

bilit

y to

win

0

10

20

30

40

50

60

70

100 150 200

Optimal price

250 300

50

1.0

0.8

0.6

0.4

0.2

0100 150 200 250 300

50

350

300

250

200

150

100

50

0100 150 200 250 300

Price ($)

Price ($)

Price ($)

Figure A.1: This graph illustrates the pricing model: the optimal price isdetermined by combining the profit function and the probability to win.

the contribution to profit, as a function of price x, isgiven by

Profit�x� = �1− ��x − c − D� x > 0� (1)

Let ��win at price x� = p�x�, x > 0 denote the proba-bility to win that customer’s business at price x, usu-ally referred to as the pricing curve; then the expectedprofit is given by

Expected_Profit�x�

= ��win at pricex� ×Profit�x�

= p�x��1− ��x − c − D�� x > 0� (2)

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


p�x� is typically a decreasing function of x (i.e., thehigher the price, the lower the chance that the cus-tomer will accept it). For most commonly used func-tional forms of p�x� (i.e., linear, polynomial, logistic,power function), Expected_Profit�x� has a local max-imum at xo, which is a solution of the following 1-Doptimization problem:

maximize Expected_Profit�x��

subject to x > 0�(3)

The strategy of offering price xo for each piece ofbusiness will maximize the expected profit in the longterm (note that the optimal price could be differentfor each corresponding piece of business). Figure A.1illustrates this model.

AcknowledgmentsMany people were instrumental in the development of thesystem described in this paper. The efforts of Marriott’sRevenue Management Systems Development team wereexceptional. We also acknowledge our user base for enthu-siastically receiving this new system and the Global Rev-enue Management business team for its guidance duringthis project. We are grateful to our support organization forworking out postimplementation issues. The guidance andoversight provided by our management team, especiallyNell Williams, Padmanabh Yardi, and John Whitridge, wereinstrumental to our success. We would also like to acknowl-edge the operations research expertise of Bill McDaniel and

David Nehme. We are grateful for valuable comments fromJeff Camm, Howard Finkelberg, and René Carmona duringthe preparation of this manuscript.

ReferencesBelobaba, P. P. 1987. Air travel demand and airline seat inventory

management. Doctoral dissertation, Massachusetts Institute ofTechnology, Cambridge.

Bijmolt, T., H. Van Heerde, R. Pieters. 2005. New empirical gener-alizations on the determinants of price elasticity. J. MarketingRes. 42(2) 141–156.

Cross, R. G. 1997. Revenue Management: Hard-Core Tactics for MarketDomination. Broadway Books, New York.

Cross, R. G., J. A. Higbie, D. Q. Cross. 2009. Revenue management’srenaissance. A rebirth of the art and science of profitable rev-enue generation. Cornell Hospitality Quart. 50(1) 56–81.

Hill, T., P. Lewicki. 2005. Statistics: Methods and Applications. StatSoft,Tulsa, OK.

Ingold, A., I. Yeoman, U. McMahon. 2001. Yield Management: Strate-gies for the Service Industries, 2nd ed. Thomson Business Press,London.

Marriott, Jr., J. W., K. A. Brown. 1997. The Spirit to Serve Marriott’sWay. HarperCollins Publishers, New York.

Orkin, E. B. 1998. Wishful thinking and rocket science: The essentialmatter of calculating unconstrained demand for revenue man-agement. Cornell Hotel Restaurant Admin. Quart. 39(4) 15–19.

Overby, S. 2005. The price is always right. CIO Magazine(February 15) 40.

Phillips, R. L. 2005. Pricing and Revenue Optimization. Stanford Uni-versity Press, Stanford, CA.

Talluri, K. T., G. J. van Ryzin. 2004. The Theory and Practice of RevenueManagement. Springer, New York.

Zeni, R. H. 2001. Improved forecast accuracy in airline revenuemanagement by unconstraining demand estimates from cen-sored data. Doctoral dissertation, Rutgers, The State Universityof New Jersey, New Brunswick.

Dow

nloa

ded

from

info

rms.

org

by [

149.

164.

92.2

11]

on 2

5 M

arch

201

6, a

t 12:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

informs is located in maryland, usa publisher: … et al.: marriott international increases revenue...

Documents