Revenue Management Saves National Car Rental

M. K. GERAGHTY 2788 Deep Woods WayMarietta, Georgia 30062

ERNEST JOHNSON National Car Rental

7700 France AvenueBloomington, Minnesota 55435

In 1993, National Car Rental faced liquidation. General MotorsCorporation (National's parent) took a $744 million chargeagainst earnings related to its ownership of National Car RentalSystems. National faced liquidation, with the loss of 7,500 jobs,unless it could show a profit in the short term. National initi-ated a comprehensive revenue management program whosecore is a suite of analytic models developed to manage capacity,pricing, and reservation. As it improved management of thesefunctions. National dramatically increased its revenue. The ini-tial implementation in July 1993 produced immediate resultsand returned National Car Rental to profitability. In July 1994,National implemented a state-of-the-art revenue managementsystem, improving revenues by $56 million in the first year. InApril 1995, General Motors sold National Car Rental Systemsfor an estimated $1.2 billion.

In the late 1980s, the car rental industry were eroded. Automobile n:\anufacturers

was in turmoil. Low profit margins were purchased almost all of the major car rentalsubsidized by tax credits. When these tax companies and, in the early 1990s, floodedcredits disappeared, the low profit margins them with cheap fleet deals. These cars

Copyright £ 1997, Institute ior Operalionn Research INDUSTRIES—TRANSPORl ATiON/SHIPPINGand the Management Sciences DECISION ANALYSIS—APPLICATIONS0092-2102/97/2701 /0]07S05,00

INTERFACES 27: 1 January-February 1997 (pp. 107^127)

GERAGHTY, JOHNSON

came with large manufacturer cash incen-tives and could be disposed of quickly andeasily, as often as every four months, bysimply returning them to designated auc-tions. The car manufacturers placed moreemphasis on using their car rental subsid-iaries to soak up excess production than toproduce profits.

This excess supply in the marketplace ledto low pricing. Several major competitors,the price leaders, paid undue attention tomarket share and made emotion a variablein the pricing equation. Companies still usevery low pricing during periods of low de-mand. These are the rates quoted in Sabreon February 9,1996 for a weekend rental ofa subcompact car at Greensboro, NorthCarolina for pickup on February 29,1996:Alamo $14.99National $15.99Budget $16.95Avis $16.99

Hertz $16.99These prices include unlimited mileage.

A comparison of these rates with the costof renting a tuxedo underscores the fre-quent irrationahty of pricing. In the early1990s, economic conditions and improve-ments in design and production quality im-proved demand for American-made cars.The manufacturers dramatically raised thecosts of cars to their car rental companies.

These market pressures, combined withthe fact that the car rental industry wasslow to apply technology, precipitated anindustry in crisis. By comparison, the air-line industry has successfully demonstratedhow to apply the technology of revenuemanagement in a service industry withhigh equipment and labor costs. Airlinesregularly sell cheaper seats to customers

willing to accept booking and travel restric-tions, such as advance payment, Saturdaynight stopovers, and penalties for no-showsand cancellations.

The major car rental companies dependlargely on corporate customers. They con-tract at fixed rates with companies whohave numbers of employees who travel.Demand peaks for rental cars midweek,forcing all companies to regularly turndown customers. The business customer,who typically travels on these days, pays afixed corporate rate. This leads to a largeexcess fleet that is idle on weekends. Thecar rental industry allows price-sensitiveleisure customers to book multiple reserva-tions with no prepayment required. Thereare rarely penalties for cancellations or no-shows. Customers arriving as much as 12hours after the specified time of reservationare given the reserved car at the reservedrate. These policies result in no-shows thatsometimes exceed 50 percent of reserva-tions. This is a major problem for the in-dustry, which must maintain highutilization to make a profit.National Car Rental Background

Before National began using revenuemanagement, it struggled with the samechallenges as its competitors. But other fac-tors made it critical for National to changequickly. National's business was predomi-nantly composed of corporate customers,who rented cars midweek. National's strat-egy focused on these business renters andneglected the leisure customers. For severalyears, starting in 1987, National had no sig-nificant advertising campaign. It plannedits fleet in one-year cycles, and made veryfew changes in fleet deployment to meetchanging customer demand.

INTERFACES 27:1 108

NATIONAL CAR RENTAL

National had three legacy systems tobuild on: The newly-developed Vehicle In-formation System (VIS) accurately trackedthe fleet. National's Reservation System(RES) was efficient at booking reservations,and the Expressway System (NEX) pro-vided the most rapid rental-and-return pro-cess in the industry. By contrast, pricingchanges were manual and extremely timeconsuming. Changes were keyed into Na-tional's rates system and then keyed inagain to the airlines' computerized reserva-tion systems (CRS). Setting pricing was ashared responsibihty. City managers, mar-keting, regional VPs, senior management,and the pricing group all shared input,with no single person ultimately responsi-ble for a location's pricing. Inventories inthe CRS were controlled by field managerswith no sophisticated system advisingthem when to increase or restrict availabil-ity. No demand forecasts existed at eitherthe city or the corporate level.Revenue Management at National CarRental

In February 1992, several Nafional execu-tives identified two key issues: (1) Nationalwas turning down large numbers of cus-tomers when cars were available to meettheir needs; (2) competitors were raisingtheir leisure prices as the date of rental ap-proached, while National's pricing re-mained stagnant. National formed the ratesautomafion team (RAT) with a limited mis-sion to determine whether, during periodsof high demand, Nafional could raise itsprices in the seven booking days beforerental. It selected a limited number of citiesfor a pilot test, for which all processes weremanual. It quickly determined that it couldraise prices and increase revenues without

eroding customer satisfacfion. Roll-out ofthese processes in the organizafion was metwith sfiff resistance. "We can't do this;we'll confuse our customers," and "If wedo this, we'll confuse ourselves" were typi-cal comments. Deciding that it needed helpbreaking through these barriers, Nafionalasked Aeronomics Incorporated, a revenuemanagement consulfing company, to evalu-ate its unrealized revenue opportunifies.

National was at a critical juncture. Gen-eral Motors had mandated that the com-pany either become immediately profitable,so that it could be sold, or be liquidated.National had already undertaken cost-cut-ting measures. It had to make more moneywith the existing operafion. Larry Ramaekers,assigned by turnaround specialists Jay Alixand Associates, led the turnaround andacted as president. "We decided to go for arevenue-based turnaround as opposed to acost-cutting turnaround" said Ramaekers11995].

Senior management agreed to conduct aneeds assessment with Aeronomics Incor-porated between January and April 1993.The mission was to understand National'sbusiness, quanfify revenue potential, rec-ommend organizafional structure and staff-ing requirements, define automation re-quirements, estimate costs, provide cost/benefit analysis, and priorifize an imple-mentafion plan. The assessment idenfifiedopportunities for increasing revenue andwas presented to senior management inApril 1993. National's owner. General Mo-tors, agreed that implementing a revenuemanagement program would be the keyimpetus to National's turnaround and com-mitted over $10 million to design and builda revenue management system (RMS), ac-

January-February 1997 109

GERAGHTY, JOHNSON

quire the necessary hardware, makechanges to legacy systems and bridge themto RMS, and build a dedicated revenuemanagement department. The departmentwould comprise 30 specialists, focusing allof their talents and energies on generatingrevenue. The application would be phasedin beginning July 1993 and would be rolledout to all locations by early 1995. The ex-tremely rapid development of RMS and itsimmediate implementation to control Na-tional's iargest demand centers was the sin-gle most important factor in keeping Na-tional alive.

Revenue management achieves its reve-nue gains by applying analytic models andmethodologies to a planning horizon. Byconsistently managing capacity, price, andbooking requests in a manner that im-proves revenue per car (RPC), revenue perday (RPD), and utilization levels, a com-pany can make and sustain revenue im-provements. There are a couple of basicprerequisites for applying revenue manage-ment in a new industry. Perishability is oneof the most important prerequisites asWeatherford and Bodily 11992] discuss intheir paper on perishable-asset revenuemanagement. The unit of inventory at Na-tional is the car rental day, which is lost ifit is not utilized. Another prerequisite is asegmentable market. (For a discussion ofhow revenue management capitalizes onconsumers' differential willingness to pay,see Cross [1986].)

The rental car problem exhibits a similarstructure to the airline problem, and tech-niques developed to solve the airline prob-lem have been particularly useful. For ex-ample, allocating airline seats at discountprices translates to planning to upgrade a

customer's rental car. Overbooking and res-ervation control are common to any indus-try that allows advanced bookings. Thereare important differences. Airlines can tar-get low rates precisely at underutilized ca-pacity. In the car rental situation, the prob-lem is more complex, Rate cuts designed tostimulate demand on low utihzation daysmay increase demand on a day when ca-pacity is constrained and compound theproblem. Managing the problem of dayswhen supply is constrained by controllingthe length of rentals is a more effective so-lution, but it requires surgical precision.Conversely, it is reasonably straightforwardto increase RPD by increasing rental rates.However, the rental car market is ex-tremely competitive. A price move thatmakes the company more expensive thanits competitors can damage utilization lev-els. A high RPD is not worth much if mostof the fleet is sitting on the lot.Information Systems

The revenue management system wasdeveloped jointly by Aeronomics Incorpo-rated, a revenue management firm; EDS,National's information services provider;and National. It is central to the flow of in-formation at National Car Rental. EDS im-plen:\ented a comprehensive set of datalinks with existing information manage-ment systems. In particular, the link be-tween RES and RMS is unparalleled in theindustry. It is a continuous transaction-level data feed of all advanced booking ac-tivity, including availability and bookingrestrictions. The continuous feed approachprovides up-to-the-minute booking levels,forecasts, and system recommendations torevenue managers through the RMS graph-ical user interface. Transactions include

INTERFACES 27:1 110

NATIONAL CAR RENTAL

bookings, cancellations, turndowns, andshoppers. Turndowns are booking requeststhat the company did not accept because ofavailability controls or booking restrictions.Shoppers are booking inquiries that did notconvert to booking requests (Figure 1).Revenue Management Organization

A major issue during the design phase ofRMS was to determine whether a decen-tralized or centralized solution would havethe largest revenue impact, short and longterm. A decentralized organization wouldhave been the least painful solution cultur-ally, because city operations managers con-trolled inventories (that is, reservations sys-tem inventories) and leisure pricing. But

decentralization would cause a number ofdifficulties:—City managers would not make revenuegeneration the highest priority, becausetheir most immediate problems are cus-tomer service and vehicle maintenance.—Recruiting and training personnel andequipping city offices would be very ex-pensive with a long lead time.—The revenue manager would be a gener-ahst and would be assigned to "burningproblems" not related to revenue genera-tion.—Pricing practices would not be consistentacross locations.—Managers might be parochial concerning

Inventory Control

ReservationsActivity

ExpresswayFleetPost Arrivals

Figure 1: RMS synthesizes information from four principal information systems. Expressway pro-vides current 0eet levels to support the capacity-management model and post-arrival data, such asno-shows and walk-ups, for the forecast of day-of-arrival activity. RES provides transaction-levelinformation on booking activity. RMS availability and length-of-rent controls are communicatedto RES after review and action by a revenue manager. The Rates system maintains current rate-level information. RMS recommends rate adjustments and provides an interactive rate updateinterface to the Rates system. Availability and rates are available on a number of airline CRS(central reservation systems). RES updates the CRS whenever availability controls or bookingrestrictions change. Rates from RMS update CRS rates at regular intervals throughout the day.

January-February 1997 111

GERAGHTY, JOHNSON

fleet disbursement.—Revenue managers would vary in levelsof skills in the field.

The decision to create a centrally locatedteam addressed all of these issues and al-lowed for the rapid change needed for im-mediate impact. Corporate revenue manag-ers were given responsibility for pricingand inventory management at the locationlevel. This allowed the revenue manage-ment department to share its expertise andinformation directly with marketing, corpo-rate sales, fleet, strategic analysis, and sen-ior management.

Traditional revenue management organi-zations in the airline, hotel, and car rentalindustries have left a dichotomy betweenthe inventory and pricing functions. This isa result of the problem structure. Inventoryis usually controlled at the level of a non-stop flight leg, a room night, or a rentalday. Pricing is focused at market segmentsthat are often multi-leg, multi-night, ormulti-day. However, pricing decisionsmade independently of fleet availabilityand customer demand retard revenue gen-eration. The decision to vest National's rev-enue managers with control of both pricingand inventories represented a big change inculture for National and pioneered newterritory for a revenue managementorganization.Revenue Management System

The National RMS supports three pri-mary business functions: capacity manage-ment, pricing, and reservations control. Thecapacity management function targetshigh-valued fleet utilization. Pricing en-hances corporate revenues through sensi-tivity to consumer price tolerance.Reservations control maximizes revenues

by accepting or rejecting booking requestsbased on length-of-rent controls. A sophis-ticated set of forecasts of demand and con-sumption patterns supports the analyticmodel (Figure 2).

RMS functions as both an automated de-cision management system and an interac-tive decision support tool. Overnight pro-cesses execute the forecasting and analyticmodels to generate recommendations con-cerning availability, rate, and length-of-rentcontrol. Revenue managers reviewthousands of recommendations each dayand make, accept, reject, or override deci-sions based on their knowledge of currentmarket conditions and forecasted demand.Revenue priority indicators assist work-flow management by pointing out thegreatest revenue opportunity.

On-reni Fcsi

LOR Fcst

Capacity Mgl

Res Control

Pricinj;

Figure 2: Historic and cuiTent demand from themarketplace is used to produce the demandforecasts that support RMS's analytic models.The capacity-management process determinesavailability levels from the on-rent forecast.Pricing recommendations complement theavailability settings. The reservations controlprocess uses more detailed length-of-rent(LOR) forecasts to generate booking restric-tions. The combination of availability controls,price adjustments, and booking restrictionschanges booking pace. The new booking paceresults in new forecasts and systemrecommendations.

INTERFACES 27:1 112

NATIONAL CAR RENTAL

Revenue managers have a wealth of sup-port tools at their fingertips during the re-view process. The multiple document inter-face allows them access to reports withhistorical data and demand and revenueforecasts at all levels of aggregation. Theycan cut and paste information into spread-sheets, word processors, or electronic fax.One of the more striking features of RMS isthe degree of interaction it permits be-tween the revenue manager and the sys-tem models. A refresh feature accesses themost recent data on the reservations sys-tem and recomputes forecasts based onintraday booking-pace profiles. Recom-puted forecasts generate new system rec-ommendations, which the revenue man-ager can compare to the overnightrecommendations as part of the reviewprocess. An analyst who disagrees withthe forecast can override values. The re-fresh feature will revise the RMS recom-mendations according to the user-sup-plied forecasts.

The logical flow of the models withinRMS matches the logical flow of marketingactivities within the company. Results fromeach model depend on the output of logi-cally precedent models. Users can adjustoutput of any of the models for what-ifanalysis. All downline models will thenproduce new recommendations. For exam-ple, the analyst can experiment with differ-ent rate levels. New length-of-rent controlswill reflect the change in the relative valua-tion of different demand elements. If theanalyst wants to try different availabilitylevels, the pricing model will generate newrates. New rates will cause the system torecommend new length-of-rent controls(Figure 3).

ForecastingA comprehensive set of demand and rev-

enue forecasts supports the analytic mod-els. Demand levels are forecasted at twoprimary levels of aggregation: length-of-rent and on-rent. Both forecasts representunconstrained demand, which is the num-ber of cars that can be rented if there are nocapacity restrictions. The length-of-rentforecast is a forecast of unconstrained de-mand for each potential length of rent, foreach arrival day in the planning horizon.National intensively manages the bookingprocess at least 60 days in advance of dayof pickup. The system generates forecastsfor all days within this horizon. It uses de-mand levels to derive optimal reservationscontrols. The term on-rent refers to thenumber of cars in use on a specific date. Itcombines cars that are picked up on thatday with cars that are already in use. Ca-pacity management and pricing modelsrely on the on-rent forecast.

The demand forecasting methodology forall levels of aggregation is based on a com-bination of long-term and short-term fore-casting. The long-term forecast is a time-series model with seasonality factorsderived from spectral analysis of historicseasonality. Demand during special eventsdoes not distort seasonality. The analystwith the best knowledge of the market de-fines special events and can override thesystem-generated factors (Figure 4).

Curry [1993] pioneered the use of Kalmanfiltering for revenue management. The vari-able gain approach has several benefits. Sys-tem initialization uses the same processes asthe nightly update. The initial gain is set touiut value and the initialization process ad-justs the gain level as it works its way

January-Eebruary 1997 113

GERAGHTY, JOHNSON

InventoryOptinuzation

iRES

Figure 3: Business processes, analytic models, and revenue manager expertise are tightly inte-grated in RMS. Each of the ovals represents a point where analysts can adjust system forecastsand recommendations. If the analyst requests a refresh, RMS updates down-line results to reflectthe user adjustments. Underlying the decision-making process are expectations of consumer be-havior. RMS provides demand forecasts at the on-rent and length-of-rent levels. The analyst canrefresh the forecast with real-time booking levels from the reservations system. The capacitymanagement models produce availability recommendations, which can also be adjusted. A re-fresh at this stage produces new rates and length-of-rent controls. Finally, rate revisions can resultin new MLRs (minimum length-of-rent controls). The analyst reviews system recommendationsbefore sending them to the electronic distribution channels.

through the available historical data. Another

benefit of the variable gain approach is that

the revenue manager can adjust the respon-

siveness of the forecast temporarily to re-

spond to expected changes in the market-

place. Once the change has taken place, the

Kalman filter gains tend toward the steady

state smoothing constant.

In generating the short-term forecast, the

system considers the offset of the actual

reservations level from a booking curve.

The booking curve represents the rate at

which bookings are expected to accumulate

in the reservation system. This offset pro-

vides information about the current book-

ing pace, that is, the rate at which bookings

are accumulating. The short-term forecast is

the expected change in the number of

bookings. The Hnal demand forecast is a

combination of the long-term and short-

term forecasts. The long-term forecast pro-

vides stable predictions early in the plan-

ning horizon. As more information

becomes available from the actual booking

INTERFACES 27:1 114

NATIONAL CAR RENTAL

8 10 12 14 ^ 16 18 20 22 24Time

Figure 4: One attractive feature of variable gains is proactive response. The analyst often has ad-vance knowledge that a significant change in the market is about to occur. By proactively increas-ing the gain, the analyst can ensure that the demand forecast responds to the new demand levelquickly.

behavior, the short-term model dominatesthe forecast of final unconstrained demand.

The system also generates additionalforecasts of day-zero activity, such as walk-ups and no-shows. Day-zero is the day acar is picked up by the customer. Day-zeroactivity includes reservations that do notmaterialize, that is, no-shows, and requestsfor cars that do not come through the reser-vations process, that is, walk-ups. No-shows are a big management problem. Theno-show rate is a day-zero effect, but itaffects down-line rental days by reducingexpected on-rent demand and creating ex-tra availability.

Walk-ups, on the other hand, introducesignificant opportunity. A manager canavert an impending oversale situation by

turning away walk-up demand. Walk-upsrepresent a significant revenue opportu-nity during periods of high demand andlow product availability. By predictingwalk-up activity, managers can set asideinventory. Walk-ups during these periodsrepresent an opportunity to achieve highrevenue per day. Conversely, managerscan stimulate walk-ups through aggres-sive pricing to compensate for underutili-zation identified late in the booking pro-cess. The management of day-zeroactivity is the responsibility of the fieldmanagers. Hand-off of the booking pro-cess from corporate revenue managers tofield staff on day-zero requires constantcommunication and information sharing.RMS contributes to this process with

January-February 1997 115

GERAGHTY, JOHNSON

distributed electronic reports.Capacity Management

RMS converts available capacity into rev-enue. Capacity management is the first stepin the process. It includes fleet planning,planned upgrades, and overhooking.Fleet Planning

To plan fleet levels. National identifieshow much of the available fleet should beat each inventory location to meet expecteddemand. An inventory location is a geo-graphical area which shares a pooled fleet.It may include several different physical lo-cations. For example, Minneapolis may bean inventory location consisting of two citystations—Minneapolis airport and a hotelstation. Capacity is managed at the inven-tory location level because fleet can easilybe moved to meet demand at the variousstations. In contrast, pricing is managed atthe physical location level as a means offurther segmenting the market. National re-lies on long-term forecasts of demand foreach inventory location in its fleet plan-ning. It matches current fleet and expectedfleet adjustments to the demand forecaststo determine the placement of cars.

National plans its fleet in three stages: forthe short term, it looks at a five-day hori-zon; for the medium term, it considers 60days in the future; and for the long term, itlooks over the coming 18 months. Short-term planning targets fleet movements be-tween locations, accelerates or retards turn-backs to manufacturers, regulates car-salesactivities, and resets one-way pricing to ef-fectively place cars in the proper locationsto satisfy demand. In mid-term fleet plan-ning, managers consider the same actionsas in the short-term process, but they alsoredirect new car deliveries, acquire new

fleet made available at the last minute bythe manufacturer, and move cars by truckor train for longer distances. RMS providestactical forecasts, notifies the analyst of pe-riods when capacity is dangerously lowand automates key components of the tacti-cal fleet-management process.

The involvement of the revenue manage-ment department in long-term fleet plan-ning has been instrumental in developingNational's annual budget for charge daysand revenue per day for each market seg-ment within each city. (A charge day is arental day that genuinely generates reve-nue.) This has contributed dramatically toNational's success because the RMS fore-cast focuses on unconstrained demand andhas helped break the pre-RMS fleet-plan-ning paradigm. In the past. National wouldplan its long-term fleets based on historicalrental patterns, which restricted growth.With accurate information illustrating un-constrained demand. National was able toincrease its fleet in a cost-effective mannerto capture a much larger volume of profita-ble business with resulting increases inrevenue and market share.Planned Upgrades

Revenue management exploits the rela-tionship between segmenting the marketand generating revenue. Firms can achievedifferential pricing for commodities by in-stituting fences, such as advanced purchaserestrictions, that capitalize on each marketsegment's willingness to pay. In this way afirm presents a variety of products, all ofwhich are based on a single commodity.The fences discourage revenue dilution be-cause the lower-valued products have re-strictions that are unacceptable to thehigher-valued market segments. National

INTERFACES 27:1 116

NATIONAL CAR RENTAL

$100 Unrealized Revenue Potentialfrom Diluted Demand

Price

$50

50

Unrealized Revenue Potentialfrom Unaccommodated Demand

100

Demand

Unrealized Revenue Potentialfrom Diluted Demand

Price

40 60

Demand

80

Unrealized Revenue Potentialfrom Unaccommodated Demand

100

Figure 5: If fences can be found that effectively segment the market, the firm can take advantageof revenue opportunities that arises from differential pricing. It can maximize revenues by settingprices at Ihe willingness-to-pay level of each market segment. The business benefits from the in-creased revenues, and the consumer benefits because the commodity is available lo a broadermarket. Revenue management has been an important driver in Ihe expansion of National'scustomer base.

January-February 1997 117

GERAGHTY, JOHNSON

Luxury

Midsize

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Economy

Figure 6: The planned-upgrades model uses a modification of the classical EMSR (expected mar-ginal seat revenue) heuristic that has evolved from Littlewood's [1972] and Belobaba's [1989] re-search. The cumulative demand distribution is multiplied by average revenue to get expected-marginal-revenue curves for each class. Inventory is protected for each class until its marginalrevenue falls below the revenue available from lower classes. Bookings are accepted for a classuntil they exhaust all inventory not protected for higher classes.

has instituted competitively priced weeklyrates, which require the customer to keepthe car over Saturday, for price-sensitiveleisure customers, fencing out the businesstraveler who wants to be home on theweekends (Figure 5).

Market segmentation is also inherent inthe different car types. The demand for cartypes falls into definite market segments.Business renters typically demand midsizecars. Low-valued leisure customers prefereconomy cars with low rates. More valu-able leisure customers want specialty vehi-cles, such as minivans or four-wheel-drivevehicles. The drawback of accomplishingmarket segmentation with different prod-ucts is that a car rental company cannotdirectly substitute one vehicle type for an-

other. There is a definite upgrade hierarchythat it needs to manage. The planned-up-grades model fits demand into this hierar-chy in a way that minimizes revenue dilu-tion while maximizing utilization. Itallocates inventory to booking classes basedon demand for the booking classes, thenumber of vehicles that are acceptable torenters in this class, and the expected reve-nue associated with each class (Figure 6).

In a fully commoditized environment,where all inventory is interchangeable,firms usually adopt a nesting approach.Nesting allows the most valuable bookingclass access to all available inventory. Sub-sequent booking classes are allowed accessto inventory nominally set aside for thenext lower classes. The planned-upgrades

INTERFACES 27:1 118

NATIONAL CAR RENTAL

hierarchy interferes with a simple nestingapproach. It is inappropriate to put higherclass customers in economy cars. However,if the company restricts economy bookingsto economy cars only, it would miss a largerevenue opportunity. Typical fleeting pol-icy is to acquire more large and midsizecars than expected demand requires andfewer economy level cars. The difference inthe costs of different sized cars is smallenough that the advantages derived fromhaving enough high-valued inventory tomeet the high-valued, high-demand peri-ods justify the extra cost. Also, few econ-omy customers complain about getting abetter car than they booked for the samerate. The planned-upgrade model decideshow many high-valued vehicles to makeavailable to lower booking classes (Table 1).

Fleet in the higher classes that is not re-quired for that class's demand enters theavailable pool. Classes with more demandthan fleet extract fleet from the availablepool of higher classes to cover their excessdemand requirements. The result is a set ofallocations of the number of cars availablefor rent in each booking class. The actual

planned-upgrades model takes account offorecast variability and expected revenuesto compute the marginal revenue for eachclass (appendix). The availability calcula-tion is based on the relative marginal reve-nues.

Planned-upgrades activity produces achange in availability in the reservationssystem because customers are driving carsin a class different from that which theybooked. When an economy customer ar-rives at the rental counter and drives awayin a midsize car, the availability for midsizedemand is decreased for each day the car ison rent. This is because the rental contrib-utes to the current on-rent value for mid-size cars even though it is an economyrental. This planned-upgrade effect is par-ticularly pronounced on booking days closeto the day of pickup. The model handlesthe change in availabihty by revising theon-rent allocation to allow for the numberof bookings from other classes it expects toimpact the current class.Overbooking

The airlines have instituted a creative so-lution to the problem of oversales, which

Class

Fullsize 4 doorFullsize 2 doorMidsizeEconomySubcompact

Fleet

23050

40010020

Forecast onRent Demand

16560

37040

165

Available Pool

65

3060

Excess Demand

10

145

Available to Sell

16560

37040

165

Table 1: This table illustrates a simple deterministic case of the planned upgrades process.Available pool represents cars that are available to lower-class bookings. It is computed by amarginal revenue heuristic. Excess Demand represents the fleet requirement for a class withinsufficient inventory to meet demand. The revenue manager influences these values throughsystem parameters that provide a throttle to the flow of cars between classes. Available to Sell isthe number of bookings National is willing to accept in each car class. The system generatesAvailable to Sell by borrowing cars from the next higher classes so that it satisfies demand.

January-February 1997 119

GERAGHTY, JOHNSON

occur when more passengers show up thancan be accommodated. When a flight isoverbooked, the airlines in effect hold anauction, offering rewards to any passengerswilling to give up their seats and take alater flight. Unfortunately, this is not possi-ble in the car-rental business: customersflow to our counters at different times, ex-pecting and receiving immediate service.They would find it intolerable to be herdedinto an enclosed area, forced to wait until acertain time, and then to take part in anauction to determine those willing to waitfor a vehicle until a later time. Thus, it iscritical that our forecast and planning beextremely accurate during peak periods(Figure 7).

The overbooking model revises the re-sults of the capacity management processto account for the impact of no-shows andcancellations. It may be regarded as a map-ping from demand space to reservationsspace where demand space represents ac-tual materialized demand on the day ofpickup and reservations space is the num-ber of bookings required, at a given num-ber of days prior to arrival, to achieve thatdemand level. The overbooking processproduces an adjusted, sometimes calledoverbooked, availability allocation for eachcar class. The overbooking model identifiesoptimal overbooking levels by balancingthe expected cost of an oversale against theopportunity cost of an unrented car, subjectto service-level constraints (appendix). Na-tional identifies acceptable oversale risklevels at the corporate level. Individual lo-cations set target utilization levels that ad-just available capacity to compensate forthe resulting oversale rate and for such op-erational issues as car turnaround time and

III1992 1993 1994 1995

Figure 7: Overbooking compensates for cancel-lations and no-shows, which are part of the ad-vance booking process. More aggressive over-booking policy results in fewer cars sitting onthe lot but increases the instances of "Reserva-tions-No Car." National's record of manage-ment of Res/No Car incidents led the industryin 1992. Unfortunately, tumdowns were highand utilization was not optimal. In 1993, withGM's direction to tum the company around,utilization improved, revenues improved, butwithout the tools provided by RMS, Res/NoCar incidents rose. When the revenue manag-ers and field managers had the use of RMS (inlimited locations) in luly 1993, National wasnol only able to continue to reduce tumdownsand improve utilization, but also to reduce ser-vice problems. The graph illustrates the greatadvances made in 1994 and 1995, after the fullcorporate-wide implementation of the systemand department.

discrepancies in checkout/return time of

day. (For a discussion of the benefit

of overbooking see Smith, Leimkuhler,

and Darrow [1992].)

Pricing

The car-rental customer population is

made up principally of two segments: cor-

porate and leisure. A corporate customer

generally books close to the date of rental,

is inflexible in rental and return dates and

times, does not shop competitors exten-

sively, is unvdlling to prepay, is not willing

to stay over a Saturday night, and most im-

portant, expects to pay an authorized fixed

rate, negotiated and reimbursed by his or

INTERFACES 27:1 120

NATIONAL CAR RENTAL

her company. The leisure customer, on theother hand, is often willing to stay over aSaturday night, books in advance, is some-what flexible on time of pickup, for exam-ple, willing to wait until noon on Thurs-days to qualify for cheaper weekend rates,is willing to prepay rental charges andshops competitors extensively looking forthe best value.

The National pricing model links ratelevels with availability and consumer book-ing activity to achieve revenue and utiliza-tion objectives. Our initial analysis of Na-tional's rate behavior indicated thatcompetitive positioning was the determin-ing factor in its pricing decisions prior tothe revenue management program. Attimes of low demand, sensitivity to com-petitor behavior is crucial. Utilization levelscan suffer drastically from poor rate posi-tioning in the marketplace. During high-demand periods however, the firm can loselarge revenue opportunifies by followingcompetition-based pricing rules that causethe firm to exhaust inventory by acceptinglow-valued bookings. RMS implements ademand-based pricing policy. We devel-oped a simple, but extremely effecfivepricing model to support this poUcy (ap-pendix).

The pricing model recommends in-creased or decreased rates based on on-rentdemand for a each arrival date. The ratesare designed to encourage maximum utili-zation of rentable capacity. Therefore if re-maining demand plus current on-rents ex-ceeds rentable capacity, the model willincrease the rate to extract high-valued con-sumers from the total demand. If remain-ing demand plus current on-rents is belowrentable capacity, it will reduce the rates to

sfimulate demand. An elasticity model re-lates historic rate and demand variability.The discrepancy between the demand fore-cast and the target utilization indicates therequired change in booking pace. The elas-ficity model provides a rate adjustment thatwill induce this change.

The model relies on the revenue analystto provide an appropriate base rate foreach individual product. The base-rate levelembodies the analyst's expertise about suchareas as the price tolerance of the marketsegment and the degree of competition atthe locafion. When the analyst disagreeswith a system-generated rate, he or she willeither reject it or override it in the processof reviewing the recommendations. Whenthis happens, the pricing model recalibrates

A comparison of these rateswith the cost of renting atuxedo underscores theirirrationality.

around a new base rate. The system infersthe new base rate by working backwardsfrom the actual rate the analyst sent to thereservafions system and the demand-based rate offset recommended by theelasticity model. The system maintains de-mand and availability information at ahigher level of aggregation than price.RMS-generated price recommendationscause groups of rates to move together.This simplifies rate management by main-taining consistent differentials. The analystadjusts rate differentials by overriding thecurrent system recommendaHon. The re-computed base rate will maintain the new

January-February 1997 121

GERAGHTY, JOHNSON

differential. The pricing model is morethan just a price-management system thatmimics analyst behavior. By making fre-quent adjustments in response to marketbehavior, it extracts the maximum revenuepotential from each market segment overthe course of the booking process. Gallegoand van Ryzin 11994] suggest that one pos-sible reason for revenue management'ssuccess is its ability to capitalize on statis-tical fluctuations.

Traditional revenue management modelscapitalize on a reduction in customers'price sensitivity later in the booking pro-cess, primarily by using restrictions on ad-vance purchases. This kind of market seg-mentation has been difficult to implementin the car rental industry, but rate premi-ums for late booking are an alternatemeans of capturing at least some of this

revenue. The pricing model supports rate-premium profiles that offset base-ratelevels by different amounts depending onthe number of days left before pickup (Fig-ure 8).

The National RMS bases its rate recom-mendations on forecasts of demand. Attimes when on-rent demand exceeds avail-ability, the pricing model extracts the mostvaluable customers from this mix by dis-couraging lower-valued demand. Whencompetitors undercut National's price attimes of high demand, they end up withthe low-end customers and leave thehigher-valued, later-booking customers forNational. RMS allows National to break the"competition paradigm" by recommendingwhen it can price above the competitionand when it should price extremelycompetitively.

$54

$45

$28

Average Daily Rate

0 15

Days before Pickup

30

Figure 8: Before RMS, NaHonal would set a base price, for example, $45 per day, and it wouldremain in effect regardless of demand or how far in advance customers would book. Price-sensitive leisure customers would search for a lower rate. The price-insensitive late-bookingcustomer would have been willing to pay more, causing revenues lo be diluted.

INTEREACES 27:1 122

NATIONAL CAR RENTAL

Reservations Inventory ControlThe capacity management and pricing

models work closely together to manageon-rent demand. The availability alloca-tions fronn the planned upgrades modelprovide a degree of reservations control butdo not provide adequate protection for con-strained days. The pricing model tries tocompensate by adjusting rate levels toachieve utilization targets. Both of thesemodels are limited because they function atan aggregate demand level. By controllinglength of rent, National can more preciselytrade off between demand elements thatare competing for inventory (Figure 9).

Demand forecasts for each length-of-rentcategory, revenue forecasts based on sys-tem rate recommendations, and remainingon-rent capacity provide input for a mathe-matical programming model that generatesminimum length-of-rent restrictions foreach arrival day (appendix). The first phaseof the model solves a deterministic linear

program to identify the length-of-rent cate-gories on each arrival day that provide thegreatest revenue. (Williamson 11988] dis-cusses optimization for reservation control.)To implement the LP recommendations.National would need a reservations systemwith the ability to switch availability onand off for each possible length of rent oneach arrival day. The existing reservationscontrols at National allow specification ofminimum length-of-rent controls for eacharrival day. Therefore, the second phase ofthe reservation-control model degrades thefull-pattern solution to a set of minimumlength-of-rent values and constrained arri-val-day indicators. The degradationalgorithm reconciles conflicting open/closerecommendations within the full-patterncontrol by weighting each recommendationby the amount of demand it impacts andits proximity to a future constrained day.

For example, the reservation system pro-cesses a booking request for a four-day

Available

On Rent

Figure 9: National fits demand for each length-of-renl into remaining capacity using a determinis-tic linear program. The result is a set of allocation recommendations for each demand element.The linear program's allocations are used to determine minimum length-of-rent (MLRs) controlsthat can be implemented on the reservations system. The MLRs protect arrival days with con-strained capacity while building up utilization on shoulder days (arrival days with availabilityadjacent to constrained days).

January-February 1997 123

GERAGHTY, JOHNSON

rental arriving on Sunday. It tests availabil-ity for all days the booking request re-quires. Once this test is passed, the length-of-rent specified for the arrival day isconsidered. The reservation system has afive-day minimum length-of-rent specifiedfor Sunday rentals, but it does not automat-ically reject the booking. It needs to check ifthe booking request demands inventoryfrom a constrained day. The reservationsystem finds that Wednesday is a con-strained day so it rejects the booking. Nextthe reservation system processes a two-daybooking request with arrival on Sunday.Because this does not impact Wednesday'sinventory, the reservation system acceptsthe booking. The constrained-day indicatorfunctions as a maximum length-of-rentcontrol. A booking must have a length-of-rent that fits between the arrival date andthe next constrained-day indicator, or elsehave a length-of-rent greater than the speci-fied minimum. In this way, the reservationsystem protects constrained inventorywhile building up utilization on the shoul-der days (days with inventory either side ofa constrained day).

The National RMS sets the standard forrevenue management in the car rental in-dustry. It is the first implementation of in-tegrated capacity management, pricing, andlength-of-rent control. The length-of-rentoptimization, which uses revenue forecastsand detailed length-of-rent forecasts by ar-rival date, is also unique.Impact of Revenue Management

When it comes to evaluating revenuemanagement performance. National CarRental has an advantage over many of itscompetitors. Traditional revenue opportu-nity models use demand untruncation

strategies to estimate lost demand. Thereare two principal types of lost demand: .tumdowns are reservations requests thatwere rejected due to revenue managementcontrols, and shoppers are customer inquir-ies that do not result in reservations re-quests. Tumdowns are key indicators ofthe effectiveness of capacity managementand reservations control. Shoppers provideuseful insight into the effectiveness of ratelevels. National's reservation system tracksturndowns and shoppers at the transactionlevel. By combining the reservations thatactually occurred with appropriate turn-down and shopper transactions, it can com-pile an accurate reservations history. Anal-ysis of the reservations history with perfecthindsight provides an estimate of the totalrevenue potential in the marketplace. Con-versely, National can estimate the revenuethat would be realized in the absence ofrevenue-management controls. We subtractthe revenue from the no-controls scenariofrom the actual revenue realized to get the

Demand peaks for rental carsmidweek.

revenue impact of revenue management atNational. National Car Rental also uses thisprocess to evaluate the impact of individualrevenue-management controls. For exam-ple, it assesses the impact of overbookingby comparing the results of the reservationprocess with the historic overbooking levelsto the results of that process with inventorylevels set at the fleet level.

The integration of the revenue manage-ment system and the revenue managementdepartment have catalyzed change in the or-

INTERFACES 27:1 124

NATIONAL CAR RENTAL

ganization. It has supported more flexiblefleet strategies and tactics. It has led in thedevelopment and implementation of newproducts (Saturday-night keep rates, pre-paid rates, guaranteed rates). It has helpedNational to give customers a higher level ofservice, especially our late-booking corpo-rate clients. It has upheld our abihty to servecustomers who suffer broken reservationsby our competitors. But most important, inits first year of coast-to-coast deployment,the operational revenue management de-partment allowed the creation and realiza-tion of $56 million in incremental revenue.APPENDIX: Length-of-Renl Control

Reservations are accepted or rejectedbased on length-of-rent controls. Optimallength-of-rent controls are determined by arevised simplex algorithm with upperbounds for the following formulation:

maximize

N L

1 = 1 ; = I

subject to

I S 6k,xt,i s C,

liik -\- j > i

0 otherwise

whereN = the number of arrival days in the

planning horizon,L = the number of length-of-rent catego-

ries,I = |i: 1, . . . , N) arrival days within the

planning horizon,k = [k: I, . . ., N\ arrival days that impact

day I availability,/ = {): 1, . . . , L) length-of-rent categories.

x,i = the decision variable for arrival day iand length-of-rent category /,

fj,,, = mean remaining demand for length-of-rent category j ,

C, = remaining capacity for arrival day i,and

r,, = expected revenue for length-of-rentcategory; on arrival day i.

The LP solution is degraded to maximumand minimum length-of-rent recommenda-tions by a voting scheme that weights theLOR recommendations for each arrival dayaccording to its proximity to a future con-strained day.Planned Upgrades

The planned upgrades algorithm is basedon a heuristic that was developed for allo-cating airline reservations inventory on asingle fiight leg (see Belobaba 119891). ThisEMSR (expected marginal seat revenue)heuristic computes protection levels foreach booking class. A protection level is thenumber of cars that should be reserved forthe demand in the current class. The opti-mality conditions for a constrained revenuemaximization problem are as follows:

dir,

whereR is the revenue function,k represents the expected marginal revenueof the last car allocated to each class, and7r, is the protection level for class /.

Once the EMSR heuristic determines pro-tection levels for each class, the planned-upgrades algorithm determines the avail-ability in each class. The availabilitynumber that appears on the reservationssystem is the sum of the availability for fu-ture bookings and the current bookings forthe car class. Availability for car classeswith excess demand is computed as thefleet in that class plus any cars availablefrom higher classes to cover the excess de-mand. The availability for car classes withexcess fieet is computed as the fleet less

January-February 1997 125

GERAGHTY, JOHNSON

any cars used by lower car classes. Fleetnot used by lower car classes is returned tothe availability of the original car class. Thefinal avail-to-sell number is the sum ofavailability, current bookings, and adjust-ments for planned upgrades that are ex-pected to pick up a car in the next 24hours.Pricing

The pricing model recommends ratechanges in response to the bookings foreach arrival date, relative to the expectedbooking pace. The magnitude of these rec-ommended changes depends on the re-sponsiveness of demand to rate changes, orthe demand elasticity. The basic assump-tion behind the model is that the historicalvariance in demand is correlated with theelasticity: the greater the variance, the moreresponsive demand is to price changes.

The rate changes are limited to a maxi-mum range; for example, between 80 per-cent and 120 percent of the PR, base price.The slope of the rate-response function de-pends on a, the variance in demand. Thisrelationship can be expressed as an inversedemand function of the following form:

AH

P = 1.2P« -la

where P and Q represent price and de-mand, respectively, and ^ represents themean demand level. This in tum implies ademand elasticity given by

where en represents the own-price elasticityof demand. It is clear that, holding otherfactors constant, an increase in a will in-crease the elasticity of demand and inducesmaller price adjustments for a given de-sired change in demand.

Expected booking pace is maintained at ahigher level of aggregation (inventory loca-tion) than prices, and there is an opera-tional requirement to maintain constant ab-

solute price relationships across programsand city stations. Therefore the percentageprice adjustments returned by this modelare converted to absolute dollar changesbased on a median rate for the inventorylocation.Overbooking

The overbooking model identifies opti-mal overbooking levels subject to servicelevel constraints. The optimal overbookinglevel is the- point at which the marginal over-sale cost is balanced against the marginalincrease in revenue due to overbooking.A = authorization, that is, maximum ac-ceptable on-rent bookings,C = capacity,S = number of on-rent cars,p(S\A) = probability density function ofthe on-rent demand for a given authoriza-tion level,U = number of empty cars,O = number of oversales,OS.^Cost =-- cost of an oversale, andSpoilage^Cost = opportunity cost of anempty car.

The expected number of empty cars isgiven by

E(U\A)= f (C ~ S)piS\A)dS.

The expected number of oversales is givenby

E(O\A) = f iS - C)p(S\A)dS.

The optimum expected revenue for authori-zation level occurs at the minimum value of

Spoilage Cost*EiU\ A) -\- OS-Cost*E(0\A).

Since expected oversales increase and ex-pected unused cars decrease with respect toauthorization, we can find a global mini-mum by increasing authorization from ca-pacity until the value of this equation startsincreasing. The authorization is constrainedabove by the following service-level re-

INTERFACES 27:1 126

NATIONAL CAR RENTAL

quirement, which is set at the corporate Flight Transportation Laboratory, MIT,level: Cambridge, Massachusetts.

1 - I p(S\A)dS ^ Maximum probability

of one or more oversales.

ReferencesBelobaba, Peter P. 1989, "Application of a proba-

bilistic decision model to airline seat inven-tory control," Operations Research, Vol. 37, No.2 (March-April), pp. 183-196.

Cross, Robert C. 1986, "Strategic selling: Yieldmanagement techniques to enhance revenue,"presentation to the Shearson Lehman Broth-ers, Inc. 1986 Airline Industry Seminar, KeyLargo, Florida, February 14.

Curry, Renwick E. 1993, "Kalman filtering andexponential smoothing," presentation to Air-line Group International Federation Opera-tions Research Societies Reservations andYield Management Study Group, May, Syd-ney, Australia.

Curry, Renwick E. 1990, "Optimal airline seat al-location with fare classes nested by originsand destinations," Transportation Science, Vol.24, No. 3 (August), pp. 193-204.

Gallego, Guillermo and van Ryzin, Garret 1994,"Optimal dynamic pricing of inventories withstochastic demand over finite horizons," Man-agement Science, Vol. 40, No. 8 (August), pp.999-1020.

Littlewood, Kenneth 1972, "Forecasting and con-trol of passenger bookings," Airline Group In-ternational Federation Operations ResearchSocieties Symposium Proc. 12, pp. 95-117.

Ramaekers, Lawrence 1995, "National CarRental Systems, Inc.," SCORECARD™, TheRevenue Management Quarteriy, first quarter,pp. 2-3.

Smith, Barry C; Leimkuhler, John F.; andDarrow, Ross M. 1992, "Yield management atAmerican Airlines," Interfaces, Vol. 22, No. 1(January-February), pp. 8-31.

Weatherford, Lawrence R. and Bodily, S. E.1992, "A taxonomy and research overview ofperishable-asset revenue management: Yieldmanagement, overbooking and pricing," Oper-ations Research, Vol. 40, pp. 831-844.

Williamson, Elizabeth L. 1988, "Comparison ofoptimization techniques for origin-destinationseat inventory control," Report FrL-R88-2,

January-February 1997 127