Vol. 42, No. 1, JanuaryFebruary 2012, pp.

4557
ISSN 0092-2102 (print) ISSN 1526-551X (online)
http://dx.doi.org/10.1287/inte.1110.0620
2012 INFORMS
THE FRANZ EDELMAN AWARD
Achievement in Operations Research
Retail Price Optimization at InterContinental
Hotels Group
Dev Koushik
Intercontinental Hotels Group, Atlanta, Georgia 30346, dev.koushik@ihg.com
Jon A. Higbie
Revenue Analytics, Atlanta, Georgia 30339, jhigbie@revenueanalytics.com
Craig Eister
Intercontinental Hotels Group, Atlanta, Georgia 30346, craig.eister@ihg.com
PERFORM
SM
with price optimization is the rst large-scale enterprise implementation of price optimization in
the hospitality industry. The price optimization module determines optimal room rates based on occupancy,
price elasticity, and competitive prices. The approach used is a major advancement over existing revenue man-
agement systems, which assume that demands by rate segments are independent of price and of each other.
As of this writing, over 2,000 InterContinental Hotels Group (IHG) hotels use the price optimization module;
all IHG properties will eventually use it. To date, price optimization has achieved $145 million in incremental
revenue for IHG. At full rollout, we anticipate that this capability will generate approximately $400 million
per year.
Key words: hotel pricing; price optimization; revenue management; price elasticity; competitor rates.
I
nterContinental Hotels Group (IHG) is the worlds
largest hotel group based on number of rooms.
Through its various subsidiaries, IHG owns, man-
ages, leases, or franchises over 4,500 hotels and more
than 650,000 guest rooms in nearly 100 countries
and territories worldwide. It owns a portfolio of
well-recognized and respected hotel brands, including
InterContinental Hotels, Hotel Indigo, Crowne Plaza
Hotels and Resorts, Holiday Inn Hotels and Resorts,
Holiday Inn Express, Staybridge Suites, and Can-
dlewood Suites. It also manages the worlds largest
hotel loyalty program, Priority Club Rewards, which
has 52 million members worldwide. Approximately
85 percent of IHGs hotels are franchised, 14 percent
are managed, and 1 percent are owned.
Each hotel, including corporate-owned and man-
aged properties, is responsible for its own prot and
loss, essentially operating as an independent busi-
ness. Each hotels revenue manager is responsible
for optimizing that hotels revenue performance by
undertaking key revenue strategies with respect to
pricing and inventory management. They include
demand forecasting, inventory control management
(overbooking and length-of-stay (LOS) controls), price
execution (rate implementation and adjustments), and
collaboration with the hotels general manager on
strategy and business planning. Some hotels have an
on-site dedicated revenue manager, titled a director of
revenue management (DORM); other hotels are part
of a corporate revenue management services group,
45
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
46 Interfaces 42(1), pp. 4557, 2012 INFORMS
which manages pricing and inventory on behalf of the
hotels and is generally located remotely.
A corporate revenue management (RM) team sup-
ports the hotel revenue managers by providing sys-
tems, strategy, and a support organization of regional
DORMs and geographic divisional vice presidents
(e.g., for North America and the Asia Pacic regions).
IHGs globally distributed RM organizational struc-
ture is the norm for the hotel industry; however, it dif-
fers from the highly centralized RM structure in the
airline industry. The complex organizational struc-
ture of hotel RM presents signicant challenges with
respect to training, adoption, and consistent execution
of RM strategies and system use.
The application of RM in the hotel industry was
adapted from airline industry RM systems, which
the industry began to implement in the 1980s (Cross
et al. 2009). Since their inception, hotel RM systems
have opened and closed rate products, the prices of
which are predetermined via manual processes, with-
out analytics. These systems assume that demand
by rate segment is independent, an assumption that
is not true and can lead to a downward spiral in
rates when demand is soft (Cooper et al. 2006). This
approach is similar to that of the early airline models
(Smith et al. 1992), which open and close fare classes
under the assumption that demands by fare class are
independent.
PERFORM
SM
is a Web portal through which more
than 4,000 users worldwide access IHGs RM sys-
tem and related tools. Like the RM systems at other
major hotel enterprises and prior to implementing
price optimization, PERFORM optimized availability
and LOS inventory controls based on the assump-
tion of independent demand. The deterministic model
described by Baker and Collier (1999) is the most com-
mon formulation used in practice. Like that used by
some other hotel RM systems, the PERFORM yield
management optimization model was a variant of the
deterministic model with stochastic demand.
The growth of Internet booking channels starting
about 2000, the deepening travel recession starting
about 2001, and the tragic events of September 11, 2001
combined to drive hotel RM systems to incorporate
pricing as well as inventory yield techniques (Cross
et al. 2009). Hotel occupancy rates fell by 1520 per-
cent at leading hotel groups (Cross et al. 2009), and US
hotel prots fell by $642 million (Bowers and Freitag
2003). Soft demand for hotel rooms also lessened the
benets of the PERFORM system because the ben-
ets of yield management models primarily derive
from tightening inventory controls when demand is
strong. Internet booking channels created increasing
price transparency, allowing consumers to compari-
son shop multiple hotels to nd the best deal. Price
transparency and the need to drive demand con-
tributed to the erosion of rate fences (restrictions),
which are essentially qualications on bookings that
support segmented demand and pricing. The erosion
of rate fences undermined the RM assumption of
independent demand.
The hotel industry has traditionally divided de-
mand into two broad segments: group and transient.
The group segment includes conferences and corpo-
rate events for which a hotel contracts with a group
to commit large blocks of rooms for a specic period.
The transient segment represents all individual book-
ings. The objective of hotel RM systems is to opti-
mize revenues for the transient segment. Although
many RM systems also include a group yield mod-
ule, transient and group segments are managed sepa-
rately. Lee et al. (2011) divided the transient demand
into retail and negotiated segments. Negotiated seg-
ments include corporate special rates for large cus-
tomers (e.g., IBM and HP). These rates are typically
xed and are not subject to dynamic price changes.
Most also have last-room availability clauses; thus,
they are not subject to the inventory controls that RM
systems generate. Only the retail segment is subject
to the full range of pricing and inventory controls.
Lee et al. (2011) further segmented retail demand into
restricted and unrestricted segments. In a study of
20062007 hotel demand, these authors demonstrated
that rates paid by unrestricted retail customers do not
tend to increase as the day of arrival approaches and
that restricted rates actually tend to decrease. This
observation is contrary to the long-held belief that cus-
tomer willingness to pay increases as the day of arrival
approaches. A cornerstone of yield management and
RM systems is the assumption that higher-booking
customers book late in the booking cycle. Lee et al.
(2011) assertand the authors of this paper agree
that segmenting hotel demand into group and retail
segments better aligns with how consumers view
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
Interfaces 42(1), pp. 4557, 2012 INFORMS 47
hotel products and that the studys ndings seriously
challenge the assumption of independent demand by
rate segment and the assumption that willingness to
pay increases as the day of arrival approaches.
The decline in hotel demand, the rise of Internet
booking channels and price transparency, and chal-
lenges to underlying RM assumptions drove changes
in hotel RM workow. IHG adapted to this changing
environment by revamping its pricing strategy, shift-
ing the focus from an inventory allocation approach
to a pricing focus. It implemented a rational pricing
structure within which restricted retail discount rates
were tied to the unrestricted best exible rate (BFR).
A uniform rational rate structure facilitated dynamic
pricing.
IHG, without an automated capability to opti-
mize prices, undertook a process of educating its
hotel staffs on the need to ex their BFRs based on
demand. Because of this strategic shift, IHG property-
based DORMs were spending, on average, more
than 30 percent of their time gathering competitive
price intelligence. This intelligence included competi-
tor rates, which they found on the Internet or through
third-party sources, including TravelClick and Rubi-
cons MarketVision reports. DORMs changed rates
through IHGs central reservations system, HOLIDEX
Plus. HOLIDEX Plus is a mainframe system, which
was not designed to facilitate frequent rate changes;
its pricing mechanism is cumbersome and time con-
suming. As a result, pricing analysis was ad hoc,
response to competitive actions slow, and execution
inconsistent. Forecasting represented 30 percent of a
DORMs time, much of it to adjust forecasts for rate
changes in the DORMs own property and in compet-
itive rates. DORMs spent only 20 percent of their time
performing more strategic analysis and business plan-
ning and less than 10 percent in managing inventory
controls. The desired division of time among tasks
is 40 percent in forecasting, 40 percent in strategy,
10 percent in pricing, and 10 percent in inventory con-
trol. Corporate RM realized that the DORMS needed
system support for pricing (including price-adjusted
forecasting) to better use their time, to improve the
quality of pricing decisions, and to facilitate improved
execution of pricing best practices. A key part of the
solution was a price optimization capability that opti-
mizes prices for the retail segment, considers com-
petitor rates, and generates a price-adjusted forecast
based on price sensitivity. A new price optimization
module in PERFORM would support pricing analysis,
recommend prices, adjust forecasts based on IHGs
own and competitive prices, and automate price exe-
cution. New reporting and a new interface to execute
prices in HOLIDEX Plus would be essential.
Building the Business Case
Designing and implementing the price optimization
capability was a major cross-functional effort. Corpo-
rate RM led the analysis and design, IHGs informa-
tion technology (IT) group developed the new screens
and reports, and a training group developed and
delivered a global training program. IHG corporate
and franchise hotel management needed to buy into
the changes. The executive leadership team needed to
approve the large corporate capital outlay and support
the massive change management process. However,
management had been burned many times by large
capital expenditures that failed to deliver promised
benets; therefore, the executive team wanted quan-
tiable proof that implementing the price optimization
capability would increase prots.
In the third quarter of 2006, IHG engaged Rev-
enue Analytics. The two organizations formed a part-
nership in which they jointly conducted a research
and scoping project that lasted through the design,
development, and deployment of the price optimiza-
tion module. The projects goals were to demonstrate
the feasibility of price optimization, develop an ini-
tial estimate of its potential benets, and identify other
capabilities that needed to be upgraded to enable
price optimization. The project team developed a sim-
ulation model, which estimated the theoretical ben-
ets of price optimization to be from 2.75 percent
to 6 percent revenue uplift on the retail segment.
It also determined that it needed to upgrade the
existing RM forecast algorithms. Transient forecast
errors reduced the benets of optimizing price by
1.3 percent. Group forecast errors reduced benets by
0.7 percent. Figure 1 shows the sensitivity of the price
optimization simulation to forecast error by elasticity.
Based on this analysis, the team immediately
launched projects to improve the PERFORM RM fore-
cast models.
Although the research and scoping phase required
extensive applications of analytics, a few simple
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
48 Interfaces 42(1), pp. 4557, 2012 INFORMS
Price optimization benefit average uplift for a sample of 776 properties
12.0
10.0
8.0
6.0
4.0
R
e
v
e
n
u
e

u
p
l
i
f
t

(
%
)
2.0
0.0
2.0
0.8 0.9
Elasticity
1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0
Impact of group forecast error
Impact of group and transient forecast error
Price optimization benefit
Figure 1: Transient and group forecast errors signicantly reduced the benets of price optimization.
models were decisive in communicating optimiza-
tion concepts to the chief marketing ofcer (CMO),
brand presidents, and other senior executives. The
only way to accurately communicate the form of
the price optimization model is through mathematics,
which is not the preferred method of explanation for
most senior executives. We found that the simple two-
dimensional example depicted in Figure 2 was pow-
erful in explaining concepts to the executives whose
approval we needed to fund our price optimization
project.
To help us in gaining executive approval, we
decided to build a simple interactive simulation model
in the form of a game (see Figure 3) in which the audi-
ence would try to guess what the optimal price should
be. The base-case scenario formed a business-as-usual
point of reference. For each turn of the game, competi-
tor rates, demand, and capacity varied; the object was
to guess the rate that would optimize revenue. The
game was fun, but also communicated the challenges
revenue managers faced in determining the best rate
for a single date. It reminded the audience that rev-
enue managers had to handle multiple-rate products
for 350 future arrival dates, while accounting for LOS
interactions. If senior executives could not guess the
right price in this simple game, how much more
challenging was the problem that the hotel revenue
managers faced? The pricing game was key in secur-
ing the funding we needed.
In the second quarter of 2007, we received fund-
ing approval for a high-level design and live market
test project. The market test required construction
of a working price optimization prototype, which
we would deploy to a limited number of hotels.
These hotels would use the prototype system to man-
age rates for their hotels for the duration of the
test. In addition to providing valuable feedback from
DORMs on the design, the market test would serve as
a robust measure of the achievable benets. The exec-
utive team and capital committee demanded proof
from live market tests, not merely theoretical esti-
mates from a simulation. The IHG capital committee,
which includes the most senior IHG executivesthe
CEO, the CFO, the CMO, and at least one regional
president, is responsible for releasing funding for
large investments. The high-level design would pro-
vide enough detail to enable the IT group to esti-
mate price optimization development costs. As part
of the high-level design, the combined IHG and RA
operations research (OR) teams would research and
prototype the four other price optimization models:
elasticity and price-sensitive demand forecast, LOS
price optimization model formulation and solution
algorithm, competitive rate shopping algorithm, and
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
Interfaces 42(1), pp. 4557, 2012 INFORMS 49
Contribution function
Room rate ($)
C
o
n
t
r
i
b
u
t
i
o
n

(
$
)
Unconstrained contribution
Demand
30,000 500
450
400
350
300
250
200
150
100
50
D
e
m
a
n
d

(
r
o
o
m

n
i
g
h
t
s
)

25,000
20,000
15,000
10,000
5,000

Constrained contribution
50 75 100 125 150 175 200
Figure 2: Demand is simply a linear function of price. In this example, the unconstrained optimal price is $110;
however, because the hotel capacity is only 200 rooms, the optimal constrained price is $130.
Base case information
Benchmark rates
Holiday Inn (HI) Quality Inn Courtyard Comfort Inn Best Western Price elasticity
$97.47 $79.00 $119.00 $89.00 $99.00 1.3
The pricing game
Quality Inn Courtyard
Comfort
Inn
Best
Western
Expected
demand
HI
capacity
Best
guess HI
rate
Best
guess
rooms
sold
Optimal HI
rate
Optimal HI
rooms
sold
Guess
revenue
$79.00 $109.00 $89.00 $99.00 38 40 $99 36 $91.10 40 $3,552.86
<ctrl>k to clear
Winner?
Margin of
victory
Optimizer
<ctrl>a for answer 2.6%
Holiday Inn-Highway Location-Tuesday
Figure 3: The pricing game was one of the most compelling tools we used to communicate the need for price
optimization to senior executives. We challenged the executives to guess the revenue optimal price under varying
supply, demand, and competitive pricing conditions. In the game, we show the key input parameters and the
recommended rates using a glass-box solution framework; next to the recommended price, we present key data
elements (e.g., occupancy, current IHG and competitor prices, IHG and competitor reference prices, and price
sensitivity) to validate the price recommendations.
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
50 Interfaces 42(1), pp. 4557, 2012 INFORMS
the competitive rate ll-in logic. To implement the
prototype and conduct the live market test, we had
to implement all these models, which we describe in
the Core OR Models section and the appendices.
To expedite the prototype development, we
decided to simplify the optimization to a staynight
model, which essentially assumes that all demand is
for only one night. The production system would be
an LOS model, which recognizes that guests can stay
for multiple nights. Modeling multiple-night stays
requires a network structure for the constraint matrix
to account for contention of different LOS periods for
the same room on a given night.
Figure 4 shows the prototypes structure. We imple-
mented the prototype in Excel VBA, connected it to an
Oracle database (the prototype DB), and refreshed this
database weekly from IHGs enterprise data ware-
house (EDW) and the PERFORM RM tables.
In July 2007, we deployed the rst prototype to
a hotel; eventually, we deployed it to 18 properties.
We conducted the live market test on 13 properties
over a 16-week period. To account for reservations on
Competitive rates
Perform/RM EDW
Prototype DB
PO prototype
worksheets
Figure 4: The prototype consisted of seven user screens and three screens
to allow an administrator to gather and report usage statistics.
the books and a ramp-up period, we excluded the rst
four weeks of the pilot, leaving a 12-week test period.
For each of the 13 treatment properties, we selected
1 to 4 control properties (34 control properties in all).
Variables controlled for included brand, region, prop-
erty size, and group mix. The 12 weeks prior to the
prototype roll-out were the baseline period. We con-
trolled for day of week by ensuring that the baseline
and test periods had equal numbers of each day of the
week. We assumed that seasonality was the same for
the prototype and control properties. Figure 5 illus-
trates the concepts of baseline and test periods and of
prototype and control properties.
The benet metric we used was total revenue per
available room (REVPAR), which is the total revenue
divided by the number of room nights available for
sale. Although the price optimization function only
recommended retail price changes, total REVPAR
includes group and negotiated segments for these
reasons. (1) REVPAR is the most important per-
formance metric because the executives and capital
committee understand it clearly. (2) We considered,
but rejected, transient REVPAR (transient revenue
divided by total rooms). Total REVPAR can vary
widely as the occupancy rate varies. REVPAR instabil-
ity is even more pronounced if we subdivide revenue
by group and transient segments. (3) Retail prices
indirectly inuence group and negotiated rooms sold
and rates; therefore, some price optimization bene-
ts are expected in these segments. Using a Pearsons
chi-squared test, we concluded with 99 percent con-
dence that the prototype properties outperformed
their control properties during the test period (and
relative to the baseline period). The mean improve-
ment in REVPAR was 3.2 percent.
Anecdotal feedback on the price optimization pro-
totype was also positive. For example, in response
to our request for feedback, Brian Cauwels, revenue
manager for a Holiday Inn Express in Louisville, Ken-
tucky, reported We had the highest revenue week
ever, aside from the Derby weekend, using the rec-
ommended rates of the tool. The GM [general man-
ager] became a big believer in pushing rate after he
saw the revenues from the rst night of the week.
Balazs Szentmary, revenue manager for the InterCon-
tinental Madrid, wrote Great Tool! [It] challenges
you to question your pricing practices. We collected
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
Interfaces 42(1), pp. 4557, 2012 INFORMS 51
Baseline period Test period Excluded Prototype property
Baseline period Test period Control property
Pilot launch
Time (weeks)
Figure 5: The test period was 12 weeks starting at week ve following the prototype launch. The baseline period
covered the 12 weeks prior to prototype launch.
detailed user feedback, incorporated it into the tool,
and performed additional analytics on usage statis-
tics. Figure 6 shows the utilization of the various
prototype screens. The staynight screen was the most
heavily used. This guided the design of the optimize
price screen in the production system (see Figure 9).
The calendar view was woven into the overall nav-
igation. The workbench screen proved important as
users gained familiarity with the system; however,
because the optimize price screen was so critical, we
had to add quick links into the production system to
allow the users to navigate directly to this screen.
The live market test provided a rigorous benets
estimate for the new price optimization capability.
The prototype and the high-level design enabled us
to reliably estimate the time and cost required to con-
struct the production system. Feedback from proto-
type users added weight to the quantitative benets
Reservations
8%
Bus. rules
5%
Calendar
7%
Workbench
26%
Staynight
32%
Comp. rates
11%
Analysis
11%
Figure 6: The chart shows the relative utilization of the prototype screens.
This feedback from the prototype helped guide the design of the produc-
tion system.
estimate of the prototype. The benets of the proto-
type were so substantial that some DORMs pleaded
to keep the prototype running. As a result, we were
able to build a solid business case for a production
version of the price optimization capability. In the
fourth quarter of 2007, based on the detailed benets
estimate, the support of hotel general managers, and
recommendations of the property and regional rev-
enue managers, the IHG capital committee approved
a multimillion dollar budget to develop the price opti-
mization module within PERFORM.
Production System Development,
Deployment Plan, and Revenue Uplift
Estimates from Beta Release Properties
Development began in January 2008. The OR team
implemented the market response model (MRM),
competitive rate shopping module, the rate expan-
sion module, and the core price optimization engine.
IT implemented the data model, the server that inte-
grates all modules, the user interface, job scheduling,
and the conguration of new servers for the price
optimization capability. The RM strategy team and
the OR team developed and implemented the change
management plan and the rollout plan and worked
with the training group to develop training modules.
The MRM describes the relationship between
demand and other driver variables. The competitive
rate shopping module species which future arrival
dates and LOS products should be shopped. Shop-
ping all future combinations of arrival date and LOS
would overburden the global hotel distribution sys-
tem; thus, it is not feasible. Because only a sample of
future arrival dates and LOS periods are shopped, the
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
52 Interfaces 42(1), pp. 4557, 2012 INFORMS
Server
Competitive
rates
EDW
Comp rates
MRM
Market vision
competitor rates
Rate expansion
Rate shopping Shop requests
Optimization
engine
Figure 7: The owchart shows the logical relationship between the rate-shopping module, the MRM module, the
price optimization engine, and the server in the price optimization architecture.
rate expansion module infers rates for products that
are not shopped. The price optimization engine builds
the optimization model formulation from the input
data and solves for optimal prices. Figure 7 depicts
the relationship between the core modules.
We used a decomposition approach to model
demand as a function of price; we modeled it as
independent of price and then modeled the remain-
ing variability in demand as a function of price. This
approach t the data well and aligned with the deci-
sion to leverage existing PERFORM RM modules to
the fullest extent possible. In particular, we wanted
to continue to use the existing forecasting and yield
optimization functions.
There were six fundamental reasons for continu-
ing to use the existing PERFORM modules. (1) Rev-
enue managers were already effectively using much
of the PERFORM functionality, including user screens
and reports. (2) The existing forecast, although not
price sensitive, was reasonably accurate. (3) The LOS
recommendations that the yield optimization module
generated were good. (4) Pricing decisions and yield
decisions are made at different frequencies. Inventory
controls change constantly and in real time as book-
ings are made; however, hotels prefer to change prices
less frequently. (5) Implementing price optimization
would require extensive retraining of revenue man-
agers and careful conguration of each property,
necessitating a staged rollout. Therefore, the existing
PERFORM system would have to continue to function
for other properties during the rollout. (6) Leveraging
the existing modules would accelerate delivery of the
new system.
Price optimization works in conjunction with the
existing PERFORM forecasting and yield optimiza-
tion components. The MRM modies the PERFORM
forecast at the optimal prices to make it price sen-
sitive. The price-neutral unconstrained demand fore-
cast, available capacity, and competitive rates are the
key inputs to the price optimization engine. Plugging
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
Interfaces 42(1), pp. 4557, 2012 INFORMS 53
CRSCentralized reservation system
Forecasting engine
Price optimization engine
Yield optimization engine
Price-sensitive forecast
Price-sensitive
forecast
CRS
Optimal prices
Optimal length of stay
Price-insensitive forecast
Figure 8: The forecasting engine generates a price-neutral demand forecast. The price optimization engine gen-
erates optimal prices, and a price-sensitive forecast is computed at these prices based on the MRM. The yield
optimization engine leverages the price-sensitive forecast to generate LOS controls.
the optimized rates into the MRM produces the price-
sensitive demand forecast at the new rates. After
the rates have been updated, the yield optimization
engine uses the price-sensitive demand forecast to
update the LOS inventory controls at the new rates.
Figure 8 depicts how PERFORMs forecasting, yield
optimization, and price optimization engines work
together.
RM executives insisted that the solution not be a
black box. Presenting the critical components driv-
ing the pricing recommendation with the recom-
mendations was critical. Data presented include the
three pillars of pricingcompetitive rates, forecasted
occupancy, and price-sensitivity ratings derived from
elasticity estimates. Figure 9 shows a screenshot of
PERFORMs optimize price tab, which is used to
review price recommendations and publish them to
the central reservation system. Training on the three
pillars, including a variation of the game we devel-
oped in the research and scoping phase, is a critical
prerequisite that revenue managers must meet before
using the price optimization functionality. This train-
ing and the need to carefully congure each prop-
ertys competitive setthe set of competitors whose
rate data need to be shopped (a critical input to price
optimization)are the two main reasons the price
optimization rollout had to be gradual.
Alpha testing for price optimization began in the
rst quarter of 2009. Training of regional DORMs and
revenue managers at beta test properties began simul-
taneously. Beta testing in the production environment
also began, and continued until the third quarter of
2009. Starting in the fourth quarter of 2009, price
optimization rollout began for the rest of the IHG
properties at a rate of approximately 100 per month.
In the third quarter of 2009, after beta properties had
been using the capability for several months, we con-
ducted a benets measurement study similar to that
conducted following the prototype market test. This
study showed a 2.7 percent increase in REVPAR for
the beta test properties. As of this writing, more than
2,000 properties worldwide are running PERFORMs
price optimization, and we add about 100 proper-
ties each month. The global nature of this capability
means that some core OR models for shopping com-
petitive rates must be modied to also account for
booking channels.
Core OR Models
Price optimization development involved intensive
OR modeling. Given the scope of this paper, we
cannot describe all the work in detail; however, in
this section we will outline ve key areas in which
we applied ORMRM, competitive rate shopping,
benets estimation, rate expansion and ll-in logic,
and optimization model. The IHG OR and Revenue
Analytics OR teams designed and implemented each
model and then integrated it with the server and user
interface, which IHG IT developed.
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
54 Interfaces 42(1), pp. 4557, 2012 INFORMS
Figure 9: The optimize price tab in PERFORM displays all the information needed to make the price recommen-
dations transparent. The benchmark rate is an aggregation of competitor prices. By comparing the remaining
capacity, remaining demand, competitor rates, and our current BFR, revenue managers can intuitively judge the
reasonableness of the price recommendations. As suggested by the other tabs, price optimization also provides
a capability to drill down into demand forecasts, competitive rates, current bookings, and additional pricing
analysis.
Market Response Model (MRM)
The MRM describes demand as a function of price and
other driver variables. Because conducting real-time
price experiments to estimate price sensitivity is dif-
cult in IHGs distributed environment, we decided
to use pseudo-random price experiments to mine his-
torical prices, historical demand, and historical com-
petitor rates to measure the response of the demand
changes against IHG and competitor price changes.
A key input from the hotel revenue managers was that
if they were to have greater acceptability of the price
optimization capability, modeling demand as a func-
tion of competitor rates would also be imperative.
Within the MRM module, we modeled demand
as a function of price and competitor rates to com-
pute price elasticity. We tried various segmentation
schemes and decided on a segmentation approach
that aligned with key business segments at which
elasticity estimates were signicant and that the rev-
enue managers accepted. If we could not nd statis-
tically signicant elasticity estimates, we also used a
logical hierarchical approach.
Competitive Rate Shopping
Dynamically shopping forward-looking rates of our
competitors is a critical component of the price opti-
mization module. We nd publicly available com-
petitor rates on the Internet and through third-party
sources (e.g., TravelClick and Rubicons MarketVision
reports), select a maximum of four hotels as com-
petitors, and shop each of the four competitors each
night. In specic regions (e.g., Greater China and
Asia Australasia), we also consider booking channels
in collecting shopping data. Each day, the optimiza-
tion engine uses the shopping data to optimize the
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
Interfaces 42(1), pp. 4557, 2012 INFORMS 55
rates for the next 350 days. If we were to undertake
shopping all our competitors for their unrestricted
rate products for the entire enterprise, our cost would
be millions of dollars per year, an infeasible expense.
In addition, our shop requests would ood the Web
and global distribution systems, bringing reservations
for IHG and other hotels to a halt. Our shopping bud-
get allowed us to afford only 2030 shops each day for
each competitor. Given the budget and distribution
constraints, we developed a random, stratied sam-
pling strategy that allows us to recommend shops by
blending future booking activity and historic booking
patterns. If a product has a high booking activity (sim-
ilar to some special event days), then the probability
of that product being shopped is higher. Historically,
if a product has been shopped frequently, then it is
more likely to be shopped.
We do only 2030 shops each day; however, to
recommend optimal rates, the optimization module
requires shop data for each of the next 350 days.
Therefore, we developed a reasonable approach to
ll in the dates for which we have not shopped.
This method considers day-of-week patterns, LOS
patterns, and last-shopped time stamps; lls in the
missing dates; and generates the full list of shopping
data to complete the rate data set before entering it
into the optimization engine. If not for this random,
stratied sampling approach and a novel way to ll in
missing rates, we would have had to spend millions
of dollars to acquire the necessary shop data.
Benets Estimation
Price optimization as a business capability is not
complete without measuring the revenue benets.
Both hotel revenue managers and senior executives
impressed upon us the need for a rigorous mea-
surement methodology that measured the impact of
price optimization. Our method involved compar-
ing the change in a key metric (REVPAR) for a test
period and a baseline period for the properties using
price optimization and for the control properties not
using it. We conducted statistical studies to account
for statistical signicance of such a REVPAR uplift.
In selecting the control properties, we considered sea-
sonality, brand, business segmentation mix, hotel type
(e.g., business, leisure, convention), and location type
(e.g., downtown, suburban, airport). The statistical
signicance tests ensured that the price optimization
related coefcients were signicant. This study found
a 2.7 percent increase in REVPAR for the beta test
properties.
We performed several iterations of this approach
in which we primarily addressed the logic of con-
trol property selection and the key driver variables,
which could vary by property. We then socialized this
approach with key stakeholders and the executives
who were involved in the initial test. The stakeholder
involvement and qualitative feedback from the users
was instrumental in the inclusion of PERFORM price
optimization in the 2009 IHG annual review. For the
properties that had used this module for the previous
12 months, a 2.7 percent increase in REVPAR trans-
lates to a revenue increase of $145 million. At full
rollout, we anticipate that this capability will generate
approximately $400 million per year.
Rate Expansion and Fill-In Logic
As we described in the Competitive Rate Shopping sub-
section, we could shop only a small fraction of future
competitor prices for arrival dates and LOS dates.
To determine optimal prices, we needed an estimated
price for each competitor for every arrival date and
LOS combination for the next 350 arrival dates. There-
fore, we developed an algorithm to expand the actual
competitive shops to the full cardinality of reserva-
tions products. Although we cannot share the specic
details of this algorithm, we can describe its general
principles.
If we did not shop a product on a given day, but
had shopped it in the previous few days, we infer an
observed price. We ll in the remaining holes in the
competitive rate matrix with shops of a different LOS
for the same arrival date, and ll in any remaining
holes with rates from adjacent arrival dates. Overar-
ching the algorithm is an inherent bias toward shorter
LOS dates. The rationale behind this bias is that
nearly 40 percent of bookings are for multiple nights
(the average LOS is about 2.0 days). Also, the gen-
eral tendency (although denitely not the rule) in the
hotel industry is that the rate for a multiple-night stay
is the sum of the one-night-stay rates.
Optimization Model
The core price optimization model is innovative in
the industry. Modeling demand as a function of
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
56 Interfaces 42(1), pp. 4557, 2012 INFORMS
price requires that the objective function is nonlinear.
Special reservation rules, which are unique to IHG,
require logical and integer constraints. The model
accounts for LOS patterns, signicantly increasing its
complexity relative to a staynight model. We for-
mulated the optimization model as a mixed-integer,
bilinear mathematical program. We implemented a
special optimization method that leverages CPLEX
to iteratively solve approximately 1,000 integer pro-
grams per day for each property. On average, rates
are generated for each property six times per day. The
price optimization module solves four million linear
programs each day. Appendix A provides details on
the optimization model; Appendix B provides details
on benets measurement.
Price Optimization Spawns New
Revenue Management Initiatives
The price optimization project has reinvigorated RM
at IHG. As we previously mentioned, the imple-
mentation of targeted enhancements to the existing
PERFORM system was an early outcome. Price opti-
mization also inspired a multimillion-dollar initiative
to revamp HOLIDEX. A new central reservation sys-
tem, REVOLUTION, will streamline the denition of
rate products and ensure that a rational rate struc-
ture is in place at all hotels. The price optimization
project also raised the visibility of the RM groups
forecasting expertise within the global IHG organiza-
tion. The RM group is now considered the corporate
forecasting center of excellence, and the development
of an enterprise-wide forecasting platform, predictive
demand intelligence (PDI), is a testament to that. PDI
generates a forecast that integrates with the key cor-
porate business functions of nance and marketing
and with property-based RM, including PERFORM.
By using a common forecast, IHG is better able to align
marketing and budgeting with the tactical pricing and
inventory control RM processes. We deployed a PDI
prototype in the fourth quarter of 2010; the production
system is currently under development. Also, because
we deployed price optimization to a majority of prop-
erties, we can no longer measure the uplift of price
optimization using control properties; hence, a project
is underway to construct a performance measurement
model, which will be similar to the model we devel-
oped during the initial research and scoping phase.
It will use simulation to estimate the revenue uplift
from pricing actions, and generate insights to support
continuous improvements in forecasting and pricing.
Concluding Remarks
The journey to develop PERFORM with price opti-
mization at IHG provided many lessons on how to
build a business case for a massive enterprise sys-
tem with OR models at its core. The research and
scoping project built sufcient momentum to help
us gain funding for the development of a prototype
and live market test. The live market test and rigor-
ous test and control benets measurement provided
the foundation of an unassailable business case and
funding for a multimillion-dollar software develop-
ment and business transformation project, which a
committee of IHG senior executives approved. IHG
RM, IT, and operations teams partnered to develop
and deploy the price optimization solution to a global
hotel enterprise. To date, price optimization has gen-
erated $145 million of incremental revenue for IHG
and its franchise partners.
IHGs price optimization system is already having
a major impact on the hotel industry. Other leading
global hotel enterprises are currently developing their
own LOS price optimization solutions. Carlson Hotels
is implementing a staynight price optimization solu-
tion (Rozell 2009). The methods for estimating price
response developed at IHG are applicable to many
industries. The specic problem of optimizing price
for demand based on LOS is directly transportable to
rental cars (length of rental) and airlines (origin and
destination). The IHG experience also helped inspire
the development of similar methods to optimize the
price of juice drinks in a resource-constrained supply
chain (Bippert 2009).
Appendix A. Optimization Model
Formulation and Solution Methodology
Our model is an adaptation to the hotel LOS prob-
lem that Gallego and van Ryzin (1997) proposed for
the airline network problem. The plain formulation
without business rules is listed below.
Variables

ud
: Rate for arrival date a and LOS d. (Optimiza-
tion decision variable.)
Koushik, Higbie, and Eister: Retail Price Optimization at IHG
Interfaces 42(1), pp. 4557, 2012 INFORMS 57
D
ud
= ] (
ud
, C
ud
): Demand generated for arrival
date u; LOS d is a function of the hotels rates (
ud
)
and competitor rates (C
ud
).
cos |
ud
: Room turn cost for arrival date u and LOS d.
|: Set of all resources; resources are the stay dates
with available capacity.
C(l): Set of all arrival date and LOS combinations
consuming resource l.
C
l
: Available capacity of stay date l.
The contribution function that IHGs price opti-
mization optimizes follows.
Max contribution =

ud
D
ud
(
ud
cos |
ud
),
where D
ud
=] (
ud
, C
ud
),
subject to

udC(l)
D
ud
C
l
l |

ud
0 u, d.
Because both demand and prices are decision
variables, we implemented a special optimization
method using a decomposition heuristic that lever-
ages CPLEX. We believe the decomposition heuris-
tic is better than a Dantzig-Wolfe decomposition
approach or a dynamic programming approach
because the optimal prices in the near future must be
more accurate when compared to prices farther out in
the decision horizon, especially when we look at the
booking prole of IHG guests.
Special reservation rules within the IHG business
environment require integer variables and logical con-
straints, greatly complicating the optimization model.
Appendix B. Benets Estimation
We described the benets estimation methodology in
the Building the Business Case section, depicted it in Fig-
ure 5, and described the variables for which we con-
trolled in the Production System Development, Deploy-
ment Plan, and Revenue Uplift Estimates from Beta Release
Properties section. However, we cannot underestimate
the degree of sophisticated analysis that was required
to design these controlled experiments and statistically
analyze the results. We computed REVPAR changes
from the baseline to the test period for both prototype
and beta properties. Despite our best efforts to con-
trol the variability, REVPAR changes were extremely
volatile. Therefore, we used a Pearsons chi-squared
test to test our hypothesis that price optimization
properties outperformed their control group. If the
price optimization property REVPAR change was bet-
ter (i.e., a larger increase or smaller decrease) than that
of a control property, we counted that treatment con-
trol pair as a win for price optimization; otherwise, we
counted it as a loss. Comparing the frequency of wins
and losses, we computed a chi-squared test statistic
for the hypothesis with much more power than a sim-
ple means test. For the beta properties, we observed
41 wins and 27 losses, resulting in a condence fac-
tor of 91 percent that the price optimization prop-
erties performed better than their control properties.
The test of means showed a 2.7 percent improvement
in REVPAR with a condence of 80 percent that the
improvement was greater than zero. Typically, a simu-
lation methodology is used to measure the benets for
such a capability (Smith et al. 1992). However, mea-
suring the benets using a test versus control group
of hotels is a reliable methodology. Because of the
methodologys rigor in selecting control groups and
the involvement of stakeholders in the benets mea-
surement process, we were able to explain the benets
case to the senior executives.
References
Baker, T., D. Collier. 1999. A comparative revenue analysis of hotel
yield management heuristics. Decision Sci. 30(1) 239263.
Bippert, D. 2009. Simultaneously maximizing consumer value and
company prot from beginning to end. Proc. 5th Annual Rev-
enue Management Price Optim. Conf., Atlanta.
Bowers, B., J. Freitag. 2003. Merchant model impact on 2003 U.S.
hotel prots estimated to be $1 billion. Accessed March 31,
2008, http://www.hospitalitynet.org/news/4017944.html.
Cooper, W., T. Homem-de-Mello, A. Kleywegt. 2006. Models of the
spiral-down effect in revenue management. Oper. Res. 54(5)
968987.
Cross, R., J. Higbie, D. Cross. 2009. Revenue managements renais-
sance. Cornell Hospitality Quart. 50(1) 5681.
Gallego, G., G. van Ryzin. 1997. A multiproduct dynamic pricing
problem and its applications to network yield management.
Oper. Res. 45(1) 2441.
Lee, S., L. Garrow, J. Higbie, P. Keskinocak, D. Koushik. 2011.
Do you really know who your customers are? A study of US
retail hotel demand. J. Revenue Pricing Management 10(1) 7386.
Rozell, J. 2009. Revenue management is ripe for change.
Accessed February 16, 2011, http://lhonline.com/technology/
software/revenue_management_carlson_0331/.
Smith, B., J. Leimkuhler, R. Darrow. 1992. Yield management at
American Airlines. Interfaces 22(1) 831.