|
Article Excerpt
In 2002, Travelocity lost revenue and market share to competition as the business environment changed. Travelocity and Sabre collaborated to develop the enterprise network model (ENM). The ENM combines discrete choice customer modeling with simulation and large-scale optimization to improve Travelocity's management supplier agreements, customer marketing, and product pricing. The ENM has helped Travelocity become a more effective retailer and has contributed over $54 million to Travelocity earnings, at a current rate of $43 million per year.
Key words: industries: transportation, shipping; marketing: buyer behavior.
**********
With approximately $80 billion in travel products sold worldwide in 2005, Sabre Holdings is the world's largest travel-distribution system, providing an electronic marketplace in which consumers and travel agents shop for and buy products from travel suppliers, such as airlines and hotels. In 1996, Sabre launched Travelocity, an online travel company that leverages Sabre's travel experience and technology to sell travel products directly to consumers. Travelocity's original business model was based on earning commissions from selling airline tickets and revenues from selling advertising space. Because commissions were consistent across suppliers, we were mostly indifferent to what ticket suppliers sold or who advertised on our site. Travelocity initially took the lead in Internet airline ticket sales, but as new competitors entered the space, Travelocity's share of travel bookings steadily eroded. Some new entrants established competitive advantages by introducing unique content, such as Expedia's merchant hotel program, Orbitz's Web fares, and Priceline's opaque air. Adding to Travelocity's loss of market share in 2001, the airlines eliminated or greatly reduced commissions in early 2002.
Because online retail was a new channel for the travel industry, very little by way of tools, processes, or analytics existed to support supplier-retailer partnerships or to optimize retail operations. Prior to late 2002, Travelocity could not even calculate its own airline revenues but rather relied on suppliers to inform it of earned commissions and overrides several weeks after the end of the reporting period. Travelocity was in danger of extinction if it stuck with its original model of passive consignment.
To survive in this new environment and to avoid dying as had most Internet companies, Travelocity had to become a sophisticated retailer by negotiating marketing agreements with suppliers, managing access to content, and expanding into more profitable lines of business. In May 2002, then-president of Travelocity, Sam Gilliland, asked the Sabre Research Group to help Travelocity adapt to the evolving environment and increase its sophistication as a retailer. The plan was to develop decision support by (1) understanding the online retail travel business, (2) identifying the drivers and calibrating their influence on online retail profitability, and (3) developing the capability to forecast and influence future performance. This approach led to the development of the enterprise network model (ENM).
Like many retailers, Travelocity based its value proposition on connecting suppliers and customers. Travelocity's store is its Web site, www.travelocity.com, and its business performance depends on getting the right products into the store through relationships with travel suppliers, getting the right customers into the store through branding, promotions, and advertising, and then providing accurate, compelling responses to customer's requests. Travelocity competes with online travel agencies, such as Expedia and Orbitz, distressed inventory outlets, such as Priceline and Hotwire, and supplier Web sites, such as NWA.com and Hilton.com. Because suppliers and customers have multiple relationships in this marketplace, changes in content, display, or price can dramatically affect market share and revenues.
The Sabre-Travelocity team developed the enterprise network model to improve decision making in this environment. We took advantage of the data available from the Travelocity site and from competitors. We developed customer models (Figure 1) to estimate consumer preferences for product quality, price, brand, and promotion. We developed supply models to evaluate supplier deals and manage the display of flights and fares in the store. We developed models relating to price and the management of available content. Finally, we developed marketing models to identify compelling travel opportunities and inform our customers of them.
[FIGURE 1 OMITTED]
The online retail environment is very dynamic with opportunities developing and evolving rapidly. The need to react quickly made the traditional process of describing requirements, developing models, and implementing them unworkable. We used a method of applying general principles to a direction for improvement. We began projects by using the available data and making a series of quick hits rather than by making a single long-term comprehensive effort. We based our approach on proof-of-concept projects in which we tried to quickly determine whether a problem was solvable. If the solution we created proved to be effective, scalable, and transferable, we invested in refining it further. Our initial models were often spreadsheets. We used them while we developed more robust solutions. Some projects never progressed beyond the proof-of-concept stage because the spreadsheets were appropriate. For some projects, we developed operational prototypes that had automated data feeds, graphical user interfaces, and ongoing support. This flexible and progressive process has delivered value for several years. Since 2002, the ENM has yielded over $54 million of incremental value, with a current annual run rate of $43 million. The team that developed the ENM consisted of about five people working for three years.
Customer Data and Modeling
Our initial priority was to improve our understanding of Travelocity's customers, in particular, their purchasing behavior. We needed to know how product quality, price, promotion, and placement of screen displays affect purchase behavior, which in turn affects Travelocity's market share and revenue. For example, how attractive is a $100 connecting fare for a 10:00 AM departure on a low-cost carrier versus a $200 nonstop fare at 2:00 PM on a major airline? Travelocity had already developed a click-stream database for accounting purposes and for analyzing system problems. These data were a potential goldmine of insight, but they had never been used to analyze customer shopping patterns. We initially developed a customer-choice model that predicted which displayed option a customer would purchase, given that the customer purchased something. We next focused on conversion rates, the probability the customer would buy something instead of leaving the site without making a purchase. These models worked well in our initial applications, but we learned that conversion is largely a function of what is available from our competition. We developed robots to shop selected markets on competitors' Web sites and log the results. Their data and the resulting analysis were crucial to understanding the impact of price and display on Travelocity bookings and market share.
For every shopping session (over 500,000 per day), Travelocity stores each shopping request the customer made, each screen displayed in response to the request, and the next request the customer made (after viewing the screen), all in sequential order. From this information, we can determine (1) what each customer requested (demand), (2) what was displayed to each customer (choice set), and (3) and what each customer did (action).
Discrete Choice Models
To model booking behavior, we use multinomial discrete choice (also called conditional logit). Here, one observation consists of a screen of itineraries, and the model assigns a purchase probability to each option. The variables in our choice model are fare, number of connections, carrier, elapsed time, and departure time. We compute utility score as a linear function of the attributes of each...
|
