CDNOW. Launching a company into a new sales environment involves unique challenges as to forecasting sales, markets, and buyer behavior. The music retailer CDNOW set up a business for selling music CDs over the Internet and needed a reliable tool to develop profiles for potential customers, as well as predict CD sales to a cohort of new customers.
Most music stores are located in specific areas and the cross section of customers in that particular locale can be readily assessed. The Internet does not have the boundaries of a brick and block store, so that the customer base must be interpreted through the number of people who visit the site, the number of people who make a purchase, the number of purchases each customer makes, and so forth.
Modeling tools already existed, but they were complex, and required an analyst with experience in managing large data bases. CDNOW was unwilling to invest in the high cost of developing and applying such a complex model. Instead, CDNOW sought a new and simpler model that merely allowed for data entry of total purchases and the number of new customers each week and that could be run in a standard spreadsheet environment such as EXCEL. This would allow management to review the results of the model frequently and make business judgments as the market changed.
The study was conducted in the first quarter of 1997. The model developed was a simple stochastic model of buyer behavior capable of giving managers a medium range tool for forecasting CD sales. Working with the large database, the number of customers each week could be entered, along with the number of units each customer purchased. Then, using an aggregate model, it was found that reasonable forecasts of medium-term sales could be made.
For the cohort of 23,570 people who first purchased from CDNOW in the first quarter of 1997, they extracted records of the total purchases for the 40 weeks beyond the calibration period (that is, April through December 1997). The forecasting performance of the model was evaluated against the actual purchasing numbers. The performance of the model is clearly demonstrated by the fact that the model underpredicts the cohort's first-year purchasing by less than two percent. In contrast, an alternative linear projection model overpredicts the cohort's first-year purchasing by more than 40 percent. This enabled CDNOW to design a reasonable business plan for their company which was not possible to do with the poorly calibrated regression model.