Rocket Science Retailing
Retailers and e-tailers have enormous amounts of data available to them today. But to take advantage of that data they need to move toward a new kind of retailing, one that blends the instinct and intuition of traditional systems with the prowess of information technology.
Editor's Note— Marshall Fisher of the Wharton School at the University of Pennsylvania, Ananth Raman of HBS and their colleague Anna Sheen McClelland recently completed a survey of 32 retail companies focusing on their practices and progress in four areas critical to what they call rocket science retailing: forecasting; supply chain speed; inventory planning; and gathering accurate, available data. The following excerpt from their report in the Harvard Business Review shows how some of these companies are making the most of the data available to them to improve product demand forecasting as a means toward better supply chain management.
For many of the retailers in our study, forecasting product demand is a right-brain function that relies on the gut feel of a few individuals and not on the systematic use of sales data. But it's a big mistake to overlook the opportunity to mix art and science. Retailers can significantly improve forecast accuracy simply by updating their predictions based on early sales data, tracking the accuracy of their forecasts, getting product testing right, and using a variety of forecasting approaches. Let's discuss each of these practices.
Update forecasts based on early sales data. Early product sales, appropriately adjusted for variations in price and availability, are an excellent predictor of overall sales (see the exhibit "No Need for a Crystal Ball"). In fact, retailers that exploit these data for production and inventory planning can more than double their profits—especially retailers of products with short life cycles, such as clothing, consumer electronics, books, and music.
But despite the potentially high payoff—and a commonly accepted belief among retailers that early sales are a good indicator of future sales—many of the companies we surveyed had no systems in place to exploit early sales data. One retailer, for example, ordered garments and committed specific quantities of each stock-keeping unit (SKU) to each of its stores 11 months before the product was even available to the public. Even retailers that paid attention to their early sales data updated their forecasts in an ad hoc manner when sales greatly exceeded or fell far short of original predictions.
Several companies have retailing practices worth emulating, however. Japan-based World Company and Spain-based Zara are fashion retailers whose merchants systematically examine early sales data to estimate future demand for various products. They conduct this analysis for every product at predetermined periods in its sales cycle. And the merchants follow through, immediately reordering items that look as though they may end up in short supply. Not surprisingly, World Company has achieved a gross-margin return on inventory investments of more than 300%—a substantially higher return than any other retailer we are aware of.
Dallas-based CompUSA, which sells computers and associated merchandise, has found that even one or two days of early sales data can be very useful to predict sales and replenish its inventory for PCs. Buyers monitor the sales of a certain product line soon after it is launched and update their forecasts based on those observations. They expedite orders for PCs that are selling better than expected and, when possible, they decline items that have not been shipped. This process of reading and reacting to market signals has improved CompUSA's ability to match supply with demand.
Finally, book and music retailer Borders Group uses historical sales data to customize the product assortment in each of its stores. Borders tracks sales at each store by product category. It uses its merchandise planning system to automatically adjust the inventory at a store based on sales in each product category. Thus, a store in Anchorage, Alaska, would carry a wide assortment of books about small planes because sales for such books tend to be high at that outlet, while the Boston store might stock relatively few items in this category because demand is lower there. Why don't more retailers customize their inventories? The answer, as we explain later on, lies in slow supply chains, inadequate or inaccurate data, the inability to measure stockouts and forecast error, and planning software that is inappropriate for the retailer.
Track and predict forecast accuracy. Only nine of the 32 retailers in our study said they analyzed the accuracy of their forecasts. And yet, tracking forecast errors, and understanding when and why they occur, is fundamental to improving accuracy. Even more important, knowing the margin of error on a forecast is vital to being able to react when the forecast is wrong. For example, if past forecasts for a certain product have been wrong by plus or minus 50%, when a merchant says you'll sell 10,000 of that item, that really means you'll sell between 5,000 and 15,000 units. Instead of buying 10,000, it might be smarter to buy 5,000 finished units and materials for an additional 10,000 units to be assembled quickly if early sales are strong.
World Company tracks and predicts forecast accuracy by item using the "Obermeyer method": new products are displayed in a room at corporate headquarters just as they would be in a retail store, and about 30 store employees, who are chosen to represent the company's target customers, estimate the likely success of each product. World has found that the products that generate greater disagreement among the employees are likely to have less-accurate forecasts.
Get the product testing right. An impressive 78 % of the retailers in our study test new products in a few stores before the actual product launch. But almost all the buyers said their test methods are highly unscientific and that any results that indicate that certain products will be unsuccessful are often ignored. Merchants often believe their products will sell well despite unfavorable test results; they blame the weather (bad or good), the poor choice of test sites, the inferior execution of tests, and other factors for suboptimal sales.
When a product testing method is developed with care and refined on a regular basis, the results can substantially improve forecasts. We helped develop a testing method at one apparel retailer that predicts the sales of a product based on the early sales at a few carefully selected test stores. We found that the selection of stores greatly affected the quality of the forecasts. By using historical sales data to pick a diverse group of test stores that matched varying customer preferences, we reduced forecast errors for each style and color from 30% to 9%.
Use a variety of forecasting approaches. Most companies we surveyed limit themselves to just one type of forecasting. Generally, a single forecast for each item is generated by the buyer or by a small group from merchandising. But generating multiple forecasts can be very valuable because in seeking to understand the differences in those forecasts, managers can explore the assumptions implicit in their forecasting techniques.
Take Old Navy, a division of the Gap. The company blends bottom-up and top-down forecasting approaches and then considers the results in a way worth emulating. Bottom-up forecasts are developed by merchandisers and planners who predict demand for each product based on factors such as current trends in the market, the product's "fit" with the target customer, and the complementary products that will also be offered. Top-down forecasts are developed by planners and occur independent of the bottom-up process. They are based on macroeconomic factors such as the economic growth rate and corporate growth objectives. The two approaches typically yield different results, which are reconciled during a meeting of managers from both groups. Old Navy finds that the different processes, and the ensuing discussion, lead to substantially better forecasts.
No Need for a Crystal Ball
Early sales data can help to predict demand for the life cycle of a product—particularly a fashion item. The information at right is from an apparel catalog company. The graph on the left plots actual life cycle demand against the forecasts made by a committee of four merchandisers. The graph on the right plots actual life cycle demand against forecasts based on sales observed during the product's first two weeks on the market, which accounted for 11% of the season's demand. The latter results in a forecast margin of error that is significantly less than the experts' forecast.
Excerpted from the article "Rocket Science Retailing Is Almost Here—Are You Ready?" in the Harvard Business Review, July-August 2000.
Marshall L. Fisher is a professor at the Wharton School at the University of Pennsylvania in Philadelphia.
Ananth Raman is a professor at Harvard Business School in Boston.
Anna Sheen McClelland was a research associate at the Wharton School at the time the research presented in this article was conducted.