Short story, some people just have bad tastes and marketing people should keep this in mind.
Summary. Results limited to repeat purchasing of consumer products. With total sales held constant, as the percentage of sales identified from these so-called harbinger-of-failure customers rises from 25% to 50%, the probability of success of that product decreases about 31%. If these customers have repeat purchase, the relative probability of success decrease even more (2 repeat purchases: success decrease 37%; 3 or more purchases: success decreases 56%).
Longer story? The paper describes their data set as follows:
This paper uses two datasets: a sample of individual customer transaction data, and a sample of aggregate store-level transaction data. Both datasets come from a large chain of convenience stores with many branches across the United States. The store sells products in the beauty, consumer healthcare, edibles and general merchandise categories. Customers visit the store frequently (on average almost weekly), and purchase approximately four items per trip at an average price of approximately $4 per item. The store level transaction data includes aggregate weekly transactions for every item in a sample of 111 stores spread across 14 different states in the Midwestern and Southwestern portions of the US. The data period extends from January 2003 through October 2009. We use the store-level transaction data to define new product survival, and to construct product covariates for our
multivariate analysis. We exclude seasonal products that are designed to have a short shelf life, such as Christmas decorations and Valentine's Day candy.
The individual-customer data covers over ten million transactions made using the retailer’s frequent shopping card between November 2003 and November 2005 for a sample of 127,925 customers.
The analysis defines a successful product as one that continues sales after 3 years (and a flop otherwise) based on aggregate sales. But they use data from loyalty card scans to identify groups of customers to test their hypothesis, and they limited their analysis to products that survived at least 1 year to avoid the "noise" associated with no chance to purchase products. FWIW, they do level some suspicion that the use/non-use of loyalty cards bias the data somewhat, however, the analysis was based on the following methodology...
Repeat purchase rates may therefore provide a more accurate predictor of new product success than initial adoption rates. For this reason, we use both initial adoption and repeat purchases to classify customers. Specifically, we ask whether customers who repeatedly purchase new products that fail provide a more accurate signal of new product failure than customers who only purchase the new product once...
First, we use a sample of new products to group customers according to how many flops they purchased in the weeks after the product is introduced. We then investigate whether purchases in the first 15 weeks by each group of customers can predict the success of a second sample of new products.
Note they don't attempt to predict out-of-the-gate flop products (like the Zune), they merely note that it seems like some people's tastes for consumer items that are subject to repeat purchase are likely not indicative of ultimate retail success and for some reason this is unexpectedly correlated across different products. Their conclusion is that the preferences of these harbingers-of-failure customers for items are indicative of a negative preference in the greater population and that will likely limit the ultimate growth potential of a product, and stores shouldn't waste precious shelf space on such items favored by these people.