It's worth skimming the (still incomplete) paper rather than just reading the article. The article misses a key point made in the paper: repeat purchases by 'harbingers' was found to be an informative feature, in predicting whether or not a new product would succeed.
tl;dr but did they actually show that the 'harbingers' were predictive or were the 'harbingers' just correlated with failure? It is really easy to find correlation when you start trawling through a large dataset but which all dissipate to nothing once you start using the correlations to make predictions.
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2420600