Slashdot Mirror
Algorithm Aims To Predict Fiction Bestsellers
benonemusic writes "Three computer scientists at Stony Brook University in New York believe they have found some rules through a computer program that might predict which fiction books will be successful. Their algorithm had as much as an 84 percent accuracy rate when applied to already published manuscripts in Project Gutenberg and other sources. Among their findings was that more successful books relied on verbs describing thought processes rather than actions and emotions. However, some disagree with the findings. Author Ron Hansen said style is not the key, but instead readers' interest in the topics in the book." There has been work done already on finding the formula for a hit song, and using analytics to craft a blockbuster movie.
However, the sample's study makes exactly the same mistake. They used Project Gutenberg as the source, and download counts as a substitute for sales. Sales has one measure: the number of dollars in the cash box at the end of the day. They should be measuring books on the NY Times bestseller list, or the Amazon Top 10 list, which have actually sold for money and are actually popular (fraudulently placed books aside.) And they should be comparing them against books from their own genres, or at least books that had similar attributes.
I think what they'd really find is that "books that sell well are those that are marketed well", regardless of the words they contain.
Maybe they could focus on a specific key reviewer: what does Oprah like and not like? Maybe when they cross compile the data from all the books, they will find they've only discovered Oprah's tastes. Which isn't a bad outcome, if they are ultimately trying to discover what kinds of books will be better positioned to make the author money. But I don't think they've come close to predicting fiction "best-sellers" yet.
John