Cracking the Bestseller Code part 1
This week has been all about collecting text data from which to build our new tool: The Bestseller Code. Basically the idea is to study the relationship between the structure of text (number of words, complexity of words, length of sentences, whether pages are dialogue heavy or description heavy) and the likelihood that the book will sell well.
The biggest problem with running this kind of analysis is finding reliable data with which to define "sells well". There is no comprehensive source for book sales figures and what is available is released by ISBN, meaning that every edition of the book will have different figures. The Hardcover sales are counted separately from the Paperback sales and the only way to get the full number is to collect all these separate numbers and add them together >.< This isn't really doable when you're looking at THOUSANDS of books.
It gets worse for classics: multiple publishers might release copies of older works, or a anniversary edition might come out ... more ISBNs, more uncounted data. Then estimating sales figures just becomes utter hell.
The only way to control for these factors is to get the data set as large as possible. That way patterns can be observed even with these minor inaccuracies in data.
So far we have data on 370 novels. I'd like to get at least 1,000, but the information we're looking at isn't available for every novel ever published ... so we'll see how much we can collect.
Along the way I've been peeking at the data in SPSS (which btw IBM has ruined >.<) to see if patterns are forming or if this is all a giant waste of time. Answer: Yes! So far we can already see a statistically significant relationship between Word Complexity and Flesch Readability Scores. But what's most interesting to me is that "Bestsellers" have more in common with the "Low-Midlist" than the "High-Midlist". Observe:
Here is the Word Complexity data for our "Bestsellers". The mean is 7.27.
Now we have the "High-Midlist" ... the mean is 8 although there is more variation
But down one further on the "Low-Midlist" we see word complexity fall back down to 7.65 and the Standard Deviation is back down to 1.9
Incidentally the "Worstsellers" had a complexity score of 7.88
These kind of results excite me because it fits with what I've observed of the publishing climate so far: there is a definitely "sweet spot" when it comes to how much complex language readers want. I don't believe that commercial fiction must be 'stupid' or that literary fiction just doesn't 'sell'. If that were true we would see the complexity scores slide gradually, instead we see a definite preference for easier to read fiction among most consumers but can feel the pull of the 'literary' market as well.
The relationship between sales and textflow (whether your pages are dialogue heavy or description heavy) is still unclear, although just from eyeballing the data "Worstsellers" seem to have higher likelihood of long paragraphs of text.
