Reading : “The Language that Gets People to Give: Phrases that Predict Success on Kickstarter”
Authors: Tanushree Mitra, Eric Gilbert
Summary
Mitra and Gilbert discuss the general subject, and prior research on crowd funding before elaborating on their own work. Crowd funding is when a business or project proposal uses a website to seek start up funds. Kickstarter is one such website and it does not finalize any monetary pledges until it the funding goal is met, by a deadline. Restricting their list to projects since July 2012 that were past their fund date, the authors had 45, 810 projects in their dataset. They used BeautifulSoup to scrap the html of these projects and followed a conventional bag of words model focusing on relatively common English words. Their model had over 20,000 phrases to use as predictive features.
After making their model, they used penalized logistic regression, which guards against co-linearity and sparsity by moving the co-linear coefficient’s weight to the most predictive feature. cv.glmnet, is the R implementation of this method that they used, because it also handles sparsity. Both sparsity and co-linearity were present in their data.
Their results are beautifully displayed in a figure with two phrase trees. One is of “funded” phrases that started with “pledgers will [receive, be, also, have, …]”. The other is of “not funded” phrases that started with “even a dollar [short, will, can, helps, …]”. The “funded” phrases showed man common elements, such as reciprocity (the tendency to return a favor after receiving one), and scarcity (limited supply of rare or distinct products that hold more value to pledgers).
Reflection
This makes me want to “hack” Kickstarter and make a fortune.
I read this one carefully because some of the tools and methods that the authors used are analogous to what my team would probably have to use on our project. It is interesting that the results break down into reciprocity and “money groveling”, respectively, as very strong positive and negative indicators of achieving funding. It is not especially surprising though, because it shows that pledgers have their own self interest in mind even when they are donating.
Certain markets might succeed much better in crowd funding that others. The target market has to a) have access to the internet and b) have enough extra money to bother looking at Kickstarter and c) have the time and means and interest to make use of any Kickstarter product. The successful pledge earning language in those already niche markets could be pretty different for each of the categories and sub-categories. It would be interesting to run a similar sub-analysis within the sub-markets. The authors poked into this by adding control variables, but it would be interesting to dig deeper.
Questions
Are there penalties for “over promising?” That is, if a project achieves a funding goal, it receives the funds. What if a founders of a project just kept the money and didn’t deliver anything?
Whatever happened to Ninja Baseball? That sounds awesome.
Bringing the qualitative to the quantitative is always a pain. How reliable are sentiment analysis tools? Can they be used to detect opinion?
