The power of machine learning methods to predict crowdfunding success: Accounting for complex relationships efficiently

The objective of this paper is to both demonstrate and explain the power of machine learning (ML) methods to predict crowdfunding success. The first step to achieve this objective is to compare the predictive performance of four ML methods (boosted trees, random forest, Shallow Neural Networks and Deep Neural Networks) to standard binary logit estimation using a dataset of more than 108,223 Kickstarter projects. The data used for the comparison is a set of categorical and continuous variables for each project that can be analyzed using all 5 modelling methods. The second step to achieve the objective is to show that the presence of complex relationships between the explanatory variables and outcome variable explains the power of ML methods. These relationships can be categorized as threshold-type (when a certain level of a explanatory variable is need for an effect to occur), Goldilocks-type (where a very specific level of an explanatory variable is needed for an effect to occur) and interactions (where the effect of one explanatory variable on the outcome variable is moderated by the levels of other explanatory variables). Our estimations show that these relationships are embedded in the context of crowdfunding. Because crowdfunding is a marketing activity, factors like attention, saturation, and synergy between levers are pervasive. It is because machine learning has the ability to account for these relationships (without knowing where they are in advance) that these methods are both better predictors and better managerial tools than standard binary logit models. We further show that even greater prediction power is unleashed by activating the text-analysis capabilities of ML methods. However, even when this is not possible, the value realized by employing ML approaches to predict crowdfunding success is substantial.