Improving Crowdfunding Decisions Using Explainable Artificial Intelligence

This paper investigates points of vulnerability in the decisions made by backers and campaigners in crowdfund pledges in an attempt to facilitate a sustainable entrepreneurial ecosystem by increasing the rate of good projects being funded. In doing so, this research examines factors that contribute to the success or failure of crowdfunding campaign pledges using eXplainable AI methods (SHapley Additive exPlanations and Counterfactual Explanations). A dataset of completed Kickstarter campaigns was used to train two binary classifiers. The first model used textual features from the campaigns’ descriptions, and the second used categorical, numerical, and textual features. Findings identify textual terms, such as “stretch goals”, that convey both elements of risk and ambitiousness to be strongly correlated with success, contrary to transparent communications of risks that bring forward worries that would have otherwise remained dormant for backers. Short sentence length, in conjunction with high term complexity, is also associated with campaign success. We link the latter to signaling theory and the campaigners’ projection of knowledgeability of the domain. Certain numerical data, such as the project’s duration, frequency of post updates, and use of images, confirm previous links to campaign success. We enhance implications through the use of Counterfactual Explanations and generate actionable recommendations on how failed projects could become successful while proposing new policies, in the form of nudges, that shield backers from points of vulnerability.