What drives a donor? A machine learning-based approach for predicting responses of nonprofit direct marketing campaigns

Abstract

Direct marketing campaigns are one of the main fundraising sources for nonprofit organizations and their effectiveness is crucial for the sustainability of the organizations. The response rate of these campaigns is the result of the complex interaction between several factors, such as the theme of the campaign, the month in which the campaign is launched, the history of past donations from the potential donor, as well as several other variables. This work, applied on relevant data gathered from the World Wide Fund for Nature Italian marketing department, undertakes different data mining approaches in order to predict future donors and non-donors, thus allowing for optimization in the target selection for future campaigns, reducing its overall costs. The main challenge of this research is the presence of thoroughly imbalanced classes, given the low percentage of responses per total items sent. Different techniques that tackle this problem have been applied. Their effectiveness in avoiding a biased classification, which is normally tilted in favor of the most populated class, will be highlighted. Finally, this work shows and compares the classification results obtained with the combination of sampling techniques and Decision Trees, ensemble methods, and Artificial Neural Networks. The testing approach follows a walk-forward validation procedure, which simulates a production environment and reveals the ability to accurately classify each future campaign.