| Crowdfunding platforms offer entrepreneurs the opportunity to validate their market and raise funding. Crowdfunding success-predicting models are important for both investors and entrepreneurs and can provide guidelines for a successful crowdfunding campaign. This study compares handcrafted textual features and metadata features with and against three transformer-based embeddings for predicting crowdfunding success. We extracted a dataset of 15,282 Kickstarter technology projects between 2015-2024, comprising 10,265 failed and 5,017 successful projects. We calculated 22 textual features, extracted 24 metadata features, and used three embedding approaches: RoB-ERTa, ModernBERT, and all-MiniLM-L12 v2. Our results reveal that transformer-based embeddings outperform handcrafted textual features. In addition, metadata features provided critical complementary signals. Ultimately, the highest overall predictive performance was achieved by an XGBoost model that combined the metadata features with ModernBERT embeddings. The research results challenge the assumption that transformer models inherently deliver superior performance on specialized prediction tasks and demonstrate the enduring value of domain knowledge and targeted feature engineering. In addition, the research textual models can be used within an AI-powered paraphrasing tool to iteratively refine project descriptions, story, and risks in real time, enhancing entrepreneurs’ chances for crowdfunding success. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.