Purpose The objective of this research is to evaluate a prediction modeling framework for reward-based crowdfunding success (CFS) that leverages machine learning (ML) algorithms. By exploring the determinants of project success and identifying optimal ML models, this study aims to assist micro, small and medium enterprises in designing and managing effective crowdfunding campaigns. Design/methodology/approach This study employs a theory-driven prediction modeling methodology to examine reward-based CFS using quality signaling theory. A dataset of 728 projects from mystartr.com (a Chinese platform in Malaysia) was analyzed using business intelligence (BI) technologies including web crawling, data mining and text analytics. Seven supervised ML algorithms, namely, bootstrap aggregating ensembles (BAE), classification tree, logistic regression (LR), random forest, support vector machine (SVM), XGBoost and deep learning (deep learning convolutional neural network) were evaluated using 12 variables. Python codes were run in Google Colab for ML models, including BI for better decision-making in CFS. Findings RF recorded the optimal F1-Score followed by SVM, LR and BAE. DL-CNN, CT and XGBoost models demonstrated low prediction performance. Additionally, CT and XGBoost showed overfitting issues. The important key features selected by the algorithms were the number of supporters, media pledge amount, target fundraising amount and the number of updates. These results confirm the feasibility of using ML models in crowdfunding predictions. Research limitations/implications This study focuses on reward-based crowdfunding projects in Malaysia using data from the MyStartr platform, which may limit the generalizability of the findings to other platforms or crowdfunding models. Future research could explore equity-, donation- or P2P-based crowdfunding across multiple platforms and countries to enhance external validity. Advanced deep learning models such as long short-term memory and transformers were not applied due to data limitations, particularly the lack of detailed unstructured textual data (e.g. full investor comments or project narratives). Future studies could use richer datasets to explore these models and examine factors such as media quality and cultural influences. Originality/value This study is the first to integrate quality signaling theory with BI tools and a range of supervised ML algorithms to predict reward-based CFS in Malaysia. Unlike prior local studies that focused on equity, donation or Islamic crowdfunding, this study addresses the unexplored area of reward-based predictive modeling. Using data from MyStartr, we analyzed 13 project features across 7 ML models. Our findings offer novel theoretical insights and practical tools that can help startups craft effective campaigns and aid platforms in enhancing their performance in Malaysia’s emerging crowdfunding ecosystem.
You et al. (Fri,) studied this question.