MSc Thesis

Optimising Crowdfunding Success: A BOHB-Driven Reward-Tier Strategy for Technology Campaigns

Nathalie Tuijp, Optimising Crowdfunding Success: A BOHB-Driven Reward-Tier Strategy for Technology Campaigns, Financial Engineering & Management, University of Twente, 2025

Supervisors: prof. dr. Joerg Osterrieder (University of Twente, Netherlands), prof. dr. Berend Roorda (University of Twente, Netherlands), Dr. Stefana Belbe (Babeș-Bolyai University, Romania)

Crowdfunding platforms like Kickstarter offer opportunities for entrepreneurs and creators to fund their projects. Most of these projects have a reward-tier structure to fund their campaigns. The reward-tier-structure largely influences the outcome of the campaign. Therefore, this thesis aims to optimise this structure by developing a BOHB (Bayesian Optimization and Hyperband) framework with the integration of LLaMA-based embeddings, identifying the most effective reward-tier strategies to enhance campaign success rates. The BOHB framework is specifically chosen, as it is particularly effective for high-dimensional, non-convex search spaces like those found in crowdfunding campaigns. Its adaptive resource allocation and multi-fidelity optimisation allow it to efficiently explore vast parameter spaces, identifying optimal reward strategies with reduced computational cost. To complement traditional numerical features such as funding goals, number of backers, and reward levels, the research integrates LLaMA embeddings which are incorporated in the model to predict the campaign success, giving a higher accuracy to the model. These embeddings capture the semantic richness and emotional tone of campaign descriptions and reward titles. By combining advanced hyperparameter optimisation with the use of LLaMA embeddings, the model identifies optimal reward configurations that enhance the probability of campaign success.

The study uses the publicly available Kickstarter database WebRobots.io, alongside additional data collected by a custom webscraper script. This thesis uses the data in the category Technology from the Kickstarter platform, followed by a data analysis in Python.

The basic RF model scored 0.7428 on accuracy, 0.7169 on precision, 0.7326 on recall, 0.7247 on the F1-Score and 0.8150 on AUC-ROC. The best performing model, the XGBoost BOHB model with all-MiniLM-L6-v2 Embeddings, scored 0.8180 on accuracy, 0.8177 on precision, 0.8182 on recall, 0.8180 on the F1-Score, and 0.9011 on the AUC-ROC. This work contributes to the academic understanding of crowdfunding dynamics and provides actionable insights for researchers interested in NLP in crowdfunding.