AI-Driven Whitepaper Analysis and ICO Performance: Predicting Fundraising Success, Token Trading, and Post-ICO Returns with Large Language Models
Files
Collin_27692100_Warscotte_96932400_2026.pdf
Open access - Adobe PDF
- 1.26 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- The thesis explores whether artificial intelligence language models can forecast success and future financial performance with the information presented in whitepapers for ICOs. Initial Coin Offerings are also known to suffer from severe information asymmetry, weak disclosure requirements, and limited investor protection, which is why whitepapers are an important but challenging source of information for investors. A novel dataset containing 1,202 ICOs is employed and four large language models are compared for this paper, ChatGPT-5.4, ChatGPT-5.4 Mini, ChatGPT-5.4 Nano, and Mistral Medium 3. The analysis involves three dimensions of ICO performance: token trading success, fundraising success, and abnormal buy-and-hold returns. The outputs indicate that AI- generated predictions represent essential information about token trading success and amount raised. Mistral presents the best classification performance, with the highest recall and the lowest AIC, but it also produces the largest number of false positives. Each of the four models is significantly correlated with increased fundraising amount; thus, we conclude that LLMs are able to capture whitepaper signals pertaining to investor behaviour in ICO phases. However, the implications are narrower for financial performance. Only GPT-Nano has statistically significant positive association with abnormal BHR, but such results on a very small portfolio can only be used cautiously. Broadly, the research indicates that AI language models are able to recognize signals associated with ICO attractiveness and market viability, but are more unclear as to their prognostications for longer-term investment performance.